On the prognostic and diagnostic modeling side, a new arXiv paper proposes bootstrapping-based regularization to reduce how much individual risk predictions shift between model refits, cutting instability by roughly two-thirds on the GUSTO-I and Framingham datasets while preserving discrimination. In applied work, a logistic regression model for acute kidney injury after cardiac arrest reached an AUC of 0.825 on external validation using MIMIC-IV data, and a meta-analysis pooling 68 studies on machine learning for poststroke mortality reported a pooled C-index of 0.727, alongside notable variability across individual models.
On the LLM side, a review in APL Bioengineering surveys ChatGPT's potential roles in precision medicine, from genetic sequence analysis to clinical trial recruitment support. A hybrid Bayesian-neural-LLM framework, DiagnoLLM, reached 88.0% accuracy on Alzheimer's disease detection while generating interpretable diagnostic reports. A randomized trial of 180 hospitalized patients found AI voice-cloned health education outperformed standard education on compliance and satisfaction, with content retention of 92.5% using patients' own cloned voice versus 86.7% for a physician's. A feasibility study of an LLM chatbot answering parental questions in the pediatric ICU found 99.3% of 1,225 generated sentences were accurate. Two new benchmarks target radiology documentation: one for longitudinal lung-finding timelines, another (PET-F2I) showing a fine-tuned 7B model outperforming 27 larger LLMs on PET/CT impression generation.
On agentic AI, a Johns Hopkins team's EligMeta framework narrowed 4,044 candidate gastric cancer trials to 39 relevant studies, recovering all 13 guideline-cited trials. DeepEvidence applies orchestrated agents to biomedical knowledge graphs across the discovery-to-evidence pipeline, while a separate paper proposes a governance framework for managing growing fleets of clinical AI agents. A CUHK/Stanford/Microsoft team's "Medical AI Scientist" reported higher first-run success rates than GPT-5 and Gemini-2.5-Pro on 57 self-reported research tasks.
Across all three areas, the throughline is the same: quantitative validation and governance tooling are maturing in parallel with capability claims, even as most agentic results remain self-reported and awaiting independent verification.
Sources
- Bootstrapping-based Regularisation for Reducing Individual Prediction Instability in Clinical Risk Prediction Models
- A machine learning model for prediction of cardiac arrest-associated acute kidney injury in the ICU
- Predictive Value of Machine Learning for Poststroke Mortality Risk: Systematic Review and Meta-Analysis
- DiagnoLLM: A Hybrid Bayesian Neural Language Framework for Interpretable Disease Diagnosis
- Effectiveness of AI-Assisted Patient Health Education Using Voice Cloning and ChatGPT
- PET-F2I: A Comprehensive Benchmark and Parameter-Efficient Fine-Tuning of LLMs for PET/CT Report Impression Generation
- Eligibility-Aware Evidence Synthesis: An Agentic Framework for Clinical Trial Meta-Analysis (EligMeta)
- Towards a Medical AI Scientist
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.
