Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 87 appraisals
Journal of medical Internet research
Advancing Human-Centered AI in Clinical Decision Support: Sociocognitive Human-in-the-Loop Study in HIV Care
BACKGROUND: AI-powered clinical decision support systems (CDSS) have shown promise in improving prediction, monitoring, and treatment optimization across clinical domains, including HIV care. However, translating AI outputs derived from electronic health records into clinically meaningful, trustworthy, and actionable decision support remains challenging, underscoring the need for more human-centered and socioecologically grounded CDSS design. OBJECTIVE: This study aimed to explore how we can effectively translate the outputs of machine learning models based on HIV electronic health records into a real AI-powered CDSS for HIV care. Using the human-in-the-loop method, we engaged a set of stakeholders, including HIV physicians, nurse practitioners, infectious disease pharmacists, social workers, and case managers. Stakeholders interacted with an AI-powered CDSS prototype to identify barriers and challenges to adoption, as well as to inform a more holistic and context-aware AI-powered CDSS design. METHODS: We conducted a field study at Prisma Health in South Carolina that included pre- and postsurveys, interactive usability testing sessions, think-alouds, and in-depth interviews with 16 clinicians providing HIV care between March and September 2025. We analyzed survey responses using descriptive statistics, and then transcribed and analyzed think-aloud and interview data using an etic and emic approach. RESULTS: Clinicians identified multiple challenges and design considerations for AI-powered HIV CDSS, demonstrating that clinician-AI interaction is inherently sociotechnical and embedded across multiple socioecological levels. While clinicians relied on familiar clinical indicators as cognitive anchors for interpreting AI predictions, they emphasized that social determinants of health were central to their own risk assessment and clinical decision-making. Additionally, clinicians' trust in AI is conditional and develops over time, with explainability and actionability emerging as critical factors for translating predictions into meaningful clinical interventions. CONCLUSIONS: Findings highlight the need to move beyond technically accurate predictions toward AI-powered CDSS designs that align with clinicians' cognitive practices and socioecological realities of HIV care. By extending a sociocognitive framework through empirical grounding in HIV clinical practice, this study offers design insights for developing AI-powered CDSS that are trustworthy, context-aware, and capable of supporting actionable decision-making in HIV care settings and beyond.
2 Aug 2026
Read appraisal →International dental journal
Accuracy of Large Language Models in Answering Dental Examination Questions: A Systematic Review and Meta-Analysis
INTRODUCTION: Large language models (LLMs), including OpenAI's GPT family accessed via interfaces such as ChatGPT and Microsoft Copilot, as well as non-GPT systems such as Google Gemini, are increasingly applied in healthcare and dental education. However, the accuracy of these systems in specialized tasks such as answering dental examination questions remains unclear. METHODS: This systematic review and meta-analysis evaluated LLM performance in answering dental questions. Databases searched were PubMed, Embase, Scopus, and Web of Science. Data on question type and number, LLM versions, and accuracy rates were extracted. Pooled accuracy was estimated using a random-effects model; heterogeneity and publication bias were assessed. RESULTS: A total of 39 studies were included, with ChatGPT-4 being the most frequently evaluated model. The pooled accuracy for LLMs was 63.7% (95% CI: 60.3%-67.1%), with high heterogeneity (I² = 91.5%). Subgroup analysis revealed ChatGPT-4 and Copilot (a GPT-based interface) achieved the highest pooled accuracies (∼73% and ∼75%, respectively). Direct comparisons confirmed ChatGPT-4 significantly outperformed earlier versions and some competitor models. Sensitivity analyses supported the robustness of findings. CONCLUSION: LLMs demonstrate moderate accuracy in answering dental examination questions and are currently insufficient for autonomous clinical decision-making. When their limitations are explicitly recognized, however, these systems may serve as valuable adjuncts in dental education and examination preparation. Methodological strategies such as structured prompting and retrieval-augmented approaches warrant further investigation but were not the primary focus of the present analysis.
2 Aug 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
An approximation theory perspective on machine learning
A central problem in machine learning is often formulated as follows: Given a dataset [Formula: see text] , which is a sample drawn from an unknown probability distribution, the goal is to construct a functional model f such that f(x) ≈ y for any (x, y) drawn from the same distribution. Neural networks and kernel-based methods are commonly employed for this task due to their capacity for fast and parallel computation. The approximation capabilities, or expressive power, of these methods have been extensively studied over the past 35 years. In this paper, we will present examples of key ideas in this area found in the literature. We will discuss emerging trends in machine learning including the role of shallow/deep networks, approximation on manifolds, physics-informed neural surrogates, neural operators, and transformer architectures. Despite function approximation being a fundamental problem in machine learning, approximation theory does not play a central role in the theoretical foundations of the field. One unfortunate consequence of this disconnect is that it is often unclear how well trained models will generalize to unseen or unlabeled data. In this review, we examine some of the shortcomings of the current machine learning framework and explore the reasons for the gap between approximation theory and machine learning practice. We will then review some of recent work that achieves function approximation on unknown manifolds without the need to learn specific manifold features, such as the eigen-decomposition of the Laplace-Beltrami operator or atlas construction. In many machine learning problems, particularly classification tasks, the labels yj are drawn from a finite set of values. We summarize another recent paper that establishes a deep connection between signal separation problems and classification problems, proposing that classification tasks should be approached as instances of signal separation. We conclude by identifying several open research problems that warrant further investigation.
2 Aug 2026
Read appraisal →Rheumatic diseases clinics of North America
Artificial Intelligence Regulation in the United States: Current Landscape and Implications for Rheumatology
Artificial intelligence (AI) is increasingly embedded in clinical tools used in rheumatology, including imaging interpretation, longitudinal disease monitoring, and electronic health record-based decision support. AI has moved from the periphery of biomedical research to an operational component of clinical care, increasingly embedded in electronic health records, imaging platforms, and decision support systems. In rheumatology, where care is longitudinal, AI systems offer substantial promise-but also poses distinct risks. Together, these commitments-advancing humanity, ensuring equity, engaging impacted individuals, improving workforce well-being, monitoring performance, innovating and learning, and promoting sustainability-operationalize trustworthy AI systems across a patient's care trajectory.
2 Aug 2026
Read appraisal →JMIR medical informatics
Large Language Models for Endodontic Symptom Assessment and Treatment Planning Using Image-Free Clinical Records: Comparative Evaluation Study
BACKGROUND: Accurate assessment of pulpal status is essential for achieving successful endodontic outcomes. However, direct evaluation remains inherently challenging because the pulp is surrounded by calcified tissue, necessitating reliance on clinical and radiographic examinations for diagnostic and prognostic decision-making. These procedures demand substantial clinical expertise and time, and less-experienced clinicians often face challenges that may lead to errors in diagnosis and treatment planning. Recent advancements in large language models (LLMs) offer promising opportunities to enhance clinical reasoning by facilitating the integration of evidence and supporting methodical diagnostic decision-making. OBJECTIVE: This study aimed to evaluate the clinical applicability of LLMs by comparing their text-based clinical screening performance and the clinical validity of their treatment plan responses with those of human evaluators. METHODS: Between January 2011 and December 2022, 100 clinical cases involving primary endodontic disease were randomly selected from the clinical records of outpatients who visited the Department of Conservative Dentistry or Advanced General Dentistry (AGD) at Yonsei University Dental Hospital. Four prompt types, combining 2 variables (language and role), were used as input for 4 LLMs. Both LLMs and human evaluators (AGD specialists, AGD residents, endodontic residents, and senior dental students) assessed the cases using text-based clinical records. Radiographic images were not directly provided. Screening performance was evaluated using a 0-to-2-point concordance scale, and treatment plan validity and relevance were assessed using a 5-point Likert scale. RESULTS: Among the 4 LLMs evaluated, ChatGPT achieved the highest mean concordance score on Korean-doctor prompts (mean 0.98, SD 0.82). However, this score did not reach the partially correct criterion of 1 on the 0 to 2-point scale. Clova X recorded the lowest mean score on English-patient prompts (mean 0.23, SD 0.63). Across both diagnostic categories, AGD specialists demonstrated the highest diagnostic accuracy (pulpal: 0.70; periapical: 0.65), with higher sensitivity but lower specificity than those exhibited by the other groups. ChatGPT also showed favorable performance among the LLMs, with accuracies of 0.65 (95% CI 0.55-0.74) for pulpal disease and 0.57 (95% CI 0.47-0.69) for periapical disease, which were comparable to those of AGD and endodontic residents. CONCLUSIONS: Under image-free clinical record review conditions, ChatGPT 4.0 showed relatively higher and more consistent performance in symptom screening and treatment planning compared to the other LLMs evaluated. However, its highest mean score of 0.98 (SD 0.82) did not reach the partially correct criterion of 1 on the 0 to 2-point scale. Hallucinations generated by LLMs and experience-dependent interpretation biases among human evaluators remain key challenges that require attention. Therefore, continuous clinical supervision and comprehensive user training are necessary for the safe and effective clinical application.
25 July 2026
Read appraisal →JMIR nursing
Effectiveness of Artificial Intelligence-Based Nursing Interventions for Chronic Illness Care: Umbrella Review
BACKGROUND: Artificial intelligence (AI)-based nursing interventions are increasingly being used to manage chronic illnesses; however, their definitive impact on clinical outcomes remains inconclusive, necessitating a comprehensive evidence synthesis. OBJECTIVE: This umbrella review aimed to synthesize the evidence regarding the effectiveness of AI-based nursing interventions for chronic illness care and their subsequent impact on health care outcomes in clinical settings. METHODS: We conducted an umbrella review and prospectively registered the protocol. A systematic search of 5 electronic databases (PubMed, CINAHL, Cochrane Library, Scopus, and Web of Science) was performed to identify systematic reviews and meta-analyses published in English between 2021 and 2025. The methodological quality of the included studies was evaluated using the Joanna Briggs Institute (JBI) Critical Appraisal Checklist. RESULTS: Eight high-quality systematic reviews were included, with machine learning identified as the predominant technology. Three primary outcome domains emerged: predictive, psychosocial, and hospital utilization. Due to measurement heterogeneity, the results were synthesized narratively. Our findings demonstrated that AI-based nursing interventions are effective in predicting adverse clinical events, unplanned hospital utilization, and health care costs. However, evidence regarding psychosocial outcomes remains insufficient. CONCLUSIONS: This review provides systematic evidence supporting the utility of AI in chronic illness management, particularly for improving predictive and utilization outcomes. These findings offer actionable insights for nursing leaders to integrate AI into clinical practice and education. Future research should prioritize rigorous empirical designs to further strengthen the evidence base for AI-driven nursing care.
17 July 2026
Read appraisal →JMIR cancer
Internet of Things-Enhanced Mathematical Oncology: Conceptual Framework for Adaptive Cancer Care Modeling
The Internet of Things (IoT) is transforming various industries, including health care. IoT-based systems are increasingly prevalent in consumer health applications, while intelligent or smart devices equipped with sophisticated sensors are gaining recognition for their potential to improve clinical care practice and decision-making. Cancer care is a particularly promising area for IoT applications, enabling real-time and personalized interventions. However, empirical research on the effects of IoT in this field is limited due to the complexities inherent in cancer as a dynamic disease and the paucity of IoT-generated data available for research. This presents an opportunity to apply mathematical modeling to understand the effects of IoT under various scenarios. These analytical and "in silico" mathematical approaches are instrumental with limited data. Such models support the analysis of treatment uncertainty and patient response while balancing patient preferences, clinical outcomes, and system-level constraints. Grounded in mathematical oncology and health informatics, this paper proposes a conceptual framework that integrates real-time IoT data as dynamic inputs into adaptive mathematical models to simulate cancer dynamics. By exploring applications across multiple levels of analysis, the study demonstrates how IoT-enhanced mathematical models could inform implementation and optimize oncology services, addressing a critical gap in current research.
15 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 →Dento maxillo facial radiology
JADE: jawbone lesion diagnosis and decision supporting system
OBJECTIVES: To develop and evaluate JADE, a proof-of-concept retrieval-augmented generation (RAG) diagnostic assistive system, designed to enhance large language model (LLM) reasoning for jawbone lesion assessment. This study examined whether RAG improves diagnostic accuracy and stability compared with standalone LLMs and ORAD, a supervised learning-based system. METHODS: JADE was developed as a cloud-based RAG system integrating an expert-curated oral radiology database embedded using text-embedding-3-large and indexed in Qdrant Cloud. Structured clinical inputs were encoded as prioritized vector queries. Hybrid semantic and keyword-based retrieval appended relevant evidence to prompts for differential diagnosis generation across LLM backbones. Performance was evaluated in 25 validation cases and compared with GPT-5, Claude Sonnet 4.5, DeepSeek-R1, Gemini 2.5 Flash, their RAG configurations, and ORAD. Diagnostic accuracy was analyzed using Cochran's Q test, with post-hoc McNemar's tests and Bonferroni correction. Intra-model stability and response time were assessed. RESULTS: RAG-GPT-5 achieved the highest diagnostic accuracy (20/25), followed by RAG-Claude Sonnet 4.5 (18/25), RAG-DeepSeek R1 (17/25), and RAG-Gemini 2.5 Flash (15/25). Standalone models achieved 9-13/25 correct diagnoses, ORAD achieved 17/25. No significant differences were observed among standalone models or RAG-based models and ORAD. A significant improvement was observed for GPT-5 when integrated with RAG (P = .002). RAG configurations showed higher intra-model stability, with RAG-GPT-5 achieving mean stability of 0.90 ± 0.11. Mean response times ranged from 3 to 10 s. CONCLUSIONS: JADE improved diagnostic accuracy and stability compared with standalone LLMs, underscoring value of RAG reasoning in jawbone lesion assessment and marking the first RAG application in dentomaxillofacial radiology.
3 July 2026
Read appraisal →International journal of medical informatics
The role of artificial intelligence in virtual emergency care: a systematic review
BACKGROUND: The integration of artificial intelligence (AI) into virtual emergency care represents a potentially transformative approach to healthcare delivery, yet the evidence base remains poorly characterized. This systematic review comprehensively evaluates the current state of AI applications in virtual emergency care settings. METHODS: We systematically searched eight databases (Embase, PsycINFO, MEDLINE, PubMed, Scopus, Web of Science, CINAHL, Cochrane Library) from inception through March 2025. Of 7,098 records identified and 4,935 screened after deduplication using Covidence, 8 studies met inclusion criteria following exclusion of one study lacking AI components. Studies were assessed using PROBAST + AI for risk of bias and quality assessment, TRIPOD + AI for reporting quality, and GRADE for certainty of evidence. RESULTS: The eight included studies (total participants: approximately 0.5 million) evaluated diverse AI applications including decision trees, machine learning ensembles, and graph neural networks across multiple virtual emergency contexts. Performance varied widely (accuracy 77.5-100%, sensitivity 63-100%, specificity 60% in single study reporting). All clinical studies demonstrated serious risk of bias. TRIPOD + AI compliance averaged only 36.9% (range 30.9-48.1%). GRADE assessment revealed very low to low certainty evidence across all outcomes, with no studies measuring actual clinical outcomes. CONCLUSIONS: Current evidence is insufficient to support widespread clinical implementation of AI in virtual emergency care. While preliminary results suggest potential benefits in triage accuracy and resource efficiency, critical gaps exist in validation, clinical outcome assessment, and reporting standards. Future research must prioritize prospective controlled trials with real patient data, clinical outcome measurements, and adherence to reporting guidelines.
2 July 2026
Read appraisal →JMIR nursing
Nurses' Experiences Using AI in Clinical Practice: Systematic Review
BACKGROUND: Artificial intelligence (AI) tools are increasingly used in clinical settings, yet most syntheses focus on nurses' attitudes or readiness rather than experiences after direct use in practice. OBJECTIVE: This study aimed to synthesize registered nurses' experiences of using AI in clinical practice and to identify perceived benefits, barriers, and implementation implications. METHODS: We conducted a systematic literature review of empirical studies reporting nurses' experiences of AI use in clinical settings. Searches were performed in CINAHL, Embase, MEDLINE, PsycINFO, and PubMed (last search: September 13, 2023). Moreover, 2 reviewers (AJSS and QZ) independently screened titles, abstracts, and full texts and appraised included studies using the Mixed Methods Appraisal Tool. We used thematic synthesis with a primarily deductive framework based on the Technology Acceptance Model 2, with the addition of facilitating conditions from the unified theory of acceptance and use of technology. RESULTS: In total, 20 studies met the inclusion criteria. Perceived usefulness and facilitating conditions were most frequently reported; nurses described AI as supporting decision-making, workflow efficiency, and confidence when implementation included adequate training, interoperability, and technical infrastructure. Ease of use was closely tied to interface design and training. Job relevance and output quality were rated more positively when nurses described AI as aligning with nursing tasks and as producing interpretable, reliable outputs. Common barriers included usability issues, limited integration into workflows and electronic systems, privacy and trust concerns, and inconsistent or poorly contextualized outputs. Across studies, nurses often described adoption as conditional on organizational readiness and meaningful involvement of nurses in design and implementation. CONCLUSIONS: Nurses' accounts suggest that AI may augment clinical work, but the perceived benefits reported in the included literature were contingent on workflow alignment, usable interfaces, training, and supportive infrastructure. These findings suggest potential value in involving nurses in the co-design and iterative refinement of AI tools, grounded in the consistent evidence that facilitating conditions, such as training, interoperability, and organizational readiness, are the primary determinants of whether AI is experienced positively. Prospective evaluation using nursing-relevant outcomes, such as usability, workflow integration, and trust, is needed to move beyond post hoc experiential accounts and establish what implementation conditions reliably produce benefit.
26 June 2026
Read appraisal →PloS one
A lightweight hybrid deep learning approach for fashion mnist classification with explainable attention visualization
The classification of fashion images is an essential task in the e-commerce sector, where accurate categorization improves user experience and refines product discovery. Convolutional Neural Networks (CNNs) and Transformers have demonstrated strong performance in image classification tasks due to their ability to learn complex visual features. However, deep variants of these architectures, such as VGG-19, ResNet-50, Vision Transformer (ViT), and Swin Transformer, contain tens of millions of parameters, requiring high memory and powerful GPUs for training, which makes them less suitable for low-resource and edge device environments. To address these limitations, this research proposes a lightweight hybrid architecture, TinyCNN with Linear Self-Attention (LSA), optimized for resource-constrained settings. The proposed model contains fewer than half a million parameters and is trainable on a CPU, achieving a classification accuracy of 91.47% on the Fashion-MNIST dataset. In addition, multiple Explainable Artificial Intelligence (XAI) techniques are implemented, including Self-Attention visualization, Multi-Head Attention, Attention Flow, Attention Rollout, Fixed query position attention maps, Integrated Gradients, LIME, and SHAP, to provide visual interpretability of the model's predictions and enhance transparency in its decision-making process.
25 June 2026
Read appraisal →Archives of gynecology and obstetrics
Comparing clinical decision-making between colposcopists and large language models in cervical dysplasia management: a pilot prospective multicenter study
PURPOSE: This prospective multicenter study aimed to compare the decision-making abilities of board-certified colposcopists and two commercially available large language models (LLM), ChatGPT-4o and ChatGPT-5, in cervical dysplasia management. METHODS: Twenty-three anonymized real-life patient cases with multiple-choice (MC) questions regarding treatment decisions were used to assess answer quality. Ten board-certified colposcopists and the two LLMs addressed the MC questions. The gold standard was defined by two guideline authors. LLMs were prompted to justify their responses. Concordance rates were calculated and compared across all questions and histopathological subgroups, including cervical intraepithelial neoplasia (CIN), unspecific histopathological results, and cervical cancer cases. RESULTS: Clinicians and LLMs achieved similar overall concordance rates compared to the gold standard (69.6% for clinicians, 69.6% for ChatGPT-4o, and 65.2% for ChatGPT-5). ChatGPT-5 outperformed clinicians in precancerous lesions (81.8% vs. 66.4%), while clinicians excelled in complex cases with unspecific histopathology (86% vs. 60%). Clinicians showed a tendency to overtreat low-grade lesions (CIN I), opting for more intensive surveillance. ChatGPT-4o performed better than ChatGPT-5 in cervical cancer cases, though both models struggled with these scenarios. CONCLUSION: This study highlights the potential of LLMs as decision support tools in cervical dysplasia management, particularly for straightforward cases like precancerous lesions. However, clinicians remain superior in handling complex or ambiguous cases. The tendency of clinicians to overtreat low-grade lesions may offer the potential to test the implementation of a decision support tool for those cases. While LLMs show promise, exploring open-ended clinical scenarios and integrating retrieval-augmented generation could enhance their practical application.
19 June 2026
Read appraisal →Journal of affective disorders
MDD-thinker: A reasoning-enhanced large language model for diagnosis of major depressive disorder.
BACKGROUND: Major depressive disorder (MDD) is a leading cause of global disability and poses a substantial public health burden. However, current diagnostic approaches largely rely on subjective assessments and lack the ability to integrate heterogeneous clinical and sociodemographic information. Recent advances in large language models (LLMs) offer new opportunities to support MDD diagnosis through reasoning over complex data, yet their clinical applicability is constrained by challenges related to interpretability, hallucinations, and reliance on synthetic data. METHODS: We propose MDD-Thinker, an LLM-based diagnostic system that integrates supervised fine-tuning (SFT) and reinforcement learning (RL) to enhance reasoning and interpretability under the evaluated conditions. Using the UK Biobank dataset, we constructed 40,000 structured reasoning samples and incorporated an additional 10,000 records from publicly available mental health datasets. MDD-Thinker was trained on these heterogeneous textual data and evaluated against conventional machine learning models, deep learning methods, and representative LLM baselines in terms of diagnostic performance and interpretability. RESULTS: MDD-Thinker achieved high performance in MDD diagnosis, with an accuracy of 0.8268 and an F1-score of 0.8081, showing better performance than conventional machine learning models, deep learning approaches, and representative LLM baselines on the evaluated dataset. Beyond predictive accuracy, it consistently produced structured reasoning paths that were clinically coherent, enabling transparent interpretation of diagnostic decisions in the evaluated experiments. The integration of SFT and RL contributed to notable improvements in both diagnostic reliability and reasoning quality. CONCLUSION: MDD-Thinker demonstrates the potential of reasoning-enhanced LLMs for large-scale MDD diagnosis under the evaluated settings. By jointly optimizing accuracy, interpretability, and efficiency, the proposed system provides a scalable and explainable approach for intelligent psychiatric assessment within the scope of the study, highlighting the potential of reasoning-oriented LLMs in mental health care.
15 June 2026
Read appraisal →International journal of medical informatics
Artificial intelligence applications in emergency triage for suspected acute coronary syndrome: a systematic review
BACKGROUND: Acute coronary syndromes (ACS) are time-critical conditions requiring rapid and accurate triage in the emergency department. Traditional triage may lead to delays, whereas artificial intelligence (AI) has the potential to enhance triage accuracy and efficiency. OBJECTIVE: This study aimed to synthesize and evaluate the effectiveness of AI applications in emergency triage or early clinical decision-making for suspected ACS, focusing on their impact on diagnostic accuracy, time to treatment, and operational outcomes. METHODS: This systematic review followed PRISMA guidelines. PubMed, Scopus, CINAHL, and IEEE Xplore were searched for studies published between 2020 and 2025. Eligible articles examined AI applications for triage of suspected ACS in the emergency department. Data on study design, AI models, comparators, inputs, and outcomes were extracted, and study quality was assessed using the Mixed Methods Appraisal Tool (MMAT). RESULTS: Fifteen studies from multiple countries and study designs were included. Convolutional neural networks and ensemble learning methods were the most used models. AI models generally demonstrated high diagnostic performance (AUROC 0.82-0.99), were associated with reductions in treatment times such as door-to-balloon and catheterization intervals, and showed potential to improve operational outcomes, including resource utilization and patient flow. However, findings varied across studies depending on model type, data inputs, and study design. CONCLUSIONS: AI-assisted triage for suspected ACS shows promise in supporting clinical decision-making and improving workflow efficiency. However, substantial heterogeneity and limited prospective validation suggest that findings should be interpreted cautiously. Further research is needed to confirm clinical effectiveness, generalizability, and safe implementation.
15 June 2026
Read appraisal →Journal of affective disorders
Implementing artificial intelligence (AI) to facilitate health outcomes in mood disorders: Application versus aspiration.
The burden of depressive and bipolar disorders at the individual and societal level are extraordinary and increasing. For decades, evidence-based treatments for these conditions have been established but outcomes amongst individuals with lived experience remains suboptimal. Notwithstanding calls to close gaps between current practice and best practice, there is an absence of evidence that overall health outcomes are significantly improving. Artificial Intelligence (AI) is the cornerstone of the digital revolution. Currently, AI sources (e.g., Open Source) are widely accessed by healthcare providers and persons with lived experience for search queries and decision support. The aspiration for AI-informed medical practice is to improve health outcomes by assisting in timely diagnostic detection, illness monitoring, informing treatment selection, integrating multimodality care, decreasing barriers to access and facilitating scalability to psychosocial interventions. Against this background, rigorous evidence is still needed to empirically demonstrate transformative improvement in each of the aforementioned areas. In addition, multiple ethical, technical, scientific and economic questions are not adequately answered including aspects of confidentiality and patient engagement. This short commentary endeavors to succinctly summarize the evidentiary base as it relates to the capabilities that AI offers currently, in the near and more intermediate term. The overarching aim is to provide readers with an up-to-date understanding of what aspects of AI are currently applicable versus those that are aspirational.
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 →The pharmacogenomics journal
Development and validation of an open data model for pharmacogenetics to enable semantic interoperability in clinical practice
Pharmacogenetics uses genetic testing to improve the safety and effectiveness of prescribed medicines, yet implementation at scale remains limited due to the absence of interoperable health IT solutions that integrate results into prescribing workflows. This study aimed to develop and validate open data standards for pharmacogenetic results to enable interoperability across healthcare systems. A baseline data model was constructed using the open standard openEHR by synthesising literature, genomic sequencing outputs, and international data specifications, and refined through iterative workshops with the Global Alliance for Genomics and Health. The model underwent two rounds of structured peer review involving 24 experts from 10 countries. Mapping to HL7 FHIR was evaluated using both manual and automated approaches, including the FHIR-Connect tool. The resulting standardised pharmacogenetic data model separates test results from therapeutic implications and incorporates recognised terminologies such as SNOMED CT and HGNC. It achieved international consensus and is published on the openEHR Clinical Knowledge Manager platform. Mapping to HL7 FHIR demonstrated bidirectional information flow within healthcare systems, with automated mapping enabling scalable and reusable transformations. This work provides a framework for storing and exchanging pharmacogenetic test results, supporting semantic harmonisation, interoperability, and integration with clinical decision support systems. Open data standards for pharmacogenetic test results therefore offer a foundation for scalable implementation of pharmacogenetics in routine clinical practice.
14 June 2026
Read appraisal →Journal of the American College of Surgeons
Improving Trauma Triage Accuracy with Large Language Models: A Comparison to Human Expert Decisions
BACKGROUND: Accurate prehospital trauma triage and communication determine morbidity, mortality, and system efficiency. Advancements in large language models (LLMs) offer an opportunity to improve triage and yet to be implemented in prehospital trauma triage. STUDY DESIGN: This retrospective cohort study evaluates LLM performance in trauma triage and accuracy of prehospital tele-communication. Of 410 pediatric activations at a Level I center (January 2023 to May 2025, IRB no. 00009569), 133 activations with emergency medical service recordings, human-generated trauma pages, and injury severity scores were analyzed. Audio was transcribed with OpenAI Whisper. Structured "Essential Transcripts" were generated with named entity recognition. Entity ablation tested redaction of triage parameters on accuracy. In a prospective arm, trauma surgeons reviewed emergency medical service transcripts and triaged activations pre- and post-LLM exposure. Cribari criteria defined over and undertriage. McNemar's test and 95% CIs assessed paired differences in accuracy. RESULTS: The primary endpoint: LLM undertriage demonstrated modest improvement; 4.8% (3 to 8.2) vs 5.1% (3.1 to 9.3, p = 0.73, Bonferroni p = 1). For secondary endpoints, LLM triage outperformed human clinicians: 83.5% accuracy (80.6 to 90.6) vs 78.9% (73.9 to 82.6, p < 0.01, Bonferroni p < 0.01), with overtriage 58.6% (51.4 to 73.7) vs 71.8% (p < 0.05, Bonferroni p < 0.09). "Essential Transcripts" reduced transcript length by 80.8% (p < 0.001) while preserving accuracy (81.9%, 76.6 to 87.5, p < 0.001, Bonferroni p < 0.05). Entity ablation had marginal effect on triage. In prospective evaluation, human triage accuracy improved after LLM exposure (73.7% [69.8 to 77.2]) to 75.8% ([71.9 to 79.2], p = 0.04, Bonferroni p = 0.12), significantly improving the odds of a correct triage decision (odds ratio 2.57, 95% CI 1.39 to 6.83, p < 0.01). CONCLUSIONS: LLMs achieve triage accuracy comparable to trauma staff in retrospective review of pediatric trauma. Further validation is needed to assess clinical outcomes, generalizability, and user acceptance before widespread deployment.
14 June 2026
Read appraisal →Family practice
Current state of electronic problems lists in primary care: a rapid scoping review
BACKGROUND: Electronic problem lists (PLs) are central to the problem-oriented medical record and increasingly underpin clinical decision support (CDS), interoperability, and emerging artificial intelligence applications in primary care. However, persistent concerns regarding PL accuracy, completeness, and governance limit their clinical value. OBJECTIVE: To update the foundational framework proposed by Hodge and Narus by synthesizing contemporary evidence on electronic PLs in primary care and primary care-relevant settings. METHODS: We conducted a rapid scoping review following Joanna Briggs Institute guidance and PRISMA-ScR reporting standards. MEDLINE (Ovid) was searched for peer-reviewed studies published between March 2016 and May 2026. Eligible studies examined digital PLs or extractable PL-related practices relevant to primary care. Findings were synthesized using a hybrid inductive-deductive approach and mapped to the seven Hodge and Narus themes. RESULTS: A total of 103 studies were included. Across settings, PLs were widely recognized as foundational for longitudinal care, safety, and coordination. However, incomplete, outdated, or inconsistently maintained PLs remained common. Persistent challenges included unclear ownership, workflow misalignment, and variability in what constitutes a "problem." Emerging technologies-including CDS, natural language processing, and machine learning-were increasingly used to support PL generation, curation, reconciliation, and organization, with greatest benefit when embedded within clinician-led workflows. CONCLUSION: Despite technological advances, the clinical value of electronic PLs in primary care continues to depend on governance, workflow integration, and shared accountability. Updating the Hodge and Narus framework highlights the need for sociotechnical approaches to PL improvement as PLs become integral to CDS and artificial intelligence-enabled care. Problem lists (PL) are a key part of electronic health records. They provide clinicians with a quick overview of a patient's important health issues—such as long-term conditions, past diagnoses, and current concerns—and help support safe and informed care. When PLs are accurate and up to date, they can improve communication and decision-making. However, in everyday clinical practice, many PLs are incomplete, outdated, or difficult to maintain. As electronic health records are increasingly used to support clinical decision support tools and new digital technologies, the quality of PLs has become more important than ever. Poor-quality PLs can lead to missed information, duplicated work, or unsafe care decisions. To better understand how PLs are working in modern healthcare, we conducted a rapid scoping review of research published between 2016 and 2026. We examined studies that looked at how PLs are used, how accurate they are, what challenges clinicians face in keeping them updated, and what solutions have been proposed. In total, 103 studies were included. Across these studies, several common problems were identified. Important health conditions were often missing or out of date, and clinicians were frequently unsure who was responsible for updating the PL. Time pressures and busy clinical workflows made regular maintenance difficult. Some studies explored new digital tools, such as automated text analysis, to help identify or update problems, but these approaches are still developing and require careful evaluation. Improving PLs will require clearer standards, shared responsibility among healthcare teams, and EHRs that make updating information easier. Thoughtful use of new technologies may also help reduce clinician workload while improving patient safety. This review highlights where PLs are working well and where improvements are still needed, offering guidance for future research, technology design, and improvements in primary care practice.
14 June 2026
Read appraisal →Journal of medical Internet research
Patient Cognitive Bias in Large Language Model-Supported Health Consultations: Simulation-Based Comparative Study
BACKGROUND: Large language models (LLMs) are increasingly used by patients for health information and preliminary medical advice. In patient-facing consultations, users may present explicitly stated diagnostic preferences or symptom narratives emphasizing a preferred explanation. Such cognitively biased input constrains the diagnostic context available to the model and may systematically steer its reasoning during interactive LLM-supported health consultations. OBJECTIVE: This study aimed to quantify the impact of patient cognitive bias on LLM diagnostic performance in multiturn consultations, assess the effectiveness of prompt-based mitigation strategies and decoding temperature adjustment, and evaluate a dual-system framework for improving robustness under biased interaction. METHODS: We developed a simulated patient agent to generate both unbiased and cognitively biased consultations using 1273 medical question answering dataset United States Medical Licensing Examination cases. Six widely used LLMs of varying capacities were evaluated through 3-round, multiturn dialogues, after which each model produced a final diagnostic judgment based on the complete consultation record. Diagnostic accuracy was the primary outcome. Secondary outcomes included bias-induced accuracy decline (absolute reduction in accuracy under biased vs standard consultations) and bias-influenced error proportion (proportion of incorrect responses aligned with the patient's preferred but incorrect diagnosis). Three prompt-based mitigation strategies and 4 decoding temperature settings were tested. In addition, a dual-system framework was evaluated, in which a conversational foundation LLM conducted patient interaction and history taking (System 1), while a reasoning-oriented LLM (o1-mini) generated the final diagnostic judgment (System 2). In the foundation-only condition, the same LLM performed both interaction and diagnosis. RESULTS: Across all 6 evaluated models, cognitively biased consultations led to marked diagnostic accuracy declines of approximately 7 to 39 percentage points compared with standard multiturn consultations, whereas static single-response tests and standard consultations showed comparable accuracy. Larger deteriorations were observed in lower-capacity models, with some approaching random-guess performance under bias. Errors were frequently aligned with patient bias, with bias-influenced error proportion exceeding one-third across models, indicating systematic conformity rather than random error. Prompt-based mitigation strategies and decoding temperature reduction yielded limited and inconsistent improvements and did not reliably prevent bias-induced performance loss. By contrast, the dual-system framework substantially improved diagnostic accuracy under biased conditions, producing gains of approximately 10 to 39 percentage points across most models and recovering a large proportion of the performance lost due to bias, particularly in lower-capacity systems. CONCLUSIONS: Patient-driven cognitive bias represents an underrecognized behavioral risk in LLM-supported health consultations. Common mitigation approaches, such as prompt engineering or decoding parameter adjustment, provide limited resilience. Explicitly separating conversational interaction from deliberative diagnostic reasoning through a dual-system framework enables more robust diagnostic performance under biased input while potentially preserving patient-facing dialogue fluency by retaining the foundation LLM as the conversational component, offering a scalable design strategy for safer medical AI systems.
14 June 2026
Read appraisal →World journal of urology
Assessing multiple chatGPT versions on novel content in the Taiwan urology board examination: accuracy, speed, and domain-specific performance
PURPOSE: Large language models demonstrate increasing utility in healthcare; however, their capacity to handle newly developed medical content beyond training data boundaries remains unclear. This study evaluated the performance of four ChatGPT models on 120 newly developed questions from the 2024 Taiwan Urology Board Examination, all created after the models' October 2023 training cutoff, thereby reducing the risk of prior model exposure. METHODS: Model performance was assessed using accuracy and processing time. Of 150 examination items, the first 120 newly developed questions constituted the primary analysis set; the remaining 30 previously used archival questions were excluded to reduce potential bias from prior exposure. The primary paired accuracy comparison evaluated o1-preview versus GPT-4o across all 120 questions. Additional analyses across 12 urological subspecialties and question-complexity strata were treated as exploratory. RESULTS: Across all 120 questions, o1-preview achieved 66.7% accuracy and significantly outperformed GPT-4o (55.8%; p = 0.012), although with longer processing time (19.20 vs. 14.94 s). Performance varied significantly across 12 urological subspecialties (p < 0.001). Unlike the other tested models, o1-preview showed slightly higher accuracy on high-complexity than on low-complexity questions (68.3% vs. 65.0%). On high-complexity surgical anatomy items, o1-preview achieved 80.0% accuracy, whereas GPT-4o and GPT-4o mini both scored 0.0% (p = 0.024). CONCLUSION: On this newly developed post-cutoff, text-based urology board question set, o1-preview achieved higher overall accuracy than GPT-4o, particularly on high-complexity surgical anatomy items, at the cost of longer response times. These findings suggest that reasoning-oriented models may offer advantages for selected specialty-level assessment tasks, although generalizability to other languages, modalities, and clinical settings requires further validation.
30 May 2026
Read appraisal →Renal failure
Time-updated explainable machine learning predicts short-term mortality in peritoneal dialysis patients
Objective: This study aimed to develop and validate a time-updated, explainable machine learning (ML) early-warning system for short-term mortality risk in patients with continuous ambulatory peritoneal dialysis (CAPD). Methods: Multiple supervised ML techniques were validated to stratify individuals at high risk of 6-month death, classifiers were trained, internally validated and independently temporally validated in this retrospective study of 1,484 CAPD patients in our PD center. We collected time-updated markers such as patient demographics, clinical characteristics, and laboratory data to inform the ML models and the performance of eight ML models was assessed via the area under the curve (AUC) and accuracy. The SHapley Additive exPlanation (SHAP) method was selected to further interpret the predictive models and link findings to clinical actionability. Results: In internal validation cohort, the light gradient boosting machine (lightGBM) model demonstrated best performance with AUC of 0.888 and highest accuracy of 0.879. In temporal validation cohort, the lightGBM model also exhibited good classification performance, achieving AUC of 0.850 and good accuracy of 0.874. Moreover, SHAP analysis identified several key features that contributed to the accurate prediction of the lightGBM model. Conclusions: The time-updated lightGBM early-warning system helps clinicians to early identify CAPD patients with high risk of death within six months based on easily accessible and time-updated data and provides a basis for personalized treatment. We developed and validated a time-updated, explainable ML early-warning system that accurately predicts 6-month all-cause mortality in CAPD patients using routinely, available laboratory data.The LightGBM model achieved excellent discriminative performance, with an AUC of 0.888 in internal validation and 0.850 in independent temporal validation.This model provides a scalable, low-cost early-warning tool for routine CAPD care, with a standardized risk-stratified intervention framework to guide timely, targeted care and advance care planning.
29 May 2026
Read appraisal →JMIR medical informatics
Multimodal Prediction of Renal Tumor Malignancy From Radiology Reports and Structured Electronic Health Records: Retrospective Cohort Study
BACKGROUND: Accurate preoperative prediction of renal tumor malignancy is critical for guiding decisions and reducing overtreatment, as a substantial proportion of renal masses prove benign. Although radiology assessments and structured electronic health record (EHR) data are routinely used, many tumor-specific descriptors remain embedded in free-text radiology reports and are underused due to extraction challenges. OBJECTIVE: This study aimed to develop and evaluate a multimodal pipeline that integrates structured EHR variables with natural language processing features from computed tomography (CT) radiology reports, including large language model (LLM)-extracted abnormality characteristics and transformer-based report embeddings, to improve malignancy prediction. METHODS: We conducted a retrospective cohort study using University of Florida Health Integrated Data Repository Observational Medical Outcomes Partnership-mapped EHR data from December 2011 to August 2024. Adults with renal tumors were included if they had longitudinal diagnostic documentation consistent with a renal mass and at least 1 preoperative renal CT report; final benign or malignant status served as the outcome. Structured features included demographics, comorbidities, medications, vital signs, and laboratory measurements. From the recent preindex CT report, an on-premises LLM isolated kidney-specific findings and extracted abnormality characteristics. Four locally deployed LLMs were evaluated against manual annotations of 500 reports. Kidney-specific text was encoded using pretrained biomedical transformer models, including radiology Bidirectional Encoder Representations from Transformers (BERT) variants. We evaluated unimodal baselines and multimodal early, middle, and late fusion strategies. Model development used 5-fold cross-validation within the 80% training partition; each fold-specific model was evaluated on the same independent 20% held-out test set, with performance reported as mean and SD across the 5 held-out test evaluations. The primary metric was area under the receiver operating characteristic curve (AUC). RESULTS: The final cohort included 967 patients (n=712, 73.6% malignant). In extraction evaluation, Qwen2.5-32B achieved 88.3% overall accuracy with a 100% extraction success rate and was selected for downstream feature generation. Among unimodal models, the structured clinical variable model achieved an AUC of 0.758 (SD 0.012), kidney-specific text with radiology BERT achieved an AUC of 0.746 (SD 0.058), and abnormality characteristics alone achieved an AUC of 0.716 (SD 0.015). Multimodal fusion models achieved higher descriptive performance than unimodal models. Early fusion achieved the highest AUC (mean 0.813, SD 0.008), and F1-score (mean 0.809, SD 0.030), while late fusion achieved an AUC of 0.805 (SD 0.016). Ablation and interpretability analyses suggested complementary predictive information from structured clinical variables and kidney-specific text embeddings. CONCLUSIONS: Integrating unstructured radiology report text with structured EHR variables achieved higher mean predictive performance than unimodal approaches in descriptive comparisons. Multimodal fusion, particularly early fusion incorporating radiology BERT-derived kidney-specific text embeddings, achieved the strongest discrimination, suggesting potential value of natural language processing-enabled multimodal EHR pipelines for informing preoperative risk stratification.
28 May 2026
Read appraisal →JMIR medical informatics
Usability and Usefulness of Machine Learning-Based Clinical Decision Support Software in Primary Care: Survey of Users in a Prospective Observational Study
BACKGROUND: The successful implementation of decision support systems promises to enhance high-quality care. However, the successful implementation of a clinical decision support system (CDSS) depends on user acceptance and adoption. A machine learning (ML)-based CDSS to assist primary care professionals treating urinary tract infections (UTIs) was implemented, and usability and usefulness were assessed through a questionnaire. OBJECTIVE: This study aimed to assess the system's usability by examining users' experiences with the software. A secondary goal was to assess users' attitudes toward evidence-based practice and innovation in health care. METHODS: In collaboration with the Netherlands Institute for Health Services Research (NIVEL) and Leiden University Medical Center (LUMC), Pacmed Ltd developed the CDSS. The cohort was mostly recruited at the care group level; practices within participating care groups were required to participate. Health insurers partly funded the research. Practitioners participated in the implementation study for 4 months. A survey based on the Unified Theory of Acceptance and Use of Technology (UTAUT) was sent to 263 general practitioners and assistants shortly after the implementation period. Furthermore, usage data were analyzed. RESULTS: Of the 34 participating practices that used the software, 30 (88%) submitted at least one survey response, with a mean of 2.23 responses per practice (SD 1.43). The CDSS was used throughout the pilot period, and 31 practices continued using the tool, with 9% dropping out during the first 8 weeks. Sixty-seven percent of respondents trusted the tool's output, and 73% found it understandable how the algorithm came to predictions. Sixty-five percent of respondents indicated that the information provided was useful in addition to the available guidelines, and 52% agreed that it supported their decision-making. However, many respondents were uncertain whether the tool improved patient care (46%) or patient outcomes (66%). Forty-eight percent of respondents found the software easy to integrate into their clinical workflow. CONCLUSIONS: The CDSS was perceived as trustworthy and easy to use. However, users were unable to determine whether the CDSS improved patient outcomes. In addition, the CDSS development could have benefited from including assistants as well as general practitioners more in the design phase of the software. Because assistants play an important role in UTI care, designing the software to better fit existing workflows may reduce the perceived time investment associated with using the tool. Finally, respondents reported strong motivation to contribute to further research in this field and indicated willingness to embrace change in health care delivery, which may also reflect selection bias in our sample.
28 May 2026
Read appraisal →Journal of neurology
Man vs. Machine: can ML algorithms diagnose headaches as accurately as clinicians? A systematic review.
BACKGROUND: Headache disorders are frequently misdiagnosed. We aimed to systematically evaluate the diagnostic accuracy, methodological quality, and clinical applicability of artificial intelligence (AI) and machine learning (ML) models for classifying adult headache disorders against clinician diagnoses using the International Classification of Headache Disorders (ICHD) criteria. METHODS: In this systematic review, we searched PubMed, Embase, and the Cochrane Library (January 2015-December 2025) for AI/ML diagnostic headache studies. Two reviewers extracted data and assessed risk of bias using the QUADAS-2 tool with the QUADAS-AI extension. Main outcomes were sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), and risk of bias. RESULTS: We included 74 studies encompassing 154,856 participants. Models utilized traditional ML (n = 47), deep learning (n = 18), and hybrid or rule-based approaches (n = 9). Data inputs included neuroimaging (n = 27), multimodal datasets (n = 20), neurophysiological signals (n = 17), and clinical questionnaires (n = 10). Only 4 studies performed independent external validation. Overall sensitivity ranged from 47.5% to 100.0%, specificity from 50.4% to 100.0%, and AUC-ROC from 0.658 to 1.000. Models using structured questionnaires reported realistic accuracies (74-86%), whereas neuroimaging models frequently produced near-perfect, likely overfit estimates. Most studies (65 of 74) exhibited a high risk of bias driven by artificial case-control designs, data leakage, and absent external validation. CONCLUSIONS: AI-based headache diagnostic systems demonstrate promising but highly variable accuracy. Pervasive methodological flaws-specifically severe spectrum bias, data leakage, and a lack of independent external validation-currently preclude clinical implementation. Future studies require prospective, real-world validation to safely integrate these tools into practice.
27 May 2026
Read appraisal →Journal of radiation research
Feasibility of retrieval-augmented generation for large language models with Japanese input in radiotherapy
Large language models (LLMs) have recently gained attention for their potential. However, concerns remain regarding their reliability due to limitations such as hallucinations and insufficient domain-specific knowledge. Retrieval-augmented generation (RAG) has emerged as a promising approach, enabling LLMs to reference external knowledge sources and generate accurate outputs. We aimed to clarify the potential of RAG-enhanced LLMs with Japanese input in the field of radiotherapy. This was assessed by evaluating performance on three certification examinations in Japan: the Japanese Medical Physicist Examination, the Japanese Board Examination for Radiologists, and the Japanese Board Examination for Radiation Oncologists. In this study, we constructed a RAG system named Rad-Hub, consisting of a Japanese Radiotherapy Knowledge Database (JRKD) and a retrieval framework built on Microsoft Azure. The JRKD was populated with 32 Japanese radiotherapy textbooks and clinical guidelines. We assessed its utility by inputting all multiple-choice questions from the three examinations into ChatGPT-4o, both with and without Rad-Hub, and recording the answers. They were then compared with reference answers determined by experienced medical physicists and radiation oncologists. Rad-Hub improved accuracy across all examinations. Accuracy increased from 77.0% ± 2.6% to 84.6% ± 1.5% in the Medical Physicist examination, from 74.9% ± 2.0% to 82.1% ± 1.1% in the Radiologist examination, and from 55.6% ± 4.4% to 71.7% ± 4.5% in the Radiation Oncologist examination. Performance gains ranged from 7.2% to 16.1%. These findings highlight the potential of RAG-enhanced LLMs, particularly ChatGPT-4o with Rad-Hub, for integration into radiotherapy applications, such as educational and clinical decision assistance.
27 May 2026
Read appraisal →Studies in health technology and informatics
Digital GO: A Clinical Decision-Support Dashboard for Gynecologic Oncology Tumor Board
Multidisciplinary Tumor Board (MTB) meetings are essential for coordinated oncologic decision-making but are often hindered by the complexity and fragmentation of patient data. This study presents the development of an interactive digital dashboard designed to streamline data access and visualization during ovarian cancer MTB sessions. The system leverages structured knowledge extracted from unstructured clinical reports using in-house large language model (LLM) pipelines. These outputs are harmonized within a cloud environment through dedicated processing workflows and mapped into a domain ontology, to ensure interoperability. The resulting dashboard, implemented using a dedicated functional module, enables both cohort-level exploration and patient-specific visualization, including longitudinal event timelines. Preliminary results demonstrate the feasibility of integrating LLM-based data extraction with a scalable visualization framework, offering a step toward the digital transformation of multidisciplinary oncology workflows.
24 May 2026
Read appraisal →Studies in health technology and informatics
Hybrid AI Decision Support in Healthcare: Insights into Usability and Trust
Decision Support Systems (DSS) increasingly integrate Artificial Intelligence (AI) to enhance clinical reasoning, yet adoption depends on professionals' trust and usability. This study explored how hybrid AI decision support, combining rule-based guideline logic with generative reasoning, can be designed and evaluated in practice. Using a human-centred design approach within preventive child healthcare, a prototype integrating Large Language Models (LLMs) into an existing guideline-based DSS was developed and tested. Usability and trust were evaluated with nine healthcare professionals through task observations, usability scales, and interviews. The system showed good usability and moderate trust; highest scores were for traceability, workflow integration, and autonomy, while confidence indicators and contextual reasoning needed improvement. The study demonstrates the feasibility and acceptability of hybrid AI DSS in child healthcare and provides initial design requirements for usable, trustworthy, and human-centred AI integration across clinical domains.
24 May 2026
Read appraisal →PloS one
A qualitative study to inform the design and implementation of AI-driven diagnosis: Challenges, barriers, and clinical insights of physicians
INTRODUCTION: Acute febrile illnesses are significant clinical challenge in tropical regions due to overlapping symptomatology among infections such as dengue, scrub typhus, and leptospirosis. This diagnostic dilemma leads to delays in appropriate treatment and negatively impacted patient outcome due to limited diagnostic tool availability. Recent advancement in artificial intelligence (AI) application in healthcare help in developing diagnostic aiding tool in such clinical dilemma. AIMS: Study aimed to explore physicians' perspectives on diagnostic challenges of tropical Acute-febrile-illnesses and assess perceived utility of AI as a supportive tool in clinical decision-making. METHODS: A qualitative interview was conducted using validated semi-structured interview guide using literature and expert opinion and validated by IPR method. To understand the physicians perspective on management infectious diseases in critical care, interview guide administered among ten physicians with 10-30 years of experience across different centres of India. Interviews were audio-recorded, transcribed verbatim, and thematically analyzed using ATLAS.ti software. RESULTS: Four major themes emerged: clinical experience and practice, diagnostic-challenges, diagnostic-parameters and decision-making, and AI integration in clinical-practice. However, no statistical test can be performed since study mainly focused to understand perception of physicians with respect to managing illness and hence prevalence cannot be determined. Diseases such as dengue, scrub typhus, and leptospirosis were reported as most prevalent, particularly during monsoon seasons. Key diagnostic challenges included overlapping symptoms, diagnostic delays, limited sensitivity/specificity of available tests, and cost constraints. Physicians primarily relied on clinical judgment and laboratory findings, despite the limitations. Most of the participants expressed positive attitudes toward the potential of AI as supportive tool, citing its ability to enhance diagnostic accuracy and streamline clinical workflow with proper validation. CONCLUSIONS: Growing in clinical dilemma AI has significant role in diagnosis of tropical fever with integration of the clinical data & physician perspective can successfully able to incorporate in clinical application.
23 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 →PloS one
Learning semantic similarity from sentence pairs using hybrid features centric approach and explainable siamese neural networks.
Semantic embeddings play an important role in modern natural language processing because they help models understand meaning beyond individual words. Accurate text similarity is essential for many applications such as search, automated scoring, summarization, and question answering. However, existing methods based on Term Frequency-Inverse Document Frequency (TF-IDF) or simple lexical overlap often fail when sentences differ in length, structure, or word choice. These methods are less in performance, especially when working with short or medium-length sentences where meaning is expressed in different ways. This study explores sentence-level similarity using a Siamese BiLSTM model that learns deep semantic patterns and relationships between two sentences. The model captures contextual meaning, word interactions, and paraphrastic variations more effectively than traditional approaches. Experimental results show that the proposed model achieves the highest performance among machine-learning regressors, with lower errors and improved stability. Compared to TF-IDF, cosine similarity, and feature-based regressors, the Siamese model provides more accurate judgments of semantic closeness with RMSE of 0.16 and MAE of 0.107. Feature-level analysis using TF-IDF, Jaccard similarity, and embedding distances further supports these findings. Explainable AI techniques (SHAP, LIME) confirm model transparency by highlighting meaningful semantic cues and distributing attention across important linguistic features.
23 May 2026
Read appraisal →Journal of medical Internet research
Explainable AI in Cancer Imaging: Scoping Review of Methods, Modalities, and Clinical Integration
BACKGROUND: Explainable artificial intelligence (xAI) is increasingly used in medical imaging to enhance transparency, clinical interpretability, and trust in artificial intelligence (AI)-assisted diagnostics, particularly in oncology. Evidence on how explainability is implemented, validated, and reported in cancer imaging remains fragmented. OBJECTIVE: This scoping review aimed to systematically map research applying xAI methods to radiologic cancer imaging, summarize methodological and clinical trends, and identify persistent gaps in validation and integration. METHODS: We conducted a structured search of PubMed and Scopus (final search executed on October 20, 2025), covering studies published from 2017 to December 2024. Eligible peer-reviewed articles using machine learning or deep learning were analyzed with a focus on xAI components. Data from 371 studies were extracted into predefined categories covering cancer type, imaging modality, AI model, xAI method, terminology, validation, code availability, and decision support system integration. RESULTS: Studies focused primarily on breast (112/371, 30.2%), lung (87/371, 23.5%), and brain (56/371, 15.1%) cancers, with prostate, thyroid, and liver cancers also represented. The primary imaging modalities were computed tomography (139/371, 37.5%) and magnetic resonance imaging (104/371, 28%). Deep learning was used in 70.1% (260/371) of studies, classical machine learning in 18.1% (67/371), hybrid pipeline methods for 10% (37/371), and emerging concept-, prototype-, or causal-based approaches accounted for 1.9% (7/371) of studies. Post hoc xAI methods were dominant (305/371, 82.2%), with visualization (163/371, 53.4%), and feature relevance (111/371, 36.4%) as the most common subcategories. Hybrid post hoc or inherent approaches comprised 12.1% (45/371) and intrinsically interpretable methods 5.7% (21/371). Data sources were mostly public (149/371, 40.2%) or mixed (100/371, 26.9%); 22.9% (85/371) used private institutional datasets, and 7.8% (29/371) did not report data sources. Among validated studies, expert or user-based validation was most common (104/193, 53.9%), followed by mixed methods (74/193, 38.3%), while quantitative metrics (10/193, 5.2%) and clinical knowledge-based (8/193, 4.1%) validation remained rare. Only 17.5% (65/371) of studies provided code and 12.1% (45/371) reported decision support system integration, with few achieving actual clinical deployment. CONCLUSIONS: This scoping review maps xAI implementation across multiple cancer imaging modalities, revealing methodological inconsistency and insufficient validation. Most research emphasizes visualization over quantitative interpretability, and few models are reproducible or clinically implemented. These findings provide an evidence base for researchers, clinicians, and regulators to prioritize standardization of xAI reporting, quantitative validation, and user-centered frameworks to advance trustworthy AI in oncology imaging.
22 May 2026
Read appraisal →Journal of medical Internet research
Alignment Between Cardiologists and AI-Driven Diagnostic Systems: Mixed Methods Study
BACKGROUND: The clinical value of artificial intelligence (AI)-based diagnostic systems depends not only on their accuracy but also on how well their outputs integrate with clinicians' judgments in practice. Critical knowledge gaps remain regarding diagnostic concordance between AI and clinicians in stress echocardiography interpretation, patient characteristics predicting discordance, and how cardiologists respond when AI recommendations conflict with their clinical diagnoses. OBJECTIVE: This study examined the diagnostic alignment between an AI-driven stress echocardiography system (EchoGo Pro [EGP]) and cardiologists' diagnoses of coronary artery disease (CAD), identified predictors of concordance and AI scan rejection, and explored cardiologists' decision-making strategies when disagreements arise. METHODS: We conducted mixed methods research. The quantitative study analyzed concordance between EGP and cardiologists using data from 854 participants with suspected CAD in the multicenter PROTEUS randomized controlled trial. Logistic regression identified predictors of agreement, disagreement, and scan rejection, adjusting for age, sex, smoking status, BMI, and cardiovascular risk factors (hypertension, hypercholesterolemia, diabetes, family history of CAD, and prior CAD events). To gain deeper insight into discordance, we conducted a qualitative study analyzing survey responses from 61 UK consultant cardiologists recruited via Qualtrics, exploring their perceptions of AI tools, the risks of following discordant AI recommendations, and their typical responses to AI-clinician disagreement. RESULTS: EGP and cardiologists agreed in 60% (512/854) of the cases, but agreement was significantly lower among patients with hypertension (OR 0.58, 95% CI 0.38-0.89; P=.01), diabetes (OR 0.56, 95% CI 0.35-0.90; P=.02), and pre-existing CAD (OR 0.48, 95% CI 0.30-0.77; P=.002). EGP rejected 26.1% (222/854) of the scans due to insufficient image quality, with rejection significantly more common in male patients (β=0.35; P=.03) and those with a family history of CAD. If a positive CAD diagnosis was assigned when either cardiologists or EGP identified CAD, the proportion of positive cases increased from 17.9% (153/854) to 22.1% (189/854), potentially identifying additional at-risk patients. Survey respondents (50/60, 85% male; 26/57, 46% aged 40-49 years; 39/61, 64% White) required 65% to 69% confidence in their initial diagnosis to justify disregarding contradictory AI recommendations. The survey findings revealed cardiologists treated AI recommendations as advisory rather than definitive. When facing discordance, they retained confidence in their judgment and sought corroboration through additional testing, data review, or second opinions rather than deferring to AI. Paradoxically, cardiologists with higher confidence in AI tools required greater confidence in their own diagnosis to disregard AI recommendations (β=7.73; P=.02). Cardiologists attributed discordance primarily to AI's inability to incorporate patient history, comorbidities, and broader clinical context. CONCLUSIONS: EGP shows promise as an adjunctive tool but struggles with multimorbid patients and exhibits high, uneven rejection rates. Cardiologists use AI to prompt scrutiny, not replace judgment. Future systems need to integrate wider patient data with imaging and minimize bias through representative training to avoid exacerbating inequities.
21 May 2026
Read appraisal →PloS one
Implicit bias in safety-aligned large language models: A multi-faceted evaluation of clinical decision-making and health equity
BACKGROUND: Large language models are increasingly integrated into healthcare for clinical decision support and patient communication. Although these models can pass explicit social bias tests, they may retain implicit biases-latent associations between social groups and attributes-that could influence medical judgment. OBJECTIVE: To systematically evaluate the presence, magnitude, and behavioral impact of implicit biases in large language models within the medical domain across six high-stakes categories: gender, race, socioeconomic status, health conditions, religion, and healthcare systems. DESIGN: A descriptive cross-sectional study using a multi-faceted evaluation framework. SETTING(S): Computational analysis of 10 mainstream global large language models, including proprietary models (ChatGPT-4o, Gemini-2.0-Flash) and open-source models (DeepSeek-V3, Qwen3). METHODS: We constructed 24 medical bias datasets across six categories. Bias was assessed using three methods: (1) the Large Language Model Word Association Test, a prompt-based method for revealing implicit biases; (2) the Large Language Model Relative Decision Test, a strategy for detecting subtle discrimination in situational decision-making; (3) Paired-Prompt Analysis, used to examine whether implicit associations predict discriminatory decisions. RESULTS: All 10 models exhibited systematic implicit biases (Mean IAT Bias > 0) across all categories, with the strongest biases observed in Race (Mean = 0.61) and Socioeconomic Status (Mean = 0.56). Advanced reasoning capabilities (Chain-of-Thought) did not significantly reduce bias magnitude. Crucially, stronger implicit associations significantly predicted discriminatory choices in downstream medical decision tasks (p < 0.001). CONCLUSION: Current safety alignment techniques fail to eliminate implicit biases in large language models within the medical domain. These latent associations translate into biased decision-making, posing risks for health equity. Future development must prioritize representational debiasing over superficial alignment. Furthermore, healthcare professionals must embrace a stance of "AI vigilance": they should critically evaluate algorithmic outputs as fallible "second opinions" rather than objective truths, thereby ensuring that human judgment remains the ultimate safeguard for equitable patient care.
20 May 2026
Read appraisal →Scientific reports
A human-centric fuzzy decision support system for medical diagnosis using fuzzy cognitive maps
Medical decision support requires models that remain interpretable under uncertainty while still adapting to evolving data and expert knowledge. Fuzzy cognitive maps (FCMs) are attractive in this setting because they encode concept-level relations in a transparent graphical form. However, static expert-defined FCMs are often too rigid, whereas purely data-driven updates may weaken semantic consistency and drift away from clinically meaningful relations. This paper proposes a human-centric adaptive framework for fuzzy cognitive map-based medical decision support. The proposed Human-Centric Fuzzy Decision Support System (HCFDSS) combines data-driven weight adjustment with an explicit expert-correction operator, allowing the causal structure of the map to be refined without discarding domain knowledge. The method is formulated through a two-stage procedure: an inner reasoning loop updates concept activations for fixed weights, while an outer learning loop updates the weight matrix using a regularized objective together with bounded expert intervention. Under standard smoothness and boundedness assumptions, we establish sufficient conditions for stability of the inner reasoning dynamics and convergence of the outer weight-update operator. A numerical illustration shows how expert feedback can revise clinically important edges and alter the final diagnostic activation in an interpretable way. In addition, the paper reports a reproducible empirical evaluation involving synthetic medical scenarios and three public medical benchmark datasets from the UCI Machine Learning Repository, together with sensitivity analysis for expert intervention, preprocessing validation, robustness checks under missingness and imbalance, and expanded comparisons with mainstream machine-learning baselines. The proposed HCFDSS is positioned not as a replacement for high-capacity black-box predictors, but as a transparent and editable decision-support framework in which clinicians can inspect, question, and refine the learned relations. The main contribution of the present work is therefore methodological and theoretical, while the empirical evaluation clarifies the predictive-interpretability trade-off of the proposed framework under reproducible benchmark settings.
20 May 2026
Read appraisal →European journal of pediatrics
Development and validation of an explainable machine learning model for predicting surgical intervention in pediatric intestinal obstruction
Pediatric intestinal obstruction, a critical acute abdomen condition, carries a risk of intestinal necrosis. Decisions for urgent surgery lack standardized criteria. This study aimed to develop an explainable machine learning (ML) model to predict surgical indication, reducing diagnostic variability. A retrospective study included 642 children from our center (split into training/internal validation sets) and 137 from two external centers. Surgical indication was defined based on the presence of one or more of the following: evidence of peritonitis on physical examination, radiographic signs of strangulation, or failure of conservative management for > 24 h; in cases of diagnostic uncertainty, consensus among three senior physicians was used as the adjudication standard. We built models using 35 easily obtainable features across eleven algorithms and evaluated performance via AUC and other metrics. SHapley Additive exPlanations (SHAP) provided model interpretability, and an optimal probability threshold was established for clinical stratification. At the optimal threshold determined by maximizing the Youden index, the model achieved a sensitivity of 89% and a specificity of 99% in the internal validation set. The random forest model demonstrated the highest discriminative performance. A streamlined, interpretable model using the eight most important features (including vital sign dynamics and clinical markers) showed no evidence of overfitting based on the close agreement between training and validation performance metrics. It achieved AUCs of 0.988 (internal validation) and 0.981 (external validation). The model has been translated into a convenient clinical tool.Conclusion: Our explainable ML model demonstrates strong discriminative performance in predicting the need for surgery in pediatric intestinal obstruction and is strongly associated with intestinal necrosis. By mitigating the "black-box" concern through SHAP, it offers transparent, interpretable decision support to clinicians.
20 May 2026
Read appraisal →JMIR medical informatics
Multimodal Fusion of Echocardiogram Images and Electronic Medical Records for Heart Disease Screening: Retrospective Algorithm Development and Validation Study
BACKGROUND: Echocardiography is a fundamental imaging modality for the diagnosis of heart disease (HD), but its interpretation remains operator-dependent and lacks standardized, data-driven decision support. Although artificial intelligence has improved image-based diagnosis, the added value and interpretability of integrating routinely collected electronic medical records (EMRs) with echocardiogram (ECHO) for large-scale screening remain underexplored. OBJECTIVE: This study aims to develop an explainable artificial intelligence framework that integrates multimodal data, including ECHO images and EMRs, with derived temporal and clinical features, to enhance the screening and interpretability of heart disease diagnosis. METHODS: In this retrospective single-center study, we analyzed 26,936 patients with linked ECHO and EMR records (1470 with clinically confirmed HD and 25,466 without). We constructed cross-sectional EMR features (demographics, comorbidities, and registration history) and extracted view-specific image features with a simplified Inception-v3 backbone. Five modality-specific feature vectors (4 imaging views + EMR) were concatenated and input to an extreme gradient boosting classifier within a patient-level stratified 5-fold cross-validation framework. To mitigate class imbalance during training, non-HD cases in the training folds were randomly downsampled to a 1:3 HD:non-HD ratio; held-out test folds preserved the original prevalence. Post hoc explainability was provided via Grad-CAM heatmaps for images and Shapley additive explanations analysis for EMR features. RESULTS: For the primary binary task (presence vs absence of HD), the multimodal model achieved an AUC of 0.8147 (SD 0.009) on held-out test folds, compared with 0.7785 (SD 0.014) for the image-only baseline, and 0.7343 (SD 0.009) for the EMR-only baseline. At the clinically selected decision threshold, sensitivity, specificity, positive predictive value, and negative predictive value were 84.6% (SD 1.6%), 72.4% (SD 2%), 15% (SD 2.3%), 98.8% (SD 0.3%) for the multimodal model, versus 82.3% (SD 1.9%), 73.5% (SD 2%), 15.2% (SD 2.1%), 98.6% (SD 0.3%) (image-only), and 77.5% (SD 2.1%), 67% (SD 2.3%), 11.9% (SD 1.7%), 98.1% (SD 0.4%) (EMR-only). Grad-CAM visualizations qualitatively highlighted anatomically and physiologically meaningful regions (eg, valvular structures and abnormal flow jets), while Shapley additive explanations analysis identified age, registration years, sex, and hypertension among the top EMR contributors-findings that align with established cardiovascular risk factors. CONCLUSIONS: Integrating echocardiographic images and EMR data in an explainable multimodal framework yields improved and clinically plausible screening performance for heart disease. Future work should focus on external multicenter validation, quantitative assessment of visual explanations against expert annotations, and prospective studies assessing clinical use and workflow integration.
17 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
Revisiting Rogers' Paradox in the context of human-AI interaction
People learn about the world in many ways: from conducting experiments to copying others' behaviour. The choices we make about how to learn can impact the collective understanding of a whole population, were others to learn from us. Alan Rogers developed simulations to study these phenomena-where agents could individually or socially learn amidst a dynamic, uncertain world-and uncovered a surprising result: the availability of cheap social learning yielded no benefit to population fitness over individual learning. Rogers' Paradox spawned decades of work to understand factors that favour social learning and better model human cultural development. But what happens when humans can learn from artificial intelligence (AI) systems that are themselves learning from us? We revisit Rogers' Paradox in the context of human-AI interaction and extend the simulations towards a simplified network of humans and AIs learning together about an uncertain world. We examine the impact of several learning strategies on the equilibrium of a society's 'collective world model' and assess levers available to stakeholders in human-AI interactions to change network dynamics. We then model negative feedback loops that may arise from humans learning socially from AI, and consider other open questions that could be explored in our simulation framework. This article is part of the theme issue 'World models in natural and artificial intelligence'.
17 May 2026
Read appraisal →JMIR medical informatics
A Multiassessment and Multiprofessional Agents Approach for Medical Chatbot Risk Estimation: Development and Evaluation Study
BACKGROUND: Assessing chatbot responses across 3 domains-medical, ethical, and legal-is essential to ensuring the safe use of artificial intelligence in health care. Although advancements in the use of large language models (LLMs) show significant improvements in evaluating question-answer datasets, such as multiple-choice medical exams, existing systems use general LLMs without incorporating specialized domain knowledge. They rely on standardized instructions without integrating real-world information, and ensemble methods such as majority voting fail to resolve disagreements among agents, resulting in misclassification and challenges in risk assessment. OBJECTIVE: This study aims to design, develop, and evaluate a synergistic approach for assessing risks associated with chatbot responses using multiassessment (MA) and multiprofessional agents (MPAs). METHODS: We designed and developed an approach consisting of MA and MPA, specifically initial assessment (MA1), which internalizes 3 roles and provides an initial risk estimation, and final assessment (MA3), which aims to reach a final consensus based on the previous assessments (MA1 and MA2), with each using 1 LLM. The verification assessment (MA2) incorporates an MPA or role-based LLM specialized agents for each risk domain (medical, ethical, and legal). We evaluated the proposed approach using the MedNLP-CHAT (Medical Natural Language Processing for AI Chat) corpus (N=226; 100 train, 126 test), covering baseline, enhanced prompt, embedding-based search, and retrieval-augmented generation (RAG). Primary metrics included macro F1-score and joint accuracy to evaluate system performance, along with CI and paired macro F1-score difference (Δ) as supporting metrics to assess the approach's effectiveness. RESULTS: The MA-MPA framework integrated with RAG achieved the highest average macro F1-score of 0.800 across risk domains and a joint accuracy of 76 (60.3%) correct predictions across all risk domains out of 126 question-answer pairs, with notable improvements over the best reported Eighteenth NII Testbeds and Community for Information Access Research Project (NTCIR-18) MedNLP-CHAT systems in the ethical (+0.252) and legal (+0.096) risk domains, while the medical domain showed a modest increase of +0.070. The MA approach contributed the largest gains, particularly from MA1 to MA2, with paired macro F1-score gains ranging from +0.176 to +0.214 across systems. The MPA approach performed better when integrated with MA and external knowledge, with paired bootstrap estimates showing a gain of +0.037 (95% CI 0.003-0.074) over baseline; however, joint accuracy gains were not evident (95% CI -2.9% to 7.7%), and gains relative to the enhanced prompt were small. Notably, MA alone achieved higher joint accuracy than RAG (62.7% vs 60.3%), indicating a metric-specific trade-off rather than consistent superiority across all metrics. CONCLUSIONS: The MA-MPA approach shows potential for improving risk estimation in chatbot responses. The results suggest that the framework is particularly useful for enhancing balanced overall performance, especially when combined with external knowledge, although the medical risk domain remains challenging. Furthermore, more specialized LLMs may further improve contextually grounded risk estimation.
17 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
Levels of analysis for large language models
Modern artificial intelligence systems, such as large language models, are increasingly powerful but also increasingly hard to understand. Recognizing this problem as analogous to the historical difficulties in understanding the human mind, we argue that methods developed in cognitive science can be useful for understanding large language models. We propose a framework for applying these methods based on the levels of analysis that David Marr proposed for studying information processing systems. By revisiting established cognitive science techniques relevant to each level and illustrating their potential to yield insights into the behaviour and internal organization of large language models, we aim to provide a toolkit for making sense of these new kinds of minds. This article is part of the theme issue 'World models in natural and artificial intelligence'.
15 May 2026
Read appraisal →Systematic reviews
Barriers and facilitators to the implementation and adoption of computerised clinical decision support systems: an overview of reviews
BACKGROUND: The use of computerised clinical decision support systems (CDSSs) holds considerable potential in enhancing healthcare delivery by improving patient safety, practitioner performance, and patient outcomes. However, despite the numerous reported advantages of CDSSs, their adoption remains restricted, thereby compromising the full realisation of their potential. To enable a profound evaluation and successful implementation of these systems, it is imperative to identify the reasons for their limited uptake. The objective of this study is to provide a comprehensive overview of the barriers and facilitators to the implementation and adoption of decision support systems across healthcare settings. METHODS: This study was reported in accordance with the Preferred Reporting Items for Overviews of Reviews (PRIOR) statement. A systematic search was conducted in the databases PubMed, IEEE Xplore, Scopus, and Web of Science from inception to 15 October 2024, targeting at review articles of primary studies focusing on the identification and reporting of barriers and facilitators to the implementation and adoption of CDSSs across healthcare settings. The risk of bias of the included reviews was assessed with the Risk of Bias in Systematic Reviews (ROBIS) tool, and the primary study overlap was calculated using the corrected covered area method. All data on barriers and facilitators to the implementation and adoption of CDSSs were extracted from the included reviews, synthesised through a mapping of identical factors, and ranked according to their frequency of occurrence. RESULTS: Of the 1640 records retrieved through database searches, 30 reviews were included in this overview, which summarised the findings of 721 unique primary studies. A total of 101 distinct factors influencing the implementation and adoption of CDSSs were identified and could be categorised into the overarching categories of human, technology-related, and contextual factors. The following six factors were reported most frequently: usability, usefulness/perceived benefits, organisational readiness, training, trust, and workflow integration. CONCLUSIONS: The findings of this study highlight the diversity of factors that influence a successful implementation of CDSSs and emphasise the need for their comprehensive evaluation that goes beyond the assessment of general performance aspects but takes into account human and contextual factors. SYSTEMATIC REVIEW REGISTRATION: PROSPERO CRD42024507614.
15 May 2026
Read appraisal →PloS one
Classification of soft decision-making methods via fuzzy parameterized fuzzy soft matrices and their performance-based statistical analysis in machine learning
This study provides a comprehensive evaluation and classification of 35 soft decision-making (SDM) algorithms based on fuzzy parameterized fuzzy soft matrices (fpfs-matrices). Although fpfs-matrices offer a strong mathematical framework for modeling uncertainty, there has been a lack of large-scale comparisons of their derivative SDM methods in machine learning. To address this, we used the Comparison Matrix-Based Fuzzy Parameterized Fuzzy Soft Classifier (FPFS-CMC) to benchmark these 35 algorithms across ten diverse datasets from the UCI Machine Learning Repository. The methods were thoroughly assessed utilizing metrics such as accuracy, precision, recall (sensitivity), specificity, and F1-score, and statistical significance was confirmed using the Friedman and Nemenyi tests. Our results show that SDM methods via fpfs-matrices perform competitively in classification tasks involving uncertainty. Notably, the best algorithms according to F1-scores were A19 (Rank 1), YHX14 (Rank 2), and a three-way tie for Rank 3 among VMH16, AKO18o, and A19/2. By identifying the most effective algorithms and offering a structured decision-support framework, this research provides both a theoretical reference and practical guidance for practitioners selecting SDM methods for complex machine learning challenges.
15 May 2026
Read appraisal →Journal of medical systems
When Chatbots Become Agents: The Next Phase of Healthcare AI
Healthcare artificial intelligence (AI) has moved beyond answering medical questions. In early 2026, OpenAI, Anthropic, and Google launched agentic systems that retrieve evidence, use tools, and execute multi-step workflows. These systems can integrate information across multiple clinical knowledge domains within a single workflow, but the evidence base for evaluating their reliability in clinical practice has not kept pace with deployment. This perspective argues that healthcare is adopting a technology evolving faster than anyone can evaluate, and that a brief window exists to shape integration before the architecture becomes entrenched. These systems share common architectural principles, yet only 19 prospective trials exist among over 4,600 clinical large language model studies, none evaluating agentic pipelines. The architecture is locking in, professional roles are being redefined, and builders hold a structural information advantage. Transparency, adaptive governance, and broader participation are needed before these choices become irreversible, and to ensure that the boundary between human clinical judgment and automated capability is drawn in a way that strengthens both. Every stakeholder in healthcare must engage with this shift now.
15 May 2026
Read appraisal →Heart (British Cardiac Society)
Individual hearts: computational models for improved management of cardiovascular disease
Cardiovascular disease remains a leading cause of morbidity and mortality worldwide, with conventional management often applying standardised approaches that struggle to address individual variability in increasingly complex patient populations. Computational models, both knowledge-driven and data-driven, have the potential to reshape cardiovascular medicine by offering innovative tools that integrate patient-specific information with physiological understanding or statistical inference to generate insights beyond conventional diagnostics. This review traces how computational modelling has evolved from theoretical research tools into clinical decision support systems that enable personalised cardiovascular care. We examine this evolution across three key domains: enhancing diagnostic accuracy through improved measurement techniques, deepening mechanistic insights into cardiovascular pathophysiology and enabling precision medicine through patient-specific simulations. The review covers the complementary strengths of data-driven approaches, which identify patterns in large clinical datasets, and knowledge-driven models, which simulate cardiovascular processes based on established biophysical principles. Applications range from artificial intelligence-guided measurements and model-informed diagnostics to digital twins that enable in silico testing of therapeutic interventions in the digital replicas of individual hearts. This review outlines the main types of cardiovascular modelling, highlighting their strengths, limitations and complementary potential through current clinical and research applications. We also discuss future directions, emphasising the need for interdisciplinary collaboration, pragmatic model design and integration of hybrid approaches. While progress is promising, challenges remain in validation, regulatory approval and clinical workflow integration. With continued development and thoughtful implementation, computational models hold the potential to enable more informed decision-making and advance truly personalised cardiovascular care.
14 May 2026
Read appraisal →PloS one
Using deep learning to identify inherited retinal diseases based on wide-field retinal imaging data
OBJECTIVE: To evaluate a novel image-based deep learning method for the automated identification of inherited retinal diseases (IRDs) and to explore the feasibility of predicting selected causative gene groups using a multimodal analysis of wide-field fundus autofluorescence (FAF) and pseudocolor fundus (pCF) images. DESIGN: The method was evaluated using a retrospective dataset of patient studies containing FAF and pCF images, as well as genetic tests for IRD. PARTICIPANTS: Patients with confirmed IRD for which both wide-field FAF and pCF images and genetic tests for IRD performed at Hadassah University Medical Center were included. The dataset consisted of 409 patients (330 patients with IRD with the 25 most commonly affected genes in our population and patients without IRD, and 79 patients without IRD). METHODS: Nine EfficientNet-V2-m convolutional neural networks were trained for the following three classification tasks: a binary IRD vs. non-IRD classification, and classification into two groups of five causative genes (Groups 1 and 2). For each task, three models were trained on the FAF images only, the pCF images only, and both the FAF and pCF images. The performance of the models was then evaluated and compared using 5-fold cross-validation. MAIN OUTCOME MEASURES: Accuracy, precision, F1 scores, AUC, and confusion matrices. RESULTS: The multimodal classification models that were trained on both the FAF and pCF images yielded the best results. The binary classification model had a mean (±SD) accuracy of 0.95 ± 0.01, a mean precision of 0.92 ± 0.01, and a mean F1 score of 0.90 ± 0.02. The Group 1 classification model had a mean accuracy of 0.92 ± 0.03, a mean precision of 0.93 ± 0.03, and a mean F1 score of 0.89 ± 0.03. Finally, the Group 2 classification model had a mean accuracy of 0.85 ± 0.03, a mean precision of 0.87 ± 0.04, and a mean F1 score of 0.83 ± 0.04. CONCLUSIONS: Our results indicate that determining whether a patient has IRD can be performed with high accuracy within this retrospective cohort based on FAF and pCF images using image-based deep learning classifiers. This image-based approach may assist clinicians during the patient's initial visit by providing decision support prior to genetic testing. It may also help prioritize patients for genetic workup, particularly in settings in which genetic testing is not readily available. Further prospective and external validation is required before clinical implementation.
13 May 2026
Read appraisal →Journal of medical systems
Clinical Plausibility in Large Language Model Robustness Testing for Medicine: A Scoping Review
Large language models (LLMs) show promise in medical applications, yet their translation into clinical practice requires rigorous validation. Current robustness testing often employs adversarial approaches borrowed from AI safety, raising questions about their alignment with authentic clinical scenarios. To systematically map methodologies used for robustness testing of LLMs in medical contexts and assess their clinical plausibility. A scoping review was conducted following PRISMA-ScR guidelines, searching PubMed, Embase, Web of Science, IEEE Xplore, ACM Digital Library, arXiv, and MedRxiv from January 2023 to September 2025. Two independent physician reviewers screened 5,331 articles, extracting data on testing methodologies, medical domains, expert involvement, and clinical plausibility. Thirty-three studies met inclusion criteria, predominantly from 2025 (82%). The most common robustness testing approaches were misleading prompts (49%) and adversarial prompts (39%). Only 33% of studies designed tests clearly mimicking plausible clinical scenarios. While 58% reported expert involvement, the depth of integration varied considerably. Studies predominantly addressed mixed medical domains (73%) rather than specialized fields. The emerging literature suggests that LLM robustness testing in medicine often emphasizes technical vulnerability detection, with fewer studies examining clinically plausible scenarios of routine use. Future frameworks should complement adversarial testing with clinically grounded, longitudinal, and specialty focused evaluations to support deployment-relevant inference.
13 May 2026
Read appraisal →PloS one
A modular deep learning architecture for interpretable disease prediction across tabular clinical and biometric datasets
Accurate disease prediction using clinical datasets is essential for improving early diagnosis and clinical decision-support systems; however, many existing deep learning approaches are disease-specific, computationally intensive, and difficult to generalize across heterogeneous biomedical datasets. This study addresses this challenge by proposing a unified and dataset-aware deep learning framework that enables accurate and interpretable disease prediction across diverse clinical datasets. The framework adopts a modular architecture that selects appropriate models based on dataset characteristics such as feature dimensionality, sample size, and class imbalance. It integrates multiple deep learning architectures, including MLP, one-dimensional CNN, FT-Transformer, autoencoder-based classifiers, and ensemble strategies. Robust preprocessing, fold-safe feature selection, and nested cross-validation are incorporated to ensure reliable performance evaluation. The framework is evaluated on three heterogeneous benchmark datasets: the UCI Heart Disease dataset (303 samples, 13 clinical features), the PIMA Indians Diabetes dataset (768 samples, 8 metabolic features), and the Parkinson's disease voice dataset (195 recordings, 22 acoustic features). Experimental results demonstrate competitive predictive performance relative to classical baselines across the diverse tasks. The FT-Transformer + autoencoder ensemble achieved an AUC of 0.8980 (±0.0483) for heart disease prediction, while the CNN + Autoencoder ensemble obtained an AUC of 0.8451 (±0.0270) for diabetes classification. For Parkinson's disease detection, the MLP achieved an AUC of 0.7538 with perfect specificity. Overall, all models achieved AUC values comparable to ML baselines. The study contributes a scalable and interpretable deep learning framework that improves reliability, generalization, and practical applicability for multi-disease prediction in real-world healthcare environments.
13 May 2026
Read appraisal →Medical science monitor : international medical journal of experimental and clinical research
Preliminary Evaluation of Large Language Models in Kennedy Classification and Removable Partial Denture Planning: An In Silico Study
BACKGROUND This study evaluated text-based large language models for Kennedy classification and removable partial denture planning in partially edentulous cases, where comparative evidence on accuracy remains limited. MATERIAL AND METHODS Twenty-six partially edentulous cases, defined according to the Federation Dentaire Internationale tooth numbering system and classified as Kennedy Classes I to IV, were included. Case scenarios were constructed by an experienced prosthodontist. Questions were independently submitted to 4 large language models without prior training or prompting: ChatGPT-5.2, Claude Sonnet 4.6, Gemini 3 Flash, and Perplexity. Responses were evaluated by 2 independent prosthodontists (different from the case developer) - blinded to the identity of each model - using predefined criteria for Kennedy classification accuracy and prosthetic planning consistency. Statistical analyses included intergroup comparisons and effect size estimation. RESULTS Statistically significant differences were identified among the models in both tasks (P<0.001). For Kennedy classification, the effect size was moderate (Kendall's W=0.402). Gemini 3 Flash achieved the highest mean score, followed by ChatGPT-5.2, Perplexity, and Claude Sonnet 4.6. Similarly, significant differences were observed in removable partial denture planning performance (Kendall's W=0.424); Gemini 3 Flash scored significantly higher than the other models (P≤0.001). CONCLUSIONS Large language models showed variable performance under zero-shot conditions. Although Gemini 3 Flash achieved higher scores, the moderate effect sizes warrant cautious interpretation. These models may serve as adjunctive decision-support tools in prosthodontics but cannot replace clinical judgment.
12 May 2026
Read appraisal →PloS one
DeepDRP: Dose-response predictions of drug pairs using deep learning based on data-driven feature representation and dose-response curve characteristics
Combination therapies have become a cornerstone of modern medicine, offering improved treatment outcomes and reduced side effects compared to monotherapies. However, the efficacy and safety of drug combinations depend heavily on the specific doses of each component, making the optimization of dosing regimens a crucial yet challenging task. Despite the importance of dose optimization, few computational methods address this challenge. Here, we propose DeepDRP, a novel approach that integrates two complementary models: the global model, which is trained using all available combinations, and the local model, which utilizes only samples with similar information to queries. The DeepDRP architecture comprises three main modules: the first module has three embedding networks to extract features from drugs, doses, and cell lines, and then predicts synergy using the fused knowledge. The second module constructs a graph using the input query and the set of similar retrieved training samples fed into a semi-supervised graph convolutional network to predict the synergy value. Finally, these two models are aggregated to have the final predicted value. To evaluate the proposed approach, it is applied to the NCI-ALMANAC dataset and O'Neil dataset. The obtained results denote that the proposed method achieves superior performance with respect to the other approaches.
11 May 2026
Read appraisal →Pediatric neurology
Electroencephalography-Based Machine Learning Models for Predicting Ketogenic Diet Outcomes in Pediatric Drug-Resistant Epilepsy
BACKGROUND: Ketogenic diet therapy (KDT) is an established treatment for drug-resistant epilepsy (DRE); however, methods for predicting its effectiveness remain underdeveloped. This study evaluated various machine learning (ML) models in predicting responses to KDT among DRE patients based on electroencephalography data. METHODS: Using leave-one-out cross-validation, this study evaluated 19 ML classifiers in predicting the outcomes of 90 DRE patients based on absolute and relative power, as well as functional connectivity measures (phase-locking value, phase lag index [PLI], and weighted PLI) across standard frequency bands. KDT significantly reduced seizure frequency at three and 6 months after initiation. RESULTS: The most effective classifier at 3 months was a Coarse Tree classifier trained on absolute power (recall = 0.933, precision = 0.767, F2 = 0.894, area under the receiver operating characteristic curve = 0.607). The most effective classifier at 6 months was a Gaussian Naive Bayes classifier trained on weighted PLI + relative power (recall = 0.759, precision = 0.854, F2 = 0.776, area under the receiver operating characteristic curve = 0.603). CONCLUSIONS: This study identified the most effective ML models for predicting KDT outcomes in DRE patients. The results highlight the potential of electroencephalography-based ML tools for guiding KDT treatment in clinical practice.
9 May 2026
Read appraisal →Journal of diabetes science and technology
Explainable Machine-Learning Models to Predict Weekly Risk of Hyperglycemia, Hypoglycemia, and Glycemic Variability in Patients With Type 1 Diabetes Based on Continuous Glucose Monitoring
BACKGROUND AND OBJECTIVE: The aim of this study was to develop and validate explainable prediction models based on continuous glucose monitoring (CGM) and baseline data to identify a week-to-week risk of CGM key metrics (hyperglycemia, hypoglycemia, glycemic variability). By having a weekly prediction of CGM key metrics, it is possible for the patient or health care personnel to take immediate preemptive action. METHODS: We analyzed, trained, and internally tested three prediction models (Logistic regression, XGBoost, and TabNet) using CGM data from 187 type 1 diabetes patients with long-term CGM monitoring. A binary classification approach combined with feature engineering deployed on the CGM signals was used to predict hyperglycemia, hypoglycemia, and glycemic variability based on consensus targets (time above range ≥5%, time below range ≥4%, coefficient of variation ≥36%). The models were validated in two independent cohorts with a total of 223 additional patients of varying ages. RESULTS: A total of 46 593 weeks of CGM data were included in the analysis. For the best model (XGBoost), the area under the receiver operating characteristic curve (ROC-AUC) was 0.9 [95% confidence interval (CI) = 0.89-0.91], 0.89 [95% CI = 0.88-0.9], and 0.8 [95% CI = 0.79-0.81] for predicting hyperglycemia, hypoglycemia, and glycemic variability in the interval validation, respectively. The validation test showed good generalizability of the models with ROC-AUC of 0.88 to 0.95, 0.84 to 0.89, and 0.80 to 0.82 for predicting the glycemic outcomes. CONCLUSION: Prediction models based on real-world CGM data can be used to predict the risk of unstable glycemic control in the forthcoming week. The models showed good performance in both internal and external validation cohorts.
9 May 2026
Read appraisal →Science (New York, N.Y.)
Performance of a large language model on the reasoning tasks of a physician.
More than 65 years ago, complex clinical diagnostic reasoning cases were introduced as the gold standard for the evaluation of expert medical computing systems, a standard that has held ever since. In this study, we report the results of a physician evaluation of a large language model (LLM) on challenging clinical cases across five experiments with a baseline of hundreds of physicians. We then report a real-world study comparing human expert and artificial intelligence (AI) second opinions in randomly selected patients in the emergency room of a major tertiary academic medical center. In all experiments, the LLM outperformed physician baselines and displayed continued improvement from prior generations of AI clinical decision support. Our study suggests that LLMs have eclipsed most benchmarks of clinical reasoning, motivating the urgent need for prospective trials.
5 May 2026
Read appraisal →Psychiatry research
A machine learning approach to predicting clinical trajectories in bipolar disorder
BACKGROUND: Bipolar disorder (BD) is characterized by significant heterogeneity in clinical presentation, age of onset, and diagnostic delays. Identifying clinical trajectories and subtypes is a central aim of precision psychiatry, with the goal of improving risk stratification and characterization of illness courses, enabling early recognition of high-risk patients and facilitating timely interventions that improve long-term outcomes. METHODS: This is a dual-site retrospective cross-sectional secondary analysis of 203 clinical datasets from BD patients to develop machine learning models: Random Forest, Weighted Logistic Regression (LR), and XGBoost, supplemented by statistical modelling, to predict age of onset timing (early/late onset) and illness severity. Survival analysis (Kaplan-Meier and Cox proportional hazards) investigated factors influencing diagnostic delay, while clustering identified patient subtypes. RESULTS: Single-split Weighted LR (AUC = 0.733) and nested cross-validation XGBoost (AUC = 0.729) achieved superior performance in predicting onset timing. Cluster analysis identified three distinct phenotypes: baseline low complexity, late-onset familial-severe, and an early-onset comorbid group. Survival analysis revealed that the time to diagnosis was independent of clinical risk profiles (Log-Rank p = 0.34), with the most severe cases experiencing significant diagnostic latency. CONCLUSIONS: Machine learning accurately predicts BD clinical trajectories using routinely collected data in an African cohort. Diagnostic delay did not differ significantly across clinical risk or severity strata, indicating that the measured variables did not explain variation in time to diagnosis.
4 May 2026
Read appraisal →BMJ health & care informatics
Omission and hallucination prevalence of clinical guidelines in diagnostic large language model outputs
OBJECTIVE: Meaningful assessments of how large language models (LLMs) incorporate clinical guidelines require large-scale testing over many queries. Here, we evaluate the prevalence of clinical guideline omissions and hallucinations in a large sample of diagnostic LLM outputs. METHODS: We used simulated case vignettes and zero-shot prompting to generate diagnostic outputs and rationales from GPT-4.1 and DeepSeek-V3. English case vignettes were created for hypercholesterolaemia and type-2 diabetes mellitus. Each vignette contained identical medical information, while sociodemographic characteristics varied in terms of sex, ethnicity and location. We calculated the prevalence of existing and hallucinated clinical guidelines in LLM outputs across disease, LLM and sociodemographic characteristics. RESULTS: We analysed a total of 12 197 LLM outputs, which quantifies three hazard areas: omissions (up to 97% for DeepSeek-V3 and 46% for GPT-4.1), hallucinations (up to 9%) and inconsistencies (guideline citation rate ranging from 0% to 78.39% across sociodemographic vignettes). Omission and hallucination rates were generally similar across vignettes with different sex or ethnicity data, yet were particularly sensitive to patient location. DISCUSSION: This study highlights significant variability in clinical guideline prediction across two different diseases, three different sociodemographic variables and two LLMs, even when the LLMs were instructed by identical prompts, establishing clinical guideline prediction in LLM outputs as a stochastic event. CONCLUSION: The stochastic nature of LLMs creates a unique challenge for evidence generation and clinical deployment. Being able to measure and capture this stochasticity within high-quality research designs will be a prerequisite to advancing the responsible deployment of LLMs in healthcare.
27 Apr 2026
Read appraisal →Journal of neurology
Predicting disease progression in multiple sclerosis with clinically accessible information and technology
BACKGROUND: Predicting disease progression at the individual level is essential for personalized medicine. We previously developed machine-learning tools to estimate 5-year progression risk in people with multiple sclerosis (PwMS). Such models should account for disease-modifying therapy (DMT) and objective outcome definitions. METHODS: In a retrospective multicenter case-control study, we evaluated adults with relapsing-remitting multiple sclerosis (RRMS) at baseline. Using machine-learning, we developed two complementary tools for individualized 5-year risk estimation: DAAE-M, optimized for transparency, software-neutral use, and mitigation of indication bias, and ELIE, optimized for dynamic landmark-based modeling, complex treatment histories, and mitigation of immortal-time bias. Disease progression was defined using both a clinical outcome (RRMS-to-progressive MS) and an objective outcome (late-stage confirmed progression independent of relapse activity). RESULTS: Among 34,510 people with RRMS (72.6% female, mean age = 37.1, mean disease duration = 5.8), 9.8% and 21% met clinical and objective progression criteria, respectively, over five years. Both models demonstrated good calibration across risk-groups (Brier scores 0.06-0.16). DAAE-M provided patient-level risk estimates with monotonic risk escalation across risk-groups for clinical (3.1%/11.2%/22.6%/33.0%) and objective (8.4%/14.5%/23.3%/38.8%) progression. For DAAE-M, high-efficacy DMT was associated with approximately half the progression risk compared with low-efficacy DMT (risk-ratios: 0.42-0.59; p < 0.01). ELIE also showed good calibration across risk deciles with increasing incidence for both clinical (0.3%/1.2%/1.7%/2.5%/3.7%/5.5%/7.2%/10.2%/14.3%/21.5%) and objective (0.9%/1.6%/2.5%/4.0%/5.8%/7.8%/10.2%/15.3%/20.9%/32.5%) outcomes. CONCLUSION: We developed two well-calibrated machine-learning-based tools for individualized 5-year prediction of clinically- and objectively-defined MS progression, each with distinct strengths in usability, bias handling, and treatment modeling. These findings support future tool use in personalized risk stratification and secondary prevention.
21 Apr 2026
Read appraisal →Scientific reports
End-to-end pipeline for automated heart failure diagnosis with clinical notes using SNOMED-CT
Diagnosis of heart failure is complex but crucial for patient outcomes and often hindered by the untapped potential of unstructured clinical notes. We introduce a novel end-to-end pipeline for heart failure diagnosis, leveraging electronic health records (EHR) and German clinical notes from 846 patients. Our pipeline synthesizes abbreviation disambiguation, translation of German clinical notes to English, medical entity linking to SNOMED-CT, and subsequent classification. The classification was performed using a Support Vector Machine (SVM) and compared against a fine-tuned medBERT.de neural baseline. We reduced the reliance on training data with zero-shot learning to address limitations with abbreviation disambiguation and entity linking approaches. Validation against benchmark datasets and cardiologists demonstrates high accuracy for real clinical use. Abbreviation disambiguation achieved an accuracy of up to 96.1%. Entity linking achieved competitive performance compared to state-of-the-art approaches on selected evaluation datasets. The SVM classification approach utilizing SNOMED-CT concepts and EHR data achieved an F1-score of 65.3%, on par with the medBERT.de neural baseline using clinical notes and EHR data. Despite challenges regarding limited language-specific resources and reference dataset availability for SNOMED-CT annotations in German, our pipeline demonstrates high potential for real-world clinical use and clinical decision support grounded in the standardized SNOMED-CT ontology.
21 Apr 2026
Read appraisal →PloS one
Comparison of OneChoice AI-based clinical decision support recommendations with infectious disease specialists and non-specialists for bacteremia treatment in Lima, Peru
Bacteremia is a major contributor to global morbidity and mortality, particularly in low- and middle-income countries where diagnostic delays and empirical antimicrobial misuse exacerbate resistance. This study assessed the accuracy of OneChoice®, an artificial intelligence (AI)-based Clinical Decision Support System (CDSS), in guiding antimicrobial therapy for bloodstream infections (BSIs) in Lima, Peru. A cross-sectional, observational design was used, comparing therapeutic recommendations generated by OneChoice®-based on molecular (FilmArray®) and phenotypic (MALDI-TOF MS, VITEK2) data-with the clinical decisions of 94 physicians (35 infectious disease [ID] specialists and 59 non-specialists) across 366 survey-based evaluations of bacteremia cases. Concordance between CDSS and physician decisions was analyzed using Cohen's Kappa and logistic regression. The overall concordance rate was 96.1% when considering any suggested treatment, and 74.6% for the top recommendation, with a substantial agreement (κ = 0.70). ID specialists showed significantly higher concordance (κ = 0.78) than non-ID physicians (κ = 0.61), and specialization was the strongest predictor of agreement (OR = 2.26, p = 0.001). Escherichia coli cases had the highest concordance, while Pseudomonas aeruginosa showed the lowest. The CDSS reduced inappropriate antibiotic use, particularly unnecessary carbapenem prescriptions. These findings support the utility of AI-CDSS tools in enhancing antimicrobial stewardship and standardizing care, especially in resource-limited healthcare settings.
20 Apr 2026
Read appraisal →PloS one
Pretraining effective T5 generative models for clinical and biomedical applications
This paper presents a study of the impact of corpus selection and vocabulary design on the performance of T5-based language models in clinical and biomedical domains. We introduce five different T5-EHR models, each pretrained from scratch using different combinations of clinical and biomedical corpora alongside domain-specific vocabularies. We evaluated these models across a variety of clinical and biomedical tasks to quantify the impact of pretraining data and vocabulary tokenization choices on downstream performance. Our findings reveal the importance of aligning both pretraining corpus and vocabulary with the target domain. Models pretrained exclusively on clinical data achieve superior performance on clinical tasks, while adding biomedical data contributes only marginal gains in most cases, with a few exceptions. Similarly, the choice of vocabulary significantly influences model performance, with clinical-specific vocabularies outperforming general biomedical vocabularies in tasks requiring a deeper understanding of clinical language. Also, the T5 generative models perform competitively with state-of-the-art discriminative models on several biomedical benchmarks, demonstrating strong generalization to biomedical domain. Overall, these results emphasize that task-specific selection of corpus and vocabulary is essential for optimizing model performance in clinical and biomedical natural language processing (NLP).
18 Apr 2026
Read appraisal →International journal of cardiology
Predicting functional results of percutaneous coronary intervention using machine learning modelling
BACKGROUND: Post-percutaneous coronary intervention (PCI) Murray's law-based quantitative flow ratio (μFR) is associated with long-term clinical outcomes. A tool capable of accurately predicting post-PCI μFR before intervention could support procedural planning, reduce the risk of suboptimal physiological results, and improve prognosis. OBJECTIVES: To develop and validate machine-learning models that predict continuous post-PCI μFR using only pre-procedural angiographic, physiological and clinical data, and to assess their ability to classify PCI outcomes as optimal (μFR ≥ 0.91) or sub-optimal (μFR < 0.91). METHODS: Four machine learning models were trained using pre-PCI variables. Internal bootstrap validation (1000 iterations) identified the best-performing model based on lowest root mean square error (RMSE) for continuous prediction. Predicted μFR values were subsequently used to classify PCI outcomes. RESULTS: In 343 vessels (291 patients), machine learning achieved high accuracy for continuous post-PCI μFR prediction (RMSE 0.036; 95% CI: 0.033-0.040; mean absolute error 0.030; 95% CI: 0.027-0.032; mean absolute percentage error 3.2%; 95% CI: 2.9-3.5), indicating reliable estimation of post-PCI μFR using only pre-procedural data. When the predicted μFR was used to classify PCI outcomes, performance remained clinically meaningful, with accuracy 0.72 (95% CI: 0.70-0.75), area under the curve 0.72 (95% CI: 0.69-0.74), sensitivity 0.90 (95% CI: 0.88-0.93), and specificity 0.29 (95% CI: 0.23-0.34). The high sensitivity ensures reliable upfront identification of vessels likely to achieve optimal physiology. CONCLUSIONS: Machine-learning models accurately predict post-PCI μFR and reliably distinguish optimal from sub-optimal outcomes before intervention. This approach supports personalized PCI planning and improves strategy selection.
17 Apr 2026
Read appraisal →JMIR cardio
Large Language Models in Cardiology: Systematic Review
BACKGROUND: Large language models (LLMs) are increasingly used in health care, but their role in cardiology has not yet been systematically evaluated. OBJECTIVE: This review aimed to assess the applications, performance, and limitations of LLMs across diverse cardiology tasks, including chronic and progressive conditions, acute events, education, and diagnostic testing. METHODS: A systematic search was conducted in PubMed and Scopus for studies published up to April 14, 2024, using keywords related to LLMs and cardiology. Studies evaluating LLM outputs in cardiology-related tasks were included. Data were extracted across 5 predefined domains and the risk of bias was assessed using an adapted QUADAS-2 tool (developed by Whiting et al at the University of Bristol). The review protocol was registered in PROSPERO (CRD42024556397). RESULTS: A total of 33 studies contributed quantitative outcome data to a descriptive synthesis. Across chronic conditions, ChatGPT-3.5 (OpenAI) answered 91% (43/47) heart failure questions accurately, although readability often required college-level comprehension. In acute scenarios, Bing Chat omitted key myocardial infarction first aid steps in 25% (5/20) to 45% (9/20) of cases, while cardiac arrest information was rated highly (mean 4.3/5, SD 0.7) but written above recommended reading levels. In physician education tasks, ChatGPT-4 (OpenAI) demonstrated higher accuracy than ChatGPT-3.5, improving from 38% (33/88) to 66% (58/88). In patient education studies, ChatGPT-4 provided scientifically adequate explanations (5.0-6.0/7) comparable to hospital materials but at higher reading levels (11th vs 7th grade). In diagnostic testing, ChatGPT-4 interpreted 91% (36/40) electrocardiogram vignettes correctly, significantly better than emergency physicians (31/40, 77%; P< .001), but showed lower performance in echocardiography. CONCLUSIONS: LLMs show meaningful potential in cardiology, especially for education and electrocardiogram interpretation, but performance varies across clinical tasks. Limitations in emergency guidance and readability, as well as small in silico study designs, highlight the need for multimodal models and prospective validation.
17 Apr 2026
Read appraisal →International journal of medical informatics
Rule-augmented constraint learning for semantic error detection in MIMIC-III knowledge graph
High-quality, error-free data is essential for developing reliable data-driven models, particularly in clinical decision support systems where inaccurate predictions can have serious consequences. While KGs offer a structured and semantically rich representation for clinical data, ensuring their consistency and correctness remains a challenge. Existing rule mining techniques provide solutions for the automatic extraction of logical constraints from KGs, but they often produce redundant or clinically irrelevant rules, especially when dealing with numeric or categorical literals such as age or lab values. KG constraints-rules intended to capture implausible or conflicting facts in the KG-can be used to spot semantic errors: facts that might conform to the underlying schema but contradict domain knowledge. In this work, we propose a novel framework for constraint learning in clinical KGs that identifies and transforms high-confidence rules into clinically plausible constraints. We propose two approaches, based on class disjointness and literal clustering combined with rule mining. We validate the clinical relevance of these generated rules using expert-curated constraints and large language models (LLMs). The results on the MIMIC-III clinical dataset show that rule filtering based constraint learning effectively preserves clinically meaningful rules that align with established medical knowledge. For numeric data, we achieve reliable value groupings through our clustering-based method, and the rules derived from these groupings were validated by LLMs. Their outputs confirm the clinical relevance of a portion of those discovered rules. By providing interpretable and scalable solutions to semantic inconsistencies in KGs, this study contributes to increasing the KG trustworthiness and its clinical usability.
17 Apr 2026
Read appraisal →JMIR formative research
Fact-Checking Large Language Model Responses to a Health Care Prompt: Comparative Study
BACKGROUND: Large language models use machine learning to produce natural language. These models have a range of potential applications in health care, such as patient education and diagnosis. However, evaluations of large language models in health care are still scarce. OBJECTIVE: This study aimed to (1) evaluate the accuracy and efficiency of automated fact-checking by 2 large language models and (2) illustrate a process through which a large language model might support a patient in redrafting a prompt to include key information needed for patient safety. METHODS: A parallel comparison of 2 large language models and 3 human experts was conducted. A clinical scenario was devised in which a woman aged 23 years questions the safety of retinoid treatment for acne by sending prompts to 2 large language models (GPT-4o and OpenBioLLM-70B). GPT-4o and OpenBioLLM-70B were asked to suggest improvements to the patient's initial prompt to elicit key information for clinical decision-making. After the patient sent the revised prompt to the large language models, the models were then asked to fact-check the final response. To test the generalizability of automated fact-checking, a set of 20 clinical statements on disparate topics, mostly related to drug indications, contraindications, and side effects, was developed. The large language models also fact-checked these 20 medical statements. The results were compared against the evaluations of 3 clinical experts. The outcome measures were as follows: (1) percentage of accuracy of automated fact-checking, (2) time to complete fact-checking, and (3) a binary outcome for prompt redrafting (advising the patient to revise her prompt by naming her acne medication to address safety concerns). RESULTS: For the scenario of a patient with acne, GPT-4o and OpenBioLLM-70B both had 86% agreement with the clinical experts' fact-checking. The large language models did not consistently convey the urgency of discontinuing isotretinoin treatment when pregnancy is suspected. In addition, the models did not adequately convey the importance of folic acid supplementation during pregnancy. For the set of 20 medical claims, GPT-4o fact-checking had 100% agreement with that of human experts, whereas OpenBioLLM-70B had 95% agreement. OpenBioLLM-70B diverged from human experts and GPT-4o on 1 question related to pediatric use of antihistamines. The expert fact-checks took a mean time of 18 (SD 3.74) minutes, GPT-4o took 42 seconds, and OpenBioLLM-70B took 33 minutes. The GPT-4o responses for the acne scenario had some inconsistencies but zero fabrication and no obvious omissions. In contrast, OpenBioLLM-70B omitted 1 key information item needed for patient safety. CONCLUSIONS: GPT-4o can interact with patients to improve the quality and comprehensiveness of the information contained in health-related prompts. GPT-4o and OpenBioLLM-70B can conduct efficient fact-checking that is close to the level of accuracy of human experts. Human experts need to perform additional checks for accuracy and safety.
17 Apr 2026
Read appraisal →Journal of investigational allergology & clinical immunology
Future Perspectives for Artificial Intelligence in Allergy: Advances in Diagnosis, Monitoring, and Personalized Care
Artificial intelligence (AI) is significantly transforming the specialty of allergy by offering novel tools for diagnosis, risk prediction, disease monitoring, and personalized care. This narrative review comprehensively explores key advances in the application of AI techniques-expert systems, machine learning, deep learning, and natural language processing-across various allergic diseases such as asthma, anaphylaxis, atopic dermatitis, food allergy, allergic rhinitis, eosinophilic esophagitis, and drug hypersensitivity reactions. Despite these advancements, several challenges persist, including model interpretability, external validation, and ethical and regulatory barriers. Nonetheless, AI holds strong potential to enhance diagnostic efficiency, therapeutic precision, and patient outcomes. This review presents an updated perspective on the state of the art of AI in allergy, highlighting both its achievements and ongoing challenges.
16 Apr 2026
Read appraisal →JMIR medical informatics
Classification of Cochrane Plain Language Summaries by Conclusiveness Using Transformer-Based Models and ChatGPT: Retrospective Observational Study
BACKGROUND: Cochrane plain language summaries (PLSs) aim to make systematic review findings more accessible to the general public. However, inconsistencies in how conclusions are presented may impact comprehension and decision-making. Classifying PLSs based on conclusiveness can improve clarity and facilitate informed health decisions. OBJECTIVE: This study aimed to develop and evaluate deep learning language models for the classification of PLSs according to 3 levels of conclusiveness (conclusive, inconclusive, and unclear) and to compare their performance with a general-purpose large language model (GPT-4o). METHODS: We used a publicly available dataset containing 4405 Cochrane PLSs of systematic reviews published until 2019, already classified by humans according to 9 categories of conclusiveness regarding the intervention's effectiveness or safety. We merged these categories into 3 classes based on the strength of conclusiveness: conclusive, inconclusive, and unclear. For the fine-tuning, we used Scientific Bidirectional Encoder Representations from Transformers (SciBERT), a pretrained language model trained on 1.14 million papers primarily from the health sciences, and Longformer, a transformer model designed specifically to process long documents. The script was developed using the Python programming language and the PyTorch framework. We computed evaluation metrics using the scikit-learn machine learning library and determined the area under the curve of the receiver operating characteristic (AUCROC) to measure the model performance in balancing sensitivity and specificity. We also analyzed a separate set of 213 PLSs and compared the predictions of our pretrained models with both manual verification and outputs generated by ChatGPT. RESULTS: The model based on SciBERT achieved a balanced accuracy of 56.6%. The AUCROC was 0.91 for "conclusive," 0.67 for "inconclusive," and 0.75 for "unclear" conclusiveness classes. The Longformer-based model had a balanced accuracy of 60.9%, with AUCROCs of 0.86 for "conclusive," 0.67 for "inconclusive," and 0.72 for "unclear" conclusiveness classes. Both models underperformed compared with ChatGPT, which demonstrated higher accuracy (74.2%), better precision and recall, and a higher Cohen κ (0.57). CONCLUSIONS: Fine-tuning 2 transformer-based language models showed mixed results in classifying Cochrane PLSs by conclusiveness, likely due to semantic overlap and subtle linguistic differences. Despite satisfactory internal test metrics, the fine-tuned models failed to generalize to newly published PLSs, where performance dropped to near-chance levels. These findings suggest that general-purpose large language models like GPT-4o may currently offer more reliable results for practical classification tasks in biomedical applications.
16 Apr 2026
Read appraisal →International journal of medical informatics
Clinician preferences for explainable AI in critical care: a comparative study of interpretable models and visualizations for intubation decision support
BACKGROUND: The complexity of many AI models hinders their clinical adoption because the clinicians using them do not regard them as transparent. This study addresses the lack of clinician-centered explainable AI (XAI) interfaces by designing and evaluating intuitive visual explanations for intubation prediction, testing the hypothesis that workflow-compatible designs enhance acceptance. OBJECTIVE: This study compares three, time-aware, visual explanations for XAI-based intubation prediction and evaluate their acceptance, comprehension, and perceived utility among clinicians. METHODS: We developed machine learning models to estimate the near-term risk of deterioration in the patient's condition which may lead to mechanical intubation using ICU time-series data. We generated global and local explanations using SHAP and designed three customized visual formats-a temporal force plot, a temporal bar chart, and a dual-encoded SHAP heatmap. Clinicians (n = 206) evaluated comprehension and usability using objective questions and a Likert-based survey. RESULTS: Based on 4608 critically ill patients with 10 medical variables over 7 hours of data for each patient, the Random Forest (RF) model achieved the highest area under the curve (AUC): 0.94. Furthermore, the local explanations were customized and evaluated by 206 clinicians through a survey conducted on the Prolific platform. A customized heatmap representation was selected as the visualization with the highest perceived clinical utility and alignment with clinical workflows. DISCUSSION: The reported findings support the need for explanation formats to be tailored to clinical reasoning and task context, supporting the concept of cognitive fit. The heatmap's close alignment with clinicians' mental models and its graphical integrity enhances interpretability and trust. This study demonstrates that explanation effectiveness depends on contextual relevance, rather than a universal standard, and that the presentation format itself significantly shapes clinicians' trust in XAI systems. CONCLUSION: This study advances clinical XAI by introducing a time-aware explanation framework for ICU intubation decisions. By integrating temporal trends with model reasoning, our visualizations closely align with clinicians' cognitive workflows. Rigorous clinician-centered evaluation identified the dual-encoded SHAP heatmap as the most useful and workflow-compatible visualization, highlighting the importance of explanation design alongside predictive accuracy for clinical adoption.
15 Apr 2026
Read appraisal →Journal of medical Internet research
Error Detection in Emergency Radiology Reports Using a Large Language Model: Multistage Evaluation Study
BACKGROUND: Emergency radiology requires highly accurate reporting under time constraints; yet, increasing workloads raise the risk of errors. While large language models (LLMs) show potential for proofreading in general radiology, their performance in emergency settings and non-English contexts remains unclear. OBJECTIVE: We aim to evaluate the performance of a domain-optimized LLM, DeepSeek-R1, for identifying errors in Chinese emergency radiology reports, with comparison against assessments by board-certified radiologists. METHODS: We compiled 7435 emergency reports (dataset 1; radiography, computed tomography, and magnetic resonance imaging) collected from November 2024 to April 2025. In stage 1, a total of 5 LLMs were evaluated using 200 reports. The best model, DeepSeek-R1, proceeded to stages 2 and 3, where 0-shot and few-shot learning were tested on a separate set (n=100). Model performance was compared against 12 radiologists. Stage 4 validated real-world utility on 800 verified reports. RESULTS: In subdataset 1, under stress-testing conditions, DeepSeek-R1 achieved a higher error detection rate in the few-shot setting than in the 0-shot setting (84.4% vs 60.9%, P=.003). Its performance exceeded that of radiology residents (84.4% vs 51.6% and 53.1%, respectively; both P<.05) and showed no statistically significant difference compared with senior radiologists and attending radiologists (84.4% vs 68.8%-93.8%, P=.26 to ≥.99). Compared with residents, DeepSeek-R1 detected more critical omissions (100% vs 25% and 50%; both P<.05) and other errors (92% vs 33% and 33%; both P=.02). In dataset 2, collected from independent institutions, DeepSeek-R1 achieved a detection rate of 95% under the few-shot setting. Reading time was shorter than that of human readers (92 vs 109 s). In real-world validation, DeepSeek-R1 identified 117 true reporting errors, yielding a positive predictive value of 56.5%. CONCLUSIONS: DeepSeek-R1 holds promise for improving quality control in emergency radiology reports. Its performance and efficiency support its use as an assistive proofreading tool in real-world radiology workflows.
15 Apr 2026
Read appraisal →Journal of medical Internet research
Context-Aware Sentence Classification of Radiology Reports Using Synthetic Data: Development and Validation Study
BACKGROUND: Automated structuring of radiology reports is essential for data utilization and the development of medical artificial intelligence models. However, manual annotation by experts is labor-intensive, and processing real clinical data through commercial large language models (LLMs) presents significant privacy risks. These challenges are particularly pronounced for non-English languages like Japanese, where specialized medical corpora are scarce. While synthetic data generation offers a potential privacy-preserving alternative, its effectiveness in capturing complex clinical nuances-such as negation and contextual dependencies-to train robust classification models without any real-world training data has not been fully established. OBJECTIVE: This study aimed to develop a context-aware sentence classification model for Japanese radiology reports using an entirely synthetic training pipeline, thereby eliminating reliance on real-world clinical data during the development phase. Furthermore, we sought to evaluate the generalizability of this approach by validating the model's performance on diverse, multi-institutional, real-world reports. METHODS: Japanese radiology reports (n=3104) were generated using GPT-4.1 and automatically annotated at the sentence level into 4 categories (background, positive finding, negative finding, and continuation) using GPT-4.1-mini. The synthetic data were partitioned into training (n=2670), validation (n=334), and test (n=100) sets. We fine-tuned several models, including lightweight local LLMs (Qwen3 and Llama 3.2 series) using low-rank adaptation and Japanese text classification models (Bidirectional Encoder Representations from Transformers [BERT]-base Japanese v3, Japanese Medical Robustly Optimized BERT Pretraining Approach [JMedRoBERTa]-base, and ModernBERT-Ja-130M). External validation was performed using 280 real-world reports (3477 sentences) from 7 institutions in the Japan Medical Image Database, with ground-truth labels established by board-certified radiologists. Evaluation metrics included accuracy, macro-averaged F1 (macro F1) score, and positive predictive value for positive findings (PPV_1). RESULTS: All models achieved high performance on the synthetic test set (accuracy: 0.938-0.951; macro F1-score: 0.924-0.940). Overall performance declined on the external validation dataset (accuracy: 0.783-0.813; macro F1-score: 0.761-0.790), reflecting distributional differences between synthetic and real-world reports; however, PPV_1 remained stable and high across datasets (eg, 0.957 on the synthetic test set vs 0.952 on the external validation dataset for Qwen3 [4B]). Parsing errors occurred in LLM-based approaches (19-260 sentences, 0.55%-7.48% in the external dataset). CONCLUSIONS: This study demonstrates the feasibility of developing context-aware sentence classification models for Japanese radiology reports using a training pipeline based entirely on synthetic data. The stability of PPV_1 indicates that the models successfully captured the essential clinical terminology and linguistic patterns required to identify positive findings in real-world reports, despite the observed performance degradation during external validation. This approach substantially reduces manual annotation requirements and privacy risks, providing a scalable foundation for constructing structured radiology datasets to support the development of clinically relevant medical artificial intelligence models.
11 Apr 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
The need for verification in artificial intelligence-driven scientific discovery
Artificial intelligence (AI) is transforming the practice of science. Machine learning (ML) and large language models (LLMs) can generate hypotheses at a scale and speed far exceeding traditional methods, offering the potential to accelerate discovery across diverse fields. However, the abundance of hypotheses introduces a critical challenge; without scalable and reliable mechanisms for verification, scientific progress risks being hindered rather than advanced. In this article, we trace the historical development of scientific discovery, examine how AI is reshaping established practices for discovery and review the principal approaches, ranging from data-driven methods and knowledge-aware neural architectures to symbolic reasoning frameworks and LLM agents. While these systems can uncover patterns and propose candidate laws, their scientific value ultimately depends on rigorous and transparent verification, which we argue must be the cornerstone of AI-assisted discovery. This article is part of the discussion meeting issue 'Symbolic regression in the physical sciences'.
10 Apr 2026
Read appraisal →JMIR human factors
Predicting Electronic Health Record Usability: Scoping Review of Adoption Models, Metrics, and Future Directions
BACKGROUND: Electronic health records (EHRs) play an essential role in modern health care, enabling data sharing and improving patient safety; however, even though vendors must adhere to International Organization for Standardization-related usability standards for EHR certification, persistent usability issues continue to undermine efficiency, contribute to clinician burden, and increase the risk of preventable errors. OBJECTIVE: This scoping review synthesizes existing research on EHR adoption and usability, emphasizing theoretical models, measurement approaches, factors, and analytic methods used to assess or predict usability. We identify gaps and opportunities for integrating predictive analytics and artificial intelligence (AI) to advance research and improve the usability of EHRs. METHODS: Following Joanna Briggs Institute and PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines, we systematically searched MEDLINE, Web of Science, IEEE Xplore, and Scopus library databases for studies published between January 1, 2009, and April 9, 2025. Inclusion criteria focused on empirical research using predictive methods or models related to EHR usability. Data were charted and synthesized thematically. RESULTS: Of the 2323 screened papers, 47 studies met inclusion criteria. Most research examined or predicted EHR adoption (not usability) using dominant frameworks, such as the technology acceptance model, unified theory of acceptance and use of technology, and the information system success model, which comprised usability. Factors related to usability-particularly perceived usefulness, perceived ease of use, effort expectancy, and facilitating conditions via the EHR adoption models-appeared frequently. Regression-based methods and structural equation modeling were the most common analytic techniques. No studies applied predictive modeling or AI to predict EHR usability. CONCLUSIONS: The focus of this study on the prediction of EHR usability and adoption for the past 15 years is distinctive in the literature. It extends prior usability reviews (mostly focusing on adoption, not prediction of usability). Predictive modeling for EHR usability remains underdeveloped throughout 2009 to 2025. Dominant frameworks in the EHR literature continue to prioritize predicting adoption over operational usability. These models rely heavily on self-reported, cross-sectional measures captured at a single postimplementation time point, embedding systematic bias and obscuring longitudinal usability dynamics. Despite the application of increasingly sophisticated predictive techniques-primarily variants of regression and structural equation modeling-usability has remained analytically subordinate to adoption and acceptance constructs for more than 15 years. As a result, widely used models, such as the technology acceptance model and unified theory of acceptance and use of technology, position usability merely as an antecedent to intention or use, rather than as an independent, system-level property that can be empirically measured, modeled, and predicted. Therefore, there is substantial opportunity to integrate predictive analytics, AI, and longitudinal usability measures to build dynamic models.
10 Apr 2026
Read appraisal →JMIR medical informatics
Exploring the Role of AI in Managing Treatment Recommendations for Lymphedema: International, Multidisciplinary, Multiprofessional Survey Study of Trust, Reliability, and Impact on Decision-Making
BACKGROUND: Upper and lower extremity lymphedema is a chronic, progressive condition that significantly impairs the quality of life of affected patients. Despite the recently established effectiveness of physical therapy and supermicrosurgical interventions, current guidelines frequently lag behind emerging evidence and commonly do not offer stage-specific treatment algorithms. This gap in evidence-based guidance may prompt clinicians with limited experience to seek support from large language models such as ChatGPT. OBJECTIVE: Given the potential of artificial intelligence to rapidly integrate emerging research, this study evaluated how clinicians from different professional backgrounds rate the quality and reliability of personalized lymphedema management recommendations generated by ChatGPT. METHODS: In this exploratory cross-sectional study, ChatGPT generated treatment recommendations for 6 standardized lymphedema case scenarios. An international panel of 67 participants (resident doctors, board-certified specialists, physiotherapists, and advanced practice nurses) from 34 institutions across 11 countries assessed the recommendations using a modified DISCERN questionnaire with a 9-point agreement scale ranging from 1 (completely disagree) to 9 (completely agree). Ratings were summarized as pooled means with variability measures and compared across clinician groups (residents vs board-certified physicians vs physiotherapists or advanced practice nurses) using group comparison testing. RESULTS: ChatGPT was rated most favorably for diagnostic accuracy and treatment relevance, with higher ratings among residents than board-certified physicians. Residents assigned significantly lower scores for source indication, source currency, and communication of uncertainty. Between-group differences were observed across multiple DISCERN items, consistent with systematically more critical appraisal by experienced specialists. Participants reported moderate to high trust and willingness to consider ChatGPT as a supplementary resource, with more favorable perceptions among younger respondents. CONCLUSIONS: Clinicians perceived ChatGPT as potentially useful for preliminary orientation and educational support in lymphedema management, especially for less experienced users. Despite not being blinded, lower ratings in evidence transparency and uncertainty communication, particularly among experienced specialists, suggest that current artificial intelligence outputs should not be used as stand-alone guidance. Future work should test clinically integrated, citation-grounded workflows in prospective settings and evaluate whether they improve decision quality and efficiency.
9 Apr 2026
Read appraisal →European thyroid journal
Promise and pitfalls of AI chatbots in complex decision-making for thyroid nodules and papillary thyroid cancer
INTRODUCTION: Artificial intelligence (AI) chatbots are increasingly used in medicine, but their reliability in scenarios with multiple management options is unclear. Indeterminate thyroid nodules and low- and low-to-intermediate-risk papillary thyroid carcinoma (PTC) represent such cases. METHODS: In a nationwide web-based survey, 201 members of the Hellenic Endocrine Society evaluated 12 clinical vignettes on indeterminate thyroid nodules and low- and low-to-intermediate-risk PTC. Their responses were compared with those generated by four conversational AI models (ChatGPT, Gemini, Copilot, and DeepSeek) at two time points, 11 months apart. DeepSeek was assessed only at the second time point. Chatbot outputs were assessed for agreement with endocrinologists' predominant answers, concordance with the most guideline-consistent options (American and European Thyroid Association recommendations), temporal stability, and inter-model agreement. RESULTS: Alignment between chatbots and endocrinologists' predominant responses was limited, reaching at most 25% across scenarios. In contrast, concordance with the most guideline-consistent options was higher, up to 83% (10/12 scenarios), depending on the model and time point. Across 12 scenarios, ChatGPT, Gemini, and Copilot changed their responses in 4, 7, and 5 scenarios, respectively, with some updates moving closer to, and others further from, guideline-based answers. Inter-model agreement ranged from 33 to 67%, indicating substantial variability among chatbots. CONCLUSION: AI chatbots show evolving but inconsistent performance in complex thyroid management scenarios. While guideline concordance can be relatively high, substantial variability across models, limited temporal reproducibility, and poor alignment with clinical practice highlight the need for ongoing longitudinal evaluation before safe integration into clinical decision-making.
9 Apr 2026
Read appraisal →Journal of medical Internet research
Initial Insights Into an Institutional Secure Large Language Model for Magnetic Resonance Imaging Examination Requests: Retrospective Study
BACKGROUND: Incomplete clinical details on magnetic resonance imaging (MRI) examination requests (MERs) can lead to suboptimal protocol selection. An institutional secure large language model (sLLM) with access to manually retrieved salient data from the electronic medical record (EMR) may improve request completeness and protocol accuracy across multiple MRI subspecialties. OBJECTIVE: The objective of this study was to compare clinician MERs with sLLM-augmented MERs for information quality and to evaluate the protocoling accuracy of the sLLM versus board-certified radiologists across body, musculoskeletal, and neuroradiology MRI. METHODS: This retrospective study included 608 random outpatient MRI examinations performed between September 2023 and July 2024 (body 206, musculoskeletal 203, neuroradiology 199). The cohort comprised 528 patients (mean 51.2 years, SD 19.2; range 4-93; n=279, 52.8% women, n=249, 47.2% men). MERs without EMR access were excluded. A privately hosted Anthropic Claude 3.5 model (temperature 0) augmented each MER with manually retrieved salient EMR data and, via rule-based parsing, mapped the extracted elements onto predefined institutional criteria to recommend region or coverage and contrast use. Two experienced radiologists established a consensus reference standard. Two board-certified general radiologists (Rad 3 and Rad 4) and the sLLM were compared with this standard. Clinical information quality was graded using the Reason-for-Exam Imaging Reporting and Data System (RI-RADS). Interrater reliability was quantified with Gwet AC1. Paired accuracies were compared with the McNemar test to determine whether there was a statistically significant difference. RESULTS: Interreader agreement for RI-RADS was almost perfect for sLLM-augmented MERs (AC1 0.97, 95% CI 0.94-0.99) and moderate for clinician MERs (AC1 0.43, 95% CI 0.34-0.52). Limited or deficient clinical information (RI-RADS C/D) fell to 0% to 0.7% (0/608 to 4/608) with sLLM augmentation vs 4.1% to 20.4% (25/608 to 124/608) for clinician MERs. Overall protocol accuracy was 93.1% (566/608; 95% CI 89.6-96.6) for the sLLM, 91.4% (556/608; 95% CI 87.6-95.3) for Rad 3, and 92.1% (560/608; 95% CI 88.4-95.8) for Rad 4 (sLLM vs Rad 3 P=.23 vs Rad 4 P=.40). Region or coverage accuracy was similar (sLLM: 579/608, 95.2%; Rad 3: 585/608, 96.2%; Rad 4: 573/608, 94.2%; P=.46 and P=.36). Contrast decisions were more accurate using the sLLM at 94.4% (574/608; 95% CI 91.3-97.5) vs Rad 3 at 92.1% (560/608; 95% CI 88.4-95.8; P=.027) and were not significantly different to Rad 4 at 92.9% (565/608; 95% CI 89.4-96.4; P=.16). Subspecialty analyses showed similar patterns, with the sLLM outperforming Rad 4 for musculoskeletal MRI contrast decisions (96.6% vs 91.1%; P=.006) and matching readers elsewhere. Manual review indicated that sLLM improvements arose from EMR details not listed on the MER (infection/inflammation, tumor history, prior surgery). No clinically significant hallucinations were identified in a manual review of discordant cases. CONCLUSIONS: Across body, musculoskeletal, and neuroradiology MRI, sLLM-augmented examination requests improved clinical context and enhanced contrast selection while demonstrating accuracy comparable to general radiologists for region or coverage. Integrating sLLMs into routine vetting workflows may reduce manual workload in protocol selection for more efficient, standardized protocoling.
8 Apr 2026
Read appraisal →Journal of the American Heart Association
Development and Validation of a Machine Learning Model to Predict Oral Anticoagulant Use in Stroke From Prothrombin Time-International Normalized Ratio and Activated Partial Thromboplastin Time
BACKGROUND: Timely recognition of oral anticoagulant use is critical in acute stroke but is often hampered by impaired consciousness and unavailable medication history. We investigated whether routinely available coagulation tests, prothrombin time-international normalized ratio and activated partial thromboplastin time, paired with machine learning can identify anticoagulant exposure at presentation. METHODS: We performed a multicenter diagnostic modeling study using the Japan Stroke Data Bank. Adults with ischemic or hemorrhagic stroke admitted in 2016 to 2021 comprised the derivation cohort; those in 2022 formed a temporally distinct validation cohort. Three models (random forest, extreme gradient boosting, and neural network) used prothrombin time-international normalized ratio and activated partial thromboplastin time as predictors to classify anticoagulant status: vitamin K antagonist, direct oral anticoagulant, or none. Discrimination (area under the receiver operating characteristic curve and area under the precision-recall curve), calibration, and decision curve analysis were assessed. Model outputs were implemented as bedside probability heat maps. RESULTS: Among 30 767 patients (derivation n=24 857; validation n=5910), vitamin K antagonist/direct oral anticoagulant use was 4.5%/8.2% in derivation and 2.8%/9.5% in validation. In the validation cohort, the neural network yielded the best performance: area under the receiver operating characteristic curve, 0.933 (95% CI, 0.908-0.955) and area under the precision-recall curve, 0.52 (95% CI, 0.44-0.61) for vitamin K antagonist; area under the receiver operating characteristic curve, 0.841 (95% CI, 0.822-0.858) and area under the precision-recall curve, 0.39 (95% CI, 0.35-0.43) for direct oral anticoagulant, with preserved calibration and highest net benefit across clinically relevant thresholds. Heat maps summarized class probabilities over the prothrombin time-international normalized ratio×activated partial thromboplastin time space to support rapid inference. CONCLUSIONS: A simple machine learning model using only prothrombin time-international normalized ratio and activated partial thromboplastin time classified vitamin K antagonist and direct oral anticoagulant exposure with high accuracy and practical clinical utility. This tool assists emergency decision-making when medication history is unknown, supporting the safety and timeliness of acute stroke care.
8 Apr 2026
Read appraisal →Medical science monitor : international medical journal of experimental and clinical research
AI-Assisted Dialysis Decision-Making: Assessing Agreement Between ChatGPT and Nephrologist in Initial Dialysis Indication and Prescription in Emergency and ICU Settings
BACKGROUND Artificial intelligence (AI) is increasingly explored as a clinical decision-support tool in nephrology; however, its real-world applicability for dialysis decision-making in emergency and intensive care unit (ICU) settings remains insufficiently studied. Hemodialysis initiation and prescription are complex, time-sensitive, and dynamic processes that require expert clinical judgment. MATERIAL AND METHODS This retrospective observational study evaluated agreement between AI-generated (ChatGPT) and nephrologist-made dialysis decisions in emergency and ICU settings. Adult patients undergoing first-time dialysis were included. Agreement was assessed for dialysis initiation, modality selection, and key prescription parameters. To ensure clinical relevance, continuous prescription variables were categorized into predefined ranges. Agreement was quantified using Gwet's AC1 coefficient and Cramer's V statistic. RESULTS Eighty-four patients were included. AI demonstrated 100% agreement with nephrologists regarding dialysis initiation. Overall agreement for dialysis modality selection was 92.9% (Cramer's V=0.87, P<0.001). Agreement for core dialysis prescription parameters - including blood flow rate, dialysate sodium, potassium, and calcium concentrations - was high across modalities (all P<0.001). Lower agreement was observed for ultrafiltration-related parameters, particularly ultrafiltration duration, reflecting the individualized and dynamic nature of volume management during dialysis. CONCLUSIONS AI-assisted decision support demonstrated high agreement with nephrologist decisions for initial dialysis initiation, modality selection, and core prescription parameters in emergency and ICU settings. Discrepancies were primarily confined to ultrafiltration-related decisions, underscoring the necessity of ongoing bedside clinical judgment. These findings support the role of AI as a decision-support tool rather than a replacement for clinician-led dialysis management.
8 Apr 2026
Read appraisal →Progress in biomedical engineering (Bristol, England)
Trustworthy AI in digital health: a comprehensive review of robustness and explainability
Ensuring trust in artificial intelligence (AI) systems is essential for the safe and ethical integration of machine learning systems into high-stakes domains such as digital health. Key dimensions, including robustness, explainability, fairness, accountability, and privacy, need to be addressed throughout the AI lifecycle, from problem formulation and data collection to model deployment and human interaction. While various contributions address different aspects of trustworthy AI, a focused synthesis on robustness and explainability, especially tailored to the healthcare context, remains limited. This review addresses that need by organizing recent advancements into an accessible framework, highlighting both technical and practical considerations. We present a structured overview of methods, challenges, and solutions, aiming to support researchers and practitioners in developing reliable and explainable AI (XAI) solutions for digital health. This review article is organized into three main parts. First, we introduce core pillars of trustworthy AI and discuss the technical and ethical challenges they pose, particularly in the context of digital health. Second, we explore application-specific trust considerations across domains such as intensive care, mental health, metabolic disease, and public health surveillance, highlighting how explainability, clinical validation, and human oversight support trust. Lastly, we present recent advancements in techniques aimed at improving robustness under data scarcity and distributional shifts, as well as XAI methods ranging from feature attribution to gradient-based interpretations and counterfactual explanations. This paper is further enriched with detailed discussions of the contributions toward robustness and explainability in digital health, the development of trustworthy AI systems in the era of large language models, and various evaluation metrics for measuring trust and related parameters such as validity, fidelity, and diversity, offering a roadmap for building safer and more reliable AI systems.
8 Apr 2026
Read appraisal →JMIR formative research
Evaluation of GPT-5 in Periodontitis Staging and Grading: Retrospective Observational Study
BACKGROUND: Periodontitis is a chronic gum disease affecting approximately 42% of adults aged 30 years and older in the United States. Training dental students to accurately diagnose and manage periodontitis is a critical component of dental education and clinical care. Recent advances in large language models offer new opportunities to support both domains, yet their performance in periodontal diagnosis remains largely unexplored, particularly for newer models such as GPT-5. OBJECTIVE: This study conducted an exploratory evaluation of GPT-5's ability to stage and grade periodontitis. METHODS: A total of 25 publicly available clinical cases explicitly reporting periodontitis stage and grade were identified through Google and PubMed searches. Each case description was entered into GPT-5 using a zero-shot prompting approach to assess guideline-based reasoning without exemplar conditioning. The model's predictions were compared with the published reference diagnoses. Performance was measured using accuracy, 95% CI, unweighted Cohen κ, and weighted Cohen κ. RESULTS: Across these cases, GPT-5 showed marked class-dependent performance and a tendency to overestimate disease severity. Grading performance was notably imbalanced, with high recall for grade C but substantially lower discrimination for grade B. GPT-5 achieved a staging accuracy of 68% (95% CI 48.4%-82.8%) and a grading accuracy of 77.3% (95% CI 56.6%-89.9%), with corresponding Cohen κ values of 0.454 (95% CI 11.0%-75.6%) and 0.179 (95% CI -15.8% to 63.8%), respectively. While staging performance showed fair agreement beyond chance, the low κ for grading indicates poor agreement and limited reliability in distinguishing periodontal disease severity. CONCLUSIONS: These findings suggest that although GPT-5 demonstrates potential for guideline-based periodontitis staging and grading, its current diagnostic performance, particularly for periodontitis grading, limits its use in clinical assessment and educational training. Meaningful application in periodontal diagnosis and training will require substantial improvements in reliability and rigorous validation in larger, more diverse, and prospectively collected datasets.
7 Apr 2026
Read appraisal →JMIR formative research
Performance of DeepSeek V3, DeepSeek R1, ChatGPT 4o, and ChatGPT o1 on the National Health Professional and Technical Qualification Examination (Intermediate Level) in China: Comparative Analysis
BACKGROUND: In recent years, large language models (LLMs) have undergone swift cycles of refinement and iteration. However, in the realm of clinical medicine, different LLMs' capability of logical reasoning and disease diagnosis needs further investigation. OBJECTIVE: The aim of our study was to evaluate the performance of 4 different LLMs in the National Health Professional and Technical Qualification Examination in China. METHODS: A total of 398 multiple-choice questions of 5 different question types were integrated within the examination with respect to the diagnosis or care of cases. These questions were categorized into different cardiology subspecialties and different clinical disciplines. DeepSeek V3 and R1 were accessed through an application programming interface, while ChatGPT 4o and o1 were queried via its public chat-based interface. We offered the same prompts instructing LLMs to assume the role of a physician and provide answers with explanations at the beginning of each conversation. We assessed different LLMs' performance by the accuracy in the responses to the multiple-choice questions. For the first 3 examination sections, McNemar test was used to compare the accuracy among the models, with post hoc pairwise comparisons performed using partitions of chi-square method and Bonferroni correction (significance set at P<.008). For the fourth section involving partially credit scoring, one-way ANOVA was performed to compare the mean scores among the models, with statistical significance set at P<.05. RESULTS: Both DeepSeek V3 and R1 showed superior performance in the first 3 sections of the Chinese National Health Professional and Technical Qualification Examination, achieving an overall performance of 93% and 93.6%, respectively. ChatGPT 4o and o1 achieved accuracies of 73.3% and 69%, respectively (all P<.001 compared with DeepSeek V3). For the fourth section, the performance of all 4 LLMs markedly declined compared to their results in the preceding sections. Particularly, in the section of gastroenterology and hematology, DeepSeek V3 achieved the highest accuracy, while R1 ranked first in cardiology and neurology. ChatGPT o1 achieved the highest accuracy in the topic of coronary artery disease, with no statistical significance. CONCLUSIONS: DeepSeek V3 and R1 showed remarkable potential in facilitating clinical decision-making in the Chinese professional examination, with both outperforming ChatGPT 4o and o1. Nonetheless, future research should continue evaluating their economic efficiency and susceptibility to hallucination.
7 Apr 2026
Read appraisal →Lifetime data analysis
Q-learning via deep learning-based Buckley-James method for non-linear censored data
In healthcare, personalized treatment strategies are vital for improving patient outcomes, especially under right-censored survival data. We propose Dynamic Deep Buckley-James Q-Learning, a novel counterfactual Q-learning algorithm that integrates deep learning with the Buckley-James method to simultaneously address censoring and nonlinear modeling challenges. By explicitly capturing complex, nonlinear interactions between covariates and treatment effects, the algorithm robustly estimates optimal dynamic treatment regimes. Leveraging a counterfactual framework, we define and estimate potential survival outcomes under hypothetical treatment sequences, enabling unbiased Q-function estimation even in the presence of time-dependent covariates and right censoring. The algorithm maximizes the expected imputed survival reward under these counterfactual scenarios. Simulation studies and real-world data analysis demonstrate its superior performance in predictive accuracy and treatment decision-making, offering a powerful framework for individualized care in complex clinical settings.
6 Apr 2026
Read appraisal →Genetics
Mondo: integrating disease terminology across communities
Precision medicine aims to enhance diagnosis, treatment, and prognosis by integrating multimodal data at the point of care. However, challenges arise due to the vast number of diseases, differing methods of classification, and conflicting terminological coding systems and practices used to represent molecular definitions of disease. This lack of interoperability artificially constrains the potential for diagnosis, clinical decision support, care outcome analysis, as well as data linkage across research domains to support the development or repurposing of therapeutics. There is a clear and pressing need for a unified system for managing disease entities-including identifiers, synonyms, and definitions. To address these issues, we created the Mondo disease ontology-a community-driven, open-source, unified disease classification system that harmonizes diverse terminologies into a consistent, computable framework. Mondo integrates key medical and biomedical terminologies, including Online Mendelian Inheritance in Man (OMIM), Orphanet, Medical Subject Headings (MeSH), National Cancer Institute Thesaurus (NCIt), and more, to provide a comprehensive and accurate representation of disease concepts with fully provenanced and attributed links back to the sources. Mondo can be used as the handle for curation of gene-disease associations utilized in diagnostic applications, research applications such as computational phenotyping, and in clinical coding systems in clinical decision support by pointing the clinician to the numerous knowledge resources linked to the Mondo identifier. Mondo's community-centric approach, stewarded by the Monarch Initiative's expertise in ontologies, ensures that the ontology remains adaptable to the evolving needs of biomedical research and clinical communities, as well as the knowledge providers.
5 Apr 2026
Read appraisal →Scientific reports
Potential of large language models for rapid clinical information support: evidence from acute kidney injury knowledge testing.
Large language models (LLMs) are increasingly applied in clinical contexts; however, to our knowledge, their performance in acute kidney injury (AKI) has not been directly compared with physician performance on standardized clinical knowledge assessments. We conducted a cross-sectional study to evaluate the performance of 13 publicly available LLMs and 123 volunteer participants at the 131st Annual Congress of the German Society of Internal Medicine in Wiesbaden, Germany. Both groups completed an identical AKI knowledge assessment consisting of two case vignettes and 15 single-best-answer multiple-choice questions. LLMs achieved a mean score of 13.5 out of 15 (90%), with several models reaching a perfect score, while human participants averaged 7.3 out of 15 (48.7%). Only 16.3% of participants scored 11 points or higher. As an illustrative example, ChatGPT-4o completed the test in approximately 0.5 minutes, whereas humans required a mean of 7.3 minutes. These findings demonstrate that LLMs substantially outperformed a heterogeneous group of medical professionals in AKI knowledge assessments and did so with markedly greater efficiency. While this highlights the potential of LLMs as rapid, cost-effective tools for clinical knowledge support, their role in real-world patient care remains undetermined, and human clinical judgment remains essential to ensure safe, context-sensitive, and patient-centered care.
5 Apr 2026
Read appraisal →Critical care explorations
Physician Gestalt Compared With Artificial Intelligence Model to Predict Intubation in Critically Ill Patients
IMPORTANCE: Accurate prediction of intubation in critically ill patients could enable interventions that improve patient outcomes. However, the performance of intensive care physicians compared with machine learning (ML) models remains unknown. OBJECTIVES: To investigate intensive care physicians' ability to predict the need for intubation within 24 hours and compare their performance against an established ML model, Vent.io. DESIGN, SETTING, AND PARTICIPANTS: This prospective observational study in two ICUs surveyed intensivists to test their ability to predict the need for intubation of adult patients under their care. Physician predictions of intubation were then compared with predictions from Vent.io. MAIN OUTCOMES AND MEASURES: Primary metrics include prediction sensitivity, specificity, and descriptive statistics for both physicians and ML model. Generalized linear mixed models investigated the fixed effect of the predictor (physician vs. Vent.io) on both sensitivity and specificity while accounting for the random effects from different physicians. Similar modeling was used to investigate the relationship between physician confidence and correctness. RESULTS: Overall, physicians were quite confident in their predictions of intubation with a median score of 8 (on a 0-10 point scale, with 0 being not at all confident and 10 being extremely confident) out of the 302 surveys administered. Sensitivity was 0.190 and 0.714 for physicians and Vent.io, respectively. Specificity was 0.960 and 0.673 for physicians and Vent.io, respectively. Generalized linear mixed modeling showed that physician confidence was associated with greater odds of correctly predicting intubation outcome (odds ratio [OR] 1.49; 95% CI, 1.22-1.84; p < 0.001). Vent.io had significantly greater odds of being correct when patients required intubation compared with physicians (OR 18.68; 95% CI, 1.87-186.31; p = 0.013). However, intensive care physicians outperformed Vent.io at correctly predicting when patients did not require intubation (OR 24.80; 95% CI, 13.22-46.52; p < 0.001). CONCLUSIONS AND RELEVANCE: Although predictive performance compared with human experts is promising, Vent.io needs real-time testing in a randomized clinical trial to determine if its deployment can improve clinical outcomes.
4 Apr 2026
Read appraisal →European journal of pediatric surgery : official journal of Austrian Association of Pediatric Surgery ... [et al] = Zeitschrift fur Kinderchirurgie
Latest Developments in Artificial Intelligence and Machine Learning Models in General Pediatric Surgery
Artificial intelligence (AI) and machine learning (ML) models rapidly transform health care with applications ranging from diagnostic image interpretation, predictive modeling, personalized treatment planning, real-time intraoperative guidance, and outcome prediction. However, their implementation in general pediatric surgery remains limited due to the rarity and complexity of pediatric surgical conditions, small and heterogeneous datasets, and a lack of formal AI training and competencies among pediatric surgeons.This narrative review explores the current landscape of AI and ML applications in general pediatric surgery, focusing on five key conditions: appendicitis, necrotizing enterocolitis, Hirschsprung's disease, congenital diaphragmatic hernia, and biliary atresia. For each, we summarize recent developments, including the use of AI in image analysis, diagnostic support, prediction of disease severity and outcome, postoperative monitoring, and histopathological evaluation. We also highlight novel tools such as explainable AI models, natural language processing, and wearable technologies.Recent findings demonstrate promising diagnostic and prognostic capabilities across multiple conditions. However, most AI/ML models still require external validation and standardization. The review underscores the importance of collaborative, multicenter research based on joint datasets as well as targeted AI education for pediatric surgeons to fully explore the benefits of these technologies in clinical practice.AI and ML offer significant potential to improve pediatric surgical care, but broader implementation will require multicenter collaboration, a robust dataset, and targeted AI education for pediatric surgeons.
4 Apr 2026
Read appraisal →The American journal of gastroenterology
A Large Language Model Assistant for Summarizing Hepatology Referral Documents
INTRODUCTION: To use a large language model (LLM) to create accurate and useable summaries of patient records for clinicians triaging new hepatology referrals. METHODS: We developed a comprehensive list of data elements required to triage a new hepatology referral and engaged in an iterative prompt engineering process to instruct an LLM to extract relevant data from patient referral documents. The final prompt was used on 50 original patient records from June to July 2025 to generate corresponding artificial intelligence (AI) summaries, which were assigned to 2 providers to review and triage according to their usual process. We assessed time to triage original vs AI files and accuracy of the AI files. A linear mixed-effects model was used to determine an adjusted time ratio comparing the time to triage AI files vs original files. RESULTS: AI-generated summaries were significantly shorter than original files (median [interquartile range] 2 [2-3] vs 23 [10.2-38.8] pages, P < 0.001). AI summaries had high accuracy (median [interquartile range]: 94.6% [86.5%-97.3%]) with a low hallucination rate. Use of the AI summaries led to a 60% reduction in triage time (adjusted mean triage time of 37.2 seconds for AI files vs 94.2 seconds for original files, P < 0.001). DISCUSSION: The use of an LLM led to significantly reduced document length, maintained an appropriate level of accuracy, and led to a significant decrease in clinician time to review the patient record. Future steps involve creating a fully automated workflow that is integrated into the electronic health record for widespread use.
4 Apr 2026
Read appraisal →Pediatric radiology
Data mining in pediatric radiology in the era of artificial intelligence
Data mining is the systematic process of extracting useful knowledge from large multimodal datasets and is increasingly enabled by artificial intelligence (AI) methods. Pediatric radiology is a natural field for data mining because multimodal data sources, including images, reports, metadata, and electronic health records, together capture rich information on anatomy, disease, treatment, and outcomes. In the current era, the boundaries between data mining and AI are increasingly blurred. AI assists in key steps of the mining workflow through automated labeling, information extraction, and representation learning, while data mining provides the high-quality curated datasets that underpin model performance, generalizability, and safety. This review, therefore, examines both domains together, emphasizing their interdependence in the pediatric context. We describe core concepts and workflows of data mining in pediatric radiology, including data collection, linkage, annotation, analysis, validation, and governance, and outline how modern AI tools such as deep learning, large language models, multimodal fusion, and federated learning support advanced pattern discovery across limited and heterogeneous pediatric datasets. We summarize current and emerging clinical applications across diagnosis, prognosis, radiation dose monitoring, operational analytics, reporting safety nets, and continual learning. We then discuss current challenges related to data quality and standardization, ethics, regulation, workflow integration, resource disparities, sustainability, and explainability. Finally, we highlight future perspectives, including synthetic data generation, foundation models, structured reporting, and pediatric-focused ethical frameworks that aim to enable safe, transparent, and equitable integration of AI-driven data mining to improve outcomes in children.
4 Apr 2026
Read appraisal →Kidney360
Artificial Intelligence in Critical Care Nephrology: Current Applications, Emerging Techniques, and Challenges to Clinical Integration
Artificial intelligence (AI), including machine learning, deep learning, reinforcement learning, and generative AI, has the potential to advance critical care nephrology (CCN) by enhancing prediction accuracy, improving diagnostic capabilities, supporting clinical decision making, and streamlining workflow processes. Current applications in CCN include AKI prediction, nephrotoxin surveillance, intradialytic hypotension forecasting, and AI-guided dialysis and continuous KRT management, with performance often exceeding traditional models. However, the effect on patient-centered outcomes such as mortality, dialysis dependence, and cost-effectiveness remains uncertain. Emerging techniques, such as conformal prediction for calibrated risk estimates, causal inference for intervention modeling, and reinforcement learning for adaptive ultrafiltration, show promise in enhancing reliability, interpretability, and individualized care. Generative AI and large language models extend these applications to clinical documentation, reasoning, and patient education, while raising new challenges, including hallucinations, regulatory oversight, and clinician trust. Persistent barriers such as data heterogeneity, limited external validation, alert fatigue, and economic constraints hinder broad adoption. This review synthesizes the current evidence and outlines four priorities for advancing AI in CCN: ( 1 ) rigorous multicenter validation focused on clinical outcomes, ( 2 ) integration of uncertainty quantification and causal modeling into AI tools, ( 3 ) development of clinician-centered interfaces that minimize cognitive load, and ( 4 ) establishment of transparent, adaptive regulatory and governance frameworks. Realizing the promise of AI in CCN will require multidisciplinary collaboration, fairness and generalizability testing, and sustainable implementation strategies that align technologic innovation with measurable improvements in patient care.
1 Mar 2026
Read appraisal →Ageing Res Rev
AI-driven aging digital twins: A roadmap for clinical translation in precision geriatrics
Aging and related diseases account for ⁓70 % of global mortality with declining physiological homeostasis and cognitive function increasing susceptibility to neurodegenerative disorders, cardiovascular diseases and cancer. Treatment is hampered by imprecise diagnostics and limited therapeutic options. Precision medicine involving artificial intelligence (AI)-powered Aging Digital Twin (ADT) technology which integrates multi-source heterogeneous data via advanced AI and deep learning has great potential. Virtual doppelgangers enable real-time health monitoring, predictive analytics and personalized interventions to facilitate early disease detection and optimize risk assessment and treatment strategies. Challenges remain and include age-stratified biases in training datasets and variations in data quality for older populations which may limit model generalizability. Critical concerns regarding data privacy, algorithmic transparency and clinical validation of ADT models remain unresolved. The current review examines the transformative potential of AI-enabled ADTs in precision medicine for aging and highlights technical and ethical limitations.
1 Jan 2026
Read appraisal →