Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 123 appraisals
Clinical gastroenterology and hepatology : the official clinical practice journal of the American Gastroenterological Association
Artificial Intelligence Tools for Gastrointestinal Research: A Practical Guide
Artificial intelligence (AI) tools, including large language models (LLMs) such as ChatGPT, Claude, and Gemini, are increasingly used in biomedical research. However, practical guidance for integrating these tools into the manuscript writing workflow remains limited. This narrative review provides a structured, step-by-step guide for gastrointestinal (GI) researchers and clinicians seeking to use AI responsibly across the stages of manuscript preparation.
2 Aug 2026
Read appraisal →Clinical spine surgery
An Introduction to Machine Learning for the Practicing Spine Surgeon
The applications of new and emerging technologies in spine surgery are constantly expanding. Specifically, machine learning algorithms have seen a rise in utilization in clinical research, allowing for interpretation of large datasets that have the capability of experiential learning. The goal of this work is to present a guide for surgeons to better understand model design, key takeaways, and common pitfalls related to machine learning to ensure accurate and appropriate interpretation of analytical findings in their practice.
2 Aug 2026
Read appraisal →Rheumatic diseases clinics of North America
Sources of Bias in Clinical Artificial Intelligence and Applications in Rheumatology
Rheumatology machine-learning models are limited by preexisting, technical, and emergent biases; the interaction of data constraints, design choices, and real-world clinical workflows, rather than from isolated technical errors. Across the model lifecycle, optimization objectives can encode patterns of care, access, and documentation, producing hidden subgroup failures that are obscured by aggregate performance metrics. Given the heterogeneity of rheumatic disease and current disparities in care delivery, addressing bias requires deliberate design choices before, during, and after a model is built, as well as a commitment to transparency, and sustained oversight.
2 Aug 2026
Read appraisal →Journal of the American Medical Informatics Association : JAMIA
Explainable machine learning in healthcare: methods, interpretation, and applications for clinical research
OBJECTIVES: To provide a practical and methodologically grounded overview of explainable machine learning (XML) approaches in healthcare, with emphasis on their interpretation and application in clinical research and decision support. By moving beyond traditional predictive models, this primer aims to foster trust, transparency, and informed clinical decision-making, ultimately bridging the gap between data science and medical practice. MATERIALS AND METHODS: We present a structured review of commonly used XML methodologies, including global and local interpretability tools such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) plots. For each method, we explain the underlying mechanism at a high level, visualize representative outputs, and provide structured guidance on interpretation, appropriate use, and limitations, illustrated using the publicly available Heart Disease dataset. RESULTS: XML techniques provided intuitive visual and quantitative insights into how predictors influence model predictions. Global methods characterized population-level feature effects, whereas local methods revealed patient-level contributions useful for individualized interpretation. Our worked examples demonstrate how XML outputs can identify nonlinear relationships, detect interaction effects, and reveal heterogeneity in predicted risk across patients, addressing key challenges in translating ML predictions into interpretable outputs for clinical research. DISCUSSION AND CONCLUSION: XML tools offer valuable interpretability for ML models and support more transparent and accountable ML applications in clinical research. By providing a methodologically grounded overview alongside practical implementation examples and structured guidance on each method's strengths and limitations, this primer helps bridge the gap between advanced ML methodology and clinical applicability. Thoughtful adoption of XML approaches may facilitate better understanding, communication, and critical evaluation of ML predictions in healthcare research, ultimately supporting evidence-based clinical decision-making.
2 Aug 2026
Read appraisal →Artificial intelligence in medicine
Artificial intelligence language models for medical text analysis: A systematic review
Medical text records serve as essential repositories of patient information, providing a foundation for informed clinical decision-making, accurate diagnosis, reliable prognosis, and effective treatment planning. Recent advancements in Artificial Intelligence (AI), particularly in Natural Language Processing (NLP) and Machine Learning (ML), have positioned AI-driven language models as powerful tools for analyzing, classifying, and generating medical textual data. In this systematic literature review, an initial search retrieved 548 records published between 1 January 2000 and 1 July 2024. After rigorous screening based on predefined inclusion and exclusion criteria, 22 original research articles were included. The review highlights substantial progress in applying advanced architectures such as Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformers (GPT) to medical text processing tasks. These models consistently outperform conventional NLP and ML approaches, achieving superior results in disease classification, automated clinical documentation, and predictive analytics. However, critical challenges persist, including the limited availability of clinically validated datasets, variability in data preprocessing protocols, insufficient external validation, and the lack of interpretable AI frameworks, all of which collectively hinder clinical trust and large-scale adoption. Future research should prioritize the development of hybrid AI systems that integrate multimodal data sources (text, imaging, and structured records), incorporate explainable AI mechanisms, and adhere to standardized reporting frameworks. Addressing these methodological gaps will be pivotal in enhancing the reliability, clinical applicability, and impact of AI language models, thereby advancing evidence-based medicine, personalized treatment strategies, and overall patient care.
2 Aug 2026
Read appraisal →Current opinion in pediatrics
Artificial intelligence in pediatric endocrinology: clinical applications, governance, and future directions.
PURPOSE OF REVIEW: Artificial intelligence has transitioned from theoretical promise to practical implementation across medicine including pediatric endocrinology. This review examines artificial intelligence applications across diabetes care, obesity management, thyroid disorders, growth and puberty, and bone age assessment followed by discussion of limitations and future directions. RECENT FINDINGS: Artificial intelligence-driven tools ranging from machine learning algorithms to multimodal large language models enable personalized risk prediction, diagnostic support, and treatment optimization. Human-in-the-Loop frameworks enhance safety by integrating computational efficiency with clinician oversight. In diabetes, artificial intelligence supports glycemic forecasting, automated insulin delivery, complication screening, and digital twin modeling for therapeutic discovery. In obesity, thyroid disease, growth disorders, and bone age assessment, artificial intelligence enhances diagnostic accuracy, risk stratification, and clinical workflow efficiency. However, widespread implementation raises concerns regarding algorithmic bias, data privacy, transparency, regulatory oversight, and health equity. Addressing these limitations through explainable artificial intelligence, federated learning, and continuous performance monitoring is essential. SUMMARY: Artificial intelligence should be viewed as an augmentative tool that enhances clinical judgment in pediatric endocrinology. Preparing the next generation of pediatric endocrinologists will require intentional integration of artificial intelligence literacy into training programs, equipping clinicians to critically evaluate algorithms, interpret outputs responsibly, and participate in the ethical development of future artificial intelligence-enabled tools.
2 Aug 2026
Read appraisal →Journal of the American Medical Informatics Association : JAMIA
Alert fatigue measurement in clinical decision support: a systematic review
BACKGROUND: Alert fatigue is defined as alert dismissals due to excessive or irrelevant alerts and is frequently cited as a barrier to clinical decision support system use and impact. However, the criteria for determining the presence or absence of alert fatigue are poorly defined. The objective of this systematic review of systematic reviews was to identify operationalized definitions and measures of alert fatigue or alert-related metrics. METHODS: Systematic reviews reporting at least one alert-related metric or measure/operationalization of alert fatigue for physician-directed electronic alerts were included. The Cochrane Library, Embase, and PubMed were searched from database start to 2024. The Revised Assessment of Multiple Systematic Reviews was used to assess study quality and risk of bias. Data were synthesized narratively and with descriptive statistics. RESULTS: A total of 22 studies were included in the review. Studies reported between 1 and 11 alert metrics. Studies were most often of medium quality. Reporting of primary study characteristics was frequently judged to be insufficient. Only one article reported an operational definition of alert fatigue. The most common alert metrics were quantity, override rate, and acceptance rate. DISCUSSION: Alert fatigue measurement methods are not clearly or consistently defined in systematic reviews related to alert fatigue in clinical decision support. Reporting of other primary study characteristics is often limited. We recommend that future efforts use a significant, sustained decrease in appropriate alert response rates from an established baseline as a measure of alert fatigue.
2 Aug 2026
Read appraisal →Rheumatic diseases clinics of North America
Toward Artificial Intelligence-driven Clinical Decision Support Tools in Rheumatology
Clinical decision support systems (CDSS) have the potential to enhance rheumatology practice by assisting with differential diagnosis, treatment decisions, and predicting patient outcomes. Rheumatic conditions are complex diseases largely diagnosed clinically rather than with a predefined set of clinical, laboratory, or imaging findings. Newer artificial intelligence (AI)-driven CDSS, known as machine-learning-based systems, may be better suited for applications to rheumatic conditions compared to knowledge-based systems, which rely on a set of predefined rules. This report reviews existing studies on CDSS in rheumatology and highlights the potential benefits and pitfalls of incorporating AI-driven CDSS into rheumatology practice.
1 Aug 2026
Read appraisal →Surgery
Artificial intelligence to revolutionize surgical decision-making is still just around the corner.
Medical artificial intelligence, especially large language models, has engendered both excitement and unease across the medical community, promising improved surgical diagnostic accuracy, perioperative decision-making, and patient safety amidst concerns of considerable bias and its implications on the future of human expertise. The evolution of artificial intelligence clinical decision support originated in the surgical field through early systems like AAPHelp, which leveraged Bayesian reasoning to diagnose acute abdominal pain. Despite the initial promise of artificial intelligence clinical decision support, attempts to develop more robust diagnostic tools largely fell short in the 1980s, with instruments unable to adequately diagnose and manage complex presentations, a lack of transparency in their decision-making processes, and limitations in scope. It was not until the 2010s that clinical decision support began to show renewed potential as a meaningful adjunct in surgical practice, largely driven by advancements in machine learning techniques and wider access to large clinical data sets. New artificial intelligence models, particularly those utilizing large language models, now demonstrate impressive capabilities, from predicting the risk of postoperative complications for individual patients to streamlining clinical documentation and beyond. However, challenges persist regarding transparency, bias, reliability, and integration into clinical workflows. Despite these hurdles, artificial intelligence-based tools like large language models represent an exciting new chapter in surgical decision-making. Built on a long history of clinical decision support systems in surgery, these technologies hold great promise to meaningfully augment surgical practice and improve care for patients.
1 Aug 2026
Read appraisal →International urology and nephrology
Acute kidney injury prediction and prognostication using machine learning
Acute Kidney Injury (AKI) is a major health concern with high costs and poor outcomes, partly due to late diagnosis. This paper reviews the application of artificial intelligence (AI) and machine learning (ML) to overcome these limitations by enabling earlier and more accurate prediction and prognostication of AKI. Many studies on AI and ML models used to predict and prognosticate AKI were included in the review. The focus was on models that analyze complex datasets, including real-time data streams like novel biomarkers and continuous vital signs, to achieve earlier and more accurate predictions than conventional methods. ML models show high predictive accuracy for AKI onset and outcomes across various clinical settings, including intensive care units, sepsis, and postoperative and postcontrast situations, with key findings like the successful integration of real-time data to reflect the evolving nature of kidney injury. AI and ML offer a powerful, proactive solution for AKI management. By leveraging diverse data, these technologies can significantly improve patient outcomes and reduce healthcare costs. While their promising performance warrants further exploration, successful clinical integration will require user-friendly platforms and continued validation.
1 Aug 2026
Read appraisal →JMIR mHealth and uHealth
Machine Learning Frameworks for Wearable-Based Stress Modeling in Naturalistic Settings: Scoping Review
BACKGROUND: Stress, as commonly recognized, is an integral part of modern life and can significantly affect both mental and physical health. While substantial advancements have been made in measuring physical fitness through wearable devices, the detection and assessment of mental stress remain in their early stages. OBJECTIVE: The objective of this paper is to review recent studies of wearable-based stress detection in naturalistic settings, with a specific focus on characterizing machine learning frameworks inspired by the model card approach. METHODS: This review was conducted using the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist. A total of 353 articles were identified through searches in databases such as PubMed, MEDLINE, ScienceDirect, IEEE, ACM Digital Library, Web of Science, and Embase. Studies were considered eligible if they collected data from healthy adults in naturalistic settings using wearable devices and used machine learning models for stress detection. RESULTS: A total of 34 articles met the eligibility criteria, including 11 conference papers, 22 journal articles, and 1 preprint published between 2017 and 2024. From these studies, we analyzed key machine learning modeling decisions such as problem formulation, ground truth determination, and machine learning algorithms. Additionally, we examined the major contributions of each study, focusing on the challenges they addressed and the solutions they proposed. Based on these findings, we proposed a model card framework for reporting machine learning-based, wearable-based stress detection. CONCLUSIONS: This scoping review highlights recent trends in machine learning models for stress detection and measurement using wearable signals. It underscores the need for improved standardization in reporting practices for datasets and key machine learning decisions, as well as the importance of addressing critical challenges associated with data collection in real-world settings. We hope this review will support and strengthen ongoing research efforts, promote knowledge sharing, and promote collaboration among researchers-ultimately advancing the field as a community.
1 Aug 2026
Read appraisal →PloS one
Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach
BACKGROUND: Early identification of diabetic retinopathy (DR), which is a primary cause of vision impairment globally, is a crucial phasis for effective intervention and treatment. Traditional screening workflows rely on manual diagnosis by ophthalmologists, which remains the gold standard but can be time-consuming and subject to variability due to human factors. To support and enhance the screening process, artificial intelligence (AI)-based tools have shown promise in automating DR detection, particularly with recent advances in deep learning. However, medical images with long-range dependencies and spatial linkages can be challenging for CNN-based algorithms to handle. METHODS: This paper proposes a Vision Transformer (ViT)-based model, specifically using a Compact Convolutional Transformer (CCT), for early automated detection of DR. The model uses self-attention techniques to improve feature extraction and classification performance; combining three main stages: the CCT tokenizer, transformer encoder, and sequence pooling. The proposed approach was trained on public datasets (EyePACS and APTOS 2019) and evaluated against state-of-the-art deep learning architectures. RESULTS: Our experimental findings demonstrate that ViT performs among the best in the current state of the art with an overall accuracy of 97% and F1-scores above 0.95 across all DR severity levels. Our system is primarily designed for the pre-screening stage of diabetic retinopathy workflows, enabling rapid and reliable identification of potential DR cases for further clinical evaluation. CONCLUSION: These results highlight the potential of transformer-based designs in medical picture analysis, as well as the implications for telemedicine and e-health solutions in real-time, especially in cases of low-resource settings.
28 July 2026
Read appraisal →Biomedical physics & engineering express
BMSNet: a boundary-guided multi-scale polyp segmentation network
Accurate polyp segmentation in colonoscopy images is essential for computer-aided diagnosis and early colorectal cancer screening. However, this task remains challenging due to large variations in polyp size, shape, and texture, as well as blurred boundaries and low contrast between polyps and surrounding mucosa. To address these challenges, we propose boundary-guided multi-scale polyp segmentation network (BMSNet), a boundary-guided multi-scale network for accurate polyp segmentation. First, an Scale-Aware Modulation Meets Transformer-T backbone initialized with officially released pretrained weights is adopted to extract multi-level features efficiently. Then, a boundary prediction module is introduced to generate preliminary boundary priors through bidirectional feature interaction and differentiable Canny-based auxiliary supervision. Based on the generated boundary information, we design a boundary-guided feature enhancement module to refine feature representations. This module combines frequency-domain enhancement, multi-scale feature interaction, and boundary-guided modulation. Finally, a feature fusion unit is introduced to progressively integrate multi-scale features and generate the final segmentation prediction. Extensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed method. On ClinicDB, Kvasir, ColonDB, ETIS, and CVC-300, BMSNet achieves Dice scores of 0.943, 0.920, 0.828, 0.825, and 0.913, respectively. These results indicate that BMSNet achieves competitive segmentation performance with moderate computational complexity under the evaluated experimental setting.
28 July 2026
Read appraisal →Physics in medicine and biology
Anatomically adaptive feature-wise linear modulation for deep learning-based low-dose CT denoising
Objective.Low-dose computed tomography (LDCT) reduces radiation dose but, introduces heterogeneous noise due to different photon attenuation based on anatomical tissue. Most deep learning techniques assume uniform noise in LDCT and perform equal noise removal across different regions, leading to sub-optimal performance across different tissues. This work aims to design a physics-based framework that explicitly models region-dependent noise characteristics to improve LDCT noise removal.Approach.We propose an anatomically adaptive noise reduction framework. The proposed anatomically adaptive feature-wise linear modulation (FiLM) model consists of a U-Net architecture that integrates two complementary units: the global FiLM unit in the encoder, which modifies features globally based on global image statistics to remove overall noise, and the local FiLM unit in the decoder, which modifies features locally based on the type of anatomical tissue. This dual design enables the modeling of overall image noise characteristics in addition to the removal of local noise associated with each anatomical tissue.Main results.The model performance was evaluated using AAPM-Mayo Clinic LDCT dataset, and the trained model tested on the TCIA dataset. The proposed model outperformed all competing methods. Local noise analysis showed that noise removal was consistent across different anatomical regions, achieving 37.14% in the lung, 48.75% in soft tissue, and 35.18% in bone. Visual results also confirmed significant improvements in noise removal, preservation of structural details, and reduction of non-residual distortions. Furthermore, the model demonstrated its ability to generalize under domain shift.Significance.This work presents a framework for anatomically adapted noise removal by linking feature modification with the physical properties of noise in LDCT through a dual-modulation process for both general and tissue-related noise. The model achieves a balance between noise removal and preservation of anatomical detail, making it a robust approach to LDCT noise removal.
28 July 2026
Read appraisal →Scandinavian journal of pain
Multimodal deep learning fusion for automatic pain detection in cancer patients
INTRODUCTION: Since pain is a multidimensional and subjective experience, pain assessment remains challenging. With advances in artificial intelligence (AI), automatic pain assessment (APA) systems offer a valuable opportunity for objective pain evaluation. However, most approaches focus on a single modality. In this proof-of-concept study, exploring multimodal fusion strategies in a controlled experimental setting, we present a deep learning framework for multimodal fusion that combines facial, acoustic, and textual information to improve APA in cancer patients. METHODS: A multimodal dataset was created from video-recorded interviews with oncologic patients. In Phase I, audio, video, and transcripts were segmented at the sentence level and temporally aligned using the Eudico Linguistic Annotator (ELAN) to ensure frame-level correspondence across modalities. In Phase II, modality-specific features were extracted: Facial Action Units from OpenFace, acoustic descriptors (MFCCs, chroma, spectral contrast, and Mel-spectrogram) from a dedicated speech-processing pipeline, and sentence-level textual embeddings from ITA-BERT. During training, the most effective analytical strategy was chosen through knowledge transfer approaches. The ELAN-assisted annotation pipeline streamlined expert labeling. Two architectures were implemented and compared: bimodal autoencoder fusion models and a transformer-based model with pairwise cross-modal attention. These models were trained and evaluated using subject-independent and stratified 5-fold cross-validation. To address the lack of independence between segments, a strictly subject-independent cross-validation strategy was adopted. RESULTS: Knowledge transfer using pretrained large-scale models outperformed traditional feature-based approaches and was applied to multimodal pain detection. Multimodal models achieved performance comparable to the strongest unimodal modality (text), while showing improved balance across modalities, suggesting potential complementary effects. Both multimodal architectures demonstrated high accuracy in distinguishing between pain and non-pain classes. The bimodal autoencoder achieved stable results across folds, with a mean accuracy of about 80 % and balanced error distribution. The pairwise transformer with cross-modal attention achieved similar performance, with smooth training and validation loss curves. No evident divergence between training and validation loss curves was observed across folds, suggesting stable behavior within the cross-validation setting. However, subject-level overfitting cannot be excluded given the limited sample size. CONCLUSION: Multimodal fusion enhances system robustness by integrating complementary signals. Despite limitations and the need for improvement, multimodal deep learning strategies can support the detection of observable pain-related expressions.
26 July 2026
Read appraisal →BMJ health & care informatics
Benchmarking large language models for de-identification of electronic health record notes
OBJECTIVES: The rapid evolution of large language models (LLMs) and their growing application in clinical text processing have created an urgent need for reliable de-identification mechanisms. While LLMs show promise in identifying sensitive health information (SHI), their capabilities require rigorous evaluation. This study aims to conduct a comprehensive benchmarking analysis of various LLM-based, traditional rule-based and hybrid de-identification methods. METHODS: Our benchmark analysis used five datasets (i2b2-2006, MIMIC-2008, i2b2-2014, i2b2-2016 and OpenDeID v1) from different countries. We developed three baseline and eight LLM-based models. The experimental setup encompassed nine different settings using various combinations of training and testing sets to assess model robustness and cross-dataset performance. RESULTS: In the baseline models, the approach trained on the combined corpus of all five datasets (setting 3) significantly outperformed the other settings, achieving a strict F1 micro-average score of 0.8172. Regarding LLM-based models, the supervised fine-tuning approach using the same combined configuration (setting 9) achieved the highest performance with a strict F1 score of 0.9447. DISCUSSION: The harmonisation of corpora ensured standardised data formatting and SHI management across five diverse datasets, highlighting the necessity for uniform categorisation to enhance the reliability of de-identification results. CONCLUSIONS: Our findings indicate that while fine-tuned LLMs offer superior accuracy, the observed performance variability across heterogeneous electronic health record sources poses significant technical challenges. Real-world implementation must address these inconsistencies to overcome the ethical and technical hurdles associated with deploying LLMs for handling sensitive health data.
23 July 2026
Read appraisal →PloS one
Textual overlap rather than domain alignment: A comparative study of fine-tuning strategies for specialised machine translation with large language models
General-purpose large language models (LLMs) may struggle in specialised machine translation, but the conditions under which fine-tuning improves translation performance remain unclear. This study compares full-parameter fine-tuning (FPFT) and parameter-efficient fine-tuning (PEFT) for Chinese-English political discourse translation using a purpose-built corpus and the Qwen3-14B model. Translation performance was assessed on three 50-item test sets using BLEU, ROUGE-L F1, METEOR, and BERTScore F1, together with BLEU pass-rate likelihood-ratio G2 tests, paired t-tests, and paired Cohen's dz for item-level score differences. The results reveal a clear contrast between unseen in-domain evaluation, maximum-overlap benchmarking, and semantically related but non-fine-tuned evaluation. On Test Set A and Test Set C, neither fine-tuning strategy produced a statistically significant BLEU pass-rate advantage over the base model, and paired tests across the continuous metrics did not show consistent fine-tuning gains. On Test Set B, which was sampled from the fine-tuning corpus, both fine-tuned models substantially outperformed the base model across all four metrics, with FPFT achieving the highest scores and PEFT providing a more computationally efficient alternative. These findings indicate that textual overlap between training and deployment data, rather than broad domain similarity alone, strongly conditions the observed benefit of fine-tuning. The study offers an empirically grounded framework for selecting fine-tuning strategies in specialised machine translation.
22 July 2026
Read appraisal →International ophthalmology
Machine learning-based model for identifying liver injury in patients with thyroid-associated ophthalmopathy
PURPOSE: To explore the relationship between thyroid-associated ophthalmopathy (TAO) and liver injury and establish a model for identifying liver injury by using machine learning so as to provide an effective diagnostic tool for liver injury in patients with TAO. METHOD: A single-center retrospective study was conducted to collect the clinical data of 318 patients with TAO admitted to a hospital from 2016 to 2022. The patients were divided into a TAO liver injury group (104 cases) and TAO normal liver function group (214 cases) according to whether the patients had normal liver function. The multivariate binomial logistic regression model was used to analyse the risk factors for liver injury in patients with TAO. Feature selection was performed using the random forest algorithm. The research data were divided into a training set and a test set at a ratio of 6:4. Taking whether accompanied by liver injury (0 = no, 1 = yes) as the outcome variable, models were established based on logistic regression, random forest, support vector machine, and decision tree. The performance of the models was evaluated using metrics, including sensitivity, specificity, positive predictive value, negative predictive value, Youden index, and accuracy. RESULT: The random forest method was used for feature screening, and variables ranked among the top 10 according to either mean decrease accuracy or mean decrease Gini were selected for model construction. In five-fold cross-validation, the RF model showed the highest accuracy of 0.937 and AUC of 0.977. In the test set, the RF model also showed good discrimination, with an AUC of 0.973, and the SVM model showed the highest AUC of 0.986, while SVM and LR achieved the highest accuracy of 0.914. CONCLUSION: This study shows that a classification model based on machine learning can effectively identify the risk of liver damage in patients with TAO.
18 July 2026
Read appraisal →PLoS medicine
How to benchmark medical AI agents
Medical artificial intelligence research is shifting from single-task models toward multimodal large language model-based agents for complex clinical workflows, requiring benchmarks that assess clinical reasoning, process safety, and resource stewardship rather than final outputs alone.
12 July 2026
Read appraisal →PloS one
Air quality index prediction using machine learning regression models: A comparative analysis
Air plays a vital role in human life, and poor air quality can lead to respiratory infections. Given the significant impact of air quality on people's health, monitoring and assessing air quality is crucial. With advancements in machine learning (ML) and artificial intelligence (AI), we now have extensive tools to measure the Air Quality Index (AQI). Air quality is influenced by various pollutants, including carbon monoxide (CO), nitrogen dioxide (NO2), ozone (O3), and sulfur dioxide (SO2), which are prevalent in highly polluted areas and contribute to a wide range of illnesses. Particulate matter, such as PM2.5 (particles with an aerodynamic diameter of less than 2.5 µm) and PM10, poses additional health risks. To address these concerns, this study focuses on predicting AQI values for major cities in Pakistan, specifically Karachi and Peshawar, using four prominent ML algorithms: Random Forest (RF), Gradient Boosting (GB), Linear Regression (LR), and Ridge Regression (RR). The results indicate that the models effectively predicted AQI using evaluation metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R2). This research's novelty lies in using the latest AQI datasets for Karachi and Peshawar and applying standard scaling for AQI normalization. Additionally, the study compares evaluation metric results across different cities, highlighting the importance of using standard scalers to achieve optimal model performance. This research underscores the value of advanced ML techniques for accurate AQI prediction and analysis.
10 July 2026
Read appraisal →Clinical infectious diseases : an official publication of the Infectious Diseases Society of America
More Than a Message: Death by a 1000 'Chats'.
Secure electronic messaging within the electronic health record (EHR), such as Epic Chat, has transformed communication in hospitals by enabling rapid, asynchronous communication. Yet for Infectious Diseases (ID) consultants, the constant barrage of messages comes at a cost: fragmented attention, disrupted clinical reasoning, and potential patient harm. Messages that are repetitive, trivial, or casually worded can also fuel miscommunication and even medicolegal risk. As this type of messaging becomes increasingly prevalent in clinical practice, we must establish clearer boundaries and a cultural shift that values purposeful communication. Practical strategies include defining what requires chat versus direct phone or in-person conversation, batching responses, protecting focus time, and teaching messaging etiquette. To optimize patient care and protect clinician well-being, we must be intentional about when and how we communicate.
9 July 2026
Read appraisal →Journal of hematology & oncology
Emerging artificial intelligence advances in oncology: latest updates from the 2026 AACR annual meeting.
Advances presented at the 2026 American Association for Cancer Research (AACR) Annual Meeting highlight a shift from standalone artificial intelligence (AI) models to integrated, agentic systems across oncology. Platforms such as Synapse enable large-scale data coordination, supporting interoperable and reproducible research. Building on this foundation, conversational and multi-agent AI tools (e.g., DrBioRight, GP CoPilot, Isabl AI Agent) allow natural language interaction with multimodal cancer data, lowering technical barriers. Agentic frameworks for real-world data (RWD) transformation, including clinical document abstraction, cohort extraction, and social determinants of health (SDOH) analysis, demonstrate high accuracy and scalability, while self-critical systems improve reliability. Clinically, AI shows growing impact through validated imaging biomarkers, enhanced trial matching, and scalable cohort identification. In parallel, multi-agent systems are accelerating therapeutic discovery, including CAR-T development and immunotherapy target identification. Collectively, these advances position AI as an active, collaborative partner in cancer research and precision oncology.
5 July 2026
Read appraisal →Langenbeck's archives of surgery
The role of large language models in the writing of surgical reviews: fact or fantasy?
BACKGROUND: Amidst the current enthusiasm concerning artificial intelligence and its possible application in the composition of different kinds of scientific and non-scientific written documents, we evaluated the usage of artificial intelligence for writing surgical short reviews. METHODS: In order to assess the formal and content quality of AI-generated texts compared to human written texts, ten AI-based text generators (five chatbots and five content creators) and four surgeons in training received the same prompt for a short scientific article on a liver surgery theme. All texts were anonymized and subsequently evaluated by three experienced liver surgeons based on a pre-defined scoring scheme, as well as for quality of references and readability according to readability indices. Furthermore, all texts were tested for plagiarism using PlagScan. RESULTS: Overall percentage of correct assessment for AI/non-AI generation by experienced surgeons lay at 78.57%. Human written text had a mean word count of 1054 versus 874 in AI-generated text, with a higher mean Flesh Reading Ease Score (FRE, 26.2 ± 5.1 versus 17.7 ± 6.1). References were PubMed-listed in 100% for human versus 46% for AI-generated text, with only one non-human text reaching 100% formally correct citation of references. PlagScan found 6.4%±1.3 mean resemblance to existing texts for human versus 7.6%±4.5 for AI-generated text. DISCUSSION: Overall, AI could already mislead experienced scientific surgeons in 26.7% of cases into believing it to be human. However, formal requirements, especially considering referencing, are still in great need of improvement with only one of AI-generated articles fulfilling our quality requirements.
4 July 2026
Read appraisal →Journal of biomedical informatics
A comprehensive survey of AI agents in healthcare
OBJECTIVE: This survey aims to systematically map the rapidly evolving landscape of AI agents in healthcare. It addresses the critical need to adapt general-purpose agentic frameworks characterized by autonomy, planning, and tool use to the high-stakes, safety-critical constraints of medical decision-making and patient care. METHODS: We conducted a comprehensive review of over 200 recent studies, synthesizing literature from major academic databases. We developed a holistic taxonomy that traces the full lifecycle of healthcare agents, analyzing perception modalities, core technical architectures, and evaluation protocols specific to autonomous systems. RESULTS: The review presents a quantitative landscape analysis showing exponential growth in the field. We structure the domain into three pillars: (1) Perception of multi-modal clinical data (e.g., EHR, imaging, genomics); (2) Agent Capabilities, including tool use, reasoning, memory, and multi-agent collaboration; and (3) an Application Ecosystem organized by stakeholder roles (clinicians, patients, researchers, and administrators). Additionally, we categorize evaluation frameworks, and discuss the deployment readiness of current systems across technical, evidentiary, and governance dimensions. Finally, we identify challenges for advancing healthcare agents from controlled evaluation toward real-world clinical integration. A continuously updated repository of related papers is available at https://github.com/AgenticHealthAI/Awesome-AI-Agents-for-Healthcare. CONCLUSION: AI agents offer significant potential to enhance healthcare through autonomous reasoning and workflow integration. However, current research remains largely concentrated in benchmark and controlled evaluation settings, and the translation into clinical practice will require advances in reliability, privacy protection, governance, and operational integration.
4 July 2026
Read appraisal →Korean journal of radiology
Agentic Artificial Intelligence for the Automated Generation of Accurate Summary Podcasts of Radiology Research Papers
OBJECTIVE: To evaluate whether a custom agentic artificial intelligence (AI) pipeline can overcome the limitations of general-purpose large language model tools, when compared with a generic commercial tool (Google NotebookLM [NBLM]), for generating podcast-style summaries of radiology research articles. MATERIALS AND METHODS: Twenty-two PDF-format original research articles published in the April 2025 issue of Radiology were processed using our Programmable, Phoneme-Aware PDF-to-Podcast Pipeline (P5) and NBLM to generate 44 audio episodes. P5 utilizes a multi-agent workflow for script generation, quality assurance, pronunciation enhancement, and audio synthesis. Four radiologists from a pool of 25 (7 generalists and 18 specialists) were randomly assigned to evaluate each blinded audio episode, yielding 176 total evaluations. The primary outcomes were the number of hallucinations (factual errors) per episode and the percentage of hallucination-free episodes. Secondary outcomes included the number of inappropriate statements, mispronunciations, and flow disruptions; the composite quality score (Quality Assessment of Educational Podcasts [QAEP]); the key results coverage score; and overall listener preference. Data were analyzed using generalized linear mixed models. RESULTS: The P5 method produced significantly fewer hallucinations per episode compared with NBLM (mean, 0.32 vs. 0.93; P < 0.001) and a higher proportion of hallucination-free episodes (71.6% [63/88] vs. 56.8% [50/88]; P = 0.013), consistently across generalists and specialists. P5 demonstrated significantly fewer mispronunciations (mean, 0.11 vs. 1.62; P < 0.001) and flow disruptions (mean, 0.19 vs. 1.06; P < 0.001) per episode, as well as higher mean QAEP composite scores (4.62 vs. 4.28; P < 0.001), compared with NBLM, consistently across generalists and specialists. Raters preferred P5 over NBLM in 72.7% (64/88) of comparisons (P = 0.003). CONCLUSION: Our custom agentic AI pipeline generated podcast-style summaries of radiology research articles with significantly higher quality and greater listener preference than the generic commercial tool.
3 July 2026
Read appraisal →Journal of the American Medical Informatics Association : JAMIA
Engineering biomarker representations of vital signs data enhances deep learning mortality prediction
OBJECTIVES: We evaluated bidirectional long short-term memory models for predicting inpatient mortality using different approaches to processing vital signs data collected during the initial 24 h of intensive care unit (ICU) admissions. MATERIALS AND METHODS: We compared 3 vital-sign representations: (1) raw data recorded every 5 min, (2) preprocessed data averaged hourly, and (3) preprocessed data using biomarker representations that extends a digital oximetry biomarker toolbox of PhysioZoo software, applied to blood pressure, heart rate, temperature, respiratory rate, and SpO2. RESULTS: Across 2 large ICU datasets, HiRID and eICU, models trained on the frequency-normalized representation achieved higher discrimination and lower Brier scores than those trained on raw 5-min and hourly averaged data. DISCUSSION: The use of biomarker representations of vital signs yielded the largest improvements in discrimination and overall probabilistic performance reflected by lower Brier scores for predicting inpatient mortality by deep learning. CONCLUSION: Thus, we recommend using a similar approach to vital signs preprocessing for time-series predictive models.
3 July 2026
Read appraisal →Journal of medical Internet research
Transformation Versus Innovation in Digital Health Care and the Future of Clinical AI
Digital innovation is frequently presented as the key to transforming health care. In this News and Perspectives article, JMIR Correspondent and academic physician Boon-How Chew reports on the distinction between and the direction of transformation and innovation, reflecting on the future of clinical AI and what lasting change in health care ultimately requires.
3 July 2026
Read appraisal →Physics in medicine and biology
Deep learning based time series analysis for breathing phase prediction in phase-gated proton therapy
Objective. Phase gating is a critical technique to mitigate tumor motion during radiotherapy, particularly in spot-scanned particle therapy where internal motion can interfere with dynamic spot scanning patterns and, simultaneously, introducing substantial range uncertainties. However, the current commercial state of the art in real-time phase prediction is challenged by patient-specific breathing variability as well as detection and delivery system latencies. This leads to suboptimal efficiency, mis-timed radiation delivery and requires frequent manual intervention. This study aims to improve phase prediction accuracy using deep learning (DL)-based time series forecasting to enable more accurate dose delivery.Approach. Retrospective breathing waveforms from 69 proton therapy patients, sampled at 30 Hz, were labeled with inspiratory peaks and assigned subjective regularity scores (four levels). DL models with various architectures were trained using waveform amplitude to predict current or future breathing phases. Model performance was evaluated using mean squared error, phase binning accuracy, and timing deviation for radiation on/off events. Bayesian optimization was used for hyperparameter tuning. Results were compared between models and against a commercial algorithm currently in clinical use.Main Results. The curated waveform dataset included 165 242 s for training, 24 057 s for validation, and 30 322 s for testing, with an additional 40 604 s from separate patients for extended validation. The long short-term memory and temporal fusion transformer models significantly outperformed the commercial algorithm, improving phase prediction accuracy by nearly 20% and reducing timing deviations across all regularity levels.Significance. DL-based time series forecasting may substantially improve breathing phase prediction accuracy over current clinically available methods, offering a more precise and reliable approach to phase-gated radiation delivery.
2 July 2026
Read appraisal →International journal of medical informatics
Artificial intelligence in gastroenterology clinical practice: Scoping review of large language model applications.
BACKGROUND: Advances in artificial intelligence have brought renewed attention to tools that can work with the large amount of written information generated in clinical practice. Among these, large language models (LLMs) stand out for their ability to interpret and generate medical text in a flexible, context-aware way. Gastroenterology, like many specialties, produces a wide range of narrative and semi-structured data, and this has encouraged researchers to explore how LLMs might help clinicians manage everyday tasks. Recent studies have examined their potential contributions to patient education, communication between care teams and patients, decision support, and routine documentation, reflecting a growing interest in how these systems might fit into real-world clinical workflows. OBJECTIVE: This scoping review aimed to map current applications of LLMs in gastroenterology clinical practice, including subspecialty focus, study designs, model types, and reported outcomes. METHOD: Following PRISMA guidelines, a systematic search was conducted in PubMed, Scopus, and Web of Science for studies published between January 2022 and August 2025. Eligible studies included original research assessing LLM applications in gastroenterology clinical practice. Data were extracted on subspecialty, application domain, LLM type, data source, and outcomes. We employed thematic analysis to address our primary research question. RESULTS: 73 out of 2895 studies identified in the initial search met the inclusion criteria. Six subspecialties and six application domains emerged from our review. Hepatology (20/73 studies, 27.3%) and endoscopy (17/73 studies, 23.2%) were the most represented subspecialties. The most frequently investigated application domains were patient education and communication (38 studies) and decision support and clinical guidance (24 studies). Most studies were simulation-based or literature-based cases, although an increasing number have used real-world clinical data, particularly in recent years. The majority evaluated general-purpose models such as GPT-3.5 and GPT-4, with some incorporating retrieval augmentation or fine-tuning. Reported outcomes varied by application domain and included measures of accuracy, concordance, completeness, relevance, safety, reliability, usability, user satisfaction, efficiency, time savings, and educational value. Commonly described limitations included variable reliability, incomplete responses, and challenges in generalizing from simulated to clinical settings. CONCLUSIONS: Research on LLMs in gastroenterology has expanded across multiple subspecialties and application domains. Current evidence is primarily based on simulation studies, with limited but growing evaluation using real-world clinical data. Further work is needed to assess performance in prospective and applied clinical contexts.
2 July 2026
Read appraisal →The Journal of prosthetic dentistry
Performance of large language models conducting systematic review tasks in prosthodontics.
STATEMENT OF PROBLEM: Systematic reviews (SRs) are time-consuming and resource-intensive processes. Whether large language models (LLMs) can improve the process is unclear. PURPOSE: The purpose of this study was to evaluate the accuracy and reliability of 4 LLMs (GPT-4, Gemini, Claude, and Elicit) in performing SR tasks (full-text screening, data extraction, and risk of bias assessment) at 3 different sequential periods of time (0, 15, and 30 days). MATERIAL AND METHODS: A comprehensive systematic search was conducted across 5 databases in December 2024, with 59 articles evaluated for screening (2 used for pilot) and 31 for data extraction (2 used for pilot). A 3-pronged prompting strategy was used, including persona-based initialization, few-shot learning, and structured population, intervention, control, outcome (PICO) criteria. Performance was assessed through 3 repeated evaluations at 2-week intervals by measuring accuracy and reliability using standard metrics (accuracy, precision, F1-score, sensitivity, and specificity) against expert assessments, data extraction quality on a 0 to 5 scale, and risk of bias agreement via the Cohen kappa, with statistical analysis using Kruskal-Wallis and Dunn post hoc tests (α=.05). RESULTS: In full-text screening, Claude achieved the highest sensitivity at 97%, while Claude and Elicit both showed strong overall performance with 86% accuracy and 87% F1-scores. All models maintained sensitivity above 90%. For data extraction, GPT-4 consistently performed best with median scores of 5.0, while Claude and Gemini showed similar capabilities. Significant differences only appeared in labeling and modeling tasks during Week 1 (P=.04). Risk of bias assessment agreement with experts varied from 55% to 90% across different criteria. CONCLUSIONS: LLMs show potential for SR efficiency (especially for data extraction) but require human oversight because of variable performance across models and tasks.
2 July 2026
Read appraisal →Journal of pediatric health care : official publication of National Association of Pediatric Nurse Associates & Practitioners
Systematic Review of Studies Investigating Infant Feeding Difficulties: Focus on Machine Learning Applications
INTRODUCTION: This study aims to systematically review studies using machine learning for infant feeding difficulties and identify the clinical applicability and limitations of such approaches. METHODS: A literature search was conducted in accordance with the PRISMA guidelines. Studies on the use of machine learning for infants with feeding difficulties were analyzed. RESULTS: We screened 1,104 studies, and 10 were eligible for inclusion. These studies applied machine learning to measure physiological signals during feeding, such as the number of sucking, swallowing and breathing events, to classify or predict feeding difficulties. The most common algorithms in the analyzed studies included support vector machines (SVMs), k-nearest neighbors (KNNs), and decision trees (DTs), which generally achieved high classification accuracy. DISCUSSION: Although all included studies used sensor-based data, only four directly applied machine learning models. Future research should expand the direct applications of machine learning, standardize measurements, and validate algorithms to improve clinical utility.
2 July 2026
Read appraisal →Current opinion in psychiatry
Opportunities and risks of large language models in digital interventions for substance use disorders
PURPOSE OF REVIEW: Large language models (LLMs) are increasingly integrated into digital mental health tools, yet their role in substance use disorder (SUD) interventions remains poorly understood. This review synthesizes emerging evidence on the opportunities and risks of applying LLMs across the digital SUD care continuum. RECENT FINDINGS: Studies report promising applications in early detection, personalized support, continuous monitoring, and relapse prevention. LLMs demonstrate capacity to extract substance-use signals from natural language, generate supportive and motivational responses, and interpret narrative data for patient-reported outcomes. However, risks are substantial. LLMs can produce inaccurate or hallucinated content, may reinforce stigma or demographic bias, and can generate misleading or potentially unsafe advice. Privacy concerns are amplified in SUD contexts, where sensitive data are often managed outside regulated healthcare systems. Existing regulatory frameworks such as the EU AI Act or U.S. device regulations, do not yet provide clear governance for anonymous, AI-supported SUD interventions. SUMMARY: LLMs have potential to expand scalable, low-threshold support for SUDs, but their safe deployment requires validation, bias mitigation, transparent data governance, and robust human oversight. Evidence remains preliminary, and clinical integration should proceed cautiously.
2 July 2026
Read appraisal →Journal of biomedical informatics
Clinical document metadata extraction: A scoping review.
OBJECTIVES: Clinical document metadata, such as document type, structure, author role, medical specialty, and encounter setting, is essential for accurate interpretation of information captured in clinical documents. However, vast documentation heterogeneity and drift over time challenge harmonization of document metadata. Automated extraction methods have emerged to coalesce metadata from disparate practices into target schema. This scoping review aims to catalog research on clinical document metadata extraction, identify methodological trends and applications, and highlight gaps warranting further investigation. METHODS: We followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines to identify articles from Ovid MEDLINE, Ovid EMBASE, Scopus, Web of Science and external sources that perform clinical document metadata extraction, either primarily as a methodology study, secondarily as a feature for a downstream application, or for analysis. We initially identified and screened 342 articles published between 2011 and 2025, then comprehensively reviewed 77 we deemed relevant to our study. RESULTS: Among the 77 articles included in our full text review, 49 were methodological, 22 used document metadata as features in a downstream application, and 6 analyzed document metadata composition. We observe myriad purposes for methodological study and application types. Available labelled public data remains sparse except for structural section datasets. Methods for extracting document metadata have progressed from largely rule-based and traditional machine learning with ample feature engineering to transformer-based architectures with minimal feature engineering. DISCUSSION AND CONCLUSION: Clinical document metadata extraction research has accelerated over recent years. The emergence of large language models has enabled broader exploration of generalizability across tasks and datasets, allowing the possibility of advanced clinical text processing systems. We anticipate that research will continue to expand into richer document metadata representations and integrate further into clinical applications and workflows.
2 July 2026
Read appraisal →Medical image analysis
A review of deep learning-based Unsupervised Anomaly Detection in brain MRI
The manual assessment of brain Magnetic Resonance Imaging (MRI) scans can be labor-intensive and time-consuming for radiologists. Deep Learning methods have demonstrated the potential to aid this process. However, their effectiveness relies on the availability of large, annotated data sets. Unsupervised Anomaly Detection (UAD) presents a promising alternative, offering the potential to identify and localize anomalies without per-pixel annotations. Instead, a normative distribution is learned using healthy data, enabling the identification of abnormalities as deviations. This allows UAD methods to detect abnormalities that were unseen during training. This appealing feature has led to numerous studies proposing innovations and novel approaches. In this work, we provide a review of the literature and systematically collect and compare the proposed approaches. We observe that UAD has made significant advancements in brain MRI analysis. However, individual approaches are often evaluated in different contexts, i.e., changes in acquisition parameters, pre- and post-processing, and anomaly scoring. This variability makes it challenging to assess which models perform best, underscoring the need for comprehensive comparative studies concerning the specific context of MRI scans. Our collection, featuring public data sets, research studies, and open implementations, is available at our GitHub repository https://github.com/FinnBehrendt/Unsupervised-Anomaly-Detection-in-Brain-MRI.
1 July 2026
Read appraisal →Physica medica : PM : an international journal devoted to the applications of physics to medicine and biology : official journal of the Italian Association of Biomedical Physics (AIFB)
Advancing radiotherapy with deep learning: A review of dose prediction models
BACKGROUND: Artificial intelligence (AI), especially deep learning algorithms, holds significant promise in transforming radiation therapy by enabling more precise, personalized, and efficient treatment planning. In particular, deep learning models have demonstrated their utility in critical tasks such as dose prediction, helping to streamline complex processes and improve patient outcomes. METHODS: This paper provides a comprehensive review of recent research on deep learning applications for radiotherapy dose prediction, highlighting the architectures, datasets, and methodologies used. RESULTS: The review highlights the advancements in accuracy, efficiency, and robustness achieved by deep learning-based dose prediction models. These approaches help reduce treatment planning time and minimize human error, thereby improving clinical workflows. CONCLUSION: Deep learning offers promising solutions to enhance radiotherapy dose delivery. Future research should focus on developing the generalized dose prediction models, multi-institutional validation frameworks, explainable AI approaches for clinical trust, real-time adaptation capabilities, uncertainty quantification methods, and strategies for handling rare anatomical variations.
1 July 2026
Read appraisal →IEEE transactions on pattern analysis and machine intelligence
A Survey on Interpretability in Visual Recognition
Visual recognition models have achieved unprecedented success in various tasks. While researchers aim to understand the underlying mechanisms of these models, the growing demand for deployment in safety-critical areas like autonomous driving and medical diagnostics has accelerated the development of eXplainable AI (XAI). Distinct from generic XAI, visual recognition XAI is positioned at the intersection of vision and language, which represent the two most fundamental human modalities and form the cornerstones of multimodal intelligence. This paper provides a systematic survey of XAI in visual recognition by establishing a multi-dimensional taxonomy from a human-centered perspective based on intent, object, presentation, and methodology. Beyond categorization, we summarize critical evaluation desiderata and metrics, conducting an extensive qualitative assessment across different categories and demonstrating quantitative benchmarks within specific dimensions. Furthermore, we explore the interpretability of Multimodal Large Language Models and practical applications, identifying emerging trends and opportunities. By synthesizing these diverse perspectives, this survey provides an insightful roadmap to inspire future research on the interpretability of visual recognition models.
1 July 2026
Read appraisal →Journal of medical Internet research
A Futures Framework for Clinical AI Governance: Anticipating Emerging Risks, Shifting Roles, and Regulatory Challenges
This viewpoint develops the futures framework for clinical artificial intelligence governance (FF-CAIG), a conceptual and anticipatory framework for organizing emerging governance challenges in clinical artificial intelligence (AI). Although life cycle-oriented oversight is increasingly reflected in clinical AI regulation and institutional governance, existing approaches remain more developed for near-term validation, current-state assurance, and retrospective risk detection than for longer-horizon sociotechnical change. This gap is increasingly relevant, as AI systems become more complex, adaptive, and autonomous, and as they become more deeply embedded in care relationships and accountability structures. FF-CAIG is grounded in 3 futures methodologies: the 3 horizons model, scenario planning, and causal layered analysis. It is operationalized through an emerging clinical AI risk taxonomy that links these methods to governance domains. Its practical outputs include horizon classification, risk-domain mapping, scenario stress-testing findings, accountability-chain mapping, and horizon-scaled minimum governance actions for deployment or continued use. Applied across near-term, transitional, and longer-term horizons, the framework proposes cross-horizon priorities, including stronger predeployment equity evaluation, clearer life cycle accountability, clinician AI oversight competencies, and safeguards for increasingly autonomous or AI-mediated care systems. We illustrate FF-CAIG through 3 representative clinical AI deployment patterns and discuss its limitations, including differential compliance burdens, risks of overdocumentation, variable feasibility across jurisdictions, and the need for empirical validation. FF-CAIG is intended not as a prescriptive policy instrument or validated assessment tool, but as a structured analytic approach for regulators, health system leaders, developers, and researchers seeking prospective and systems-oriented approaches to clinical AI governance.
30 June 2026
Read appraisal →The Behavioral and brain sciences
Fair contracts for artificial intelligence?
This comment explores resource-rational contractualism (RRC) as a framework for artificial intelligence (AI) alignment and reasoning. It assesses RRC's capacity to capture the insights of philosophical contractualism and canvasses its extension to AI alignment. It suggests that a modified version of RRC, centered on impartial deliberation, can be used to identify principles for alignment and to guide efficient AI moral cognition.
27 June 2026
Read appraisal →Journal of medical Internet research
Vision-Based Artificial Intelligence Technologies for Epilepsy Monitoring: Scoping Review and Taxonomy Development Study
BACKGROUND: Artificial intelligence (AI) technologies for vision-based epilepsy monitoring are advancing rapidly in health care. Despite growing research using various video data sources and analytical approaches, no comprehensive framework exists to classify these technologies. OBJECTIVE: This scoping review aimed to develop and validate a taxonomy for AI technologies in vision-based epilepsy monitoring and to characterize visual AI approaches in epilepsy care. METHODS: Using an extended taxonomy development framework, we developed the taxonomy in 5 iterative cycles, drawing on theory and practice. We conducted a scoping review, market analysis, and applicability evaluation with market-ready solutions. We searched Scopus, Web of Science, and PubMed, including MeSH (Medical Subject Headings) terms; the final search was completed on January 16, 2026. We included primary studies from 2013 onward on AI-based or machine learning-based monitoring or prediction of epileptic seizures in humans using visual data. We excluded reviews, non-English publications, nonepilepsy studies, studies focused only on electroencephalography or wearables, animal studies, and pre-2013 publications. Evidence was charted through narrative and tabular synthesis and descriptive frequency analysis. In line with scoping review guidance, we did not conduct a meta-analysis or critical appraisal. To assess validity and practical relevance, 9 domain experts evaluated the taxonomy using a Delphi technique. RESULTS: We included 40 original studies. Study analysis yielded 16 dimensions, including data acquisition source, tracking target, image processing, classifier type, performance metrics, environment, seizure classification, data privacy, and user interface. Expert feedback added 4 further dimensions, including communication mode and information purpose. The final taxonomy comprises 23 dimensions with 102 characteristics. The review identified structural evidence gaps across settings, evaluation maturity, and reporting practices. Detection and classification in stationary settings predominated, whereas predictive approaches and real-time feedback were limited. Deep learning detection methods were common, but performance reporting was inconsistent, and patient-facing functionalities were limited. Privacy safeguards and standardized metrics were often incompletely reported, reducing comparability and maturity assessment. The taxonomy translates these patterns into guidance for benchmarking, procurement evaluation, user interface, and explainable AI design. We synthesized 5 main findings and 10 implications for research and practice. Key challenges concern standardization, seizure prediction, and real-time applicability. CONCLUSIONS: Vision-based AI technologies for epilepsy monitoring are still dominated by proof-of-concept and pilot evaluations, indicating a gap between technical feasibility and deployment-ready systems. This scoping review presents an implementation-oriented taxonomy integrating application context, system architecture, visual analysis, AI models, performance reporting, and feedback design into a single classification framework. Unlike prior work that mainly maps methods or data sources, the taxonomy provides a shared structure for consistent system-level characterization and comparison across studies and emerging solutions. It may support benchmarking, implementation-focused evaluation, procurement, and translation into clinical and home settings.
25 June 2026
Read appraisal →BMJ health & care informatics
Clarifying the relationship between biomedical and health informatics and digital health: expert perspectives.
The rise of digital health (DH) has sparked debates about its relationship with the long-established field of biomedical and health informatics (BMHI), with some questioning the need for new academic programmes in DH since much of the likely content is already covered in BMHI training programmes. At the International Medical Informatics Association panel, 'Shaping the Future of Biomedical and Health Informatics in the Digital Health Era' (MEDINFO 2025), only one-third of participants clearly distinguished the two fields, highlighting conceptual ambiguities and the influence of DH narratives. This article compiles the perspectives expressed by the panellists, all experienced BMHI scholars, presenting their distinct viewpoints on how the field can maintain its scientific identity while engaging with emerging DH initiatives.
24 June 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
AI agents are sensitive to nudges.
Large language models (LLMs) are increasingly deployed as autonomous agents that make choices and use tools on behalf of users. Yet, we have limited evidence about how their decisions are shaped by their environment. We adapt a human decision-making task to test leading LLMs under four forms of choice architecture: defaults, suggestions, information highlighting, and "optimal" nudges derived from a resource-rational model of human choice. We treat human behavior as a baseline for predictable sensitivity to such interventions. Across models and prompting strategies, LLMs often depart substantially from this baseline. They sometimes pay excessive costs to acquire information, sometimes ignore available information, and, most crucially, are far more responsive to nudges than humans, such that weak cues that slightly shift human behavior have larger effects on model choices, toward both better and worse payoff outcomes. Chain-of-thought prompting and in-context human data do not reliably stabilize behavior. Recent reasoning-optimized LLMs can, in some configurations, restore more human-level sensitivity to nudges, but do so inconsistently and at substantial computational cost. These results point to an important and largely neglected safety concern: LLM agents can be behaviorally brittle under subtle changes in choice architecture, even in the absence of adversarial settings.
23 June 2026
Read appraisal →Medical education online
Cross-cultural adaptation and psychometric evaluation of the artificial intelligence learning intention scale (AILIS) among medical sciences students: a methodological study
BACKGROUND: This study aimed to adapt and psychometrically evaluate the Artificial Intelligence Learning Intention Scale (AILIS) among Iranian medical sciences students. METHODS: This methodological psychometric study was conducted among 800 medical sciences students at Babol University of Medical Sciences, Iran. The AILIS was translated and culturally adapted into Persian using a standardized cross-cultural adaptation process. Construct validity was examined through exploratory factor analysis (EFA) and confirmatory factor analysis (CFA). Reliability was assessed using Cronbach's alpha, McDonald's omega, and the intraclass correlation coefficient (ICC). Convergent and discriminant validity were evaluated using composite reliability (CR) and average variance extracted (AVE). Measurement invariance across gender and structural stability were assessed using multi-group confirmatory factor analysis and exploratory graph analysis (EGA). RESULTS: EFA supported a four-factor structure consisting of Epistemic Capacity, Psychological Attitudes, Facilitating Environments, and Psychological and Behavioral Outcomes, explaining 44.55% of the total variance, with factor loadings greater than 0.40. Confirmatory factor analysis indicated an acceptable model fit, supporting the proposed factor structure (χ²/df = 3.42, RMSEA = 0.078, CFI = 0.932). Internal consistency and test-retest reliability were satisfactory, with Cronbach's alpha ranging from 0.854 to 0.904, McDonald's omega from 0.854 to 0.899, and ICC from 0.750 to 0.860. Composite reliability values were adequate, while AVE values indicated acceptable but partially limited evidence of convergent validity. Discriminant validity was generally supported. In addition, multi-group analysis demonstrated measurement invariance across gender, and exploratory graph analysis confirmed the structural stability of the four-factor model. CONCLUSION: The Persian AILIS is a valid and reliable instrument for assessing AI learning intention among medical sciences students. Beyond its psychometric robustness, the instrument may support curriculum planning, evaluation of AI-related educational interventions, and evidence-based decision-making in medical sciences education by providing a standardized measure of students' intentions to engage with AI learning.
20 June 2026
Read appraisal →International journal of medical informatics
Machine learning approaches to identifying neurodevelopmental disorders using social media data: a systematic review
BACKGROUND: Neurodevelopmental disorders (NDDs), such as autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD), often emerge in early childhood and can significantly impact social, academic, and emotional functioning. Early identification is critical to improving long-term outcomes, yet traditional diagnostic processes are time-consuming and resource-intensive. As social media becomes an integral part of daily life, user-generated content offers a novel source of behavioral and linguistic data that may support early detection. Advances in artificial intelligence (AI) and machine learning (ML) now make it possible to analyze these large-scale data streams for clinical insights. METHODS: A comprehensive search was performed across five databases-PubMed, Scopus, Web of Science, IEEE Xplore, and ACM Digital Library-from inception up to July 2025. Studies were included if they used AI or ML methods to analyze social media data for detecting NDDs. Data extraction focused on platform type, targeted disorder, dataset size, ML technique, and diagnostic performance. RESULTS: Nineteen studies met the inclusion criteria. Most focused on ASD and ADHD, using platforms such as Reddit, Twitter, YouTube, and Facebook. ML models achieved moderate to high classification performance, with F1-scores ranging from approximately 0.48 to 0.89 depending on the data type and disorder. Video-based models showed particular promise in identifying nonverbal behavioral markers. However, research on other NDDs remains limited, and methodological heterogeneity, small sample sizes, and ethical challenges persist. CONCLUSION: AI-driven analysis of social media data holds significant promise for scalable, non-invasive screening of neurodevelopmental disorders. While current work largely focuses on ASD and ADHD, future research should extend to underrepresented NDDs and address concerns related to data validity, bias, and privacy. With continued advancement, these tools may serve as valuable complements to traditional diagnostic methods.
15 June 2026
Read appraisal →International journal of medical informatics
Human in the loop artificial intelligence in healthcare: applications, outcomes, and implementation challenges
BACKGROUND: The integration of artificial intelligence in healthcare has transformed clinical practice and research methodologies. However, concerns regarding algorithmic accountability, interpretability, and safety have necessitated human oversight in AI systems. Human in the loop artificial intelligence represents a collaborative paradigm where human expertise and machine intelligence converge to enhance decision making while maintaining ethical standards and clinical safety. AIM: This review synthesizes current evidence on human in the loop AI in healthcare delivery and research, examining implementation frameworks, clinical outcomes, comparative advantages over fully automated and clinician-only approaches, and challenges. METHOD: A comprehensive narrative review was conducted using PubMed, Scopus, Web of Science, and IEEE Xplore databases covering studies from 2018 to 2025. Data were thematically synthesized to identify patterns, frameworks, and outcomes. This narrative approach enables comprehensive conceptual synthesis across diverse HITL-AI applications and contexts. RESULTS: Human in the loop AI demonstrates significant applications across diagnostic imaging, clinical decision support, patient monitoring, drug discovery, and research data analysis. Evidence indicates improved diagnostic accuracy, reduced medical errors, enhanced patient safety, and increased clinician trust compared to both automated AI and traditional approaches. Implementation requires EHR interoperability, clear liability frameworks, adaptive training protocols, and quantum-safe cryptographic security. Challenges include workflow integration, regulatory gaps for adaptive systems, and sustainability concerns. CONCLUSION: This review advances the field by synthesizing cross-domain implementation patterns, mapping collaboration models to risk-stratified contexts, identifying regulatory gaps for adaptive systems, and proposing future directions including post-quantum cryptographic integration, AI-driven adaptive architectures, and multi-center scalability frameworks for optimizing human-machine collaboration in healthcare.
15 June 2026
Read appraisal →Scientific reports
AIDx: a locally deployable AI system for physician clinical decision support
The dynamic environment of medicine, particularly in settings such as the Emergency Department, challenges physicians with an influx of patient data and the need for swift diagnosis and treatment under high-pressure conditions. Numerous AI models have been developed for medical applications, yet their integration into hospital systems remains limited because they often do not align with existing clinical workflows. To address these challenges, I developed AIDx, an AI-powered system designed to help physicians streamline clinical decision making, improve diagnostic support, and provide an integrated platform for AI-assisted analysis. Unlike many standalone systems, AIDx is designed for interoperability with electronic health record (EHR) systems and can leverage supplementary tools such as updated medical knowledge bases and auxiliary AI models. At its core, AIDx includes AIDx-Copilot, a large language model fine-tuned on de-identified EHR data and optionally grounded using retrieval-augmented generation (RAG) from open medical references. This study is a text-only benchmark assessment. I evaluated AIDx-Copilot using the MultiMedQA benchmark suite under a unified, deterministic, single-pass protocol (temperature=0; no voting). Across nine MultiMedQA subsets (MedQA, PubMedQA, MedMCQA, and medically focused MMLU subjects), AIDx-Copilot achieves a mean accuracy of 83.61% (SD 7.37). I report per-dataset Wilson 95% confidence intervals to quantify uncertainty from finite test set sizes. To isolate component contributions, I conducted ablation experiments comparing the base model, the fine-tuned model without retrieval, and the fine-tuned model with RAG enabled. EHR-based fine-tuning accounts for the majority of the performance gain (+17.8 percentage points over the base model on average), while RAG provides a modest additional benefit (+0.4 points on average) that varies by dataset. A qualitative error analysis of 200 incorrectly answered items identifies knowledge gaps (41.0%) and reasoning errors (38.0%) as the dominant failure modes. Comparative numbers for larger proprietary systems are provided only as context in the supplementary material because they may use different prompts and inference settings. The deployment configuration (quantized weights and on-premises serving) supports fast inference and local deployment. Under the quantized configuration, the system achieves a median latency of 0.84 seconds per query and fits within 28.1 GB of VRAM across two commodity GPUs. The primary results are based on public benchmark evaluations without RAG; I did not perform clinical, user-study, or real-world validations. In addition, while AIDx is designed to support fully local operation, specific deployments may choose to use third-party embedding or vector-search services; this paper documents the data-flow boundary and a fully local configuration path.
14 June 2026
Read appraisal →Blood advances
Accurate identification of sickle cell disease cases in a large genotyped cohort using electronic health record data
The US Food and Drug Administration recently approved sickle cell disease (SCD) gene therapies that require a contemporaneous cohort to compare their outcomes with those of individuals who did not receive gene therapy. Previously, we developed an automated algorithm to create a contemporaneous cohort of children and adults with SCD using electronic health record data. We tested the hypothesis that the updated Vanderbilt University Medical Center (VUMC) algorithm can identify individuals with SCD with sensitivity and specificity of >95%. In the VUMC health care system, we identified a cohort of 33 141 adults and children of primarily African ancestry who underwent β-globin gene sequencing, with a 1.4% prevalence of SCD (473/33 141). For SCD, the performance of the VUMC algorithm using International Classification of Diseases (ICD) and laboratory data showed a sensitivity of 97.6% (462/473), specificity of 99.9% (32 663/32 668), positive predictive value of 98.9% (462/467), and negative predictive value of 99.9% (32 663/32 674). Three SCD phenotypes account for 98.5% of the individuals in the cohort. The cohort had sensitivities of 96.8% (272/281), 93.7% (136/145), and 85% (34/40) for hemoglobin SS (HbSS)/Hb Sβ0-thalassemia (HbSβ0), HbSC, and HbSβ+, respectively. The cohort had a specificity of 99.9% across all 3 phenotypes: HbSS/HbSβ0, HbSC, and HbSβ+, respectively. We have developed an updated algorithm using ICD and laboratory data that accurately identifies individuals with SCD within a large health care system, and distinguishes HbSS/HbSβ0, HbSC, and HbSβ+.
9 June 2026
Read appraisal →Bioorganic chemistry
Multi-relational knowledge graph for drug-drug interaction prediction via dual aggregation and collaborative optimization
Identifying potential drug-drug interactions is crucial in clinical care and new drug development, as mutual interference between drugs can lead to adverse reactions. Recently, computational methods have been widely employed for predicting DDIs. However, these approaches often encounter high computational complexity during the process of node feature representation and frequently overlook the interactivity and polysemy of entities. Effectively handling heterogeneous information remains a significant challenge. To cope with these limitations, we propose a dual aggregation and collaborative optimization learning framework，named MRACO，for DDIs prediction based on a multi-relational knowledge graph. MRACO utilizes structural information from these graphs, efficiently learning interactivity information across various relationship types. Additionally, MRACO employs dual aggregation operations to explicitly encode and aggregate multi-type information, capturing diverse semantic relationships of drug nodes. Moreover, MRACO simplifies complex computational processes by utilizing the collaborative loss optimization function, eliminating redundant information and enhancing model robustness. By integrating deep-level interactive information, MRACO leverages rich contextual data from knowledge graphs, enhancing prediction stability. Experiments demonstrate the effective of MRACO feature extraction capabilities in multi-relational networks, highlighting its understanding of the underlying mechanisms of DDIs.
7 June 2026
Read appraisal →Biomedical physics & engineering express
Stabilizing neuromorphic ECG processors via adaptive fractional fatigue
Spiking neural networks (SNNs) deployed on wearable devices can exhibit runaway firing when processing noisy electrocardiogram (ECG) signals, increasing switching activity and dynamic power and destabilising power regulation. To suppress redundant firing while preserving ECG classification performance, we introduce the adaptive fractional fatigue (AFF) SNN, which equips each hidden neuron with a power-law weighted memory buffer of its past membrane states, thereby forming a fractionally-inspired activity trace that drives an adaptive threshold: as the accumulated trace rises with sustained firing, the effective threshold increases and suppresses further spikes, implementing an intrinsic negative feedback loop. Ten independent seeds were evaluated with a fixed random train/test split of2969/743heartbeat windows (lengthT=180;3712balanced windows total;1856normal,1856arrhythmia) drawn from three publicly available Massachusetts Institute of Technology-Beth Israel Hospital records, with spikes per sample as the primary energy proxy and test accuracy and the area under the receiver operating characteristic curve as secondary endpoints. Spikes/sample decreased from2384.990±537.719(integer-order) to891.481±227.654(AFF), a62.62%reduction (bootstrap 95% confidence interval[54.00%,69.35%]), which was statistically significant (pairedt-testp=0.000015two-sided; Wilcoxon signed-rankp=0.001953two-sided; paired effect sizedz=-2.662). Accuracy was preserved (78.923±1.725%vs79.206±1.541%; pairedt-testp=0.227761), and cross-seed spike variance decreased by a factor of5.58×. AFF suppresses noise-driven hyperactivity and stabilises event traffic, delivering a statistically significant reduction in spikes/sample without accuracy loss and improving predictability for low-power neuromorphic ECG deployment.
30 May 2026
Read appraisal →American Society of Clinical Oncology educational book. American Society of Clinical Oncology. Annual Meeting
Artificial Intelligence in Oncology: Clinical Applications, Challenges, and Opportunities
Artificial intelligence (AI) is reshaping cancer research and clinical oncology by enabling large-scale analysis of complex biomedical data. Although early AI efforts focused on single-modality tasks such as imaging interpretation or pathology classification, contemporary oncology increasingly requires integrated reasoning across heterogeneous inputs, including imaging, digital pathology, and genomics. Multimodal artificial intelligence (MMAI) models represent a major evolution by fusing these data types into unified predictive systems capable of informing diagnosis, risk stratification, treatment selection, and disease monitoring. This manuscript reviews the state of AI in oncology and highlights emerging foundation models within the field. There will be a focus on histopathology, and how this can be incorporated into risk assessment models. There will also be examples of MMAI utility in radiology and spatial proteomics. Additionally, we will discuss ethical and legal implications of MMAI in modern oncology, and recommendations for responsible application within clinical oncology practice.
30 May 2026
Read appraisal →Scientific reports
Automated RECIST tumor response classification through prompt-guided large language models
This study investigates whether an entirely offline, general-purpose large language model (LLM) can reliably automate the classification of routine radiology reports according to RECIST (Response Evaluation Criteria in Solid Tumors) guidelines, focusing on how different prompting strategies support accurate, privacy-preserving tumor response assessment without additional model fine-tuning. An offline, in-house implementation of LLaMA-3.3 (70B) was used to classify real-world CT imaging reports from oncology patients. Reports were authored following RECIST-structured reporting but had outcome labels programmatically withheld prior to processing. Three prompting strategies-zero-shot, few-shot, and chain-of-thought prompting-were tested to guide the model in assigning RECIST categories: Baseline (BL), Complete Response (CR), Partial Response (PR), Stable Disease (SD), and Progressive Disease (PD). Model outputs were benchmarked against original expert labels using accuracy, precision, recall, and F1 scores. Across all tested prompting strategies, the LLaMA-3.3 model achieved strong classification performance. The best results were obtained with chain-of-thought prompting, reaching micro F1 scores of 0.81 across all RECIST categories. Overall model predictions aligned well with human expert assessments. Operating entirely offline within hospital infrastructure, the system preserved full compliance with stringent data privacy requirements. Prompt-driven large language models can accurately classify tumor response categories from real-world radiology reports in a scalable, reproducible, and privacy-preserving manner. Offline LLM deployment, combined with optimized prompting strategies, offers a promising approach for automating structured oncology report interpretation, potentially enhancing consistency and efficiency in clinical decision support workflows.
29 May 2026
Read appraisal →BMJ health & care informatics
Machine learning-based prediction of a high-risk kidney function trajectory class after acute kidney injury
OBJECTIVES: Monitoring kidney function after acute kidney injury (AKI) hospitalisation is essential for identifying patients at risk of rapid progression. This study developed a machine learning (ML) model to predict membership in a high-risk longitudinal estimated glomerular filtration rate (eGFR) trajectory associated with subsequent kidney replacement therapy (KRT) initiation. METHODS: We conducted a cohort study of 88 632 adults with ≥2 post-discharge eGFR measurements between 2010 and 2017. Joint latent class mixed models were used to identify dynamic eGFR trajectories from 665 499 measurements. Patient-level data splitting (80/20) was strictly enforced. Extreme gradient boosting (XGBoost) and random forest models were trained using clinical features within 3 months after discharge to predict membership in the highest-risk eGFR trajectory class. RESULTS: Three eGFR trajectories were identified: Steady Low (high-risk phenotype), Slow Decline and Recovery. The Steady Low group showed the strongest association with KRT. The 10-variable XGBoost model demonstrated excellent discrimination (AUROC 0.974, sensitivity 0.913, specificity 0.919). The eGFR slope from discharge to 3 months was the most influential predictor of high-risk progression. DISCUSSION: Although the study was limited to a single healthcare system, this ML framework identified patients at risk of rapid kidney function decline using routinely available clinical data. The model supports early risk stratification after AKI and may enable more timely kidney-protective interventions. External validation, calibration assessment and benchmarking against simpler eGFR-based rules remain important next steps. CONCLUSION: A 10-variable XGBoost model accurately identifies patients with deteriorative eGFR trajectories (Steady Low) early after discharge, providing a practical framework for personalised post-AKI kidney care.
29 May 2026
Read appraisal →Journal of medical Internet research
Adaptive Fast-Slow Large Language Model Framework for Multidimensional Classification of Prenatal Ultrasound Reports: Comparative Study
BACKGROUND: Phenotype-driven prenatal diagnosis relies on the precise correlation between ultrasound findings and genetic outcomes; however, this process is hindered by the unstructured nature of clinical ultrasound reports. While large language models (LLMs) hold the potential to address this challenge, their specific application in this domain remains systematically underexplored. OBJECTIVE: To establish an effective LLM implementation framework for the clinical multidimensional classification of prenatal ultrasound reports, we evaluated the open-source DeepSeek-V3.2 family on real-world anomalous reports-covering both factual and subjective categories-while integrating retrieval-augmented generation (RAG) and chain-of-thought (CoT) reasoning. METHODS: From a cohort of 4256 pregnancies, we extracted 254 reports with fetal anomalies. We comprehensively evaluated both the high-speed base model (DeepSeek-V3.2-B) and the reasoning-enhanced model (DeepSeek-V3.2-R) across all 5 classification dimensions, comprising 4 factual extraction tasks-primary classification, standardized terminology, anatomical system, and abnormality count-and 1 subjective severity assessment. We further explicitly evaluated the efficacy of RAG for the subjective tasks. Finally, to validate the clinical utility of this approach, we performed a correlation analysis between the expert-validated multidimensional phenotypic profiles and definitive genetic outcomes derived from amniocentesis. RESULTS: While V3.2-B achieved high efficiency in factual tasks (accuracy and F1-score >90%), it underperformed in subjective severity grading (56.6% accuracy), exhibiting a recall of 0 for minor anomalies. Crucially, while RAG significantly improved both models' performance on internal retrieval datasets (P<.05), this benefit did not generalize to external test datasets (P>.05). In contrast, the V3.2-R model utilizing CoT reasoning achieved superior robustness (86% accuracy and F1-score=0.75) on external data without RAG; notably, introducing RAG to V3.2-R degraded performance to 81%, suggesting potential noise interference. Clinical validation against amniocentesis outcomes confirmed that accurate multidimensional phenotypic profiles significantly stratified pathogenic genetic risks. CONCLUSIONS: The rapid base models are efficient for factual classification, and RAG enhances performance on data similar to the knowledge base, whereas CoT is indispensable for subjective assessment. Within the constraints of our dataset and current retrieval implementation, CoT proved more robust than RAG for subjective assessment. However, this finding is specifically tied to our experimental setup and should not be generalized as a universal conclusion. We recommend clinically adopting this adaptive "fast-slow" LLM framework to efficiently perform the multidimensional classification of prenatal ultrasound anomalies. This privacy-preserving, locally deployable solution provides a scalable path to accelerate phenotype-genotype research and optimize invasive diagnostic decision-making.
29 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
Reproducible workflow for online artificial intelligence in digital health
Online artificial intelligence (AI) algorithms are an important component of digital health interventions. These online algorithms are designed to continually learn and improve their performance as streaming data are collected on individuals. Deploying online AI presents a key challenge: balancing adaptability of online AI with reproducibility. Online AI in digital interventions is a rapidly evolving area, driven by advances in algorithms, sensors, software and devices. Digital health intervention development and deployment is a continuous process, where implementation-including the AI decision-making algorithm-is interspersed with cycles of re-development and optimization. Each deployment informs the next, making iterative deployment a defining characteristic of this field. This iterative nature underscores the importance of reproducibility: data collected across deployments must be accurately stored to have scientific utility, algorithm behaviour must be auditable and results must be comparable over time to facilitate scientific discovery and trustworthy refinement. This article proposes a reproducible scientific workflow for developing, deploying and analysing online AI decision-making algorithms in digital health interventions. Grounded in practical experience from multiple real-world deployments, this workflow addresses key challenges to reproducibility across all phases of the online AI algorithm development life cycle. This article is part of the theme issue 'Statistical workflow'.
29 May 2026
Read appraisal →Journal of medical Internet research
Large Language Model-Generated Patient Instructions for Prescriptions in Primary Health Care: Preclinical Algorithm Validation
BACKGROUND: The application of generative artificial intelligence to simplify medication use instructions has the potential to enhance people's health by improving treatment adherence. OBJECTIVE: We evaluated the performance of large language models (LLMs) in generating medication usage instructions to complement prescriptions in primary health care. METHODS: This randomized, blinded experimental preclinical study used prescription-inducing scenarios, assigned to 62 health care professionals, to validate instructions generated by LLMs during electronic prescriptions. The instructions were generated by ChatGPT-4.0 (OpenAI), Llama3.1-8B (Meta), and Llama3.1-8B-RAG (Meta) using retrieval-augmented generation based on patient information leaflets. Performance metrics assessed adequacy, completeness, clarity, language simplification, usefulness, and errors in the generated instructions, with scores to analyze overall and individual metrics. RESULTS: The 3 models yielded high overall scores for producing qualified instructions (ChatGPT-4.0: median 88.4, IQR 22.8; Llama3.1-8B: median 66.5, IQR 50.9; Llama3.1-8B-RAG: median 79.9, IQR 34.4; Kruskal-Wallis test P=.003). Llama3.1-8B-RAG received evaluations with similar overall scores to ChatGPT-4.0 (post hoc test, P=.05) and similar to Llama3.1-8B (post hoc test, P=.44). ChatGPT-4.0 outperformed Llama3.1-8B (Bonferroni test, P<.001). Regarding specific domains, Llama3.1-8B-RAG received scores equivalent to those of ChatGPT-4.0 for adequacy (mean 6.24, SD 2.3 vs mean 6.82, SD 2.1; post hoc test, P=.54); completeness (mean 5.94, SD 2.2 vs 6.55, SD 1.9; post hoc test P=.38), clarity (mean 5.77, SD 2.4 vs mean 6.68, SD 1.9; post hoc test P=.09), and usefulness (mean 5.42, SD 2.4 vs mean 5.96, SD 2.2; post hoc test P=.63). ChatGPT-4.0 received higher scores in the language simplification criterion than Llama3.1-8B-RAG (mean 7.05, SD 1.5 vs mean 5.44, SD 2.6; post hoc test P<.001). Interrater variability in assigning scores ranged from 4.2% (n=3) to 85.8% (n=6) among primary health care professionals. Instructions leading to incorrect use of the medication had similar frequency among the models(ChatGPT-4.0: n=15, 22.7%; Llama3.1-8B: n=19, 22.8%; Llama3.1-8B-RAG: n=19, 22.8%; chi-square test P=.71). The frequencies of hallucination were similar (ChatGPT-4.0: n=7, 10.6%; Llama3.1-8B: n=9, 13.6%; Llama3.1-8B-RAG: n=6, 9.1%; chi-square test P=.67). CONCLUSIONS: The open-source LLM enhanced with external information presented similar performance to the closed-source model, except for ChatGPT4.0, which was superior in language simplification of messages. LLM generation demonstrated potential for instructing patients on medication use. Nonetheless, the introduction of this innovation into the electronic prescribing workflow demands prescriber validation for human oversight of the technology and requires a strategy for LLM performance governance. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): RR2-10.12688/verixiv.1359.1.
27 May 2026
Read appraisal →Critical care explorations
External Validation, Recalibration, and Extension of a Prediction Model of Early Acute Kidney Injury in Critically Ill Children Using Multicenter Data
BACKGROUND: Acute kidney injury (AKI) is common in critically ill children and is associated with high morbidity and mortality. Risk prediction models designed for clinical decision support implementation can facilitate early identification and proactive mitigation of AKI risk. Existing models have primarily been validated using single-center data, partly because of the lack of appropriately detailed multicenter datasets. OBJECTIVE: To determine the performance of a single-center model to predict new AKI at 72 hours of ICU admission in children across two multicenter datasets and refine this model to improve prediction performance while maintaining acceptable alert burden. DERIVATION AND VALIDATION COHORTS: We analyzed two datasets: the Pediatric Learning Health System Network-Virtual Pediatric Systems (PEDSNET-VPS) dataset, created through the linkage of PEDSnet electronic health record (EHR) extraction with VPS (LLC, http://www.myvps.org), and the PICU Data Collaborative dataset, created through EHR extraction and harmonization from eight participating institutions. We divided each dataset into a derivation and test split. PREDICTION MODEL: We first recalibrated an existing single-center model and measured discrimination (area under the receiver operating characteristic curve [AUROC] and area under the precision-recall curve [AUPRC]) and performance at multiple cutpoints. We next added features available at 12 hours of ICU admission, optimizing by precision and recall. We measured discrimination and performance at multiple cutpoints and identified the features contributing most to the risk score. RESULTS: In total we analyzed 186,540 ICU admissions. We found early AKI by serum creatinine criteria within 72 hours of admission in 2.2-2.7%. Initial recalibration of an existing single-center model demonstrated poor discrimination (AUROC 0.65-0.78; AUPRC 0.10-0.12). Following the addition of new features, the model had higher AUROC (0.80-0.88) and AUPRC (0.13-0.22). CONCLUSIONS: In this first use of two new multicenter datasets, we found improved performance in a model designed using features available at 12 hours of ICU admission, balancing sensitivity and precision to predict patients at risk for AKI development.
26 May 2026
Read appraisal →PloS one
A hybrid BiLSTM and rule-based system for integrated diabetes prediction and personalized guidance
Diabetes is a common chronic disease that needs early diagnosis and proper management to avoid severe complications. While current Artificial Intelligence (AI) tools generate predictive information, they often lack an integrated element for post-diagnosis support in order to fill in this critical gap in patient self-management. This research proposes and validates a hybrid System which aims to bridge this gap. The methodology is based on a novel, fused dataset (PIMA and Type 2 Diabetes) that was carefully preprocessed following a leakage-safe protocol in order to increase generalizability. The system architecture is a combination of two different critical components: Strong Bidirectional Long Short term Memory (BiLSTM) model for prediction and rule based engine for creating personalized lifestyle recommendations. In order to validate the efficacy of the BiLSTM model, seven traditional machine learning (ML) models and standard deep learning (DL) models have been comparatively tested, in which BiLSTM model has demonstrated a better generalization and prediction performance. Rigorous 10-fold cross validation was used to validate the system, which came up with an accuracy of 84.02%, precision of 87.89%, and recall of 80.50%. This research concludes that by successfully combining the high-performance predictive engine and real-time guidance module, it is possible to develop a holistic clinically relevant tool to close the loop between diagnosis and proactive self-management.
25 May 2026
Read appraisal →Studies in health technology and informatics
Interoperability Skills for Digital Health: A Reference Model from the XiA Project
As healthcare systems advance through digital transformation, achieving interoperability across Health Information Systems (HIS) relies on skilled professionals who enable effective integration and data exchange. Under the scope of the Xpanding Innovative Alliance (XiA) project, this article introduces a Reference Model for Interoperability Skills that defines the core skills needed by professionals to support interoperable digital health ecosystems. Through a systematic literature review, the model provides a structured view of competencies across professional profiles, grounded in international best practices. The proposed model serves as a strategic tool for workforce development, curriculum design, and organizational capability assessment. By promoting a shared understanding of interoperability, it supports the creation of a skilled, adaptable, and collaborative workforce capable of advancing integrated, data-driven healthcare across Europe.
24 May 2026
Read appraisal →JMIR human factors
Emotion-Adaptive Large Language Model-Driven Clinical Decision Support: User Evaluation of the Empathic Clinical Decision Support System Framework for Trust and Explainability
BACKGROUND: The increasing prevalence of cannabis use has motivated researchers to develop computational behavioral models that predict usage patterns and related health impacts in naturalistic environments. However, the opaque nature of many artificial intelligence (AI) systems limits users' ability to interpret outputs and undermines trust. Existing explainable artificial intelligence techniques often remain overly technical and do not account for the confusion or frustration clinicians may experience when interpreting complex model explanations. OBJECTIVE: We propose and evaluate an empathic clinical decision support system (empathic-CDSS) that integrates large language models (LLMs) with real-time emotion recognition and explainability modules. The goal is to provide transparent, adaptive, and emotionally attuned explanations that enhance interpretability, user confidence, and trust in AI-assisted clinical decision-making. METHODS: Our empathic-CDSS integrates explainable artificial intelligence, causal inference, and affective computing within an interactive LLM-driven framework. Users' affective states were inferred using the circumplex model of affect, which characterizes emotions along 2 dimensions: valence (the degree of pleasure or displeasure) and arousal (the level of physiological activation or energy). Based on the captured emotional signals, the system dynamically adjusts the tone, structure, style, and complexity of its explanations generated by a fine-tuned LLM, enabling personalized and plain-language explanations. A total of 33 participants with diverse medical and technical backgrounds engaged with the system through guided evaluation tasks and postsession assessments to evaluate 6 dimensions of user-centered questions: usability, personalization and relevance of insights, clarity and comprehensibility, system benefits, satisfaction, and trust and reliability. RESULTS: Our empathic-CDSS effectively generated personalized and transparent explanations that revealed the causal reasoning behind model predictions while enhancing users' emotional engagement and trust in the system's decision logic. Continuous, affect-based feedback enabled the system to dynamically adapt explanation delivery to individual user needs. Participants reported significantly improved usability, clarity, satisfaction, and trust compared with a baseline clinical decision support system without emotion adaptation. Improvements were also observed across additional evaluation dimensions, including personalization and perceived system benefits, supporting the feasibility and added value of integrating empathy-aware communication into AI-driven clinical decision support. CONCLUSIONS: This study introduces a transparent, trustworthy, and emotionally adaptive framework for AI-assisted prediction and clinical decision support. By uniting causal reasoning, affective sensing, and LLM-based natural-language explanations, the empathic-CDSS offers a novel direction for developing emotionally intelligent and user-centered AI systems, with potential applications in behavioral monitoring and personalized interventions, including cannabis use and related health domains.
23 May 2026
Read appraisal →BMJ open
Impact of COVID-19-related data drift on machine-learning prognostic models predicting 30-day opioid-related emergency department visits, hospitalisation or mortality: a population-level administrative data study in Alberta, Canada
OBJECTIVE: To develop machine-learning (ML) models during the COVID-19 pandemic and adjacent time periods to evaluate the impact of data drift on model performance. DESIGN: This prognostic study used population-level administrative health data to develop ML prediction models. SETTING: Alberta, Canada during 2019-2023. PARTICIPANTS: All patients over 18 who received at least one opioid dispensation from a community pharmacy within the province of Alberta between 2019-2023. EXPOSURE: Each opioid dispensation served as the unit-of-analysis. MAIN OUTCOMES/MEASURES: Opioid-related outcomes were identified from linked health administrative datasets. Light Gradient Boosting-machine models were developed on pre-pandemic, pandemic and endemic data and temporally validated on 2023 data (pre-pandemic model was also validated on 2020-2021 data) to predict the risk of emergency department visit, hospitalisation or mortality within 30-days of an opioid dispensation. We described key feature distributions across the study time period and changes in model prediction performance on the validation sets using relevant metrics. RESULTS: Among 1.2 million study participants representing over 13 million opioid dispensations, there were 59 809 (2.1%), 134 402 (2.4%) and 62 143 (2.3%) events reported in the pre-pandemic (2019), pandemic (2020 and 2021) and endemic (2022) time periods, respectively (estimated 2023 validation set pre-test probability of 2.8%). Notable differences in key features were observed in the 2020-2021 model relative to other years. In the 2023 validation set, discrimination performance was highest for the pre-pandemic and endemic models compared with the pandemic model (0.81, 0.83, 0.74, respectively). A similar trend regarding changes from pre-test to post-test probabilities in higher categories of predicted risk (23%, 40%, 16%) was observed. 2020-2021 had the lowest discrimination performance (0.71) and uninformative post-test probabilities (<10%). CONCLUSION: COVID-19 pandemic health data contributed to significant ML drift. Although ML approaches allow for quick re-training to mitigate drift, health regulators should approach ML prediction with caution when using pandemic-times data.
23 May 2026
Read appraisal →Journal of medical Internet research
Benchmarking Large Language Models and Prompt Engineering Strategies in Microsatellite Instability Cancers: Evaluation Study
BACKGROUND: The reliability of general-purpose large language models (LLMs) for complex clinical tasks in specialized domains such as microsatellite instability (MSI) cancers remains critically uncharacterized. The absence of a domain-specific benchmark to evaluate and guide the optimization of their capabilities across diverse clinical tasks poses unevaluated risks to patient safety. OBJECTIVE: This study aimed to develop and validate Microsatellite Instability Cancer Benchmark (MSIC-Bench), a novel, two-tiered benchmark for MSI cancer, covering both consensus and frontier knowledge. Using this framework, we aimed to systematically assess LLM performance across various prompting strategies, identify task-specific weaknesses, and reveal effective pathways for performance improvement. METHODS: We developed MSIC-Bench, a 511-question benchmark derived from clinical guidelines and a curated knowledge base. Three state-of-the-art LLMs (GPT-4o, [OpenAI], Gemini 2.5 Pro [Google], and Claude Opus 4 [Anthropic]) were evaluated using 4 prompting strategies, including vanilla, chain-of-thought, reflection of thoughts, and retrieval-augmented generation (RAG), under both multiple-choice and open-ended modalities. Performance was assessed on accuracy, safety (honesty), error composition, and token usage. RESULTS: LLMs demonstrated a significant "scaffolding effect," with accuracy dropping substantially in open-ended scenarios. For non-RAG strategies, the primary failure mode was an internal knowledge deficit. The integration of RAG proved to be the most effective intervention. A domain-aligned RAG strategy not only significantly improved accuracy in complex decision-making tasks but also fundamentally shifted the system's primary bottleneck from knowledge deficits to retrieval failures. In terms of safety, RAG induced a favorable shift from high-risk fabrications to safer refusals, though this introduced a safety-utility trade-off in the form of false refusals. Notably, our hybrid-RAG configuration, which combined both knowledge sources, demonstrated the most robust and generalizable performance across all tasks. CONCLUSIONS: Current LLMs lack the specialized knowledge required for reliable application in MSI oncology. A well-designed RAG architecture is the pivotal intervention to address this gap. However, its success is not automatic; it transforms the nature of system failure, making retrieval precision and knowledge base quality the new critical determinants of performance and safety. Our findings establish a clear directive for developing trustworthy clinical artificial intelligence: focus must shift toward optimizing the retrieval component and curating high-quality, comprehensive knowledge sources. MSIC-Bench provides a robust framework to guide these future efforts.
22 May 2026
Read appraisal →Journal of medical Internet research
Large Language Models in Colorectal Cancer Care and Clinical Decision Support: Systematic Review
BACKGROUND: Colorectal cancer (CRC) is a leading cause of cancer morbidity and mortality worldwide. The complexity of guideline-concordant care and unstructured clinical data has driven demand for decision-support tools. Large language models (LLMs) show promise for processing clinical data and patient-provider communication, yet evidence is fragmented, and a CRC-specific synthesis across the full care continuum is lacking. OBJECTIVE: This systematic review evaluates the current applications, performance determinants, and clinical implications of LLMs across the continuum of CRC care. METHODS: Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses), we searched 6 databases (PubMed, Embase, Web of Science, Scopus, CINAHL, Cochrane) through April 1, 2026. Eligible studies were peer-reviewed original investigations of LLMs on CRC tasks with extractable outcomes; reviews, editorials, and abstracts were excluded. Two reviewers assessed quality with QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies-2), PROBAST (prediction model risk of bias assessment tool), and ROBINS-I (Risk of Bias in Nonrandomized Studies - of Interventions). Data on model types, applications, prompts, input/output formats, and outcomes were analyzed descriptively, with narrative synthesis per synthesis without meta-analysis (SWiM) guidelines. RESULTS: Of 8880 records, 37 studies met inclusion criteria (2023-2026), mostly from China and the United States, with GPT series most frequently evaluated. Overall risk of bias was low in 10/37 studies (27.0%), moderate in 14/37 (37.8%), unclear in 7/37 (18.9%), and high or serious in 6/37 (16.2%). Problematic domains included outcome measurement, intervention classification, patient selection, and lack of blinded assessment. LLMs showed utility in automating data extraction from clinical texts, supporting patient education, aiding diagnosis, and assisting clinical decision-making, with emerging visual interpretation and multimodal capacities. Domain-specific and multimodal models showed advantages over general-purpose models in certain tasks. Performance was significantly influenced by prompt design, from zero-shot queries to fine-tuning. Despite efficiency and outcome benefits, challenges persist regarding methodological quality, data privacy, and generalizability. CONCLUSIONS: This review provides an integrative framework synthesizing evidence across study designs and LLM categories in CRC care. Unlike prior reviews addressing gastroenterology broadly or limited to one design, it covers the full CRC continuum and, for the first time, comparatively evaluates general-purpose, domain-specific, and multimodal LLMs, clarifying how prompt engineering and heterogeneous metrics shape outcomes. Although findings support LLMs' clinical potential, results must be interpreted cautiously, given low overall evidence quality. Most studies lacked safeguards against bias-blinded assessment, confounder adjustment, or prospective multicenter validation. Substantial heterogeneity across tasks, LLM types, prompts, reference standards, and outcomes means reported advantages cannot be generalized. Future work should prioritize real-world integration via prospective multicenter validation, robust privacy frameworks, and rigorous human oversight. Amid rising global CRC burden and health care disparities, this review informs clinical translation, equitable scaling, and policy on LLM deployment.
22 May 2026
Read appraisal →JMIR mental health
Automated Safety Testing and Reporting Application for Conversational Safety Monitoring of Generative AI Tools for Mental Health: Development and Validation Study
BACKGROUND: Artificial intelligence (AI)-based conversational tools are rapidly expanding within mental health care as a means of increasing access and scalability. At the same time, these systems introduce distinct safety risks arising from both user disclosures (eg, self-harm ideation) and inappropriate or inadequate AI responses. OBJECTIVE: This study aimed to develop and evaluate the Automated Safety Testing and Reporting Application (ASTRA), an external system intended to identify clinically relevant risk behaviors across entire AI-mediated mental health conversations. METHODS: ASTRA was tested on a dataset of 100 synthetic therapeutic conversations written by licensed clinicians to reflect risk behaviors and harmful responses between users and AI tools. Conversations varied in length and included both subtle and overt risk behavior examples across 8 predefined categories. Human coder consensus ratings served as the reference standard. ASTRA's classifications were evaluated across 2 prompt iterations using standard diagnostic performance metrics and agreement statistics. RESULTS: ASTRA demonstrated consistently high concordance with expert human ratings across all categories. Accuracy exceeded 0.90 for all risk behavior categories examined, with specificity uniformly high and sensitivity varying by category (range 0.55-1.00). Agreement beyond chance was substantial to almost perfect between ASTRA and human raters (κ=0.65-1.00). Detection of user self-harm indicators was particularly accurate, even in conversations where risk was expressed subtly. CONCLUSIONS: In this initial validation study, ASTRA reliably identified multiple forms of mental health-related risk behaviors at the conversation level. These findings support the feasibility of independent safety monitoring systems as a complement to AI tools used in mental health contexts and underscore the need for further evaluation using larger and real-world datasets.
21 May 2026
Read appraisal →Renal failure
Machine learning-based risk prediction model for postoperative acute kidney injury in surgical patients with chronic kidney disease (CKD): development, validation, and SHAP-based explainability
This study aimed to develop machine learning models to predict postoperative acute kidney injury (AKI) in surgical patients with pre-existing chronic kidney disease (CKD), a serious complication associated with high morbidity and accelerated disease progression. Using data from 3,851 surgical patients with CKD from the MIMIC-IV database, nine supervised machine learning models were developed to predict AKI, as defined by the KDIGO criteria. Postoperative AKI occurred in 24.2% (932/3,851) of the patients. The XGBoost model demonstrated the best predictive performance, achieving an AUC of 0.844, accuracy of 0.791, sensitivity of 0.79, and F1-score of 0.73, followed by Random Forest and LightGBM. The calibration curve showed a good agreement between the predicted and observed risks. For interpretability, SHAP analysis was used to identify SOFA score, estimated glomerular filtration rate (eGFR), systolic blood pressure, albumin, and phosphorus as the most influential predictors. In conclusion, machine learning models, particularly XGBoost, provide accurate and interpretable predictions of postoperative AKI in this high-risk population. The integration of explainable AI facilitates transparent risk stratification, which may support personalized perioperative care and the development of real-time clinical decision support systems.
21 May 2026
Read appraisal →JMIR medical informatics
Machine Learning Prediction Model for Dyslipidemia and Its Association With Atherothrombotic Events in 3 Independent Cohorts From South Korea, Japan, and the United Kingdom: Algorithm Development and Validation Study
BACKGROUND: Dyslipidemia is a multifactorial and complex condition that warrants investigation through advanced analytical approaches such as machine learning (ML). Few previous ML studies predicting dyslipidemia have been validated across multiple international populations. OBJECTIVE: This study aimed to develop an ML model to predict the 5-year incidence of dyslipidemia using routinely collected health examination data. To ensure generalizability, the model was externally validated in populations from South Korea, Japan, and the United Kingdom. Furthermore, the clinical relevance of the model-derived risk was evaluated by examining its association with atherosclerotic outcomes, including acute myocardial infarction and cerebral infarction. METHODS: This study was conducted using 3 independent, large-scale, population-based cohorts. The discovery cohort from South Korea (n=471,650) was used for model training and internal validation, while 2 validation cohorts from Japan (validation A; n=7,255,685) and the United Kingdom (validation B; n=408,725) were used for external validation. We evaluated various ML-based models using 23 features extracted from regular health screening data to predict the new onset of dyslipidemia within 5 years. Shapley Additive Explanations values were calculated to assess feature importance. To ensure the robustness of the proposed ML model, we evaluated the risk of atherothrombotic events (acute myocardial infarction or cerebral infarction) based on the model probability (tertiles; T1, T2, and T3) using a Cox proportional hazards model. RESULTS: In the discovery cohort, soft-voting ensemble learning with Light Gradient Boosting Machine and categorical boosting exhibited performance metrics of area under the receiver operating characteristic curve (AUROC) of 0.783, precision of 37.9%, and area under the precision-recall curve of 0.469. The model showed moderate discriminatory performance in the external validation cohorts (cohort A: AUROC 0.744; precision 27.2%; and cohort B: AUROC 0.687; precision 5.07%). Shapley Additive Explanations value analysis identified smoking, alcohol intake, and physical activity as the most important features for predicting dyslipidemia. Finally, a higher model probability (T3 vs reference) was pronounced with an increased risk of acute myocardial infarction (adjusted hazard ratio 2.34, 95% CI 1.84-2.97) and cerebral infarction (adjusted hazard ratio 2.43, 95% CI 2.19-2.71). CONCLUSIONS: This multinational study developed and validated an ML-based model using routine health checkup data to predict the 5-year risk of new-onset dyslipidemia, which was also associated with atherosclerotic events.
21 May 2026
Read appraisal →Scientific reports
Quantum machine learning for predicting anastomotic leak: a clinical study
Anastomotic leak is a life-threatening complication following colorectal surgery. This study benchmarks Quantum Neural Networks (QNNs) against hyperparameter-tuned classical models (logistic regression, multi-layer perceptrons, boosting algorithms) for anastomotic leak prediction. Using a 200-patient clinical dataset strictly bounded by a priori medical constraints, we simulated QNNs with ZZFeatureMap encoding and EfficientSU2/RealAmplitudes ansatze under realistic hardware noise. To ensure statistical reliability, performance metrics were averaged across 10 independent optimization runs. The EfficientSU2-BFGS configuration achieved the highest mean AUC of [Formula: see text], while RealAmplitudes with CMA-ES maximized Average Precision ([Formula: see text]). Crucially, at a fixed, clinically necessary sensitivity of [Formula: see text], specific QNN configurations achieved significantly higher specificity (up to [Formula: see text]) and Negative Predictive Value (up to [Formula: see text]) compared to classical models (maximum [Formula: see text] and [Formula: see text], respectively), effectively minimizing false positives. However, classical models maintained superior probability calibration for continuous risk stratification. We conclude that QNNs offer robust discriminative performance for clinical screening, warranting further validation on larger, independent cohorts.
20 May 2026
Read appraisal →BMJ quality & safety
Retrospective validation study of a large language model approach to screening for intraabdominal surgical site infections for quality and safety reporting
Surgical site infections (SSIs), particularly intra-abdominal (IAB) infections, are challenging to identify and remain a resource-intensive focus of infection prevention programmes. Current automated screening measures rely on discrete data from the electronic health record (EHR), such as microbiology results, diagnosis codes and/or return to the operating room. This approach has poor specificity, and therefore surveillance methods depend heavily on additional manual chart review by trained infection preventionists. Large language models (LLMs) offer an opportunity to improve surveillance by synthesising complex clinical documentation alongside structured data elements.We evaluated the performance of a locally hosted LLM (gpt-35-turbo-16k) to improve IAB SSI screening using perioperative clinical notes and microbiology results. The model analysed documentation across the perioperative period (3 days before through 30 days after surgery) to generate case-level SSI summaries and likelihood assessments. We compared the performance of this tool against the current EHR-based screening workflow.Among 1977 abdominal surgical cases, including 56 with confirmed IAB SSIs, the LLM screened 104 cases as high risk, identifying all infections (negative predictive value (NPV) 100%) and achieving a positive predictive value (PPV) of 53.8%. In contrast, the EHR-based workflow identified 288 cases for further review, with a PPV of only 19.4% and the same NPV of 100%. Analysis of 57 224 notes required ~107 million tokens, translating to approximately USD 0.05 per case.An LLM-based approach to SSI surveillance has the potential to substantially improve efficiency while remaining highly accurate and cost-effective. By reducing reliance on manual chart review, this strategy could allow infection preventionists to shift their attention from surveillance toward quality improvement and patient safety initiatives.
20 May 2026
Read appraisal →BMJ open
What are the views of cancer care administrators and clinicians in England on the use of a machine learning clinical decision support system (ML-CDSS) to predict patients' risk of hepatic and renal deterioration during chemotherapy? A qualitative study.
OBJECTIVE: To explore administrators' and clinicians' views on the factors that influence their use and adoption of a machine learning clinical decision support system (ML-CDSS) to predict patients' risk of hepatic and renal deterioration during chemotherapy. METHODS AND ANALYSIS: This was a qualitative study that used purposive sampling. 18 participants with administration and clinical backgrounds working in cancer care in England were recruited. Qualitative data were collected by conducting semi-structured interviews and a focus group. Data were analysed thematically using the framework method to identify key themes. RESULTS: Participants acknowledged that monitoring blood chemistry is a core component of chemotherapy as it helps clinicians assess patient fitness and treatment response. The ML-CDSS was perceived as a potentially valuable tool for identifying patients at increased risk of hepatic and renal deterioration, supporting clinical decision-making and enhancing care efficiency. However, several concerns were raised regarding its potential implementation in practice. Participants questioned clinicians' willingness and capacity to integrate the tool into their existing workflows. Participants also believed it was important to demonstrate the ML-CDSS's sensitivity, specificity and validity in accurately predicting patients' risk to build clinicians' trust in the tool, demonstrating evidence of its efficacy and effectiveness in practice. CONCLUSION: Administrators and clinicians recognised the potential benefits of the ML-CDSS to enhance the delivery of chemotherapy by identifying patients at risk for hepatic and renal deterioration. Successful adoption in practice depends on building trust with the tool by being transparent in its development, its effectiveness and impact. Future work should demonstrate the ML-CDSS being used in practice to generate real-world evidence.
20 May 2026
Read appraisal →Clinica chimica acta; international journal of clinical chemistry
Exploring clinical feasibility of zero-shot learning in a large language model for immunofixation electrophoresis image interpretation
INTRODUCTION: Immunofixation electrophoresis (IFE) is the gold standard method for detecting and typing monoclonal immunoglobulins, but its interpretation requires expert evaluation and remains susceptible to subjectivity and variability. Recent advances in multimodal large language models (LLMs) have enabled zero-shot visual reasoning without task-specific training. This study aimed to evaluate the feasibility and diagnostic performance of zero-shot multimodal LLMs for automated interpretation of IFE images. MATERIALS AND METHODS: A dataset of 487 immunofixation electrophoresis images representing ten semantic classes (AK, AL, GK, GL, K, L, LGL, MK, ML, and NONE) was retrospectively collected and annotated by expert laboratory specialists. Two multimodal LLMs, ChatGPT-5.2 and Gemini3 Pro, were evaluated using a zero-shot visual question answering (ZS-VQA) framework without additional training or fine-tuning. Each image was analyzed using two prompt configurations: a simple prompt and a detailed prompt with structured diagnostic guidance. Model performance was assessed using precision, recall, F1-score, and confusion matrix analysis. RESULTS: Gemini outperformed ChatGPT across most classes, particularly with detailed prompts, achieving high F1-scores in clinically relevant monoclonal categories such as MK (89.80), AK (84.75), and GK (71.54). Detailed prompting consistently improved recall and F1-scores, underscoring the importance of prompt design. In contrast, ChatGPT showed lower recall, F1-scores, and classification consistency. Both models performed less effectively in visually subtle patterns. CONCLUSIONS: Zero-shot multimodal LLMs, particularly Gemini, show promising potential for interpreting IFE images without task-specific training. However, performance variability and limitations in certain classes indicate that further optimization and validation are required before clinical implementation.
18 May 2026
Read appraisal →Computers in biology and medicine
A unified multimodal framework for chest X-ray retrieval and disease prediction for clinical decision support
Recent advances in medical imaging and natural language processing enable new opportunities for automated diagnostic support. Chest X-rays (CXRs) remain the most common imaging modality for screening pulmonary, cardiovascular, and systemic diseases; however, the growing volume of studies and free-text reports can overwhelm clinicians. We propose a unified multimodal retrieval and prediction framework that jointly leverages DICOM-format CXRs and radiology reports by projecting visual and textual features into a shared semantic space. Trained with contrastive and multi-label objectives, the system supports disease classification, case-based retrieval, and explainable AI. Experiments on the Open-I dataset demonstrate strong performance, achieving a Macro AUROC of 0.95 and Macro F1-score of 0.71 across 22 diagnostic categories. The retrieval module attains high ranking quality (MRR and nDCG >0.93) with sub-millisecond query latency. Quantitative explainability analysis further shows strong agreement between attention- and gradient-based attribution maps (Pearson ρ≈0.92), supporting trustworthy clinical decision-making.
17 May 2026
Read appraisal →Computer methods and programs in biomedicine
Integrating multimodal data and deep learning for functional assessment and rehabilitation prediction after cerebral hemorrhage
BACKGROUND: Intracerebral hemorrhage (ICH) is a leading cause of long-term disability, particularly in China. Post-stroke motor recovery exhibits considerable heterogeneity, presenting substantial challenges for clinicians in establishing realistic goals. While integrative AI paradigms have shown success in other complex medical domains, existing models often rely on single-modality data, limiting their predictive accuracy and clinical utility. This study aims to develop and validate a multimodal predictive model integrating CT imaging, clinical data, and rehabilitation assessments to simultaneously predict motor recovery and global rehabilitation outcomes following cerebral hemorrhage. METHODS: We conducted a retrospective study involving 739 patients (315 for motor function prediction and 424 for rehabilitation outcome assessment) who received rehabilitation therapy after cerebral hemorrhage. To predict motor function, we constructed a late-fusion deep learning model leveraging 3D-DenseNet for CT neuroimaging and Multi-Layer Perceptron (MLP) for clinical and laboratory features. To predict rehabilitation outcomes, a Gradient Boosting Decision Tree (GBDT) model was developed and validated using 5-fold and 10-fold cross-validation, comparing it against other machine learning algorithms, including SVR, Random Forest and AdaBoost. Model performance was assessed using metrics including AUC and R². Additionally, univariate and multivariate regression analysis were performed to identify significant factors influencing motor recovery and rehabilitation outcomes. RESULTS: A total of 739 patients were included. The multimodal fusion model achieved an AUC of 0.856 (95 % CI: 0.741-0.971) and an F1 score of 0.897 (95 % CI: 0.819-0.975), significantly outperforming the imaging-only (AUC: 0.833) and clinical-only (AUC: 0.749) models. For rehabilitation outcome prediction, the GBDT model achieved an R2 of 0.849 (95 % CI: 0.803-0.887), demonstrating superior stability and accuracy over other models. Additionally, multivariate analysis revealed that serum albumin (ALB), neutrophil percentage (NEUT%), triglycerides (TG), and thrombin time (TT) were independent predictors of motor recovery, while age, admission mBI, and time to start rehabilitation significantly influenced functional outcomes. CONCLUSION: This study confirms that a multimodal deep learning framework integrating routinely available CT imaging and clinical biomarkers provides high predictive value for simultaneously forecasting motor recovery and global functional outcomes after ICH. This proof-of-concept approach offers a reproducible, data-driven tool for early risk stratification, facilitating the formulation of individualized rehabilitation strategies and optimizing resource allocation in clinical workflows.
17 May 2026
Read appraisal →JMIR formative research
Complication Risk Classification in Children and Adolescents With Type 1 Diabetes: Interpretable Machine Learning Study Based on Saudi Clinical Guidelines
BACKGROUND: Complication risks in children and adolescents with type 1 diabetes (T1D) can lead to serious health outcomes if not detected early. Despite the availability of clinical data, there remains a gap in interpretable tools that support risk stratification in this age group, particularly in alignment with local clinical guidelines. OBJECTIVE: The purpose of this study is to develop a clinically interpretable model that classifies the risk levels of T1D complications-acute, chronic, and low-using real-world data and expert clinical rules derived from the Saudi Diabetes Clinical Practice Guidelines. METHODS: A pediatric T1D dataset comprising of 306 patients was preprocessed through structured cleaning and feature engineering. Risk labels were constructed using Saudi Diabetes Clinical Practice Guidelines-derived rules. Feature selection was performed using a hybrid approach that combined the SHAP (Shapley Additive Explanations) analysis with exhaustive feature selection. A decision tree model was trained and optimized via cross-validation, using the F1-score as the primary performance metric. RESULTS: The final model achieved a high mean F1-score of 0.9876 with a low variance of 0.0189, using only 5 clinical features: BMI, hypoglycemia, disease duration, hemoglobin A1c, and impaired glucose metabolism. These features were consistently ranked as the most influential. The resulting decision tree offered a transparent logic path, enhancing its clinical interpretability and usability. CONCLUSIONS: This study demonstrates that a simple and interpretable model, guided by national clinical guidelines, can effectively predict the risk levels of T1D complications in children and adolescents. Its strong performance, clarity, and reliance on a small number of clinically meaningful features make it a promising candidate for integration into clinical decision support systems. This supports a shift toward predictive and personalized diabetes care.
17 May 2026
Read appraisal →PloS one
Multi class photoplethysmography-based deep model for cardiovascular disease classification
BACKGROUND: cardiovascular disease is the leading global cause of mortality. Photoplethysmography (PPG), widely embedded in consumer wearables, offers a scalable diagnostic modality. However, prior approaches are often constrained by handcrafted features, binary classification, and poor generalizability, limiting their clinical impact. METHODS: A deep hierarchical convolutional neural network (CNN) was designed to extract both morphological and rhythmic characteristics directly from raw photoplethysmography (PPG) signals. The architecture employs progressively structured convolutional filter hierarchies to capture multi-scale signal features. To enhance signal stability and training efficiency, a dual-stage normalization strategy was implemented, consisting of Z-score standardization followed by Min-Max scaling. In addition, batch normalization and dropout regularization were incorporated to improve model generalization and reduce the risk of overfitting. The proposed framework was trained and evaluated on a multi-source dataset comprising 612 patients and 2,448 annotated PPG segments distributed across six diagnostic classes: atrial fibrillation (AF), heart failure (HF), acute coronary syndrome (ACS), cerebral vascular accident (CVA), deep vein thrombosis (DVT), and normal sinus rhythm (NSR). RESULTS: The model achieved an overall accuracy of 93.48%, a macro-average F1-score of 0.9386, and a Cohen's Kappa of 0.8968, indicating "almost perfect" agreement. AF and HF were detected with flawless precision and recall (1.000), while ACS achieved high sensitivity (recall 0.964). Errors were primarily confined to physiologically related conditions (e.g., ACS vs. CVA). Inference efficiency was demonstrated with <5 ms per segment on consumer-grade hardware, confirming feasibility for real-time applications. CONCLUSION: The proposed framework advances beyond lightweight but underpowered or overly complex models by combining representational depth with computational efficiency. Limitations include the need for external, multi-center validation and explainability integration. This study establishes a robust foundation for PPG-based, multi-class cardiovascular diagnostics, supporting clinical decision support and next-generation wearable health technologies.
17 May 2026
Read appraisal →European heart journal
Foundation models for electrocardiogram interpretation: clinical implications.
BACKGROUND AND AIMS: The 12-lead electrocardiogram (ECG) remains a cornerstone of cardiac diagnostics, yet existing artificial intelligence (AI) solutions for automated interpretation often lack generalizability, remain closed source, and are primarily trained using supervised learning (SL), which requires extensive labelled datasets and may limit adaptability across diverse clinical settings. Self-supervised learning (SSL) can potentially overcome these limitations by learning robust representations from unlabelled data. To address these challenges, this study developed and compared two open-source foundational ECG models: DeepECG-SL, a supervised multilabel ECG model, and DeepECG-SSL, a self-supervised model. METHODS: Both models were trained on over 1 million ECGs using a standardized preprocessing pipeline and automated free-text extraction from ECG reports to predict 77 cardiac conditions. DeepECG-SSL leveraged unlabelled data through self-supervised contrastive learning and masked lead modelling before fine-tuning for downstream tasks, while DeepECG-SL was trained directly on labelled diagnostic data in an end-to-end fashion. Performance was evaluated across seven private, multilingual healthcare systems and four public ECG repositories, with assessment of fairness by age and sex, and investigation of privacy vulnerabilities as well as memory and compute requirements. RESULTS: DeepECG-SSL achieved micro-averaged area under the receiver operating characteristic curves (AUROCs) across all 77 cardiac conditions for ECG interpretation of 0.990 [95% confidence interval (CI): 0.990, 0.990] on the internal dataset (MHI-ds), 0.981 (95% CI: 0.981, 0.981) on external public datasets (UKB, CLSA, MIMIC-IV and PTB), and 0.983 (95% CI: 0.983, 0.983) on external private datasets (UW, UCSF, JGH, NYP, MGH, CSH and CHUM), while DeepECG-SL demonstrated AUROCs of 0.992 (95% CI: 0.992, 0.992), 0.980 (95% CI: 0.980, 0.980), and 0.983 (95% CI: 0.983, 0.984), respectively. Fairness analyses revealed minimal disparities (true-positive rate and false-positive rate difference <0.1) across age and sex groups for both models. DeepECG-SSL demonstrated superior performance on limited-data digital biomarker tasks, with the largest improvements in long QT syndrome (LQTS) genotype classification (AUROC 0.931 vs 0.850, P = .026, n = 127 ECGs) and 5 year atrial fibrillation risk prediction (AUROC 0.742 vs 0.734, P < 0.001, n = 132 050 ECGs), while achieving superior performance in left ventricular ejection fraction ≤40% classification (AUROC 0.926 vs 0.917, P < 0.001, n = 25 252 ECGs) and comparable performance in LQTS detection (AUROC 0.767 vs 0.735, P = 0.117, n = 934 ECGs). CONCLUSIONS: This study establishes SSL as a promising paradigm for ECG analysis, particularly in settings with limited annotated data, enhancing accessibility, generalizability, and fairness in AI-driven cardiac diagnostics. By releasing model weights, preprocessing tools, and validation code, this work aims to support robust, data-efficient AI diagnostics across diverse clinical environments and questions.
16 May 2026
Read appraisal →JMIR mental health
Large Language Models and Their Applications in Mental Health: Scoping Review
BACKGROUND: Large language models (LLMs) are poised to transform mental health care, offering advanced capabilities in diagnosis, prognosis, and decision support. Since their inception, numerous mental health-focused LLMs have emerged in the scientific literature, reflecting the growing interest in leveraging these models across various clinical applications. With a broad range of models available, diverse optimization strategies, and multiple use cases, reviewing the current landscape is critical to understanding where future impact lies. OBJECTIVE: This study aimed to conduct a scoping review investigating the use of LLMs in mental health across diagnostic, prognostic, and decision support tasks. METHODS: We screened 3121 papers from PubMed, Scopus, and Web of Science for studies published between January 2023 and October 2025, using terms related to LLM and mental health. After removing duplicates, 2 reviewers (MCL and WWBG) independently screened the studies, with a third (JJK) to resolve conflicting opinions. We extracted and synthesized information on the models, use cases, datasets, and adaptation methods from selected papers. RESULTS: In total, 41 papers were selected. Many studies included evaluations on OpenAI's GPT series applications: GPT-4 (24 studies, 58.5%) and GPT-3.5 (16 studies, 39%). Others included Bidirectional Encoder Representations from Transformers-derived models (9 studies, 22%), LLaMA (8 studies, 19.5%), and RoBERTa-derived models (6 studies, 14.6%). While all studies initially applied out-of-the-box LLMs, several adapted them through few-shot learning or fine-tuning to better align with specific research goals. The most common use case was in diagnostics (31 studies, 75.6%), while the most common target condition was depression (11 studies, 26.8%). While many studies reported superior performance of LLMs, only a minority of studies (13 studies, 31.7%) validated LLM performance against clinician assessments using real patient data, with the majority relying on proxy outcomes such as clinical vignettes, examination questions, or social media posts. CONCLUSIONS: Despite rapid growth and diversity of LLM applications in mental health, the field remains nascent and exploratory. Future developments must emphasize consistent model adaptation procedures to ensure safety and clinical workflow alignment. Models must also be evaluated on robust evaluation criteria by using standardized protocols and real clinical outcome measures.
16 May 2026
Read appraisal →European heart journal
Deep learning electrocardiogram model for risk stratification of coronary revascularization need in the emergency department
BACKGROUND AND AIMS: Identification of patients with acute coronary syndrome requiring coronary revascularization can be challenging due to inconclusive electrocardiogram (ECG) findings or biomarker results. A deep learning model to detect ECG patterns associated with revascularization likelihood was developed, aiming to guide further assessment and reduce diagnostic uncertainty. METHODS: A convolutional neural network model was trained on 144 691 ED visits from a US cohort (60 ± 19 years; 53% female; 0.6% revascularization), tested on a separate test cohort (n = 35 995), and benchmarked against clinician ECG interpretation and cardiac troponin T (TnT). External validation was performed for the outcomes revascularization and Type 1 myocardial infarction (MI) on 18 673 ED visits from Europe (55 ± 21 years; 49% female; 1.5% revascularization; 1% Type 1 MI). Primary performance metric was area under the receiver operating characteristic curve (AUROC). RESULTS: In the test cohort, the model achieved an AUROC of 0.91 (95% confidence interval [CI] 0.91-0.91), outperforming clinician ECG interpretation (AUROC 0.65, 95% CI 0.54-0.76) and conventional cardiac TnT (AUROC 0.71). In the external validation cohort, ECG model AUROC was 0.81 (95% CI 0.81-0.82) for revascularization, and 0.85 (95% CI 0.84-0.85) for Type 1 MI, compared with 0.70 (95% CI 0.57-0.83) and 0.74 (95% CI 0.56-0.92) for clinician interpretation, and 0.85 and 0.87 for high-sensitivity (hs)-TnT, respectively. The ECG model had higher specificity but lower sensitivity compared with high-sensitivity-troponin T. CONCLUSIONS: The model was able to detect revascularization and Type 1 MI with competitive performance, suggesting a potential role to complement current clinical assessment.
16 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
A sentence is worth a thousand pictures: can large language models understand hum4n L4ngu4ge and the W0rld behind W0rds?
The current generation of large language models (LLMs) has been linked to claims about human-like linguistic performance, and their applications are hailed both as a step towards artificial general intelligence and as a major advance in understanding the cognitive and even neural basis of human language. To assess these claims, first, we analysed the contribution of LLMs as theoretically informative representations of a target cognitive system versus atheoretical mechanistic tools. Second, we evaluated the models' ability to see the bigger picture through top-down feedback from higher levels of processing, which requires grounding in previous expectations and past world experience. We hypothesize that since models lack grounded cognition, they cannot take advantage of these features and instead solely rely on fixed associations between represented words and word vectors. To assess this, we ran a novel leet task (l33t t4sk), which requires decoding sentences in which letters are systematically replaced by numbers. In line with our hypothesis, the results suggest that humans excel in this task, whereas models struggle. We interpret these results by identifying the key abilities that are still missing from the current state of development of these models, which require solutions that go beyond increased system scaling. This article is part of the theme issue 'World models in natural and artificial intelligence'.
15 May 2026
Read appraisal →BMJ open
Prediction of Oswestry Disability Index and Numeric Rating Scale scores after lumbar spine surgery: machine learning model development and fairness assessment
BACKGROUND: One-third of patients operated for degenerative conditions in the lumbar spine do not report substantial improvement after 12 months. Most previous outcome prediction models are classifiers. This constrains nuances in prediction and use for decision support. OBJECTIVES: To develop and test models for the prediction of continuous outcome scores and retrieval of similar patients' outcomes, and to evaluate the models' fairness. SETTING: Norwegian public and private specialist healthcare. PARTICIPANTS AND DATA SOURCE: All cases recorded with an elective operation for lumbar disc herniation (LDH, n=18 377) or lumbar spinal stenosis (LSS, n=24 540) in the Norwegian Registry for Spine Surgery from 1 January 2007 to 23 May 2023. OUTCOME MEASURES: All outcomes were patient-reported 12 months after the operation. The primary outcome was the Oswestry disability index (ODI), modelled on a scale ranging from 0 to 100. Numeric Rating Scale scores (range 0-10) for back and leg pain were secondary outcomes. MODEL BUILDING AND PERFORMANCE: We selected 22 predictors recorded preoperatively by patients and clinicians based on Shapley Additive Explanations values. Data were split into 80%/20% training/test samples for LDH and LSS. Six machine learning methods for regression, that is, with a continuous outcome (extreme gradient boosting (XGBoost), Gaussian process regression, gradient boosting regression, artificial neural networks and linear regression), were trained for both conditions using fivefold cross-validation. We report the magnitude and distribution of errors as mean absolute error (MAE) with 95% CIs, and explanatory power as the coefficient of determination (R2). Fairness and calibration were assessed with violin and calibration plots of error. We developed a patient-similarity function that uses a K-nearest neighbour model to retrieve the individual outcomes of the 50 most similar patients and evaluated it by calculating L1 distances (Manhattan distances) across subgroups. RESULTS: XGBoost regression performed best for both conditions. The models showed good calibration and predicted ODI with MAE 11.32 (95% CI 11.00 to 11.63) and R2 0.27 (95% CI 0.24 to 0.29) for LDH and MAE 12.05 (95% CI 11.76 to 12.32) and R2 0.31 (95% CI 0.28 to 0.34) for LSS. The MAEs for back and leg pain were 2.09 (95% CI 2.04 to 2.15) and 1.95 (95% CI 1.90 to 2.00) for LDH and 2.33 (95% CI 2.28 to 2.38) and 2.13 (95% CI 2.08 to 2.16) for LSS. All models were fair with differences in error between subgroups for sex, age, education level and native language. In the patient-similarity function, distances at baseline were evenly distributed across subgroups. CONCLUSIONS: Our machine learning models predicted continuous outcomes with MAEs close to the SEs of measurements. The models were fair across sociodemographic subgroups. We succeeded in developing a patient-similarity function which supplements the predictions.
15 May 2026
Read appraisal →Nature communications
A deep learning ECG model for identification and localization of occlusion myocardial infarction.
Rapid identification and localization of an acute coronary occlusion are vital to prevent myocardial damage, yet reliance on ST-segment ECG criteria misses many acute occlusion myocardial infarctions (OMI) and triggers unnecessary acute angiographies. Here, we present a trained and validated deep learning model using 540,372 emergency ECGs paired with definitive catheterization outcomes. The model has a C-statistic of ≥0.95 for OMI and ≥0.87 for non-OMI infarctions and can localize culprit lesions in the three main coronary branches, which can guide the angiographer. Performance is similar across age, sex, and ECG hardware subgroups. Obviating dependence on ST-elevations and troponins, this model for the identification and localization of OMI has the potential to shorten the time to reperfusion of an acute coronary occlusion and save resources. Because human oversight of OMI detection on the ECG is limited, randomized clinical trials with patient-relevant outcomes are warranted.
15 May 2026
Read appraisal →JMIR medical informatics
Development Process of a Clinical Decision Support System for Empiric Antibiotic Therapies in Patients With Sepsis: Case Study
BACKGROUND: Antibiotic therapies are the main treatment for bacterial infections, but growing antibiotic resistance is a major global health threat, severely impacting patients with sepsis. Rapid selection of the most effective antibiotic therapy is critical for survival and for preventing further resistance. Physicians must consider numerous factors for proper empiric treatment selection. A clinical decision support system (CDSS) aims to support physicians in this process, facilitating rapid and targeted therapy. OBJECTIVE: The purpose of this work is to explore the extent to which the realization of a CDSS is possible based on the data available to us and to document insights gained during the development of a foundational model designed to assist physicians in determining empiric treatment options for patients with sepsis. In this regard, rather than aiming to develop a CDSS for clinical application, we highlight the importance of close interprofessional collaboration between scientists from various disciplines and analyze the effects of data quality and quantity on the performance of our statistical models. METHODS: Empirical scientists conducted interviews with medical practitioners to acquire the medical knowledge required to develop sound statistical models. We developed and applied 2-step cross-sectional, as well as time-series classification models, to carefully preprocessed data of patients with sepsis admitted to the intensive care unit of a German hospital. RESULTS: We identified several factors as crucial information for valid decisions on empiric therapy for treating patients with sepsis. These include the patients' core data, especially the infection focus. To prevent further resistance, individual risk factors such as travel history and professional background should be considered. The evaluation of a therapy's effectiveness is mainly based on the patient's general condition and blood values such as procalcitonin and interleukin 6. One key factor in the acceptance of a CDSS is the explainability of the results produced by the applied methods. Our models demonstrated mainly weak predictive ability for all considered empiric antibiotic therapies. However, they are not yet suitable for use in clinical practice, especially as they are based on prescribing habits rather than on optimal treatment decisions. CONCLUSIONS: This work highlights the importance of interprofessional collaboration between medical experts and model developers, ensuring that data quality and clinical relevance are central to the process. It emphasizes the urgent need for high-quality, comprehensive data to overcome challenges such as data discontinuity and improve model performance, particularly through enhanced digitization in health care. This feasibility study will facilitate future efforts to develop a CDSS for treating patients with sepsis and to translate it into clinical use.
15 May 2026
Read appraisal →Physiological measurement
A teacher-student deep learning framework for enhanced clinical screening of heart failure from 12-lead electrocardiograms
Objective.Identifying heart failure (HF) from electrocardiograms (ECGs) is challenging due to the lack of definitive features. This study aims to develop a deep learning-based clinical decision support system, CTTSnet, for accurate and automated HF screening using ECGs.Approach.We propose a novel teacher-student framework, CTTSnet, which synergistically integrates a transformer as the teacher and a convolutional neural network as the student. This architecture combines global context-awareness with efficient local feature extraction. The model was trained and evaluated on a large-scale, real-world dataset of 27 018 ECGs and further validated on two external public datasets to assess generalizability.Main results.CTTSnet achieved an AUROC of 0.941 on the primary clinical dataset, surpassing strong baselines. The model attained an accuracy of 87.3%, with a recall of 88.6% and specificity of 85.8%. External validation on two additional public datasets further confirmed the model's generalizability.Significance.CTTSnet provides a scalable, high-accuracy tool for HF triage from routine ECG recordings. Its potential clinical value lies in assisting early identification of patients at risk of HF, thereby helping reduce missed diagnoses and facilitating timely referral for further evaluation.
14 May 2026
Read appraisal →PloS one
Comparing machine and deep learning models for pediatric anxiety classification using structured EHRs and area-based measures of health data
OBJECTIVE: This retrospective, case-control study with internal validation evaluates the performance of machine learning (ML) and deep learning (DL) models in classifying pediatric patients at risk for anxiety disorders using structured electronic health records (EHRs) and area-based measures of health (ABMH). The aim is to enable proactive care by monitoring potential anxiety onset across developmental stages. METHODS: We trained a series of ML models (Logistic Regression, Decision Tree, Random Forest, K-Nearest Neighbors, XGBoost) and DL models (LSTM, GRU, RETAIN, Dipole) using structured EHR data from 30-day windows prior to diagnosis. Two datasets were used per age group: one with structured EHR data only, and another including both EHR and ABMH data. ML models were trained using short-term cross-sectional features, while DL models leveraged full longitudinal patient histories. Performance was assessed using AUROC, AUPRC, PPV, NPV, F1 score, and accuracy. Due to differences in input scope, model performance reflects both algorithmic and temporal design differences and is not intended as a direct comparison between ML and DL. RESULTS: ML models offered strong baseline performance, with XGBoost achieving AUROC scores of 0.817 (EHR) and 0.816 (EHR+ABMH) for 8-year-olds. Adding ABMH features did not significantly improve performance. DL models, particularly RETAIN and Dipole, achieved the highest AUROC values (e.g., Dipole: 0.853 with EHR, 0.857 with EHR+ABMH for 8-year-olds), outperforming other DL and ML models within their respective design constraints. CONCLUSION: Both ML and DL models successfully identified likely anxiety onset using structured EHR data. DL models using longitudinal data achieved the highest performance, while XGBoost provided a robust ML baseline. The minimal impact of ABMH features highlights integration challenges, and performance variation across ages emphasizes the need for age-stratified modeling approaches.
14 May 2026
Read appraisal →JMIR medical informatics
Scalable Identification of Clinically Relevant Chronic Obstructive Pulmonary Disease Documents in Large-Scale Electronic Health Record Datasets With a Lightweight Natural Language Processing Model: Retrospective Cohort Study
BACKGROUND: The widespread adoption of electronic health records has resulted in the generation of large volumes of clinical notes. Learning algorithms and large language models can be trained on these resources, but they are susceptible to noise-irrelevant or noninformative data. This sensitivity can lead to significant challenges, including performance degradation and the generation of inaccurate predictions or "hallucinations." This study addresses a critical challenge in clinical informatics: efficiently filtering millions of documents for relevance before advanced language model processing, particularly in resource-constrained environments. OBJECTIVE: We present a novel framework for determining document relevance in clinical settings using a chronic obstructive pulmonary disease (COPD) dataset. METHODS: We developed a novel framework using weak supervision and domain-expert heuristics to generate "silver standard" labels for training data and gold standard expert-annotated labels, creating 2 datasets to optimize the model during the development phase and subsequent testing phase. Various text representation techniques (bag of words, term frequency-inverse document frequency, lightweight document embeddings, compression-based features, and Unified Medical Language System concept extraction) were evaluated. These representations were used to train random forest, extreme gradient boosting, and k-nearest neighbor classifiers. Models were optimized on a small expert-annotated dataset and evaluated on a held-out test set. RESULTS: The combination of lightweight document embedding with a random forest classifier demonstrated the best performance, achieving a precision of 0.73, recall of 0.86, and F1-score of 0.80 (95% CI 0.76-0.87) for identifying relevant COPD documents. This significantly outperformed baseline heuristics (precision=0.70; recall=0.38; F1-score=0.50, 95% CI 0.43-0.56) and other tested methods. CONCLUSIONS: Our study presents a novel framework for identifying COPD-relevant clinical documents using lightweight embedding and machine learning. This approach effectively filters pertinent documents, enhancing information retrieval precision. The framework's scalability and minimal annotation needs make it promising for diverse health care applications, potentially optimizing clinical outcomes through efficient document selection for data-driven decision support systems.
14 May 2026
Read appraisal →Journal of medical Internet research
AI-Based Automation for Medication Reconciliation: Scoping Review
BACKGROUND: Medication reconciliation (MedRec) has the potential to improve patient safety by enhancing the continuity of medication information across settings. MedRec involves 3 core tasks: the creation of a best possible medication history, the identification of medication discrepancies among medication lists, and the resolution of medication discrepancies. While artificial intelligence (AI) has the potential to improve MedRec, existing reviews have not identified the ways in which researchers have used AI to facilitate MedRec tasks and their constituent subtasks or the level of automation achieved. OBJECTIVE: This scoping review aimed to map how previous research has applied AI to MedRec tasks and subtasks and assess the extent of automation achieved. METHODS: We searched MEDLINE, Embase, Web of Science, IEEE Xplore, and Compendex in June 2024 for studies that used AI to support a MedRec task or subtask, excluding entirely rule-based tools or studies focused on other aspects of medication management. After screening 2345 unique records, we conducted backward citation searching of studies included at the full-text stage, identifying an additional 795 unique records. We used a 4-stage model of human information processing as a structural lens to guide our considerations of automation, mapping the core tasks of MedRec onto this model. RESULTS: A total of 94 studies met the inclusion criteria. All studies addressed subtasks related to the creation of a best possible medication history. Only 2.1% (n=2) of the studies also addressed the identification of discrepancies. Thus, the highest stage of automated information processing achieved was information analysis, although most studies (92/94, 97.9%) only automated information acquisition steps. Most studies (67/94, 71.3%) used free-text clinical notes from the electronic health record, although a significant proportion (21/94, 22.3%) used images of pills or images of other medication-related items. Studies using text-based data used a variety of machine learning methods (eg, recurrent neural networks, conditional random fields, support vector machines, and transformers), whereas those that leveraged images typically used convolutional neural networks. Most studies (61/94, 64.9%) used publicly available data from benchmarking datasets (eg, n2c2 2022) and were strictly model development studies, with only 1.1% (n=1) being usability studies. CONCLUSIONS: This is the first review to consider the role of AI in the automation of MedRec tasks, offering a basis for prioritizing future development efforts. Current applications of AI to automate MedRec tasks are preliminary, with most work focusing on the extraction of medication information and limited to proof-of-concept model development. Future work should consider addressing infrastructural barriers to the AI-based automation of MedRec tasks (eg, data incompleteness in sources of medication information) and exploring approaches to automate discrepancy resolution. Beyond developing models, there is also a need to implement them in tools and evaluate them in real-world contexts.
13 May 2026
Read appraisal →JMIR mHealth and uHealth
Contribution of Longitudinal Mobile Health Measures in the Dynamic Track of Patients With Major Depressive Disorder: Multiple Centers, Prospective Cohort Study Using Functional Data Analysis and Machine Learning
BACKGROUND: Continuous follow-up for patients with major depressive disorder (MDD) is essential for treatment decisions and a better prognosis. There remains limited evidence regarding the critical issue of depression variation trajectory prediction using mobile health (mHealth) measures. Moreover, the temporal dynamics of mHealth measures have not been fully modeled in previous studies, and the poor patient adherence to mHealth records poses great challenges to the dynamic feature modeling. OBJECTIVE: This study aimed to examine the contribution of mHealth measures in predicting depression variation trajectory for patients with MDD, with full consideration of the temporal dynamics of mHealth measures. METHODS: A total of 229 patients with MDD from a multiple-center, prospective cohort were included. A 12-week follow-up was conducted involving the collection of the Hamilton Depression Rating Scale (HAMD-17), along with patient-reported outcomes (Immediate Mood Scaler and Altman Self-Rating Mania Scale) via mobile devices and sleep duration through wearable wristbands. We used functional data analysis to extract dynamic features from the sparse mHealth records, rather than aggregating the data to a single scalar summary measure through collapsing over time. Subsequently, 3 machine learning models were applied to predict the depression variation trajectory classes based on the baseline characteristics and these extracted dynamic features. RESULTS: Based on the variation of HAMD-17 scores within 12 weeks, the participants were labeled into 4 classes through the k-means algorithm. The classes included stable decline (n=93), fluctuate decline (n=44), fast decline (n=60), and delayed and fluctuate (n=32), in light of the shape of depression trajectories. With both baseline features and dynamic features of the mHealth measures, accuracy rates for the overall data were 54.35%, 60.87%, and 56.52%, for the stable decline patients were 78.95%, 84.21%, and 73.68%, for the nonstable decline patients were 59.26%, 62.96%, and 70.37% based on the 3 machine learning models, respectively. The results were significantly superior to the prediction obtained without mHealth measures (with an overall accuracy below 50%) and only showed a marginal reduction in accuracy relative to the ideal prediction with assessment obtained from clinical visits. Moreover, in the construction of the most accurate prediction model, dynamic features of the Immediate Mood Scaler, the Altman Self-Rating Mania Scale, and sleep duration emerged as the most influential predictors, ranking first, third, and fourth, respectively, in terms of their relative importance. CONCLUSIONS: Longitudinal mHealth measures show potential in depression variation trajectory monitoring for patients with MDD even under poor patient adherence. Our work provides practical help in alleviating the follow-up burden for patients with MDD and validates the effectiveness of mHealth measures in clinical applications.
13 May 2026
Read appraisal →Pediatric surgery international
Development of a machine learning-based prognostic prediction model and a web-based tool for pediatric hepatoblastoma: a Surveillance, Epidemiology, and End Results (SEER) database study
BACKGROUND: Prognostic evaluation of pediatric hepatoblastoma (HB) remains challenging due to the low accuracy of traditional risk stratification models. This study aims to leverage the U.S. Surveillance, Epidemiology, and End Results (SEER) database to develop and compare several machine learning-based survival models and create an online prediction tool for individualized survival probability estimation in children with HB. METHODS: A total of 614 pediatric HB patients diagnosed from 2000 to 2021 in the SEER database were identified and randomly split (7:3 ratio) into a training set (n = 429) and an internal hold-out validation cohort (n = 185). This split was selected to retain sufficient data for model development while preserving an adequate validation cohort for performance assessment. Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was used for feature selection, and three prognostic prediction models (Cox regression, random survival forest (RSF), and a deep learning-based survival (DeepSurv) neural network) were developed using the training data. Model performance was evaluated by Harrell's concordance index (C-index), time-dependent area under the receiver operating characteristic (ROC) curve (AUC) at 1, 3, and 5 years, and decision curve analysis (DCA). We also used SHapley Additive exPlanations (SHAP) analysis to interpret the RSF model. Based on RSF-derived risk scores, patients were classified into high-, intermediate-, and low-risk groups for survival analysis. Ultimately, a web-based tool was developed to enable real-time prediction of 1-, 3-, and 5-year survival probabilities using individual patient characteristics. RESULTS: In the validation cohort, the RSF model achieved the highest C-index (0.745), outperforming the DeepSurv (0.720) and Cox regression models (0.709). The RSF also yielded favorable 1-year, 3-year, and 5-year AUC values, and DCA indicated a greater net benefit across clinically relevant threshold probabilities. SHAP analysis highlighted distant-stage disease, surgical treatment status, and tumor extent as key factors influencing survival predictions. Kaplan-Meier curves stratified by the RSF risk groups showed significantly different survival outcomes among high-, intermediate-, and low-risk patients (log-rank P < 0.001 for all comparisons). The finalized online tool allows users to input patient characteristics and obtain estimated 1-year, 3-year, and 5-year survival probabilities, which may assist individualized prognostic assessment as an adjunctive reference. CONCLUSION: The RSF-based prognostic model showed favorable predictive performance compared with the Cox and DeepSurv models in this retrospective SEER-based cohort of pediatric HB. With interpretable model outputs and an accessible web-based interface, the model may serve as an adjunctive tool for individualized risk assessment. However, external validation and prospective evaluation are required before routine clinical implementation.
13 May 2026
Read appraisal →PloS one
Machine learning-based model for predicting recanalization in isolated distal deep vein thrombosis and analysis of predictors
BACKGROUND: Isolated distal deep vein thrombosis (IDDVT) is common, yet tools for predicting poor recanalization remain limited. We aimed to develop and compare machine learning models for predicting poor recanalization in patients with IDDVT and to identify the most informative predictors. METHODS: A total of 1600 patients with IDDVT were retrospectively enrolled. The dataset was randomly divided into a development set (n = 1280) and an independent test set (n = 320) using stratified sampling. Six predictive models were developed and compared: logistic regression (LR), support vector machine (SVM), random forest (RF), multilayer perceptron (MLP), extreme gradient boosting (XGBoost), and a Voting Ensemble. Model training and hyperparameter tuning were performed in the development set using five-fold stratified cross-validation, and optimal classification thresholds were determined using the Youden index. Model performance was evaluated by discrimination, calibration, and classification metrics, with 95% confidence intervals estimated by bootstrap resampling (10,000 iterations). SHAP analysis was applied to interpret the final model. RESULTS: In the independent test set, all models showed acceptable to strong discrimination, with AUC values ranging from 0.808 to 0.908. XGBoost achieved the best overall performance, with an optimal threshold of 0.183, an AUC of 0.908 (95% CI, 0.855-0.952), a Brier score of 0.077 (95% CI, 0.058-0.096), an accuracy of 0.900 (95% CI, 0.866-0.931), a precision of 0.650 (95% CI, 0.529-0.767), a recall of 0.803 (95% CI, 0.686-0.906), an F1-score of 0.717 (95% CI, 0.615-0.806), and a specificity of 0.918 (95% CI, 0.884-0.950). The calibration intercept and slope of the XGBoost model were 0.149 (95% CI, -0.192 to 0.454) and 1.410 (95% CI, 1.098-1.809), respectively, indicating acceptable overall calibration. SHAP analysis identified D-dimer rate, provoking-factor-related variables, anticoagulant use, and age group as the most influential predictors. CONCLUSION: Among six candidate models, XGBoost showed the best overall performance for predicting poor recanalization in patients with IDDVT. This study establishes an interpretable machine learning-based prediction framework focused specifically on poor recanalization in IDDVT and highlights the contribution of dynamic laboratory information, particularly D-dimer rate. The model may support early risk stratification and individualized follow-up planning, but external validation is required before routine clinical implementation.
13 May 2026
Read appraisal →Journal of medical Internet research
Rheumatic Digital Twin: Proposed Machine Learning-Based Multimodal Framework to Inform Clinical Decision-Making
Rheumatic diseases are chronic, immune-mediated conditions characterized by significant heterogeneity in presentation and disease course. However, current clinical approaches often rely on snapshot-based assessments that fail to capture the complex longitudinal evolution of these conditions. To address these limitations and support the implementation of precision medicine, we present the design for the Rheumatic Digital Twin, a novel, modular conceptual framework intended to integrate heterogeneous multimodal data, ranging from electronic health records and clinical notes to imaging and omics, into a dynamic, computational representation of the patient journey. Our theoretical architecture addresses challenges related to data silos and variable availability of data modalities through a multistage approach that envisions the use of domain-specific foundation models to independently process distinct data modalities. To effectively model the temporal progression inherent in chronic diseases, the proposed design utilizes Transformer architectures, leveraging self-attention mechanisms to treat patient events, such as lab results or medication changes, as sequential data tokens. We describe how these unimodal representations would subsequently be fused via joint embedding techniques to construct a shared, multimodal representational space. Envisioned to function analogously to a recommender system, the Rheumatic Digital Twin framework is modeled to map patients into a latent space where proximity reflects clinical and biological similarity. By identifying "nearest neighbors," historical patients with comparable trajectories, the system aims to enable in silico cohorting, theoretically allowing clinicians to forecast key clinical events, predict treatment responses, and identify likely disease courses based on the outcomes of similar peers.
11 May 2026
Read appraisal →PloS one
Development and evaluation of an ontology for non-invasive respiratory support in acute care
Managing patients with respiratory failure increasingly involves non-invasive respiratory support (NIRS) strategies to support respiration, often preventing the need for invasive mechanical ventilation. However, despite the rapidly expanding use of NIRS, there remains a significant challenge to its optimal use across all medical circumstances. It lacks a unified ontological structure, complicating guidance on NIRS modalities across healthcare systems. In this study, we introduced NIRS ontology to support knowledge representation in acute care settings by providing a unified framework that enhances data clarity and interoperability, laying the groundwork for future clinical decision-making research. We developed NIRS ontology using the Web Ontology Language (OWL) and Protégé to organize clinical concepts and relationships. To enable rule-based clinical reasoning beyond hierarchical structures, we added Semantic Web Rule Language (SWRL) rules. We evaluated logical reasoning using a sample of 6 patient scenarios and used SPARQL queries to retrieve and test targeted inferences. The ontology has 145 classes, 11 object properties, and 18 data properties across 949 axioms that establish concept relationships. To standardize clinical concepts, we added 392 annotations, including descriptive definitions based on controlled vocabularies. SPARQL queries successfully validated all test cases (rules) by retrieving appropriate patients' outcomes: for instance, a patient treated with HFNC (high-flow nasal cannula) for 2 hours due to acute respiratory failure may avoid endotracheal intubation. This NIRS ontology formally represents domain-specific concepts, including ventilation modalities, patient characteristics, therapy parameters, and outcomes. SPARQL query evaluations across clinical scenarios confirmed the ontology's ability to support rule‑based reasoning and therapy recommendations, providing a foundation for consistent documentation practices, integration into clinical data models, and advanced analysis of NIRS outcomes. In conclusion, this proof-of-concept NIRS ontology demonstrates how clinical concepts can be formally represented and queried, offering a foundation for future validation, EHR integration, and evidence-based refinement of NIRS practices.
11 May 2026
Read appraisal →EuroIntervention : journal of EuroPCR in collaboration with the Working Group on Interventional Cardiology of the European Society of Cardiology
Derivation and external validation of a deep learning model to predict changes in coronary plaque burden
BACKGROUND: Predicting the progression/regression of coronary plaque burden is challenging. AIMS: We aimed to develop a deep learning model to forecast changes in percent atheroma volume (ΔPAV) using intravascular ultrasound (IVUS). METHODS: We analysed data from IBIS-4 and PACMAN-AMI. Core lab measurements of plaque burden were available from IVUS pullbacks. Each model consists of a bidirectional Long Short-Term Memory (biLSTM) layer followed by two fully connected layers with one neuron each, resulting in both a classification for input progression/regression and an estimation of the ΔPAV. RESULTS: For the derivation and validation, a total of 1,960 regions of interest (ROIs) from the IBIS-4 dataset were used. The mean±standard deviation of the model accuracy was 0.85±0.02, the Matthews correlation coefficient was 0.70±0.04, and the F1 score was 0.85±0.02 for both progression and regression classes. In the testing (external validation) process with the PACMAN-AMI dataset, 5,283 ROIs were utilised. The mean ΔPAV was -0.31±5.63, for which 2,665 featured regression with a mean ΔPAV of -4.57±3.73, and 2,618 presented progression with a mean ΔPAV of 4.02±3.55, representing 49.6% of plaque progression prevalence. The predictive performance across the 100 trained models in the testing dataset showed an accuracy of 0.84, a Matthews correlation coefficient of 0.68, and an F1 score for the progression and regression classes of 0.84. CONCLUSIONS: This is the first deep learning model capable of detecting changes in plaque progression by analysing the rate of plaque burden change between adjacent frames.
10 May 2026
Read appraisal →Journal of biopharmaceutical statistics
Revolutionizing cardiovascular disease classification through machine learning and statistical methods
BACKGROUND: Cardiovascular diseases (CVDs) include abnormal conditions of the heart, diseased blood vessels, structural problems of the heart, and blood clots. Traditionally, CVD has been diagnosed by clinical experts, physicians, and medical specialists, which is expensive, time-consuming, and requires expert intervention. On the other hand, cost-effective digital diagnosis of CVD is now possible because of the emergence of machine learning (ML) and statistical techniques. METHOD: In this research, extensive studies were carried out to classify CVD via 19 promising ML models. To evaluate the performance and rank the ML models for CVD classification, two benchmark CVD datasets are considered from well-known sources, such as Kaggle and the UCI repository. The results are analysed considering individual datasets and their combination to assess the efficiency and reliability of ML models on the basis of various performance measures, such as precision, kappa, accuracy, recall, and the F1 score. Since some of the ML models are stochastic, we repeated the simulation 50 times for each dataset using each model and applied nonparametric statistical tests to draw decisive conclusions. RESULTS: The nonparametric Friedman - Nemenyi hypothesis test suggests that the Extra Tree Classifier provides statistically superior accuracy and precision compared with all other models. However, the Extreme Gradient Boost (XGBoost) classifier provides statistically superior recall, kappa, and F1 scores compared with those of all the other models. Additionally, the XGBRF classifier achieves a statistically second-best rank in terms of the recall measures.
10 May 2026
Read appraisal →International journal of medical informatics
Development of an interpretable machine learning model for predicting 4-year chronic kidney disease risk in elderly hypertensive patients
INTRODUCTION: Age and hypertension are key drivers of renal impairment, predisposing older hypertensive adults to faster kidney function decline and higher mortality. We aim to develop an interpretable machinelearning model to predict 4-year chronic kidney disease (CKD) risk in this population. METHODS: Our study incorporated 4,142 hypertensive patients from the Health and Retirement Study (HRS) 2010 and 2012 cohorts for model development and internal validation, with additional temporal validation performed within the HRS 2006 and 2008 cohorts. External validation was conducted using three distinct subcohorts derived from the China Health and Retirement Longitudinal Study (CHARLS) database. Feature selection was implemented through an integrated LASSO-Boruta algorithm, followed by model construction using eight machine learning approaches. Discriminative performance was rigorously evaluated through multiple metrics, including receiver operating characteristic (ROC) curve analysis, accuracy, sensitivity, specificity, and Brier score. The optimal model underwent interpretability analysis via SHapley Additive exPlanations (SHAP) to elucidate decision-making mechanisms and was subsequently deployed as a web-based clinical prediction tool. RESULTS: Using a combined LASSO-Boruta strategy, we identified nine routinely available predictors for model development. In the training set, SVM achieved the highest AUC (0.735), closely followed by XGBoost (0.734); notably, in the temporal validation cohort, XGBoost was the only model with an AUC > 0.700 (0.702). Overall performance metrics derived from confusion matrices, together with Brier scores, suggested that XGBoost provided a favorable balance between sensitivity and specificity while maintaining acceptable probabilistic calibration. Calibration curves further suggested that XGBoost showed relatively stable agreement between predicted and observed risks across datasets, supporting its selection for subsequent SHAP-based interpretation and web deployment; SHAP identified age as the leading contributor to CKD risk. CONCLUSIONS: We developed an interpretable model using routine clinical indicators to predict 4-year CKD risk in elderly hypertensive adults, with applicability across Asian and Caucasian populations.
9 May 2026
Read appraisal →Computer methods in biomechanics and biomedical engineering
Modified mutual information feature selection algorithm to predict COVID-19 using clinical data
The COVID-19 pandemic has profoundly impacted health, emphasizing the need for timely disease detection. Blood tests have become key diagnostic tools due to the virus's effects on blood composition. Accurate COVID-19 prediction through machine learning requires selecting relevant features, as irrelevant features can lower classification accuracy. This study proposes Modified Mutual Information (MMI) for feature selection, ranking features by relevance and using backtracking to find the optimal subset. Support Vector Machines (SVM) are then used for classification. Results show that MMI with SVM achieves 95% accuracy, outperforming other methods, and demonstrates strong generalizability on various benchmark datasets.
6 May 2026
Read appraisal →Medical image analysis
Robust fine-grained echocardiographic view classification with supervised contrastive learning
Accurate classification of echocardiographic views is fundamental for automated cardiac analysis. However, clinical practice relies on a large, heterogeneous set of fine-grained acquisitions that introduce substantial inter-observer variability. Existing studies have primarily focused on limited view sets, often collapsing specialised views into broad categories, which limits their clinical relevance. We address this limitation by introducing TTE47, the first publicly available benchmark comprising 47 clinically meaningful views annotated independently by three experts. This dataset enables the rigorous quantification of inter-observer agreement and establishes a foundation for reproducible, clinically relevant evaluation. To tackle the dual challenges of subtle inter-class distinctions and structured label variability, we propose a novel supervised contrastive learning framework incorporating a tailored loss function. Our method outperforms cross-entropy and standard supervised contrastive baselines, achieving leading performance among evaluated methods on TTE47 and surpassing prior work on TMED-2 without dataset-specific pretraining, using a model pretrained on TTE47. Beyond accuracy, we introduce clustering-based metrics, Detection Rate and Label Recovery Precision, that measure semantic coherence and the model's ability to resist annotation variability. Results show that the learned feature space aligns more strongly with underlying anatomical structure than with any single annotator's style, enabling resilience to label shifts and maintaining robustness comparable to human-level disagreement. By integrating multi-expert evaluation, robust representation learning, and interpretable feature-space analysis, this work establishes a scalable and clinically relevant framework for fine-grained echo view classification. Our findings highlight the potential of contrastive pretraining to standardise interpretation, mitigate subjectivity, and enhance the reliability of AI-assisted echocardiography in diverse clinical settings.
6 May 2026
Read appraisal →Annual review of biomedical engineering
Physics-Informed Machine Learning in Biomedical Science and Engineering
Physics-informed machine learning (PIML) is emerging as a potentially transformative paradigm for modeling complex biomedical systems by integrating parameterized physical laws with data-driven methods. Here, we review three main classes of PIML frameworks: physics-informed neural networks (PINNs), neural ordinary differential equations (NODEs), and neural operators (NOs), highlighting their growing role in biomedical science and engineering. We begin with PINNs, which embed governing equations into deep learning models and have been successfully applied to biosolid and biofluid mechanics, mechanobiology, and medical imaging, among other areas. We then review NODEs, which offer continuous-time modeling, especially suited to dynamic physiological systems, pharmacokinetics, and cell signaling. Finally, we discuss deep NOs as powerful tools for learning mappings between function spaces, enabling efficient simulations across multiscale and spatially heterogeneous biological domains. Throughout, we emphasize applications where physical interpretability, data scarcity, or system complexity make conventional black-box learning insufficient. We conclude by identifying open challenges and future directions for advancing PIML in biomedical science and engineering, including issues of uncertainty quantification, generalization, and integration of PIML and large language models.
5 May 2026
Read appraisal →Cancer medicine
Predicting Prognosis for Gastric Cancer Patients Receiving Neoadjuvant Treatment With Body Composition-Based Deep Learning
BACKGROUND: This study sought to develop an innovative body composition (BC)-based deep learning (DL) model to precisely evaluate survival in gastric cancer (GC) patients undergoing neoadjuvant treatment (NT). MATERIALS AND METHODS: This retrospective study included GC patients undergoing NT from two centers. CT images both pre-NT and post-NT were preprocessed, focusing on the automatic segmentation of subcutaneous fat, visceral fat, and skeletal muscle regions using TotalSegmentator. Delta Radiomics features were extracted using Pyradiomics. After feature fusion and selection, the optimal model is Naive Bayes (Rad model). A hybrid DL model was developed by combining ResNet18 and Transformer networks for feature extraction. The Clinic_Rad_DL model was constructed by combining clinical features, radiomic signatures, and DL signatures. The ExtraTree classifier was used for the Clinic_Rad_DL model, while a separate Cox regression model was developed for survival analysis using the same features. RESULTS: A total of 356 patients (mean age, 59 ± 10 years; 264 males [74.2%]) were enrolled and divided into training, validation, and test sets in a 7:2:1 ratio. The DL model outperformed the Rad model. The Clinic_Rad_DL model outperformed both Rad model and DL model, with AUC of 0.915, 0.890, and 0.890 in training, validation, and test sets, respectively. The Cox proportional hazards model showed C-index of 0.806, 0.803, and 0.819, effectively stratifying patients into high- and low-risk groups with significant survival differences. CONCLUSION: The study developed and validated a BC-based DL model to predict survival in GC patients undergoing NT, offering potential for personalized treatment strategies in clinical practice.
5 May 2026
Read appraisal →Resuscitation
Deep learning-based ROSC prediction and ECG phenotyping in out-of-hospital cardiac arrest
INTRODUCTION: Electrocardiogram (ECG) signals during cardiac arrest contain detailed information on cardiac rhythm characteristics and have been associated with resuscitation outcomes. We aimed to develop a convolutional neural network (CNN) model to predict return of spontaneous circulation (ROSC) and identify novel ECG phenotypes in patients with out-of-hospital cardiac arrest (OHCA). METHODS: This retrospective cohort study used Korean OHCA Registry and ECG data from Seoul emergency medical services between July 2021 and December 2023. Adult patients with nontraumatic OHCA who had prehospital ECG signals were included. Five-second ECG segments obtained during resuscitation were analyzed. A one-dimensional CNN was developed to simultaneously predict the probability of ROSC within 2 min and to classify the concurrent ECG rhythm as shockable or non-shockable. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). ECG phenotypes were identified by applying K-means clustering to feature representations extracted from the trained CNN. RESULTS: 3452 patients (median age, 73 years [interquartile range, 61-81]; 34.4% female) were included. The AUCs for predicting ROSC within 2 min and shockable rhythm in the test set were 0.921 (95% confidence interval [CI], 0.897-0.945) and 0.983 (95% CI, 0.979-0.988), respectively. Five ECG phenotype categories with varying morphologies and ROSC probabilities were identified. Phenotype 1 consisted of shockable rhythms with high ROSC probability (30.4%), whereas Phenotype 2 showed shockable rhythms with low ROSC probability (4.8%). Phenotypes 3 and 4 included pulseless electrical activities with relatively high (5.2%) and lower (0.5%) ROSC probabilities, respectively. Phenotype 5 primarily consisted of asystole or near-asystole rhythms. Transitions between ECG phenotypes were associated with CPR quality. CONCLUSIONS: The CNN-based model accurately predicted ROSC and shockable rhythm from ECG signals and identified five novel ECG phenotypes in OHCA. These findings can enable accurate ROSC prediction and individualized prehospital resuscitation for patients with OHCA.
4 May 2026
Read appraisal →IEEE transactions on bio-medical engineering
Multi-Stage Respiratory Sound Analysis: Confidence-Driven Wheeze and Crackle Detection
OBJECTIVE: Accurate detection of adventitious respiratory sounds, such as wheezes and crackles, is essential for diagnosing and managing respiratory conditions. This study introduces a multi-stage, confidence-driven framework for automated pediatric auscultation analysis, performing a three-way classification of normal, wheeze, and crackle sounds to improve diagnostic accuracy. METHODS: We develop a comprehensive pipeline integrating anomaly-specific segment selection, segment-level classification, and confidence-based fusion. Our contrastive variational recurrent neural network (CVRNN) enhances feature extraction, while a confidence-weighted aggregation strategy refines final predictions. The system is validated using a diverse pediatric dataset from 742 subjects (aged 1-59 months) from seven countries. RESULTS: The multi-level framework is evaluated across three stages. The anomaly-specific segment selection achieves 98.47 $\%$ recall, identifying adventitious regions. Next, segment-level classifiers improve sensitivity, achieving balanced accuracies of 72.15 $\%$ (wheeze) and 68.1 $\%$ (crackle). This performance surpasses state of the art systems on the same dataset and demonstrates enhanced balanced performance in detecting both crackle and wheeze sounds, which present different challenges to automated systems given their markedly different acoustic profiles. Finally, the confidence-driven fusion outperforms traditional aggregation methods, yielding a final three-way classification of 62.12 $\%$. CONCLUSION: Our confidence-based multi-stage approach enhances automated respiratory sound classification by prioritizing high-certainty segment predictions, aligning with expert physician annotations. SIGNIFICANCE: This framework advances computer-aided respiratory diagnostics, improving early detection and monitoring of pediatric respiratory conditions. By integrating expert-inspired segmentation with machine learning-driven confidence estimation, it has potential to enhance clinical workflows and screening for pulmonary diseases, particularly in resource-limited settings.
4 May 2026
Read appraisal →Artificial intelligence in medicine
CL-MHAD: Contrastive Learning-based Multi-Hypergraph Aggregation and Diffusion model for prescription recommendation
Multiple syndrome-based prescription recommendations are significant for personalized diagnosis and treatment in Traditional Chinese Medicine (TCM). However, it remains a challenge to effectively extract and fuse multi-dimensional knowledge in herbs for accurate syndrome-based prescription recommendations. To address it, we propose CL-MHAD, a contrastive learning-based multi-hypergraph aggregation and diffusion model for prescription recommendation. Focusing on prescriptions, herbs, properties, and dosages, this model presents a multi-view hypergraph reconstruction mechanism, which employs a consistent Top-K neighbor strategy to generate affinity and feature matrices. Subsequently, a diffusion-enhanced method is proposed, utilizing multi-step propagation and weighted mapping to capture high-order relationships. Then, we propose a cross-view contrastive learning strategy to enhance discrimination and present a topology-aware random walk augmentation approach to alleviate data sparsity. An integrated loss function combining classification, contrastive, and augmentation losses is designed to optimize model performance. We evaluate CL-MHAD in a real-world clinical setting using Gastrointestinal Diseases (GID) as a case study. Extensive experiments, including comparative evaluations, ablation studies, and parameter sensitivity analyses, show that CL-MHAD achieves overall superior performance compared with representative baselines across multiple evaluation metrics. Specifically, the model achieves performance gains of 1.27% to 24.67% over representative baselines across all datasets. We further evaluate the model through dosage-syndrome interaction and data-sparsity robustness tests, with results demonstrating the effectiveness of our weighted fusion strategy and the model's superior stability. This study offers an effective solution and a promising paradigm for accurate syndrome-based prescription recommendations in TCM.
4 May 2026
Read appraisal →IEEE transactions on bio-medical engineering
Extending Multiscale Characterization of Heart Rate Variability via Deep Learning for Mortality Risk Prediction
OBJECTIVE: To improve mortality risk prediction from heart rate variability (HRV) signals by capturing nonlinear scaling patterns often overlooked by traditional linear analyses. METHODS: This study combines detrended moving average (DMA) analysis with convolutional neural networks (CNNs). DMA curves were computed from 2-hour overlapping windows of 24-hour Holter ECG recordings in 916 survivors and 70 nonsurvivors. A CNN was trained to extract features from these curves and benchmarked against models using traditional HRV and clinical features. RESULTS: The CNN achieved an ROC-AUC of 0.72 and an adjusted hazard ratio of 2.129 for daytime recordings, outperforming standard models. Two patient groups emerged based on DMA scaling patterns. Group 1, with dominant short-term scaling, exhibited reduced slopes in nonsurvivors, suggesting impaired autonomic adaptability. Group 2 showed earlier transitions between short- and long-term behavior, where reduced long-term slopes more strongly predicted mortality. Integrated gradients analysis identified key timescales in the DMA curve driving model predictions. CONCLUSION: DMA combined with CNNs enhances HRV-based mortality risk stratification and reveals distinct physiological scaling patterns associated with survival outcomes. SIGNIFICANCE: This study highlights the potential of DMA and CNNs in improving mortality risk stratification and providing mechanistic insights into HRV dynamics, with implications for personalized health monitoring.
4 May 2026
Read appraisal →American Society of Clinical Oncology educational book. American Society of Clinical Oncology. Annual Meeting
Harnessing Artificial Intelligence for the Management of Patients With GI Cancers
Artificial intelligence (AI) has the potential of reshaping GI oncology by enabling more nuanced interpretation of complex clinical, imaging, and molecular data, while supporting more timely and patient-centered decisions. This article synthesizes perspectives across the GI cancer continuum, beginning with a framework for context-aware AI that emphasizes metadata, multimodal integration, and lifecycle quality and safety as foundations for trustworthy tools that clarify, rather than conceal, uncertainty. Next, AI in endoscopy is highlighted as an example in clinic practice, focusing on computer-aided detection and diagnosis systems that not only increase adenoma detection rates but also raise questions about surveillance burden, real-world effectiveness, and the balance between skill enhancement and potential deskilling of endoscopists. Another section explores how AI can help GI oncologists design, prioritize, and implement highly innovative clinical trials-particularly multi-omic and imaging-driven approaches-while envisioning a future in which far more patients participate in trials that align with their goals and values. The final section reviews emerging AI-enabled clinical trial matching pipelines, including large language model-based retrieval and prescreening tools that operate on real-world electronic health record and protocol data, and discusses challenges related to bias, privacy, explainability, and workflow integration. Together, these contributions argue that the greatest impact of AI in GI oncology will come from deliberately aligning technical capabilities with highly relevant patient-centered clinical questions, ethical governance, and implementation strategies that expand access to trials and improve outcomes for patients with GI malignancies.
2 May 2026
Read appraisal →JMIR medical informatics
Deep Learning Model Using Transfer Learning for Detecting Left Ventricular Systolic Dysfunction: Retrospective Algorithm Development and Validation Study
BACKGROUND: Artificial intelligence-augmented electrocardiogram (AI-ECG) models for detecting left ventricular systolic dysfunction (LVSD) often exhibit degraded performance in patients with comorbidities. OBJECTIVE: This study aimed to introduce and validate a recalibration method using longitudinal patient data to enhance prediction accuracy and simulate its clinical utility for ongoing monitoring. METHODS: We conducted a multicenter, retrospective cohort study using data from 2 hospitals in Korea. A dataset of paired transthoracic echocardiograms (TTEs) and electrocardiograms (ECGs) matched within a 2-week interval was constructed, separating pairs into baseline (first for each patient) and follow-up assessments. In addition to conventional supervised learning, we developed a patient-wise recalibration strategy that incorporated historical left ventricular ejection fraction measurements and prior AI-ECG outputs to adjust for future predictions, thus empirically mitigating confounding effects. Pretraining was also implemented to enhance the model's performance. RESULTS: The recalibrated 12-lead DeepECG LVSD model achieved an area under the receiver operating curve of 0.956 (95% CI 0.946-0.965) for internal validation and 0.940 (95% CI 0.936-0.945) for external validation of follow-up TTE-ECG pairs. The uncalibrated 12-lead DeepECG LVSD model also showed modest performance, with an area under the receiver operating curve of 0.953 (95% CI 0.941-0.965) in the internal validation and 0.947 (95% CI 0.943-0.951) in the external validation when tested on baseline TTE-ECG pairs. Recalibration yielded statistically significant improvements in the 12-lead DeepECG LVSD models (P<.001), with enhanced and more balanced performance across all clinical subgroups. CONCLUSIONS: Patient-wise recalibration improved accuracy and consistency across various comorbidities by mitigating performance degradation and bias. This broadens the application of AI-ECG for LVSD detection from low-risk screening to high-risk longitudinal monitoring.
27 Apr 2026
Read appraisal →Biomedical physics & engineering express
Impact of multimodal information on the estimation of fetal growth indicators using machine learning regression models
Accurate estimation of fetal growth indicators such as birth weight, birth length, and gestational age at birth is essential for monitoring pregnancy outcomes and guiding clinical decisions. Traditional predictive models typically rely on ultrasound-based fetometric data to estimate fetal weight or length. While valuable, these models provide estimates only at the time of measurement, rather than predicting values at birth, and may overlook important clinical and sociodemographic factors that also influence fetal growth. This study aimed to evaluate whether incorporating echographic, clinical, and sociodemographic features could improve the accuracy of predicting fetal growth indicators at birth and to quantify the contribution of each variable. Data from 154 cases were collected and processed for model development (61.5% for training and 38.5% for testing), divided into three feature sets: fetometric, clinical-sociodemographic, and combined clinical, echographic, and sociodemographic data. Six regression models were developed to predict three fetal growth indicators: birth weight, birth length, and gestational age at birth. Model performances were assessed usingR2, mean absolute error (MAE), and mean absolute percentage error (MAPE). The multimodal models significantly outperformed those relying only on fetometric or clinical-sociodemographic data, with the random forest achieving the best performance for birth weightR2: 0.8991; MAE: 255.08 g; MAPE: 8.46%, birth lengthR2: 0.8679; MAE: 7.21 cm; MAPE: 2.73%, and gestational age at birthR2: 0.8886; MAE: 1.34 d; MAPE: 2.76%. Feature relevance analysis revealed that variables such as maternal height, maternal weight, placenta location, and alcohol consumption played substantial roles in prediction accuracy, alongside classic fetometric measurements such as head circumference. These findings highlight the multifactorial nature of fetal growth and demonstrate that integrating clinical and sociodemographic information enhances the performance of fetal growth prediction models, ultimately supporting improved perinatal care.
25 Apr 2026
Read appraisal →Journal of medical economics
Value frameworks for digital health technologies: a comparative analysis
AIMS: This study aimed to review and compare existing value frameworks for digital health technologies (DHTs) and identify common domains to guide their evaluation. MATERIALS AND METHODS: A comprehensive literature search was conducted across PubMed, Embase, and Google Scholar, along with citation and website searches, was conducted in August 2025. English-language publications that provided guidance on evaluating value of DHTs were included. A narrative review approach was employed to synthesize findings and identify key value domains and indicators. RESULTS: From an initial pool of 1,132 articles, four met the inclusion criteria and formed the basis of the review, supplemented by additional frameworks from relevant organizations. Six core value domains were identified: technical and security aspects, data rights and governance, clinical characteristics, economic characteristics, health inequalities, and user preferences. Five value frameworks from the United Kingdom (UK) National Health Service, Digital Therapeutic Alliance, World Bank, Institute for Clinical and Economic Review-Peterson Health Technology Institute, and UK National Institute for Health and Care Excellence were compared across these domains. A heat map analysis demonstrated variation in the inclusion of indicators, reflecting differing stakeholder priorities and perspectives. LIMITATIONS: Limitations include potential bias, subjectivity, and limited scope inherent in a narrative review, although elements of PRISMA were applied to reduce risks. Non-English frameworks and those specific to certain technologies were excluded. Additionally, frameworks for both DHTs and digital therapeutics were considered together. CONCLUSION: Six core value domains are consistently emphasized across value frameworks. However, variations across value frameworks highlights the influence of stakeholder needs and priorities. Further research is needed to refine these frameworks and support consistent and informed decision-making in DHTs adoption and implementation. Digital health technologies (DHTs) are increasingly used in healthcare, but their value is assessed in variable ways. We reviewed published studies to identify common elements. Six key areas emerged: technical and security aspects, data rights and governance, clinical and economic characteristics, health inequalities, and user preferences. While some areas are consistently included, others vary across frameworks. Our findings suggest that more consistent approaches could support clear decisions about adopting digital health tools.
25 Apr 2026
Read appraisal →Renal failure
Explainable machine learning integrating bioelectrical impedance for 6-month cardiovascular risk in peritoneal dialysis
Peritoneal dialysis (PD) is a common treatment for end-stage renal disease, yet cardiovascular disease (CVD) remains a major cause of morbidity. Inadequate fluid management can elevate CVD risk, and bioelectrical impedance spectroscopy (BIS) is increasingly used to assess fluid status. This study aimed to develop an artificial intelligence model integrating BIS measurements and medical history to predict major adverse cardiovascular events (MACEs) within six months in clinically stable PD patients. Data were stratified by MACE occurrence and subject, with 80% assigned to training and 20% to testing. Class imbalance was addressed using the synthetic minority over-sampling technique. Four algorithms - logistic regression, random forest, XGBoost, and deep neural networks - were trained using fivefold cross-validation and grid search for optimal hyperparameters. Model performance was evaluated with area under the ROC curve (AUC), calibration plots, and decision curve analysis (DCA). Feature ablation experiments compared models using all 15 features, only 11 BIS-recorded features, and only 4 medical history features. The random forest model achieved the highest performance (AUC = 0.88), with CVD history as the most influential predictor. Isotonic regression calibration improved probability alignment (brier score = 0.0451). DCA suggested potential clinical benefit. While the model relied on four medical history features, its initial sensitivity was modest (0.68). Integrating BIS features significantly enhanced diagnostic sensitivity (0.84). The random forest model, based on 15 clinically accessible features, accurately predicts the risk of MACE within six months in clinically stable PD patients, demonstrating strong discriminatory ability with a performance reaching 88%.
24 Apr 2026
Read appraisal →PloS one
Hybrid feature-selection and diversity-guided stacking framework for interpretable ensemble learning: Application to COVID-19 mortality prediction
BACKGROUND: Reliable predictive modeling in high-dimensional biomedical data requires a balance between accuracy, interpretability, and computational efficiency. However, existing ensemble methods often overlook model diversity or rely on ad hoc feature-selection approaches, which limit generalizability. This study introduces a hybrid feature-selection and diversity-guided stacking framework designed to improve robustness and scalability across clinical and other data-intensive domains. METHODS: The proposed framework integrates a hybrid feature-selection pipeline-combining Variance Inflation Factor (VIF), Analysis of Variance (ANOVA), Sequential Backward Elimination (SBE), and Lasso regression-to reduce multicollinearity and overfitting. It also employs a diversity-aware stacking strategy that constructs sub-model sets based on pairwise diversity measures (Disagreement, Yule's Q, and Cohen's Kappa) and non-pairwise metrics (Entropy and Kohavi-Wolpert). Sixteen base classifiers and five meta-learners were trained using repeated 10-fold cross-validation. The framework was evaluated using data from 4,778 hospitalized COVID-19 patients with 116 clinical and laboratory attributes, preprocessed using robust scaling and ROSE-based class balancing. RESULTS: The optimal configuration, which stacked Random Forest and XGBoost models using a Neural Network meta-learner, achieved 91.4% accuracy (95% CI: 89.8-92.8), AUC = 0.955, F1 = 0.801, and MCC = 0.746, outperforming the best individual model (AdaBoost, 90.2%). Training time (~450 s) and per-case inference time (<0.2 s) demonstrated computational feasibility. Feature-importance analysis and SHAP-based interpretation confirmed clinical relevance and interpretability. CONCLUSIONS: The hybrid feature-selection and diversity-guided stacking framework improves predictive accuracy and interpretability while maintaining computational efficiency. Although validated using COVID-19 mortality data, the approach is broadly applicable to biomedical, environmental, and engineering prediction tasks that require interpretable and scalable ensemble learning.
19 Apr 2026
Read appraisal →JMIR medical informatics
Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study
BACKGROUND: Physicians routinely document specifics of patient encounters in clinic visit notes, a critical but potentially time-consuming task. Ambient listening artificial intelligence (AI) technology is being integrated into clinical workflows to reduce documentation burden by creating draft visit notes. While this technology is promising, it is not perfect, and the potential for patient harm needs to be understood and mitigated. We developed and piloted an efficient, standardized approach to evaluating AI-generated notes for safety concerns in ambulatory care visits. OBJECTIVE: The objective of this quality improvement project was to develop and pilot an efficient, standardized, and scalable approach to evaluating AI-generated notes for safety concerns in ambulatory care visits. METHODS: During a 2-month pilot (July to August 2024), 31 physicians across multiple specialties used an ambient listening AI scribe to assist with the creation of 7545 clinic notes. A novel survey instrument was developed to assess note quality, focusing on 4 error types: accidental inclusions, accidental omissions, hallucinations, and bias. Physicians evaluated 356 (4.7%) AI-generated notes. Where an error was present, physicians rated its severity based on its potential to cause patient harm if it was not corrected, on a 0 to 5 scale. Additionally, a vendor-reported metric on the percentage of note content edited by physicians was analyzed. RESULTS: Of the 356 evaluated notes, accidental omissions were the most frequent error (n=64, 18%), followed by hallucinations (n=41, 11.5%), and accidental inclusions (n=33, 9.3%). Bias was rare (n=4, 1.1%). Most (119/142, 83.8%) errors were rated as mild to moderate (severity 1-3), with only 19 (5.3%) notes containing errors rated as posing serious or imminent risk (severity 4-5). Editing metrics across all AI-created notes showed a median of 9.0% (IQR 2.5%-21.9%) of AI-generated words were changed, with 14.9% (143/960) of notes left entirely unedited. Physician editing practices varied widely, with average percentages of AI-generated words changed ranging from 1.9% to 69.3% (median 9.0%, IQR 2.5%-21.9%). CONCLUSIONS: AI-generated clinical notes were generally of high quality, with 94.7% (337/356) free from significant errors. However, because a small number contained errors that carried the risk of serious harm if not corrected, careful clinician review of notes remains imperative. Prior to deploying an AI scribe, organizations should pilot the technology and include an efficient review process to understand the nature and type of errors common at their organization. This pilot provides a scalable model for other health systems seeking to implement AI scribe technology responsibly.
18 Apr 2026
Read appraisal →BMJ open respiratory research
Personalised management of community-acquired pneumonia and the role of Clinical Decision Support Software
Community-acquired pneumonia (CAP) is one of the most common causes of hospital admission and is associated with significant morbidity and mortality. National and international guidelines are available to guide the management of CAP, including antibiotic prescribing. However, these guidelines are often not adhered to and there is significant overprescribing of broad-spectrum antibiotics, contributing to the growing pandemic of antimicrobial resistance. Clinical Decision Support Software (CDSS) are electronic tools that use individual patient data to generate patient specific assessments or recommendations, that can then be acted on by the patient or clinical decision maker. This article reviews the evidence surrounding the initial management of CAP in hospital and considers the potential of CDSS to support CAP management.
17 Apr 2026
Read appraisal →Clinica chimica acta; international journal of clinical chemistry
Artificial intelligence and machine learning applications in dialysis: Current applications, challenges, and future directions.
Artificial intelligence (AI) and machine learning (ML) applications have emerged as transformative technologies in nephrology, particularly in dialysis care. The availability of multimodal datasets including electronic health records, hemodialysis machine data, laboratory values, and imaging has enabled the development of sophisticated prognostic, diagnostic, and treatment support tools. This comprehensive review narratively examines current AI/ML applications in dialysis, evaluates their clinical performance, and identifies future research directions. We conducted a comprehensive literature search of PubMed, Web of Science, and other major databases from 2020 to 2025, focusing on peer-reviewed studies that employed AI/ML techniques in dialysis care. Studies were categorised by application domain and analysed for methodology, performance metrics, and clinical implications. Our analysis identified five major application domains: (1) prediction and prognosis, including intradialytic hypotension (IDH) prediction with AUROC values ranging 0.89-0.95, mortality prediction achieving C-indices up to 0.83, and hospitalisation risk assessment; (2) early detection of chronic kidney disease (CKD) progression and dialysis risk; (3) clinical decision support for anaemia management and treatment optimisation; (4) vascular access monitoring with AI-driven image analysis achieving AUROC ≈ 0.96; and (5) natural language processing (NLP) applications for symptom detection. Federated learning (FL) approaches are emerging to enable multi-centre collaboration while preserving data privacy. AI/ML technologies demonstrate significant promise in enhancing dialysis care through improved prediction accuracy, personalised treatment approaches, and clinical decision support. However, widespread clinical adoption remains limited due to challenges including data privacy concerns, model interpretability issues, regulatory complexity, and the need for diverse, representative datasets.
16 Apr 2026
Read appraisal →JMIR medical informatics
Code-Based Versus AutoML Methods for Pill Recognition in Clinical Settings: Comparative Performance Study
BACKGROUND: Visual identification and verification of medications during dispensing and administration are prone to human error, particularly in high-pressure and high-volume clinical settings. Misidentification can lead to medication errors, posing risks to patient safety and placing a burden on health care systems. Recent advances in computer vision and object detection offer promising solutions for automated solid oral dosage form (pill) recognition. However, comprehensive studies comparing code-based and no-code (automated machine learning [AutoML]) approaches for pill recognition are lacking. OBJECTIVE: This study aimed to evaluate and compare the performance, cost, usability, and deployment feasibility of pill recognition models developed with Ultralytics YOLO11 and 3 cloud-based AutoML platforms (Amazon Rekognition Custom Labels, Google Vertex artificial intelligence [AI] AutoML Vision, and Microsoft Azure Custom Vision) using multiple datasets, including real-world clinical images. METHODS: Five training subsets of increasing size (1230, 3450, 7380, 14,400, and 26,880 images) from 30 commonly dispensed medications were used to train models on YOLO11 and 3 AutoML platforms. Models were evaluated on 6 datasets from different environments: clinical images from 3 hospitals, a verification dataset, a laboratory dataset, and an exhaustive testing set. Performance metrics, including accuracy, precision, recall, and mean average precision, were calculated. We evaluated the impact of training data size on performance and benchmarked training time, platform costs, and limitations. RESULTS: No single platform dominated across all test environments. On the verification dataset (optimal conditions), accuracy ranged from 80.83% (YOLO11) to 91.60% (Google Vertex AI) when trained with the full training dataset. YOLO11 showed consistent performance improvement with increasing training data (accuracy: 63.06%-80.83%) and achieved near-perfect precision and mean average precision scores (0.95-1.00). Google Vertex AI reached above 90% accuracy on 3 training subsets but showed unpredictable declines. Amazon Rekognition maintained near-perfect precision (0.92-1.00) but had the highest false negative rates (up to 0.74), missing many pills. Custom Vision demonstrated steady performance improvements (77.08%-85.62% accuracy) but lagged behind other AutoML platforms, probably due to its older YOLOv2-based architecture. On clinical datasets, accuracy fluctuated (20.62%-90%) depending on the dataset and platform. Training costs and time varied: YOLO11 (open-source), Microsoft Azure (US $9.50-US $28.60, allowed user-predefined training duration), Google Vertex AI (US $69.30 with consistent 2.5-3-hour training times), and Amazon Rekognition (US $5.43-US $43.89 with size-dependent training time scaling, reaching nearly 40 hours on the full 26,880-image dataset). CONCLUSIONS: Each platform offers distinct advantages and trade-offs: YOLO11 provides the highest flexibility and lowest platform costs but requires technical expertise, while AutoML platforms can offer high performance at a higher cost but with limited user control, introducing unpredictability. The performance variations demonstrate that successful clinical deployment requires careful platform selection based on specific performance requirements, budget constraints, and available technical resources, followed by rigorous validation using real-world, representative data to ensure patient safety in clinical workflows.
11 Apr 2026
Read appraisal →PloS one
Development and evaluation of a multimodal feature-based predictive model for radiotherapy-induced oral mucositis in nasopharyngeal carcinoma
BACKGROUND: Accurate prediction of radiation-induced oral mucositis is crucial for personalized treatment in head and neck cancer. However, developing robust predictive models utilizing high-dimensional multimodal data (CT imaging, dose distribution, and clinical features) remains challenging, particularly in cohorts with limited sample sizes. OBJECTIVE: This study aimed to rigorously evaluate and compare the multi-class predictive performance of traditional machine learning algorithms and deep learning architectures under a small-cohort setting. METHODS: Multimodal data from 108 patients were collected. A comprehensive evaluation framework incorporating nine traditional machine learning algorithms and two deep learning models (a dimensionality-reduced 1D-CNN and a multimodal 3D-CNN) was established. To ensure robust evaluation, a stratified 5-fold cross-validation was employed. Model performance was comprehensively quantified using mean ± standard deviation (SD) across multiple metrics, including the Area Under the Curve (AUC), accuracy, and Matthews Correlation Coefficient (MCC). RESULTS: Inter-rater reliability for RIOM grading was excellent (Cohen's kappa = 0.82, 95% CI: 0.73-0.91). Among traditional machine learning approaches, the Extra Trees (ET) algorithm achieved the highest discriminative capacity (AUC: 0.956 ± 0.046), while Logistic Regression (LR) demonstrated optimal overall accuracy (0.832 ± 0.155) and stability. Regarding deep learning, the lightweight 1D-CNN utilizing fused low-dimensional features exhibited highly competitive and robust performance (AUC: 0.900 ± 0.072; Accuracy: 0.732 ± 0.140). In stark contrast, the high-dimensional multimodal 3D-CNN suffered from severe overfitting and mode collapse phenomenon, yielding significantly inferior results (AUC: 0.568 ± 0.090; MCC: -0.025 ± 0.031). CONCLUSIONS: For small-cohort radiomics and dosimetric analyses, ensemble learning models (e.g., ET) and appropriately regularized linear models (e.g., LR) remain highly effective. While deep learning holds promise, high-dimensional architectures like 3D-CNNs are highly susceptible to mode collapse without massive datasets. Instead, employing feature dimensionality reduction combined with lightweight networks (1D-CNN) is a vastly superior strategy to extract reliable predictive patterns from limited clinical data.
10 Apr 2026
Read appraisal →PloS one
An interpretable deep learning framework for predictive modeling of postoperative infections in ICU patients
A significant proportion of intensive care unit (ICU) patients undergo surgical procedures, and some may develop postoperative infections. Accurately predicting postoperative infection risk and identifying key contributing factors is crucial for improving postoperative management and understanding infection mechanisms. However, this task is challenging due to the complex interplay of multiple risk factors. While machine learning models can model these intricate associations to predict postoperative infection risk, their lack of interpretability - failing to uncover each factor's impact-hinders their adoption in clinical settings. To address this difficulty, we introduced an interpretable deep neural network (DNN) model that integrates a permutation feature importance test (PermFIT). PermFIT rigorously evaluates the impact of each feature on postoperative infection risk through a rigorous statistical inference. By using only the identified important features as inputs, the DNN's predictive performance can be further enhanced. We conducted an extensive study using electronic health records (EHRs) from the Medical Information Mart for Intensive Care (MIMIC-III), a large-scale ICU EHR database. Under the PermFIT framework, our DNN model effectively identifies significant factors associated with postoperative infections while delivering the most accurate postoperative infection risk predictions. These findings highlight the clinical utility of our proposed DNN framework in managing postoperative care for ICU surgical patients, ultimately improving their health outcomes.
10 Apr 2026
Read appraisal →Journal of medical Internet research
Machine Learning in Predicting the Risk of Esophagogastric Variceal Bleeding Among Patients With Liver Cirrhosis: Systematic Review and Meta-Analysis
BACKGROUND: Liver cirrhosis (LC) can lead to several complications. Esophageal variceal bleeding (EVB) and esophagogastric variceal bleeding (EGVB) are particularly severe, leading to a high risk of mortality. Early identification of esophageal varices and esophagogastric varices is essential. Several studies have constructed prediction models for EVB and EGVB in patients with LC. However, robust systematic evidence to prove their performance is lacking. OBJECTIVE: We included original studies that developed prediction models for esophageal or gastric variceal bleeding in patients with LC under different modeling variables. This study aimed to review the predictive performance of various models for EVB or EGVB in patients with LC, providing insights into the development or updating of simplified scoring tools in the future. METHODS: PubMed, Web of Science, Embase, and the Cochrane Library were searched up to August 21, 2024, to collect original full-text studies on machine learning (ML) in the prediction of EVB and EGVB in patients with LC. The models were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Subgroup analyses were carried out based on the modeling variables. RESULTS: In total, 21 studies were included, with 7011 patients with LC, among whom 1412 (20.14%) developed EVB and 733 (10.45%) developed EGVB. The meta-analysis results suggested that the pooled c-index, sensitivity, and specificity of the prediction model for predicting EVB in the validation set were 0.85 (95% CI 0.77-0.92), 0.93 (95% CI 0.87-0.96), and 0.66 (95% CI 0.46-0.82), respectively. The pooled c-index, sensitivity, and specificity of the prediction model for predicting EGVB in the validation set were 0.89 (95% CI 0.85-0.94), 0.77 (95% CI 0.66-0.85), and 0.81 (95% CI 0.67-0.90), respectively. The subgroup analysis based on modeling variables revealed that, for predicting EVB, the c-index in the validation set was 0.84 (95% CI 0.80-0.88) for models based on clinical features, 0.82 (95% CI 0.69-0.96) for radiomics-based models, 0.78 (95% CI 0.67-0.89) for models based on radiomics and clinical features, and 0.97 (95% CI 0.95-1.00) for models based on endoscopic features. Subgroup analyses based on modeling variables revealed that, for predicting EGVB, the c-index in the validation set was 0.91 (95% CI 0.86-0.96) for models based on clinical features and 0.85 (95% CI 0.75-0.96) for models based on radiomics and clinical features. CONCLUSIONS: ML methods are feasible for predicting EVB and EGVB in patients with LC. Nevertheless, the number of included original studies is limited. In the future, more studies with larger sample sizes are needed to promote the application of ML in the early assessment of EVB and EGVB in patients with LC in clinical practice.
9 Apr 2026
Read appraisal →Journal of the American Heart Association
Predicting Recurrence and Outcomes After Stressor-Associated Atrial Fibrillation Using ECG-Based Deep Learning
BACKGROUND: Stressor-associated atrial fibrillation (AF) refers to new-onset AF that occurs with a reversible, acute stressor. Identifying individuals at highest risk for AF recurrence is essential to guide management. Although clinical factors have shown limited value, the utility of contemporary artificial intelligence (AI)-enabled models using the 12-lead ECG to estimate recurrence risk remains unknown. METHODS: We retrospectively analyzed consecutive primary care and cardiology patients with stressor-associated AF occurring during hospitalization. We quantified the cumulative incidence of recurrence accounting for death as a competing risk. We investigated the relationship between time-varying recurrence and a composite end point of AF-related adverse events (stroke, heart failure, all-cause death) using Cox models. We then developed and validated a penalized regression model to predict recurrence using clinical factors, stressor type, and AF risk estimates from a previously validated ECG-based AI model. RESULTS: We analyzed 3371 patients with stressor-associated AF (mean age, 69±12 years; 40% women). Over a median of 3.7 years (interquartile range, 1.8-7.2), the 10-year cumulative incidence of AF recurrence was 41% (95% CI, 39-44). AF recurrence was strongly associated with AF-related adverse events (hazard ratio, 2.24 [95% CI, 1.81-2.76]). A model incorporating clinical factors, stressor type, and ECG-based AI model AF risk estimates (clinical-AI) discriminated AF recurrence (area under the receiver operating characteristic curve, 0.768 [95% CI, 0.707-0.830]) favorably compared with clinical features (area under the receiver operating characteristic curve, 0.707 [95% CI, 0.642-0.772]; P<0.05). CONCLUSIONS: AF recurrence rates following stressor-associated AF are considerable and are associated with substantially higher risk of adverse cardiovascular events. Models incorporating ECG-based AI risk estimates may prioritize individuals for intensive monitoring and preventive interventions.
8 Apr 2026
Read appraisal →Hematology (Amsterdam, Netherlands)
Machine learning in the prediction of liver iron concentration and iron chelation therapy adjustment
OBJECTIVES: Iron chelation therapy titration is managed by hematologists who monitor iron levels and adjust medication dosages to achieve patient outcomes. This study developed a machine learning (ML) algorithm to predict the liver iron concentration (LIC) and chelation therapy adjustments. METHODS: This retrospective, single-centre cohort study included adult patients who underwent FerriScan MRI to obtain LIC for chelation therapy monitoring. Using an XgBoost-based ML framework, the proximal-time model (PTM) utilised clinical/drug, laboratory and MRI data features from one visit prior to the target visit, whereas the all-time model (ATM) utilised the data from all prior visits. RESULTS: 94 patients with 892 consecutive visits between January 2008 and November 2018 were included in this study. We assessed the prediction capabilities of the PTM and ATM in LIC, changes to chelation drug type and dosage changes. The PTM model was superior to the ATM model in all the experiments. When drug features were excluded, the CLICT model for predicting patient iron overload status improved to an AUROC of 0.83 [95% CI 0.75-0.91] for PTM; compared to an AUROC of 0.73 [95% CI 0.66-0.80] when drug features were included.For predicting changes in chelation type, the CLICT model showed AUROC of 0.83 [95% CIs 0.77-0.89] for PTM. There is high concordance of the agreement of hematologists with ML in not changing the chelation drug type. DISCUSSION/CONCLUSION: The ML model is a step toward creating a clinical decision support system tool for the prediction of LIC and iron chelation therapy adjustment in patients with haemoglobinopathies or hemolytic anemias.
5 Apr 2026
Read appraisal →Yonsei medical journal
Predicting Post-Induction Hypotension in Diverse Surgical Populations: A Multiclass Classification Universal Model Using Machine Learning Techniques
PURPOSE: Our study aims to develop a machine learning model that not only predicts the occurrence of post-induction hypotension (PIH) but also assesses its severity, addressing a broader patient population than previous studies which mostly focused on a single population. MATERIALS AND METHODS: In our study, we extracted data from 71473 patients aged 18 years and older who underwent general anesthesia for non-cardiac surgery at a tertiary care hospital, using the electronic medical record system for modeling. We used patient demographics, baseline and pre-induction blood pressure, preoperative laboratory tests, surgical details, and anesthetics data, focusing on predicting PIH. The severity of hypotension was assessed by integrating and calculating the integral value of hypotensive periods. We employed several machine learning techniques and evaluated their performance using accuracy, precision, F1-scores, and macro-averaged area under the curve. Additionally, SHapley Additive exPlanation values were used to determine the key factors influencing the predictions. RESULTS: A multiclass classification model, which assesses hypotension severity through the area of hypotension, surpassed the binary model with an F1-score of 0.664. Among various machine learning techniques, the eXtreme Gradient Boosting (XGBoost) model exhibited superior prediction performance, achieving an accuracy of 0.755 and an F1-score of 0.664. Models using preoperative blood pressure and demographic data outperformed those using other datasets. CONCLUSION: We found that using the XGBoost ensemble machine learning technique aids in predicting PIH before surgery, and introducing a multiclass classification model that indicates the severity of hypotension to clinicians leads to an overall enhancement in prediction performance, thereby increasing its clinical utility for real-world applications.
4 Apr 2026
Read appraisal →Heart rhythm
Machine learning-enabled assessment of risk of drug-induced QT prolongation at the time of prescribing
BACKGROUND: Many medications are associated with long QTc. Current long QTc predictors have limited generalizability and/or modest performance. OBJECTIVE: This study aimed to compare the performance of machine learning-enabled approaches to drug-induced long QTc prediction with current risk scores, including Tisdale and RISQ-PATH. METHODS: We identified patients who began QTc-prolonging medications (QTdrugs) and had follow-up 12-lead electrocardiograms (ECGs) within 1 year. Using 5-fold cross-validation, we trained XGBoost (XGB), an ECG-based deep neural network, and combined models using electronic health record data and ECG traces to predict QTc of ≥500 ms within 1 year. We assessed performance using the area under the receiver operating characteristic (AUROC) curves and compared them with corresponding Tisdale and RISQ-PATH scores. RESULTS: QTdrug records were identified for 345,371 patients, 5.7% of whom had long QTc within 1 year. In the subset with baseline ECGs available (N = 182,448; 7.7% events), both the XGB and deep neural network models demonstrated high performance (AUROC 0.869 and 0.864, respectively) but their combination yielded no significant improvement (AUROC 0.874). Therefore, focusing on the XGB model, we observed superior performance vs RISQ-PATH in the overall population (AUROC 0.859 vs 0.701), as well as Tisdale in predominantly inpatients (N = 110,558; 8.8% events; AUROC 0.855 vs 0.770). Positive predictive value was 61.5% vs 28.35% and 54.0% vs 28.3% at equivalent operating points for the XGB model vs RISQ-PATH and Tisdale, respectively. CONCLUSION: Development and retrospective validation of 3 machine learning-based models for predicting drug-induced long QTc at the time of new QTdrug starts demonstrated superior performance compared with current clinical risk calculators and may be useful tools to support medical decision making when initiating new therapies.
4 Apr 2026
Read appraisal →Journal of medical Internet research
Predictive Value of Machine Learning for Poststroke Mortality Risk: Systematic Review and Meta-Analysis
BACKGROUND: People with stroke face a high mortality risk, and an accurate prediction model is essential to the guidance of clinical decision-making in this population. Recently, with growing attention paid to machine learning (ML) in stroke care, some researchers have investigated the effectiveness of ML in predicting the mortality risk in stroke. However, systematic evidence is still lacking for its effectiveness. OBJECTIVE: This systematic review aims to evaluate the value of ML in predicting the stroke mortality risk. The findings are expected to offer an evidence-based basis for developing and assessing clinical risk prediction tools. METHODS: A search was made in Cochrane Library, PubMed, Embase, and Web of Science up to June 23, 2025, and studies that reported a complete performance of ML in predicting stroke mortality were included. Studies with only risk factors analyzed were excluded. The risk of bias of the included studies was assessed using PROBAST (Prediction model Risk of Bias Assessment Tool). Pooled risk ratios with 95% CIs and prediction intervals (PIs) were derived using the Hartung-Knapp-Sidik-Jonkman method under a random-effects model. Subgroup analyses were also conducted by model type, stroke type, patient source, and treatment background. Moreover, a metaregression was conducted on the C-index for out-of-hospital mortality at different time points to explore the influence of time factors on the model's predictive performance. RESULTS: Sixty-eight studies were included (23 predicting in-hospital mortality and 45 predicting out-of-hospital mortality), describing the development of 75 prediction models and 43 external validations. The follow-up period was 1 month to 15 years. For predicting in-hospital mortality, the external validation set had a pooled C-index of 0.727 (95% CI 0.677-0.781, 95% PI 0.521-1.000), with sensitivity and specificity of 0.64 (95% CI 0.57-0.70) and 0.74 (95% CI 0.70-0.77), respectively. For predicting out-of-hospital mortality, the pooled C-index was 0.847 (95% CI 0.808-0.887, 95% PI 0.750-0.956) in the external validation set, with sensitivity and specificity of 0.71 (95% CI 0.55-0.82) and 0.76 (95% CI 0.74-0.78), respectively. Comparatively, the overall pooled C-indexes were 0.788 (95% CI 0.766-0.810, 95% PI 0.621-0.999) and 0.812 (95% CI 0.798-0.826, 95% PI 0.693-0.952), respectively. The metaregression revealed a gradual decline in the predictive performance of the overall model and logistic regression model alone, whereas a random forest model maintained sustained performance. Age, National Institutes of Health Stroke Scale score, and stroke-related complications were the most frequently used variables for modeling. CONCLUSIONS: This is the first meta-analysis to demonstrate that ML-based prediction of stroke mortality is feasible. The performance of ML supports its role as an auxiliary tool for identifying high-risk populations, thereby optimizing clinical monitoring and resource allocation. However, due to substantial heterogeneity and a relatively high risk of bias in available studies, caution is warranted in real-world application. The effectiveness of ML may vary across settings, and external validation is recommended before broader implementation. TRIAL REGISTRATION: PROSPERO CRD420251086321; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251086321.
4 Apr 2026
Read appraisal →IEEE transactions on bio-medical engineering
Explainable Prediction of ICU Transfer in Acute Pancreatitis: A Neural Network Model With SHAP Interpretability for Clinical Decision Support
OBJECTIVE: Acute pancreatitis (AP) is a life-threatening disorder commonly observed in emergency departments. Patients with acute pancreatitis may necessitate transfer to the intensive care unit (ICU) if standard treatments prove ineffective. The development of a machine learning (ML) model that can precisely forecast the necessity for ICU admission in patients with acute pancreatitis might dramatically enhance in-hospital healthcare for patients. METHODS: For this study, data were extracted from the MIMIC-III Clinical Database Care-Vue subset (MIMIC-III v 1.4), Medical Information Mart for Intensive Care-IV (MIMIC-IV v 3.1), and eICU Collaborative Research Database (v 2.0). Fifteen features were identified through Boruta's algorithm, Least Absolute Shrinkage and Selection Operator (Lasso), and Recursive Feature Elimination (RFE) utilizing the XGBoost. RESULTS: Seven machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Neural Network (NNET), etc. were developed based on these features. The neural network approach demonstrated the highest reliability, attaining an AUC of 0.891, sensitivity of 0.852, and specificity of 0.767 in the test set. Meanwhile, an interpretability assessment of the most robust models was performed utilizing SHapley Additive exPlanations (SHAP) to prioritize feature importance and elucidate the final model. This model has been deployed on a web platform: https://icupredict.shinyapps.io/shiny6/. CONCLUSIONS: The neural network model effectively predicted whether patients with AP would be transferred to the ICU, demonstrating optimal performance. SIGNIFICANCE: We developed an interpretable ML model for prediction of ICU admission in AP patients. It provides clinicians with a reliable tool for early identification and treatment.
4 Apr 2026
Read appraisal →Journal of medical Internet research
Artificial Intelligence, Connected Care, and Enabling Digital Health Technologies in Rare Diseases With a Focus on Lysosomal Storage Disorders: Scoping Review
BACKGROUND: Rare diseases affect more than 300 million people globally, and only about 5% have approved therapies. Lysosomal storage disorders (LSDs) exemplify the diagnostic and long-term care complexity typical of rare diseases, and digital health technologies (DHTs), especially artificial intelligence (AI) and connected care (CC), are emerging tools to support LSD management. OBJECTIVE: We aimed to map and synthesize peer-reviewed and gray literature from the past decade on DHTs relevant for LSD care, with a primary analytic focus on AI-enabled and CC solutions and a contextual mapping of other enabling DHTs. Evidence distribution was charted by population, care-journey phase, and outcome domains to identify gaps, methodological limitations, and timely priorities relevant for research, clinical practice implementation, and policies. METHODS: We conducted a scoping review guided by a population, concept, context framework and operationalized through a Population, Intervention, Comparison, and Outcome (PICO)-informed data-charting structure to map study characteristics and reported outcomes, without causal or effectiveness assumptions and without risk-of-bias assessment. We searched PubMed, Google Scholar, and ClinicalTrials.gov for studies published between October 2015 and September 2024, complemented by AI-assisted discovery tools for citation extension. Reproducibility logs (search strings, run dates, filters, and stepwise counts) were maintained. Of 1751 records retrieved, 245 were included. Evidence was charted by LSD population, intervention class (AI, CC, and other enabling DHTs), outcome domains (patient, health care, and societal), and phase of the care journey. RESULTS: Among 245 included records, 92.2% (226/245) were peer-reviewed, and 7.8% (19/245) were gray literature; no completed and published randomized controlled trials or LSD-specific systematic reviews were identified, with evidence dominated by small, single-center observational studies. Overall, 40 peer-reviewed records reported AI-driven DHTs, 89 reported CC DHTs, and 144 reported other enabling DHTs (some multilabeled). Evidence was concentrated mostly in Gaucher and Fabry diseases. Nearly half of the mapped literature focused on screening and diagnosis, with fewer records addressing treatment intensification, rehabilitation, and end-of-life care. Outcomes were predominantly health care delivery performance measures, with fewer patient and societal outcomes. AI applications mainly supported diagnostic decision support, phenotyping, monitoring, tracking, and risk stratification; CC commonly involved telemedicine, remote monitoring, and patient-engagement platforms; enabling DHTs included interoperable data systems, registries, and digital infrastructures. CONCLUSIONS: The evidence base is appreciable for a niche field and reflects growing interest in AI and CC for LSD care, but heterogeneity and methodological limitations preclude inferences on effectiveness or routine implementation. This evidence map highlights relatively stronger areas and gaps, providing a structured foundation to inform timely expert consensus-building and research prioritization. Key priorities include interoperable data infrastructures and data availability, prospective multicenter evaluations, transparent reporting of algorithms and workflows, and implementation-relevant outcomes to support safe, equitable, and scalable adoption aligned with evolving European Union and global rare-disease priorities.
4 Apr 2026
Read appraisal →Current opinion in lipidology
Machine learning to predict elevated lipoprotein(a).
PURPOSE OF REVIEW: Lipoprotein(a), Lp(a), is a genetically determined, lifelong risk factor for atherosclerotic cardiovascular disease (ASCVD). Despite broad guideline support for universal one-time testing, Lp(a) measurement remains rare in clinical practice. This review summarizes recent advances in machine learning-based strategies that can enhance the efficiency, yield, and equity of Lp(a) screening. RECENT FINDINGS: To date, three studies have developed and validated machine learning models to identify individuals with elevated Lp(a) using routinely available clinical variables. The ARISE framework, derived from the UK Biobank and validated across multiple US cohorts, reduced the number needed to test by more than 50% while maintaining consistent discrimination across demographic subgroups. Additional studies have confirmed the feasibility of decision-tree and neural network models to improve case finding for elevated Lp(a) in both clinical and population-based settings. SUMMARY: Machine learning-based strategies provide a scalable means of operationalizing universal Lp(a) testing recommendations within health systems. When developed using unbiased data, externally validated, and assessed for fairness and interpretability, these models can support systematic identification of individuals with elevated Lp(a) and integration of Lp(a) measurement into routine cardiovascular risk assessment.
4 Apr 2026
Read appraisal →Transl Psychiatry
Application of artificial intelligence in schizophrenia rehabilitation management: a systematic scoping review
Artificial intelligence (AI) is increasingly used in mental health, yet its rehabilitation-oriented applications in schizophrenia have not been systematically mapped. We conducted a systematic scoping review of PubMed, Web of Science, IEEE Xplore and the ACM Digital Library (January 1, 2012-October 31, 2025; two search rounds), applying operationalized rehabilitation boundaries and excluding diagnostics-only case-control studies. We extracted data on data sources, feature engineering, model families, validation, calibration, interpretability, application domains, outcomes and implementation readiness. Eighty-three studies met inclusion criteria (median sample size 160; 55% longitudinal). Applications focused on symptom monitoring (48/83), medication management (19/83) and risk management (16/83), whereas functional training (1/83) and psychosocial support (3/83) were rarely targeted. Supervised learning predominated (53/83, 63.9%) over representation learning (20/83, 24%), most commonly using speech/text, electronic health records and smartphone sensing. Across classification tasks, the median AUC was 0.79 (IQR 0.71-0.86); relapse early-warning models showed a median sensitivity of 31.5% at 88.0% specificity. Only four studies reported external validation and three described closed-loop deployment, including one randomized trial that improved adherence. Proxy endpoints were more common than clinical endpoints, and reporting of calibration/uncertainty and fairness auditing was sparse. Overall, AI shows promise for monitoring, adherence support and relapse risk stratification, but routine-care deployment will require externally validated and calibrated human-in-the-loop decision support, privacy-preserving multimodal pipelines and pragmatic trials targeting functional outcomes and participation.
1 Mar 2026
Read appraisal →Genes (Basel)
The Evolving Role of Artificial Intelligence in Medical Genetics: Advancing Healthcare, Research, and Biosafety Management
The integration of artificial intelligence (AI) with medical genetics is transforming healthcare by addressing the analytical challenges posed by the vast complexity of multi-omics data. This review explores the synergistic convergence of these fields, highlighting AI's transformative role in enhancing diagnostic precision, enabling non-invasive molecular profiling through imaging-genetics, and advancing predictive and personalized medicine via polygenic risk scores and pharmacogenomics. AI is also emerging as a powerful generative tool in therapeutic design, accelerating drug discovery, protein engineering, and precision gene editing. However, this powerful synergy introduces significant ethical, regulatory, and biosecurity challenges, including data privacy, algorithmic bias, and the dual-use risks of AI-enabled genetic engineering. The future envisions a responsible co-evolution, with multimodal AI and the concept of the Digital Twin driving precision medicine, underpinned by interdisciplinary collaboration to ensure fairness, transparency, and societal trust. This article charts the current landscape and proposes actionable directions, emphasizing the need for robust governance to harness AI's potential while mitigating its risks for the benefit of human health.
1 Dec 2025
Read appraisal →Blood
Artificial intelligence in hematology
Artificial intelligence (AI) and its subdiscipline, machine learning (ML), have the potential to revolutionize health care, including hematology. The diagnosis and treatment of hematologic disorders depend on the integration of diverse data sources, such as imaging, pathology, omics, and laboratory parameters. The increasing volume and complexity of patient data have made clinical decision-making more challenging. AI/ML hold significant potential for enhancing diagnostic accuracy, risk stratification, and treatment response prediction through advanced modeling techniques. Generative AI, a recent advancement within the broader field of AI, is poised to have a profound impact on health care and hematology. Generative AI can enhance the development of novel therapeutic strategies, improve diagnostic workflows by generating high-fidelity images or pathology reports, and facilitate more personalized approaches to patient management. Its ability to augment clinical decision-making and streamline research represents a significant leap forward in the field. However, despite this potential, few AI/ML tools have been fully implemented in clinical practice due to challenges related to data quality, equity, advanced infrastructure, and the establishment of robust evaluation metrics. Despite its promise, AI implementation in hematology faces critical challenges, including bias, data quality issues, and a lack of regulatory frameworks and safety standards that keep pace with rapid technological advancements. In this review, we provide an overview of the current state of AI/ML in hematology as of 2025, identify existing gaps, and offer insights into future developments.
1 Nov 2025
Read appraisal →