AI in healthcare research — a few things from today’s reading — 20 September 2026
Today's LLM papers centered on documentation and diagnostic support rather than new model releases. A blinded comparison of 99 emergency department encounters found LLM-generated one-liner summaries preferred over physician-authored ones 50.5% of the time versus 38.4% for physicians. A similar prompt-engineered LLM assistant for hepatology referral triage cut clinician review time by about 60% while keeping hallucination rates low. On the diagnostic side, ChatGPT-4o and GPT-5 reached 77.1% and 57.6% partial concordance respectively against true diagnoses in 33 pediatric tumor pathology cases, and a four-version ChatGPT comparison on 50 complex oral and maxillofacial cases was benchmarked directly against experienced dentists. A separate study of general-use AI chatbots classifying impacted mandibular third molars found none reached acceptable agreement with expert consensus, a useful reminder that diagnostic imitation and diagnostic reliability are not the same thing.
On the ML/DL side, imaging models again showed strong discrimination across a range of conditions. A deep learning-enabled single breath-hold abbreviated MRI protocol for hepatocellular carcinoma matched conventional full MRI (sensitivity 77.9-88.7% vs. 84.4-92.5%) while cutting scan time by more than 15 minutes and removing the need for contrast. A CNN-based diagnostic model for laryngoscopy images, trained and externally validated across five countries, reached AUC 0.85-0.90 and outperformed both general practitioners and ChatGPT-4o. A CNN using dynamic cervical radiographs predicted spinal cord compression (AUC 0.888) as a lower-cost alternative to MRI, and a multi-signal model combining CT deep learning, radiomics, and blood immune profiling raised lung cancer diagnosis AUC from 0.69-0.73 (individual signals) to 0.81 combined. On the risk-prediction front, a random forest model beat standard clinical scores for ulcerative colitis readmission, and a SHAP-interpretable analysis of 12,041 type 2 diabetes patients showed readmission risk factors shift over time — age matters most at 30 days, inflammatory markers and length of stay at 60 days, chronic complications by 365 days.
Today's findings lean toward incremental validation work: LLMs proving useful for documentation and triage support, and imaging/ML models continuing to post strong but not universally reliable diagnostic accuracy.
Sources
- LLM-generated ED clinical summaries, blinded comparison
- LLM assistant for hepatology referral summarization (60% faster triage)
- ChatGPT-4o vs GPT-5 diagnostic concordance in pediatric tumor pathology
- Chatbots vs experts classifying impacted mandibular third molars
- DL abbreviated single breath-hold MRI protocol for hepatocellular carcinoma
- Multicenter AI diagnostic model for laryngoscopy images
- CT deep learning + radiomics + blood immune profiling for lung cancer diagnosis (AUC 0.81)
- SHAP-interpretable ML model for time-varying type 2 diabetes readmission risk
Prepared with AI assistance from newly published research. See the sources for the original studies.
