Axial spondyloarthritis (axSpA) is an inflammatory disease marked by chronic low back pain, with a global average diagnostic delay of 6.7 years. Early diagnosis is crucial for improving prognosis and reducing disability rates, yet primary care physicians (PCPs) may find it challenging to ensure timely recognition and referrals. This study developed and validated Spondyloarthritis Agents (SpAgents), an early diagnostic system based on a multi-agent framework integrating large language models (LLMs) and imaging models. The SpAgents framework includes PlannerAgent, DataAgent, ToolAgent, and DoctorAgent, supported by long-term memory for dynamic knowledge updates. We enrolled 596 patients, dividing 545 from one hospital into a training dataset (n = 359) and a validation dataset (n = 186), along with an independent cohort of 51 patients from five additional hospitals for testing. SpAgents demonstrated strong diagnostic performance, achieving sensitivity of 0.8615 and specificity of 0.8000 during validation, and 0.9375 and 0.7368 during testing. SpAgents exhibited significantly higher sensitivity (0.9400) and accuracy (0.8600) than both PCPs and junior rheumatologists, with overall performance equivalent to that of senior rheumatologists. Under SpAgents-assisted diagnosis, both PCPs and junior rheumatologists showed marked improvements in sensitivity and accuracy. SpAgents effectively enhance early axSpA identification among PCPs, offering an innovative solution to reduce diagnostic delays.

Melanoma is one of the top 5 cancer types, causes most deaths among skin cancers, and can be frequently misdiagnosed. Recent pathology image foundation models remain difficult to make accurate differential diagnosis across over forty melanocytic neoplasm histologic subtypes. Motivated by the diagnostic reasoning process of dermatopathologists, we curated a high-quality image and knowledge corpus database containing 2893 images and 1102 knowledge entries annotated by expert dermatopathologists at the University of Pennsylvania. Leveraging this multi-modal dataset, we present "Melan-Dx", a knowledge-enhanced AI framework that augments frozen pathology vision-language models through retrieval from a curated vision-knowledge database, improving differential diagnosis at both patch and whole-slide levels. Melan-Dx, at its best performance, demonstrates 0.869 accuracy for binary classification, 0.699 Top-1 accuracy among forty-class classification, 0.915 ROC AUC for few-shot WSI tasks, and 0.925 AUPRC for fully supervised WSI tasks. Across all experimental settings, Melan-Dx shows improvements up to 13.8% over linear and fully finetuned methods, 23-70.6% over zero-shot approaches and up to 8.4% improvements in whole slide image classification. These findings suggest that a query database with a knowledge-enhanced AI framework can further improve existing pathology foundation models without fine-tuning the vision backbone. The code is publicly available at https://www.github.com/zhihuanglab/Melan-Dx-code .

Accurate and generalizable cancer diagnosis from whole slide images (WSIs) remains challenging due to limited fine-grained annotations, complex tumor architectures, and domain shifts across scanners and institutions1. We introduce StructMIL, a structure-aware and prototype-driven multiple instance learning framework designed for robust and interpretable cancer detection and grading2. StructMIL integrates graph-based topological priors with histological context, employs prototype-enhanced pooling for stable and transparent predictions, and incorporates a unified domain-generalization strategy that combines contrastive alignment, adversarial confusion, and consistency regularization. Evaluated on Camelyon16 for breast cancer metastasis detection and PANDA for prostate cancer Gleason grading, StructMIL achieves state-of-the-art performance. On Camelyon16, StructMIL improves cross-center AUC by +3.2% over standard MIL baselines, reaching an AUC of 0.967. On PANDA, it improves cross-scanner Gleason grading robustness with a +7.4% Cohen’s Kappa gain compared with prior MIL models, demonstrating substantially reduced performance degradation under domain shift. StructMIL further provides interpretable prototype-based attribution maps that highlight biologically meaningful structures more reliably than conventional MIL and graph-free approaches3. By jointly improving accuracy, interpretability, and generalization across scanners and medical centers, StructMIL offers a practical and clinically aligned solution for large-scale deployment in multi-center computational pathology workflows4.

Pain is a complex multidimensional experience that integrates sensory and emotional components, presenting significant challenges for accurate assessment in clinical practice. Traditional methods of pain evaluation rely on subjective self-reporting and each individual's ability to communicate their pain experience. In light of the effect of pain on the Autonomic Nervous System, researchers are interested in developing objective assessment techniques using physiological signals. This paper outlines the latest advances in pain biomarkers and machine learning methods for assessing pain using physiological signals, highlighting the growing interest and unmet demand in this area. A comprehensive literature review was conducted, covering studies between 2014 and 2024. The discussion is organised into two areas: first, an analysis of the variations in signal feature behaviour across different pain types, and second, a review of the current state-of-the-art models for pain assessment developed using classical machine learning and deep learning techniques.

Positron emission tomography (PET)/computed tomography (CT) for myocardial perfusion imaging (MPI) provides multiple imaging biomarkers, often evaluated separately. We developed an artificial intelligence (AI) model integrating key clinical PET MPI parameters to improve the diagnosis of obstructive coronary artery disease (CAD). From 17,348 patients undergoing cardiac PET/CT across four sites, 1664 with invasive coronary angiography and no prior CAD were retrospectively analyzed. Coronary artery calcium (CAC) scores were derived from CT attenuation correction maps, and XGBoost model was trained on one site using 10 image-derived parameters: CAC, stress/rest left ventricular ejection fraction, stress myocardial blood flow (MBF), myocardial flow reserve (MFR), ischemic and stress total perfusion deficit (TPD), transient ischemic dilation ratio, rate pressure product, and sex. External validation was performed across three independent sites. In the testing cohort (n = 1278; CAD prevalence 53%), the AI model achieved an area under the receiver operating characteristic curve (AUC) of 0.83 (95% CI: 0.81-0.85), outperforming experienced physicians (0.80, p = 0.02) and individual biomarkers such as ischemic TPD (0.79, p < 0.001) and MFR (0.75, p < 0.001). Performance was consistent across sex, body mass index, and age. AI integrating perfusion, flow, and CAC scoring improves PET MPI diagnostic accuracy, offering automated and interpretable predictions for CAD diagnosis.

This study addresses the critical challenge of master-slave latency in robot-assisted telesurgery by introducing a Digital Twin Visual Assistance (DTVA) system. DTVA integrates parametric 3D modeling and virtual endoscopic visualization within a tri-layered architecture to enable real-time bidirectional synchronization. The system was evaluated on a geographically distributed robotic platform using programmable latency emulation. Results demonstrated that DTVA maintained spatial precision within 2 mm error under typical conditions and reduced peg-transfer completion time by 13.6% under 900 ms communication latency while lowering operator workload by 27.2%. Clinical validation through teleoperated radical nephrectomy under 300 ms communication latency confirmed feasibility, with all procedures completed successfully without complications and favorable perioperative outcomes. The study establishes DTVA’s capacity to mitigate latency effects and demonstrates preliminary clinical feasibility for telesurgical procedures.

Apple Watch provides continuous monitoring of physiological and behavioural health metrics, increasingly used to support health-care delivery. Yet, evidence regarding its measurement accuracy remains limited. We aimed to assess the accuracy of measurements from Apple Watch. We searched nine databases from inception to September 24, 2025, with no restrictions on language or publication type. Eligible studies validated any Apple Watch health metric against a criterion method. The primary outcome was the agreement between Apple Watch and the criterion. We included 82 studies, which assessed 14 health metrics (430,052 participants; pooled mean age 41.3 years [SD 13.3]). Bland-Altman meta-analysis showed a small underestimation of heart rate, although limits of agreement (LoA) indicated moderate measurement variability (mean bias -0.27 bpm [95% CI -0.72-0.17]; LoA -7.19 to 6.64). For atrial fibrillation detection, Apple Watch was more specific than sensitive (specificity 0.91 [95% CI 0.81-0.96]; sensitivity 0.79 [95% CI 0.61-0.90]). For blood oxygen saturation, there was low mean bias (-0.04% [95% CI -0.42-0.35]) but wide limits of agreement (-4.00 to 3.94). Accuracy for sleep and step count was moderate, whereas error for energy expenditure was inconsistent and frequently large. Measurement accuracy varied by metric, measurement conditions, and individual physiology. Longitudinal validation of key clinical metrics, including vital signs, is needed to inform clinical practice and policy. This study was registered with PROSPERO, CRD42023481841.

Breath analysis offers a promising noninvasive strategy for early cancer detection by capturing disease-specific volatile organic compound (VOC) signatures in exhaled breath. In this study, we developed a hierarchical deep convolutional neural network (HD-CNN)-based platform for dual-cancer classification using a multimodal gas sensor array. Clinical breath samples were collected from 206 participants, including 67 healthy controls (HC), 78 lung cancer (LC), and 61 gastric cancer (GC) patients. The sensor array, composed of semiconductor metal oxide (SMO), electrochemical (EC), and photoionization detector (PID) sensors, generated time-resolved signals that were converted into 2D response maps for classification. The HD-CNN model employed a two-stage structure: a coarse classifier to distinguish HC from cancer patients (CP), followed by a fine classifier to separate LC and GC. In 5-fold cross-validation, the HD-CNN achieved classification accuracies of 82.1% (HC), 84.0% (LC), and 88.1% (GC), and average AUCs of 0.89, 0.92, and 0.89, respectively. Compared to a 1D CNN, the HD-CNN demonstrated superior class separability, increased prediction confidence, and overall enhanced performance. Additionally, we evaluated multiple coarse–fine configurations and found that isolating HC in the first stage resulted in the highest overall performance. These results support the utility of hierarchical learning and multimodal sensing for robust and scalable breath-based multi-cancer screening.

Large language models (LLMs) have the potential to scale assessment in medical education, but their psychometric equivalence to expert-written items and the detectability of their origin remain uncertain. In a preregistered, single-center, blinded observational, within-subject comparison, we evaluated 24 GPT-4o-generated versus 24 human-authored topic-matched multiple-choice questions (MCQs) across radiation oncology, radiology, and nuclear medicine. Medical students (n = 82) and physicians (n = 46) completed an identical 48-item formative mock examination, with item origin masked. Item difficulty (human: mean 0.65 [SD 0.22] vs LLM: 0.67 [0.20]) and discrimination (0.27 [0.12] vs 0.29 [0.12]) did not differ significantly; participants did not identify item origin above chance (0.50). Expert ratings of appropriateness and didactic quality showed low interrater agreement (ICC = 0.07-0.18). In this expert-reviewed, human-in-the-loop workflow, the item difficulty and discriminatory power of MCQs generated with GPT-4o did not differ significantly from those of expert-authored items, and were not reliably recognized as AI-generated by examinees. These findings delineate a feasible pathway for responsibly scaling formative assessment content in imaging-focused medical education, while underscoring the need for explicit educational policies regarding oversight, transparency, and fairness.

Early and accessible detection of Alzheimer's disease (AD) remains a major clinical challenge. We developed a machine learning-based blood transcriptomic model, the Lactylation-Derived Score (LDS), from lactylation-related genes across nine AD cohorts, using a standardized pipeline with z-score normalization, random forest-based feature screening, plsRglm modeling, and 10-fold cross-validation. LDS was externally tested in seven independent brain transcriptomic datasets and clinically validated in an independent plasma cohort (n = 540); logistic regression was used to integrate LDS with plasma phosphorylated tau 181 (p-tau181) and p-tau217. LDS achieved an AUC of 0.897 (95% CI 0.849-0.934) in the Training Cohort and 0.772 (95% CI 0.729-0.815) in the plasma validation cohort, while the three-marker model (LDS + p-tau181 + p-tau217) yielded the highest diagnostic performance (AUC 0.859, 95% CI 0.824-0.893). LDS alone effectively identified AT⁺ individuals (AUC 0.861, 95% CI 0.827-0.897), and a five-gene classifier derived from LDS genes stratified amnestic mild cognitive impairment with an AUC of 0.809 (95% CI 0.714-0.836). LDS-high individuals showed neuroinflammatory activation and metabolic stress signatures, indicating that this scalable, interpretable transcriptomic model complements plasma p-tau biomarkers and supports precision digital medicine in AD.