AI in Healthcare: Research Roundup — October 8, 2026
Today's LLM additions center on documentation and communication tasks already moving into live clinical use. In a Dutch academic hospital, blinded clinicians rated 292 paired discharge summaries and found LLM-generated versions more concise than physician-written ones (median 5.00 vs 4.50, p<0.001) but somewhat less complete (4.50 vs 5.00, p<0.001), with no significant difference in correctness. A Duke University crossover study of 553 patients found GPT-4-translated discharge summaries improved comprehension across four conditions (heart failure, pneumonia, DKA, stroke), with the largest gains among patients with limited health literacy. On the evaluation side, a 16-model benchmark (GPT-o1, Gemini 2.5 Pro, Claude 3.7 Sonnet, and others) tested whether LLMs would catch and correct a deliberately incorrect physician diagnosis across 200 vignettes of commonly misdiagnosed conditions, introducing a new "diagnostic correction rate" metric. A JMIR meta-analysis of four studies (2,604 evaluations) found LLM-assisted communication raised perceived empathy substantially (SMD 1.02), though the authors note more randomized trials are needed.
On the ML/DL side, prognostic and risk models continued targeting specific disease populations. An ensemble XGBoost model using PROFUND frailty components for 30-day readmission in multimorbid patients reached only modest discrimination (AUC 0.59) versus standard LACE (0.52) and HOSPITAL (0.54) scores, while a SHAP-interpretable stacking ensemble for 30-day readmission in elderly type 2 diabetes/heart failure patients performed much better (AUC 0.867, n=870). A multicenter CONFIDENT-HFpEF model reached a C-index of 0.72 for mortality in external validation, outperforming the established PREDICT-HFpEF score. A deep learning model for acute pancreatitis severity from CT, pretrained on over 10,000 studies, hit an internal AUROC of 0.888, and a random survival forest model for transthyretin amyloid cardiomyopathy (n=850, two European registries) outperformed existing Mayo Clinic and National Amyloidosis Centre staging systems for predicting mortality and HF hospitalization.
Across both tracks, today's additions reflect models increasingly benchmarked against established clinical scores rather than evaluated in isolation.
Sources
- AI-generated vs. physician-written discharge summaries, Dutch academic hospital
- GPT-4-translated discharge summaries improve patient comprehension
- Benchmark of LLMs correcting diagnostic errors
- LLMs in physician-patient communication, systematic review and meta-analysis
- PROFUND-based machine learning for 30-day readmission in multimorbid patients
- CONFIDENT-HFpEF machine learning risk stratification
- Deep learning model for acute pancreatitis severity from CT
- Machine learning risk prediction for transthyretin amyloid cardiomyopathy
Prepared with AI assistance from newly published research. See the sources for the original studies.
