AI in healthcare: today’s research roundup — 25 October 2026

Several new studies look at how LLM-based diagnostic agents compare with physicians. In a multi-agent framework tested across three benchmarks and seven LLMs, editing case findings counterfactually to probe competing diagnoses consistently improved accuracy over standard prompting, with the largest gains on ambiguous cases. A related agent, PULSE, matched senior endocrinology specialists and beat residents on 82 case reports, and unlike the physicians its accuracy held steady even for rarer diseases. Deep-DxSearch, trained end-to-end with reinforcement learning on literature and case retrieval, raised physicians' average diagnostic accuracy from 45.6% to 69.1% across 150 real cases. A pooled meta-analysis of 42 studies (101 effect sizes) found LLM assistance gives physicians a small but significant accuracy boost (Hedges' g=0.22), while a four-model comparison on the Japanese National Medical Examination found GPT-5, Grok-4, Claude Opus 4.1, and Gemini 2.5 Pro all scored 95.6%-97.2%, above the threshold considered textbook-level. A separate evaluation of GPT-4o-summarized discharge letters found high layout adherence (88%) but flagged medication details as the main weak point.

On the prognostic and diagnostic model side, a randomized multi-reader trial gave 12 surgical oncologists an AI prognostic tool for colorectal liver metastases and measured the change in their AUC for predicting 3-year mortality with versus without its help. A LightGBM model trained on a UAE PCI cohort and externally validated on MIMIC-IV targeted ischemic and bleeding risk after coronary intervention, while a new ensemble (LightGBM + CoxBoost) built from a Saudi diabetes registry of nearly 19,000 patients targets three-year diabetic kidney disease risk using literature-informed social and clinical variables. A sepsis cohort of over 16,000 MIMIC-IV patients produced a gradient-boosting delirium-risk model with a public online calculator. Not every model held up, though: a deep learning tool for predicting BI-RADS assessments on breast MRI underperformed radiologists in external validation (AUC 76.7% vs. 91.3%, with a notably larger specificity gap), underscoring that external validation keeps separating promising models from deployment-ready ones.

Across both threads, the pattern continues: LLM-based reasoning agents are closing in on expert-level diagnostic performance under controlled conditions, while traditional ML risk models keep delivering strong internal discrimination that doesn't always survive external validation.

Sources

1. Improving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning
2. Human-AI Co-reasoning for Clinical Diagnosis with Evidence-Integrated Language Agent (PULSE)
3. End-to-End Agentic RAG System Training for Traceable Diagnostic Reasoning (Deep-DxSearch)
4. The Effect of LLM Assistance on Diagnostic Accuracy: A Systematic Review and Meta-Analysis
5. Textbook-Level Medical Knowledge in Large Language Models: A Comparative Evaluation Using the Japanese National Medical Examination
6. Impact of an AI prognostic tool on clinician performance in colorectal liver metastases
7. Literature-informed ensemble machine learning for three-year diabetic kidney disease risk prediction (PSMMC NephraRisk)
8. Deep learning model for BI-RADS assessment prediction on breast MRI shows mixed results in external validation

Prepared with AI assistance from newly published research. See the sources for the original studies.