In machine-learning and diagnostic/prognostic models, several new studies focused on mortality and risk prediction at scale. A random forest model trained on a 7,239-patient global cirrhosis cohort reached an AUROC of 0.815 internally and 0.859 in an external validation of 28,670 US veterans, outperforming logistic regression and LASSO. A multinational validation of the AHA's PREVENT and Europe's SCORE2 cardiovascular risk equations across roughly 6.4 million people found both performed reliably worldwide, with adding albuminuria meaningfully improving PREVENT's accuracy in diabetes and chronic kidney disease. A radiomics-deep learning model combining VGG11 features with a random forest reached an external validation AUC of 0.779 predicting esophageal cancer response to neoadjuvant immunochemotherapy, a Cox-based model trained on 3,372 kidney cancer patients reached an iAUC of 0.81 for disease-free survival, and new SEER-based nomograms for early-stage triple-negative breast cancer reached C-indices above 0.87 in development but dropped to roughly 0.69–0.74 on external validation.
In large language models applied to healthcare, today's new evidence centered on benchmarking and documentation quality. Medmarks-style scrutiny continued with MLB, a new Chinese-language clinical benchmark whose top model scored 77.3% overall but dropped to 61.3% on patient-facing tasks, and PanCanBench, built from real pancreatic cancer patient questions, found LLM scores ranged from 46.5% to 82.3% with synthetic grading rubrics inflating scores by nearly 18 points. A two-phase study found GPT-4o reached 92.63% accuracy guiding patients to the correct outpatient department. On documentation, a blinded comparison of an LLM ambient scribe against physician notes found comparable quality (4.20 vs. 4.25 out of 5) but a higher hallucination rate (31% vs. 20%), while a fine-tuned Llama 3 8B model for pediatric occupational therapy notes improved clarity and completeness but did not meaningfully cut documentation time in a small pilot.
In agentic AI for medicine and health research, eight new systems targeted EHR navigation, imaging workflows, and clinical reasoning. AgentEHR's retrospective-summarization method improved performance up to 29.16% over baselines while cutting interaction errors by up to 92.3% on raw EHR decision-making tasks, and EHRNavigator, a multi-agent system splitting structured-query and note-retrieval work across agents, reached 86% accuracy on real-world patient-level questions. ClinSeekAgent showed that letting an agent actively seek its own multimodal evidence, rather than relying on pre-curated context, raised Claude Opus 4.6's F1 score from 47.5 to 62.6 on complex clinical reasoning cases. In imaging, AgentsEval proposed a multi-agent evaluator for radiology report quality, and MARL-Rad jointly trained region-specific and global agents with reinforcement learning to generate chest X-ray reports rated clinically comparable to ground truth. Rounding out the batch, Vivaldi used role-structured ED-team agents to explain physiological time series, Small Agent Group showed a cluster of small deliberating LLMs could outperform a single giant model at lower cost, and MARC v1 offered an open-source, YAML-configurable framework for traceable multi-agent clinical reasoning pipelines.
Across all three areas, the throughline is the same: new systems keep posting strong numbers on narrow, well-defined benchmarks, while the harder problems — patient-facing reliability, external validation across populations, and real-world deployment of multi-agent systems — remain the open frontier.
Sources
- Enhancement of inpatient mortality prognostication in cirrhosis with external validation
- Multinational validation of the PREVENT and SCORE2 cardiovascular risk equations
- UroPredict: machine learning model for kidney cancer recurrence prediction
- MLB: a scenario-driven benchmark for LLMs in clinical applications
- Assessing the quality of AI-generated clinical notes (ambient scribe evaluation)
- AgentEHR: autonomous clinical decision-making via retrospective summarization
- EHRNavigator: multi-agent system for patient-level clinical question answering
- ClinSeekAgent: automating multimodal evidence seeking for clinical reasoning
- MARL-Rad: multi-agent reinforcement learning for radiology report generation
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.
