AI in Healthcare: Research Roundup — October 1, 2026

On the LLM side, today's additions center on clinical decision support and triage rather than diagnosis. ChatGPT (GPT-4) showed considerable-to-almost-perfect agreement with emergency physicians managing 792 traumatic laceration cases (82.6%-91.7% accuracy), and GPT-4o feedback on clinical note quality was found non-inferior to physician feedback in a discharge-planning documentation study. A postoperative decision-support pilot in a Gambian neurosurgery ward improved management-plan quality in 60% of 75 LLM-assisted interactions, with 14 care changes attributed to the tool. Triage work was more cautionary: a new benchmark (CARE-Bench) found prompted models preserved the correct "ask for clarification" step only 33.5% of the time, and a 15-chatbot psychiatric triage study found average accuracy of 42.0%-71.8% with a net over-triage bias. A related paper argued emergency-triage LLMs need explicit decision objectives, not just better predictions, after a consumer system under-triaged over half of gold-standard emergencies.

On the ML/DL side, risk-prediction models again showed strong, disease-specific performance. LightGBM reached an AUC of 0.966 predicting multidrug-resistant infection in ICU COPD patients, and a benchmark of 10 models across eight hospital datasets found gradient-boosting methods (especially CatBoost) and the TabPFN tabular foundation model most consistently robust. A 2.5D deep learning model using multiparametric MRI predicted prostate cancer biochemical recurrence with AUCs of 0.938 (internal) and 0.935 (external validation), and a dual-task deep learning model for head-and-neck cancer radiotherapy achieved C-indices of 0.92 and 0.74 for metastasis and recurrence risk. Separately, a CNN-XGBoost hybrid improved demographic fairness in chest X-ray diagnosis without sacrificing accuracy.

Together, today's findings suggest LLM decision-support tools are approaching clinician-level agreement in narrow, well-defined tasks, while triage remains an open problem; ML/DL risk models continue to post high discrimination (AUCs above 0.93) on disease-specific prediction.

Sources

1. ChatGPT agreement with emergency physicians on laceration management
2. GPT-4 quality-improvement feedback for clinical notes
3. LLM postoperative decision support in a Gambian neurosurgery ward
4. CARE-Bench: benchmarking patient-facing LLM triage
5. One-shot emergency psychiatric triage across 15 AI chatbots
6. High-stakes decision-making with language models in emergency triage
7. Machine learning risk model for multidrug resistance in ICU COPD patients
8. Benchmark of 10 ML models for clinical risk prediction
9. 2.5D deep learning model for prostate cancer recurrence
10. Dual-task deep learning model for head-and-neck cancer radiotherapy outcomes

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