We present ResidencyRL, a reinforcement learning method for training clinical AI agents through simulated multi-turn clinical encounters. On held-out evaluations, the agent improved diagnostic accuracy by 7.0% under adversarial conditions and reduced missed red flag rates by 31%

We introduce BMC-LongCLIP, a biomedical vision-language embedding model with extended text capacity (up to 512 tokens) trained on BIOMEDICA-LongCAP, a dataset of 1M context-rich image–caption pairs. The model reduces token waste from 55 % to 2.2 % and achieves +30 % Recall@1 and faster convergence, highlighting the promise of long-context modeling for biomedical vision-language models.

We introduce BIOMEDICA, an open-source framework that transforms the PubMed Central Open Access subset into a comprehensive dataset of over 24 million image-text pairs with expert-guided annotations, enabling state-of-the-art performance in biomedical vision-language models across diverse tasks and domains.

An overview on AI advancements in medicine, including scientific discovery, clinical applications, patient engagement, and ethical considerations.

Can LLMs match the conclusions of systematic reviews written by clinical experts when given access to the same studies? To explore this question, we present MedEvidence, A human-curated benchmark of 284 questions (from 100 open-access SRs) across 10 medical specialties.

Functions to summarize DNA methylation data using regional principal components. Regional principal components are computed using principal components analysis within genomic regions to summarize the variability in methylation levels across CpGs.

We developed a computer vision model to analyze laparoscopic surgery videos, identifying fine-grained surgical actions linked to operative blood loss and surgeon experience.

Molecular data from mass spectrometry were used to build classifiers, achieving high diagnostic accuracy when compared to pathology results, highlighting its potential for real-time surgical guidance.

Cross-validation (CV) can underestimate test error variance due to correlated error estimates from using the same samples for training and testing. Nested CV mitigates this issue by providing more accurate coverage through improved error variance estimation, which this work extends to the Cox proportional hazards model.

We introduce game theoretic centrality, which integrates biological network knowledge with Shapley value from coalitional game theory to prioritize disease-associated genes. Applied to autism spectrum disorder (ASD), the approach identifies biologically relevant genes, demonstrating potential regulatory interactions and offering insights into the genetic basis of complex disorders.