AI in Healthcare: Research Roundup — September 28, 2026

On the LLM side, today's additions span clinical diagnosis, evaluation, and documentation. A randomized trial in nephrology had 97 clinicians from 9 countries complete 556 complex, open-ended case evaluations and found LLM-supported workflows improved diagnostic accuracy while AI-induced errors were rare. In clinical neuroanatomy, ChatGPT-4 showed stable diagnostic performance and precise anatomical terminology across 20 novel cases, while Gemini 2.5 was more sensitive to prompt phrasing. ClinicalGPT-R1, a reasoning-enhanced diagnostic LLM trained on 20,000 clinical records, outperformed GPT-4o on Chinese diagnostic tasks and matched GPT-4 in English. A four-way comparison of ChatGPT, Gemini, Grok, and Copilot on multiple sclerosis questions found Gemini scored highest for relevance and transparency, with Copilot trailing. On documentation, a functional evaluation framework (DistillNote) found LLM summaries of clinical notes compressed 20x still retained 97% of the diagnostic signal (AUROC 0.92 vs. 0.94) for heart failure detection, while a qualitative pilot study of LLM-assisted SOAP notes in pediatric rehabilitation found friction came from workflow mismatch and clinician trust, not language quality. A new benchmark (BRIE) testing nine LLMs on EHR information retrieval found frequent omissions of clinically important information, especially when synthesis across multiple documents was required.

On the ML/DL side, risk and diagnostic models continued to post strong discrimination on disease-specific tasks. An AutoML model for trauma patients with hemorrhagic shock accurately predicted in-hospital mortality using GCS, ISS, lactate, and fibrinogen as top predictors. A lightweight LightGBM model for breast cancer risk, trained on routine lab panels from 13,285 patients, reached an internal AUC of 0.987 (external AUC 0.83) while running in under 6 seconds. A Vision Transformer outperformed CNN architectures for patient-level classification of acute pancreatitis on CT (89.19% accuracy, AUC 0.946), and a hierarchical multiscale fusion network combining CT and clinical variables achieved high precision for preoperative esophageal cancer T-staging across three centers. A systematic review of 26 studies found AI models for MASLD risk stratification consistently reached AUROC 0.72-0.95, outperforming conventional risk scores. Elsewhere, a knowledge-augmented multimodal EHR model combining an LLM with structured data reached AUCs of 0.84 and 0.92 for 30-day readmission and in-hospital mortality, respectively.

Taken together, today's findings reinforce a consistent pattern: LLMs are increasingly tested in realistic, open-ended clinical scenarios rather than multiple-choice benchmarks, with mixed but generally positive results, while ML/DL risk and diagnostic models continue to deliver strong, disease-specific discrimination that still needs broader external validation.

Sources

1. Randomized controlled trial of LLM-assisted diagnostic accuracy in nephrology
2. Evaluation of large language models in clinical neuroanatomy
3. ClinicalGPT-R1: reasoning-enhanced generalist disease diagnosis LLM
4. DistillNote: functional evaluation of LLM-generated clinical note summaries
5. A living benchmark (BRIE) for information retrieval from electronic health records
6. AutoML prognostic model for traumatic multiple fractures and hemorrhagic shock
7. Lightweight LightGBM model for breast cancer risk prediction
8. Vision Transformer framework for acute pancreatitis classification on CT
9. AI-based models for risk prediction in MASLD (systematic review)

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