AI in healthcare research — a few things from today’s reading — 15 September 2026

Today's LLM findings leaned toward decision support and documentation rather than raw diagnosis. A retrospective cohort of 10,344 emergency department encounters found ambient AI scribes cut on-shift documentation time by 72.6 seconds per encounter, roughly 24 minutes over an 8-hour, 20-encounter shift, while a separate non-inferiority study found GPT-4o matched hematology residents on documentation quality (88.9 vs 89.6) while reducing time from 14.9 minutes to 40.1 seconds. On the decision-support side, results were mixed: ChatGPT-5 showed only fair concordance with gynecologic oncology tumor boards (kappa 0.258-0.334) and was fully consistent across repeated queries in just 37.2% of cases, and ChatGPT-4o in a urological tumour board matched final recommendations only 79.2% of the time, below the study's benchmark. A pilot study on previously misdiagnosed dermatology cases found Gemini 3 led multimodal diagnostic accuracy at 60.0%, ahead of ChatGPT-5.2 (56.7%) and Claude 4.5 Sonnet (33.3%), and a large-scale analysis of 620,699 public ChatGPT conversations found mental-health-related content prevalence swung 23-fold (0.21% to 4.90%) depending on definition alone.

ML/DL models continued to post strong, narrow discrimination on well-defined prediction tasks. An XGBoost model distinguishing malignant from benign pleural effusion reached AUC 0.988, and an explainable XGBoost model for cerebral small vessel disease achieved 88.0% accuracy with a companion nomogram. A gradient-boosting framework for in-hospital mortality in acute ischemic stroke hit external AUC 0.871 using high-granularity bedside data, and a similar interpretable GBM model for sepsis-induced myocardial injury mortality reached AUC 0.751 with a deployed web calculator. A 401-patient multicenter JASA study built a modernized prognostic score for spinal metastases using LASSO regression, addressing the fact that older scoring systems predate current oncologic therapies.

Across both threads, LLMs are increasingly being tested where they're meant to assist — documentation and second-opinion support — rather than replace clinical judgment, while purpose-built ML models keep delivering high discrimination on tightly scoped, single-outcome prediction problems.

Sources

1. Ambient AI scribes reduce ED documentation burden (10,344 encounters)
2. GPT-4o vs residents on medical history documentation quality and speed
3. ChatGPT-5 concordance with gynecologic oncology tumor boards
4. ChatGPT-4o as decision support in a urological tumour board (prospective)
5. Multimodal LLMs on previously misdiagnosed dermatology cases
6. Prevalence of mental health discussions in public ChatGPT conversations (JAMA Network Open)
7. XGBoost model for diagnosing malignant pleural effusion (AUC 0.988)
8. Interpretable ML framework for in-hospital mortality in acute ischemic stroke
9. Gradient boosting model for mortality in sepsis-induced myocardial injury
10. JASA multicenter machine learning-based prognostic scoring for spinal metastases

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