Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 10 appraisals
PloS one
Empowering rural governance with digital technology: Deep learning models for automated detection of rural buildings using remote sensing images
Building detection from drone imagery represents a transformative approach to rural governance by enabling precise spatial data acquisition for critical applications including illegal construction monitoring, disaster assessment, and cadastral mapping. However, automated detection systems face persistent challenges including extreme scale variations in rural buildings, complex background interference from vegetation and shadows leading to boundary ambiguity, and severe scarcity of high-quality annotated datasets that limit model generalization. To overcome these limitations, this study introduces an integrated framework featuring three innovative components: the Multi-scale Hybrid Attention module employs parallel convolutional pathways with channel and spatial attention to dynamically capture multi-scale features while suppressing background noise; the Dynamic Feature Pyramid Network utilizes content-aware routing to adaptively fuse hierarchical features for optimal scale-invariant representation; and the Progressive Contrastive Learning strategy leverages both labeled and unlabeled data through hard sample mining to enhance discriminability under data constraints. Extensive experiments validate the model's efficacy, achieving a mean Intersection over Union (MIoU) of 87.3%, pixel accuracy (PA) of 94.2%, and mean Average Precision (mAP) of 89.6% on the Massachusetts Buildings Dataset, substantially surpassing benchmarks like U-Net (80.1% MIoU), SegNet (78.9% MIoU), and DeepLabV3+ (82.4% MIoU), with ablation studies confirming critical module contributions (e.g., MIoU drops to 81.5% without MHA). The framework demonstrates robust cross-dataset generalization (72.3% MIoU on Chinese rural data) and effective problem resolution, establishing a scalable solution for intelligent rural governance through accurate building extraction. The dataset and code used in this study have been uploaded to the GitHub website: https://github.com/xiexie1234567890/rural_building_detection/tree/main.
14 June 2026
Read appraisal →JMIR medical education
Performance of Large Language Models on the Brazilian National Medical Education Examination: Comparative Benchmark Study
BACKGROUND: Large language models (LLMs) are rapidly incorporated into medical education and examination preparation; yet, most benchmarking evidence is derived from English-language material. Whether frontier commercial models and Brazilian Portuguese domain-specialized systems perform equivalently on high-stakes Brazilian medical examinations remains unclear. OBJECTIVE: This study aims to quantify and compare the performance of 9 frontier commercial LLMs and 1 Brazilian Portuguese domain-specialized system (Charcot, Voa Health) on 2026 Brazilian National Medical Education Examination (Exame Nacional de Avaliação da Formação Médica [ENAMED] 2026) and to describe the patterns of systematic between-model error as complementary quality signal. METHODS: All 100 items of ENAMED 2026 (99 valid after annulment) were administered to 10 frontier-panel models across 5 independent runs under identical Portuguese prompts (temperature=0; top-p=.95). Commercial models were accessed through a unified OpenRouter client layer (DeepSeek provider-pinned). The primary outcome was mean accuracy against the preliminary key; the secondary outcomes were convergence error (CE), normalized mean response time (NMRT), and intermodel agreement. Accuracy was analyzed with Shapiro-Wilk, Levene, Kruskal-Wallis (ε2), Dunn-Holm post hoc, and a binomial generalized linear mixed model with question and run random intercepts. NMRT excluded Charcot (different stack) and Grok 4 (latency outlier). A total of 7 open-weight and small language models were assessed as a small language model (SLM) substudy. RESULTS: Frontier-panel accuracy ranged from 73.74% (365/495) for GPT-4o-mini to 96.97% (480/495) for Charcot. Accuracy was nonnormal (Shapiro-Wilk, W=0.82; P<.001) with homogeneous variance (Levene P=.26). Kruskal-Wallis showed large between-model differences (H9=47.65; P<.001; ε2=0.97). Dunn-Holm flagged 8 of 45 pairs: Charcot was separable from GPT-4o-mini, DeepSeek v3.2-exp, and Grok 4, but not from the top frontier cluster. The generalized linear mixed model preserved the ranking (all comparators odds ratio<1 vs Charcot; upper CI<1 except GPT-5). A total of 9 items met the default CE criterion; item 77 showed 10-of-10 convergence, later confirmed by Instituto Nacional de Estudos e Pesquisas Educacionais Anísio Teixeira rectification, and sensitivity analysis preserved the CE set (5-20 items). Intermodel agreement was high (Fleiss κ=0.852; Krippendorff α=.852). Among 8 retained commercial models, NMRT correlated positively with accuracy (Spearman ρ=0.74; P=.04); these results are interpreted as descriptive of an orchestration-level latency-accuracy law. SLM accuracy ranged from 47.47% (Gemma 3 4B) to 82.22% (GPT-OSS 120B), with lower agreement (Fleiss κ=0.508) and 17 CE items. The best SLM-panel model lagged every frontier-panel model except GPT-4o-mini, with a 12-15 percentage-point gap against the top cluster. CONCLUSIONS: On ENAMED 2026, a Brazilian Portuguese domain-specialized system ranked first, indistinguishable from frontier commercial cluster and above subfrontier and open-weight systems. Charcot's architecture is not publicly disclosed; these findings should be interpreted as comparative black-box evidence of performance and not as mechanistic evidence of specialization. CE was stable, and it prospectively flagged 1 rectified item, supporting its use as a quality assurance screen.
31 May 2026
Read appraisal →PloS one
Transformation of artistic style and innovative design of oriental folk patterns based on AIGC Technology-A case study of Zhuxian town new year paintings from China
In response to the limited cross-domain innovation in the digitalization of traditional oriental folk art, this study takes the New Year pictures of China's Zhuxian Town as a case. It develops a collaborative technical framework of combining the Liblib platform and a LoRA model to explore AI-generated digital re-creation of traditional art. A high-quality dataset was built through a three-stage process of image acquisition, multidimensional screening, and expert review. A three-layer keywords thesaurus was constructed through literature analysis, questionnaire surveys, and semantic clustering. Using a two-stage training strategy combining pre-training and fine-tuning, along with dynamic optimization, the model accurately captures the stylistic features of Zhuxian New Year pictures. The generated outputs integrate traditional aesthetics with modern design and are applied to cultural and graphic creative products. The results demonstrate the effectiveness of the proposed framework for preserving artistic style while enabling cross-domain innovation and offer a practical technical reference for the digital inheritance and modernization of related folk arts.
30 May 2026
Read appraisal →PloS one
Research on the semantic segmentation of Thangka images via an improved PIDNet.
In the digital preservation, restoration, and research of Thangka paintings, real-time semantic segmentation plays a crucial role in rapid image analysis. However, Thangka images exhibit intricate compositions, where principal figures often blend with backgrounds, ritual objects, and intricate ornaments, leading to blurred boundaries and fine details that challenge conventional segmentation methods in balancing accuracy and efficiency. To address this, we propose an improved PIDNet-based model incorporating ECA-Pag (Efficient Channel Attention-Path Aggregation) and LGFM (Local-Global Feature Fusion Module) modules, along with PConv3 (Partial Convolution 3), enhancing feature extraction and segmentation precision. Experimental results demonstrate that our model achieves 73.28% mIoU and mB-Fscore (mean Boundary F-score) of 40.01% on a custom Thangka dataset while maintaining 109.03 FPS (Frames Per Second), ensuring both high accuracy and real-time performance. Furthermore, evaluations on the Cityscapes benchmark confirm the model's generalization capability, outperforming baseline methods. This work provides an efficient and reliable solution for Thangka image segmentation, with potential applications in cultural heritage preservation and broader computer vision tasks.
17 May 2026
Read appraisal →PloS one
Smartphone-based accurate pothole depth estimation using monocular RGB and LiDAR-guided deep learning
The maintenance of the road infrastructure is vital to ensure public safety and the efficiency of transportation, but the unpredictable and widespread occurrence of potholes often hinders continuous monitoring. Traditional detection approaches typically rely on expensive equipment or manual inspections, rendering them impractical for large-scale, real-time deployment. To overcome these limitations, we present a low-cost and scalable pothole assessment framework that leverages the built-in LiDAR sensors available in modern smartphones in combination with a hybrid deep learning pipeline. The proposed system fuses RGB imagery with LiDAR-guided refinement to achieve precise depth estimation and accurate pothole segmentation, further enhanced through post-processing for improved boundary alignment. A comprehensive dataset of 25,000 RGB-LiDAR image pairs, collected under various road and lighting conditions, was curated to train and evaluate the framework. Extensive experiments demonstrate that our approach achieves high accuracy in both depth estimation and pothole delineation, significantly outperforming existing methods. This work highlights the potential of lightweight smartphone-based solutions for real-time automated road condition assessment, offering a practical and economical tool for modern infrastructure management.
12 May 2026
Read appraisal →European journal of dental education : official journal of the Association for Dental Education in Europe
Assessment of the Large Language Models in Creating Dental Board-Style Questions: A Prospective Cross-Sectional Study
INTRODUCTION: Although some studies have investigated the application of large language models (LLMs) in generating dental-related multiple-choice questions (MCQs), they have primarily focused on ChatGPT and Gemini. This study aims to evaluate and compare the performance of five of the LLMs in generating dental board-style questions. MATERIALS AND METHODS: This prospective cross-sectional study evaluated five of the advanced LLMs as of August 2024, including ChatGPT-4o (OpenAI), Claude 3.5 Sonnet (Anthropic), Copilot Pro (Microsoft), Gemini 1.5 Pro (Google) and Mistral Large 2 (Mistral AI). The five most recent clinical guidelines published by The Journal of the American Dental Association were used to generate a total of 350 questions (70 questions per LLM). Each question was independently evaluated by two investigators based on five criteria: clarity, relevance, suitability, distractor and rationale, using a 10-point Likert scale. RESULT: Inter-rater reliability was substantial (kappa score: 0.7-0.8). Median scores for clarity, relevance and rationale were above 9 across all five LLMs. Suitability and distractor had median scores ranging from 8 to 9. Within each LLM, clarity and relevance scored higher than other criteria (p < 0.05). No significant difference was observed between models regarding clarity, relevance and suitability (p > 0.05). Claude 3.5 Sonnet outperformed other models in providing rationales for answers (p < 0.01). CONCLUSION: LLMs demonstrate strong capabilities in generating high-quality, clinically relevant dental board-style questions. Among them, Claude 3.5 Sonnet exhibited the best performance in providing rationales for answers.
6 May 2026
Read appraisal →Developmental science
Making Machine Learning Accessible for Developmental Science: The Case of Automated Face Detection
The last decade has seen rapid advancements in machine learning, significantly transforming fields like cybersecurity and healthcare. Developmental science has been slower to adopt these technologies. Yet, machine learning holds immense potential to transform this field, enabling scalable and data-driven insights into developmental processes. Broader adoption is currently hindered by challenges in algorithm selection and technical implementation. We address these barriers by focusing on an area that has reached high sophistication from a machine learning perspective while also being of significant interest to developmental scientists: face detection. Face detection is crucial for analysing visual experiences through children's dynamic, first-person views. Automatising this process allows efficient handling of large egocentric datasets, enabling well-powered studies otherwise limited by labour-intensive manual annotation. Here, we systematically evaluated 13 state-of-the-art face detection algorithms (DeepFace library) using data from two increasingly common developmental methodologies involving children under 3 years of age: head-mounted eye-tracking in more structured settings (N = 20; n = 10 4-month-olds, n = 10 8-month-olds) and head-mounted cameras in naturalistic home environments (N = 10 18-29-month-olds). Benchmarking these algorithms against manual annotations revealed that YOLOv11Face (M) and RetinaFace consistently outperformed others in terms of precision and recall, exhibiting strong concordance with manual ratings, lower error, reduced systematic deviation and robust rank-order correlations with manual annotations. To facilitate broader adoption, we introduce an accessible face detection tool (TinyExplorer Detection App), promoting efficiency, scalability, and innovation in developmental science by widening access to machine learning.
5 May 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
MMC-CS: Multi-branch multi-stage contrastive learning for self-supervised compressed sensing
In recent years, optimization-inspired networks that integrate optimization theory into deep neural networks (DNNs) have achieved remarkable success in image compressed sensing (ICS). However, current DNN-based ICS methods face two key challenges: (1) the difficulty of acquiring ground-truth (GT) data from labeled measurements and (2) the underutilization of measurement data. To address these issues, a novel self-supervised deep learning framework is proposed to solve the inverse problem in ICS. Building upon traditional self-supervised learning, our proposed method leverages measurement values through a multi-branch, multi-stage progressive cross-contrast structure, enabling effective learning of the common prior of the original image in the absence of GT data. Based on this framework, a novel Multi-branch Multi-stage Cross-contrast CS (MMC-CS) end-to-end DNN is designed, which unfolds the Proximal Gradient Descent (PGD) algorithm. Our approach further implements multi-scale co-optimization in the image reconstruction process, integrating both the image paths (IPs) and convolutional feature paths (CFPs). Additionally, the reconstruction performance across multiple spatial levels is enhanced by incorporating wavelet convolution (WTConv). Extensive experiments demonstrate that our proposed method outperforms existing self-supervised approaches, achieving an average PSNR improvement of 0.3-1.6 dB. Moreover, it shows strong potential to compete with state-of-the-art supervised methods.
4 May 2026
Read appraisal →PloS one
Adaptive Feature Fusion Gate and Gated Channel-Spatial Attention in CNN-Transformer Models for Music Genre Classification
With the rapid growth of music data, automatic music genre classification has become a critical task in music information retrieval. Traditional methods based on handcrafted features are increasingly inadequate when handling large-scale analysis. This paper proposes the Convolutional Neural Network-Gated Transformer Network (CT-GateNet), a hybrid architecture that integrates a gated channel-spatial attention mechanism with an adaptive feature fusion gating mechanism to achieve discriminative feature learning and efficient feature integration. To mitigate data scarcity, a data augmentation strategy based on a denoising diffusion probabilistic model is introduced. Experiments are conducted on three public music genre datasets: GTZAN, FMA-SMALL and FMA-Medium. The method achieves classification accuracies of 98.72%, 89.42%, and 69.07% on GTZAN, FMA-SMALL and FMA-Medium, respectively, demonstrating outstanding performance and robust generalization capabilities. These results validate CT-GateNet's effectiveness in music genre classification and provide valuable insights for audio classification research.
11 Apr 2026
Read appraisal →PloS one
Prediction of on-field rugby scrummaging contact forces from videos using artificial neural networks
Rugby is a contact sport with a high risk of injury. Scrums are a key component of the game and represent a collision event associated with serious injuries, especially to the cervical spine. Traditional methods for measuring scrum forces rely on both pressure sensors and scrum machines, but these have limitations for on-field monitoring. Thus, this study investigates the feasibility of using a machine learning approach to predict contact forces in rugby scrummaging during real-game situations, based on only anatomical landmarks of players extracted from video footage. A recurrent neural network (RNN) model was trained to predict contact forces from velocity signals. A dataset of eleven rugby teams (22 forward packs, n = 176 players) during on-field scrum engagements was used. Data augmentation techniques using a generative adversarial network and the Mixup technique were applied to increase the size of the training dataset. The RNN was trained using time-dependent data, with 2D trajectories from video landmarks as inputs and force signals from pressure sensors as outputs. The RNN´s prediction results showed good to excellent agreement between the predicted and measured time-dependent normal contact force signals, with a mean correlation of 0.95 ± 0.05. The mean normalized RMSE was 9.3 ± 4.3% and the mean normalized absolute difference in peak force was 6.7 ± 5.5%. This study demonstrated the feasibility of using RNNs to quantify contact forces in rugby scrummaging using only single-camera video, without players wearing sensors. In combination with a 2D human pose estimation model, the ANN trained in this study could support performance analysis, coaching, technique correction, and injury prediction using in-game data.
9 Apr 2026
Read appraisal →