A foundation model trained on raw, multimodal sleep recordings identified five clinically distinct risk groups that stratified long-term mortality and disease incidence beyond the apnea–hypopnea index (AHI). The…

Imaging AI is moving beyond fixed detection tasks toward transferable models, outcome forecasting, noninvasive tumor phenotyping, and harder scrutiny of commercial systems. The central question is no longer…

This week’s strongest studies pushed medical AI beyond one-off classification: pathology models that reason and travel across hospitals, speech and CT biomarkers that track change, and immune signatures…

FIND-AF 2.0 used routinely coded health data to identify a substantially higher-yield population for atrial fibrillation screening. In prospective testing, atrial fibrillation was detected in 4.5% of high-risk…

A radiograph-based deep learning model distinguished benign from potentially malignant bone tumors with high accuracy across external and prospective testing. Used as a second reader, it also improved…

A deep-learning biomarker derived from routine postoperative histology identified which patients with resected pancreatic ductal adenocarcinoma were more likely to benefit from gemcitabine or modified FOLFIRINOX. External validation…

MerMED-FM shows that one vision foundation model can interpret seven medical imaging modalities while remaining competitive with specialist models. Its clearest advantage emerged when labelled data were scarce:…

Recent Digital Health & Informatics coverage shows models becoming integrated clinical infrastructure: learning from biological signals, combining modalities, tracking disease remotely, and confronting transportability across sites.

Cardiac AI is moving beyond isolated classifiers toward reusable representations, decision-focused imaging, and lower-barrier assessment. The central question is shifting from model accuracy to where these tools improve…

PRISM2 combines whole-slide pathology with clinical dialogue, enabling prompt-based diagnostic inference without task-specific training. In retrospective testing, it matched or surpassed three specialized cancer-detection products while producing representations…