In machine-learning and diagnostic/prognostic models, several new studies targeted risk stratification for chronic conditions. A random survival forest trained on 850 patients with transthyretin amyloid cardiomyopathy reached concordance indices of 0.72–0.77, outperforming the long-standing Mayo and National Amyloidosis Centre staging systems by 2–10% in concordance and identifying a low-risk group with 85–91% three-year event-free survival. In chronic kidney disease, an ensemble of Random Forest, XGBoost, and LightGBM predicted renal function decline with an externally validated AUC of 0.89 against a Cox model's 0.82, while a separate transcriptomics-based model used gene co-expression analysis and LASSO regression to flag CKD progression genes, with validation in a mouse kidney-injury model. A random forest classifier built on 1,836 STEMI patients identified smoking, length of stay, hypertension, and ejection fraction as the top predictors of one-year unplanned readmission after PCI, and a new statistical-ML ensemble method extended the Super Learner approach to dynamic time-to-event prediction in primary biliary cholangitis.
In large language models applied to healthcare, new evidence continued to narrow the gap between LLMs and physicians on specific tasks. A medRxiv study comparing four LLMs (including GPT-4o and DeepSeek-R1) against seven physicians on 1,100 cases found no statistically significant difference in predicting 6-month mortality for post-ICU survivors, though physicians were more likely to flag their own uncertainty. A GPT-4-based pipeline for drafting hospital discharge letters within a secure EHR environment scored 89% complete and 78% accurate on clinical summaries in expert evaluation, and a separate synthetic-data pipeline (SynthMedic) generated discharge summaries rated 94.35% usable by human reviewers. A Mayo Clinic conference study tracked diagnostic accuracy across six successive GPT models on sleep-medicine vignettes, rising from 74.4% (GPT-3.5) to a plateau near 91–94% for GPT-5 and GPT-o3.
In agentic AI for medicine and health research, several new systems and surveys focused on drug discovery and autonomy limits. A review of 49 papers proposed a seven-dimensional taxonomy for evaluating healthcare AI agents and found roughly 59% of evaluation dimensions went unimplemented across the literature. A separate evaluation of six agentic drug-discovery frameworks across 15 task classes found they handle computational workflows like molecule representation well but show a substantial "reality gap" on messy, real-world pharmacology data. Two new multi-agent systems, Mozi (governed skill-graph orchestration) and AgentD (GPT-4o-based modular task execution via UniProt/ChEMBL/Semantic Scholar tools), targeted different stages of the drug-discovery pipeline, while a 300-reference survey proposed a three-level autonomy taxonomy for clinical agents and framed full self-evolving autonomy as still aspirational.
Across all three areas, today's studies point the same direction: AI systems are closing in on expert-level performance on narrowly defined tasks, but every one of them is paired with an explicit caveat about external validation, generalization, or real-world deployment still being unresolved.
Sources
- Machine learning-driven risk prediction for transthyretin amyloid cardiomyopathy
- Ensemble ML for renal function decline in chronic kidney disease
- Random forest model for one-year readmission after PCI in STEMI
- Comparing prognostic performance between LLMs and physicians
- Real-world LLM pipeline for clinical discharge summaries
- SynthMedic: synthetic discharge summary generation and validation
- Seven-dimensional taxonomy for agentic AI in healthcare
- Beyond SMILES: evaluating agentic systems for drug discovery
- Mozi: governed autonomy for drug discovery LLM agents
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.
