NPJ Digital Medicine最新文献

Large language models (LLMs) show promising diagnostic and triage performance, yet direct comparisons with healthcare professionals (HCPs) and collaborative effects remain limited. We conducted a systematic review and meta-analysis of studies (January 2020 to September 2025) comparing the diagnostic or triage accuracy of LLMs, HCPs, or their collaboration across seven databases. Studies using multiple-choice formats rather than open diagnostic generation were excluded. We extracted top-1, top-3, top-5, and top-10 diagnostic and triage accuracies and pooled results using multilevel random-effects models to account for nested observations. Of 10,398 studies screened, 50 met criteria, evaluating 25 different LLMs across diverse medical specialties. The relative diagnostic accuracy of LLMs versus HCPs progressively improved from 0.89 (95% CI, 0.79-1.00) for top-1 to 0.91 (0.83-1.00) for top-3, 1.04 (0.89-1.22) for top-5, and 1.17 (0.87-1.57) for top-10 diagnoses, with significant model variability. LLM-assisted HCPs outperformed HCPs alone, with relative diagnostic accuracy of 1.13 (1.00-1.27) for top-1, 1.11 (1.01-1.23) for top-3, 1.42 (1.16-1.73) for top-5, and 1.33 (0.94-1.87) for top-10 diagnoses. Triage accuracy was similar between LLMs and HCPs (1.01 [0.94-1.09]). These findings show potential for LLM integration but methodological flaws in studies necessitate rigorous real-world evaluation before clinical implementation.

Intracranial aneurysms (IAs) are life-threatening vascular conditions requiring accurate risk assessment to guide treatment. Hemodynamic biomarkers such as wall shear stress and oscillatory shear index are promising predictors of rupture risk but remain underused clinically due to the high computational cost of traditional CFD methods. We propose a physics-constrained graph neural network (GNN) framework trained on high-fidelity CFD data to predict full 3D, time-resolved hemodynamic fields throughout the cardiac cycle. Our model incorporates enhanced node features and physics-based constraints to capture complex spatio-temporal flow behavior in near real time. It generalizes to varying inflow conditions and unseen patient-specific geometries with no fine-tuning. Additionally, we release a benchmark dataset of 105 patient-derived aneurysm geometries with CFD fields to support the machine learning (ML) community. This is the first GNN model applied to transient 3D aneurysmal flow prediction, paving the way for rapid, AI-driven hemodynamic analysis toward risk stratification and treatment planning.

Wearable electroencephalography (EEG) devices are miniaturized, portable, and wireless systems for long-term brain monitoring, demonstrating significant potential as accessible mild cognitive impairment (MCI) screening tools based on objective neurophysiological biomarkers. However, their performance in MCI detection remains unclear, and their translation to real-world applications faces several challenges. This study aimed to comprehensively evaluate wearable EEG for MCI detection, identify key characteristics that optimize classification performance and usability, and address gaps in effective design implementation. We conducted a systematic search across seven databases, screening 1562 records and analyzing 21 studies that examined 16 distinct wearable EEG devices for MCI detection. The results revealed considerable variation in classification accuracy (range: 46-95%). A system-level analysis of the entire wearable EEG system and data flow identified seven critical factors that optimize the trade-off between diagnostic performance, portability, and affordability: (1) moderate channel density; (2) frontal and parietal electrode placement; (3) elderly-friendly multi-domain cognitive tasks; (4) adaptive signal preprocessing; (5) multi-domain feature extraction; (6) ensemble classifiers; and (7) multimodal integration. Additionally, methodological considerations for future wearable EEG-based MCI detection research include: (1) standardize MCI diagnostic frameworks; (2) increase sample diversity; (3) optimizing device usability and technical specifications; (4) standardize recording protocols; (5) harmonizing data processing pipelines; (6) validate in real-world settings; (7) assess cost-effectiveness; and (8) implement comprehensive reporting guidelines. These insights enable further translational applications of wearable EEG-based MCI detection and provide a foundation for developing user-friendly systems that could transform early cognitive impairment screening in community and primary care settings.

Primary open-angle glaucoma (POAG) screening using artificial intelligence (AI) has emerged as a transformative method to identify undiagnosed disease. African ancestry individuals are under-represented in current datasets for AI models, despite being disproportionally affected by this blinding disease. We developed a deep learning model that screens for POAG using fundus photography from Primary Open-Angle African American Glaucoma Genetics (POAAGG) subjects (n = 64,129 images, including 42,914 images from 1782 cases and 21,215 images from 682 controls). Our final diagnosis pipeline is as follows: (1) select the six most informative images from single timepoint using a Binary Classifier, (2) predict POAG probability from each image using Vision-Transformer, (3) make final POAG predictions by averaging predicted probabilities across selected images (AUC = 0.925). The model was evaluated on the REFUGE-1 dataset of Chinese ancestry individuals (AUC = 0.920). Our model has applications to POAG screening in public settings such as primary care offices, as well as low-resource settings.

The causes of endurance running-related injury (RRI) are multifactorial, yet little research has been conducted which utilizes multidisciplinary risk factors for individualized RRI prediction. This paper presents a machine learning (ML)-ready RRI weekly prediction dataset using evidence-based multidisciplinary risk factors. Risk factors in genetic single-nucleotide polymorphisms, history, muscular strength, biomechanics, body composition, nutrition, and training were collected from competitive endurance runners (n = 142), who were prospectively monitored for 12 months for RRIs, accumulating 6181 weekly samples. ML models were fitted using (i) risk factors with high-level supporting evidence, and (ii) a broader range of risk factors to establish a performance baseline. Model performance (AUC = 0.784 ± 0.014) showed moderate improvement compared to previous RRI prediction modeling. Random forest achieved the best performance (AUC = 0.781 ± 0.016, 0.784 ± 0.014), which was significantly higher (q < 0.05) than most other algorithms. Only logistic regression achieved significantly improved (q < 0.05) performance when trained using a broader range of risk factors compared to a selection of high-quality risk factors. This study introduces a reproducible methodological framework for future ML sports injury prediction research and a valuable dataset for pooling in larger-scale analytics. Comparisons among different ML methods revealed nuanced insights into the interaction between data structure and model suitability.

Rare diseases affect an estimated 300-400 million people worldwide, yet individual conditions remain underdiagnosed and poorly characterized due to low prevalence and limited clinician familiarity. Computational phenotyping offers a scalable approach to improving rare disease detection, but algorithm development is constrained by scarce high-quality labeled data. Expert-labeled datasets from chart reviews and registries are highly accurate but limited in scope, whereas labels derived from electronic health records (EHRs) provide broader coverage but are often noisy or incomplete. To efficiently leverage both sources, we propose WEST (WEakly Supervised Transformer) for rare disease diagnosis and subphenotyping from EHRs. At its core, WEST employs a weakly supervised transformer trained on a limited set of expert-validated labels and extensive probabilistic silver-standard labels-derived from structured and unstructured EHR features-that are iteratively refined across training rounds to improve model calibration. We evaluate WEST on two rare pulmonary conditions using EHR data from Boston Children's Hospital and show that it outperforms existing methods in phenotype classification, identification of clinically relevant subphenotypes, and prediction of disease progression. By reducing reliance on manual annotation, WEST enables label-efficient representation learning that supports accurate rare disease diagnosis and reveals deeper clinical insights from routine EHR data.

Isolated REM Sleep Behavior Disorder (iRBD) is a strong predictor of neurodegenerative diseases, particularly synucleinopathies. Current diagnosis requires overnight video-polysomnography (vPSG) in sleep laboratories. Limited access to vPSG and differences in sleep habits result in diagnostic challenges. Here we aimed to evaluate the feasibility of identifying iRBD from a lumbar-mounted wearable sensor in the home setting and explored night-to-night variability. Seventy-three participants (15 iRBD, 58 controls) underwent vPSG, followed by six nights of wearing a lower-back inertial measurement unit at home. iRBD participants showed distinct mobility patterns compared to controls. Machine learning models were trained on mobility features and classified iRBD with high sensitivity and moderate specificity. Performance improved with increased nights, plateauing at five nights recorded at home. Principal component analysis identified substantial differences between lab and home data. Our findings suggest that lumbar-mounted wearables can support sensitive, multi-night home-based detection of nocturnal motor patterns associated with iRBD, with potential utility as part of a staged screening approach and for enriching cohorts for further evaluation.

Brown adipose tissue (BAT) plays a key role in energy metabolism and cardiometabolic health. Its detection typically relies on 18F-FDG PET, which is costly, radiation-intensive, and impractical for large-scale screening. We propose a deep learning model to estimate regional metabolic activity in adipose tissue from standard non-contrast CT, enabling PET-like insights without radiotracers. Using paired PET/CT data from two independent cohorts, we trained a conditional Generative Adversarial Network (cGAN) to predict standardized uptake values (SUV) within adipose regions identified on CT. The network included a fat-focused loss function to enhance metabolic signal estimation. Predicted activations showed strong agreement with PET-derived values and were reproducible across anatomical regions and datasets. This method provides a radiation-sparing alternative for assessing adipose metabolic activity in clinical and research settings and it could support population-based studies of BAT, metabolic health, and disease progression using routine chest CT scans without additional imaging burden.

Prostate cancer is a leading cause of male cancer mortality, and early, accurate diagnosis is critical. Artificial intelligence (AI), including machine learning, deep learning, and radiomics, enhances detection, characterization, and treatment assessment across TRUS, mp-MRI, and PSMA PET/CT. AI models achieve high accuracy, often matching experts, improving small-lesion detection, and supporting risk stratification. Challenges remain in data quality, generalization, clinical integration, and ethics, with future prospects in multi-omics, explainable AI, and workflow-embedded decision support.

Sarcopenia assessment requires biomarkers capturing muscle-specific strength beyond single-slice measurements. We developed an automated MRI framework segmenting 27 pelvic-thigh musculoskeletal structures to investigate muscle distribution as functional biomarkers. Among 37,004 UK Biobank participants (64.5 ± 7.9 years), transformer-based segmentation achieved Dice similarity coefficient of 0.896. Dixon MRI-derived thigh muscle volume showed exceptional DEXA concordance (r = 0.936). Posterior/anterior (P/A) muscle ratio independently predicted adverse outcomes: weak grip strength (OR 1.60, 95%CI 1.45-1.77), sarcopenia (OR 1.42, 95%CI 1.13-1.78), mortality (OR 1.49, 95%CI 1.23-1.81), and falls (OR 1.12, 95%CI 1.05-1.20), all p < 0.005, while left/right asymmetry showed no associations. Automated MRI phenotyping reveals muscle distribution patterns, particularly reduced anterior compartment volume, predict functional decline independent of total muscle mass, supporting evolution toward composition-aware sarcopenia criteria.