AI chatbots can appear supportive while progressively reinforcing psychiatric vulnerabilities. A clinically validated audit of 810 multi-turn conversations shows that these risks depend strongly on the user profile,…

Recent coverage points to a harder standard for medical AI: not whether it performs, but whether it changes decisions safely, targets benefit precisely, detects threats early, and distributes…

Oncology AI is moving closer to the decision point: selecting neoadjuvant therapy, guiding perioperative care, reading immune architecture, and measuring whether the host can withstand treatment.

VirTues turns heterogeneous spatial proteomics images into transferable representations spanning proteins, cells, niches, and whole tissues. Across thousands of patients, the foundation model supported virtual staining, cross-cohort annotation,…

Medical AI is moving beyond benchmark gains toward harder evidence: actionable phenotypes, measurable patient benefit, real-world validation, failure testing, and governance that determines who can safely use these…

Explainable AI did not affect all users equally: it improved dermatological decision-making overall, yet persuasive explanations made lay people substantially more vulnerable to incorrect AI advice. Expertise—and whether…

Psychiatry AI is moving beyond diagnostic labels toward behavioral biomarkers, computational phenotypes, and widely used chatbots. The central challenge is no longer signal detection alone, but clinical validity,…

A multimodal transformer trained on millions of longitudinal clinical records detected pan-cancer disease with an AUROC of 0.956 and anticipated cancer up to one year before diagnosis. The…

EchoAI-Peds detected 28 congenital defects, abnormalities, repairs, and interventions from complete pediatric echocardiography studies, achieving a macroaveraged AUROC of 0.89 in external validation. Its substantial advantage over adult…

Daily orthopaedic trauma operating-room demand may be substantially more predictable than conventional estimates suggest. In this cohort study, a locally calibrated time-series model reduced test-set error to 1.80…