AI in healthcare research — a few things from today’s reading — 13 September 2026

Today's LLM findings clustered around clinical decision support and diagnostic reasoning. ChatGPT-4o with retrieval-augmented generation outperformed 290 practicing neurologists on guideline adherence for MS and NMOSD management (80.5% vs. lower rates), and a small open-source model, Qwen2.5-14B with RAG, approached GPT-4-turbo's performance on 98 breast cancer guideline questions. GPT-4o showed moderate appropriateness for routine cataract lens selection but struggled on complex cases, while three LLMs (GPT-4o, Grok-3, DeepSeek) showed mixed but promising results interpreting dental caries guidelines. On the research side, RaDaR, a 32B-parameter open-source reasoning model trained on real and synthetic rare-disease cases, beat the much larger 671B DeepSeek-R1 and flagged the correct rare-disease diagnosis a median 1.87 months before clinical suspicion in a retrospective cohort.

ML/DL diagnostic and prognostic models kept posting strong, narrow performance. A LightGBM model distinguishing Kawasaki disease from febrile mimics in 4,469 children hit an AUC of 0.979 (93.4% accuracy). An explainable model combining inflammatory markers and traditional risk factors diagnosed coronary artery disease in 4,656 angiography patients (AUC 0.815), and a deep learning mammography risk model outperformed breast density at stratifying future breast cancer risk across 123,091 screening exams. In oncology prognosis, a multimodal transformer fusing pathology, ultrasound, and clinical data reached a C-index of 0.840 for recurrence risk in 768 HR+/HER2- breast cancer patients, while deep learning-derived MRI subtypes improved 5-year recurrence-free survival prediction in rectal cancer (AUC 0.799 to 0.834).

Across both areas, general-purpose LLMs are increasingly being benchmarked directly against clinicians on guideline-based decisions, while disease-specific ML/DL models continue to deliver high-discrimination results on narrower diagnostic and prognostic tasks.

Sources

1. ChatGPT-4o with RAG vs. 290 neurologists on MS/NMOSD guideline adherence
2. Small open-source LLMs with RAG vs. GPT-4 for breast cancer clinical decision support
3. RaDaR: 32B reasoning LLM for rare disease diagnosis, randomized physician-assistance trial
4. LightGBM diagnostic model for Kawasaki disease (AUC 0.979, n=4,469)
5. Explainable ML model for coronary artery disease diagnosis in non-AMI patients
6. Deep learning breast cancer risk model vs. breast density for supplemental screening
7. Multimodal deep learning recurrence risk model for HR+/HER2- breast cancer (C-index 0.840)
8. Deep learning MRI radiomics subtypes and recurrence risk in rectal cancer

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