AI in healthcare research — a few things from today’s reading — 10 September 2026

Today's LLM papers cluster around emergency and urgent-care triage. A seven-month comparative study found an LLM-based ED triage architecture outperforming NLP and joint-embedding models and current nurse triage (F1 0.900, AUC-ROC 0.879), though the authors flagged severe overfitting and called for external validation before deployment. In trauma, GPT-5.2, Gemini 3.0 Flash, and Claude 4.5 Opus showed strong overall discrimination on 150 vignettes but undertriaged geriatric and blunt-trauma cases disproportionately, and 11.2% of repeated prompts shifted triage category under minor wording changes. A cardiology-specific model, TriageMaster-70B, processed acute coronary syndrome triage 39% faster than cardiologist review across over 16,000 retrospective cases, while a broader meta-analysis of LLMs for ED triage found only moderate accuracy and limited sensitivity for the highest-acuity cases — a reminder that speed and headline accuracy don't always mean the models are ready to act alone.

On the prediction side, readmission and mortality modeling was the theme. An XGBoost model predicted 12-month heart failure readmission with AUC 0.861 in a 14,131-patient multicenter ED trial, and a separate SHAP-interpreted Boruta-XGBoost model reached AUC 0.873 for readmission or death in elderly heart failure patients during the vulnerable post-discharge phase. A multimodal deep learning model combining CT imaging, biomarkers, and clinical data improved 5-year mortality risk stratification in high-risk stage II colorectal cancer, and a separate multimodal model for non-small cell lung cancer reached a C-index of 0.71 for 5-year prognosis while also identifying novel biomarkers. Not every model held up equally well: a pediatric ICU readmission model saw AUC drop from 0.70-0.73 internally to as low as 0.60 on external validation, and an Australian cardiac-surgery registry model showed only limited predictive power, underscoring how much generalizability still varies by task and dataset.

Across both threads, today's papers show narrow, well-specified models continuing to post strong discrimination, while triage- and deployment-facing LLM systems are being stress-tested harder for consistency, not just accuracy.

Sources

1. LLM-based ED triage architecture vs. NLP/JEPA models (F1 0.900)
2. Prompt fragility and undertriage risk in frontier LLMs for trauma triage
3. TriageMaster-70B for acute coronary syndrome triage
4. Meta-analysis of LLM diagnostic accuracy for ED triage
5. XGBoost model for 12-month heart failure readmission (AUC 0.861)
6. SHAP-interpreted model for elderly heart failure readmission/death (AUC 0.873)
7. Multimodal deep learning model for colorectal cancer 5-year mortality
8. Multimodal deep learning for NSCLC survival and biomarker discovery (C-index 0.71)
9. Pediatric ICU readmission model, internal vs. external validation

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