AI in healthcare research — a few things from today’s reading — 8 September 2026

A few studies worth noting from today's screening. A benchmark comparison found that general-purpose LLMs (GPT-5.2, Gemini 3.1 Pro, Claude Opus 4.6) performed better than specialized clinical AI tools, including UpToDate's own AI, on real physician queries. Separately, a comparison of four LLMs against 20 landmark NEJM trials found ChatGPT matched the published conclusions in every case tested.

Not every result favored LLMs working alone, though. A study on perioperative anticoagulation management found LLM performance modest on its own, with meaningful improvement only after pairing the model with structured clinical decision support.

On the diagnostic side, purpose-built ML/DL models continue to perform well in narrower, well-defined tasks: a deep learning model reached 96% positive predictive value diagnosing seven ocular surface diseases, an MRI-based classifier matched radiologist-level accuracy for Alzheimer's diagnosis, and a sepsis outcome model reached an AUC of 0.983.

Taken together, it's a reasonable snapshot of where things stand: LLMs are increasingly competitive on reasoning and communication-heavy tasks, while dedicated ML/DL models still lead on narrow, high-stakes diagnostic prediction.

Sources

1. General-purpose LLMs outperform specialized clinical AI tools
2. Benchmarking LLM agreement with NEJM trial conclusions
3. LLM + clinical decision support in anticoagulation management
4. Deep learning diagnostic model for ocular surface diseases
5. MRI-based Alzheimer's classifier vs. radiologists
6. Explainable ML for sepsis outcome prediction

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