Today's new risk-prediction literature spans drug safety and dementia. A machine learning model combining structured trial data with a clinical language model (XGBoost plus ClinicalModernBERT) stratified dosing-error risk across more than 42,000 ClinicalTrials.gov entries, reaching an AUC-ROC of 0.862. In type 2 diabetes, a dynamic model (ML-CVD Primary) that updates cardiovascular risk using one-year changes in risk factors was published to support ongoing surveillance rather than a single static estimate. And in Alzheimer's disease, a deep-learning approach using a Bayesian variational autoencoder reported improved polygenic risk prediction when analyses were stratified by age and APOE genotype.
On the LLM side, three new evaluations compared model performance against clinicians. ChatGPT-4o and ChatGPT-3.5-mini reached 99% diagnostic accuracy on standardized emergency case presentations, versus 92% for a human emergency physician (p=0.039), though the comparison involved a single specialist and simulated cases. GPT-4o was also tested on predicting emergency department disposition (discharge, ward, or ICU) in complex respiratory cases, reaching 76.5% overall accuracy with high sensitivity (91.9%) but low specificity (20.8%). Separately, ChatGPT-5 outscored Gemini 2.5 on the official Polish diabetology specialty exam, 78.6% versus 68.4% (p=0.0455), with both models clearing the passing threshold.
Four new studies looked at agentic and multi-agent systems in medicine and research. A proof-of-concept multi-agent pipeline (Prompt-to-Pill) was proposed spanning drug discovery through clinical trial simulation, including virtual patient recruitment. A self-evolving multi-agent clinical decision support framework (VIBEMed) added a dedicated longitudinal-feedback agent and a safety sandbox constraining model updates. A notable cautionary finding came from a 2,400-case MIMIC-CDM study of abdominal pain: assembling the best-performing individual components into a "Best of Breed" system reached only 67.7% diagnostic accuracy, below the 77.4% of a top multi-agent system, despite scoring highest (85.5%) on the lab-interpretation sub-task alone. Finally, a conversational multi-agent tool (ChatDRex) was released to let non-programmers run disease-module and drug-repurposing analyses over a biomedical knowledge graph in plain language.
Taken together, today's studies point to a shared caution: strong performance on a narrow task or a single component rarely translates automatically into reliable performance across a full clinical or research workflow.
Sources
- Early Risk Stratification of Dosing Errors in Clinical Trials Using Machine Learning
- Improved polygenic risk prediction for Alzheimer's disease and related dementias using deep learning
- Artificial intelligence vs. emergency physicians: who diagnoses better?
- Comparative Performance of ChatGPT-5 and Gemini 2.5 on the Official Clinical Diabetology Specialty Examination
- Prompt-to-Pill: Multi-Agent Drug Discovery and Clinical Simulation Pipeline
- The Optimization Paradox in Clinical AI Multi-Agent Systems
- Conversational No-code, Multi-agentic Disease Module Identification and Drug Repurposing Prediction with ChatDRex
Prepared with AI assistance from newly published research. See the sources for the original studies.
Physician, epidemiologist and researcher. DPhil fellow in Clinical Medicine at the University of Oxford.
