AI in Healthcare: Research Roundup — October 17, 2026

Today's LLM additions leaned toward evaluation and evidence-mapping rather than new deployments. A JAMA Network Open study tested six commercial LLMs (including Claude Sonnet 4/4.6, Gemini 3 Pro, and GPT-5/GPT-5 mini) on 288 emergency department encounters to screen for missed diagnostic opportunities, finding true misses in 13.5% of cases and model AUCs ranging from 0.57 to 0.82 depending on the cohort, with physicians agreeing with 81.9% of flagged cases. A separate study evaluated DeepSeek-R1 on 7,435 Chinese emergency radiology reports, reaching 84.4% error detection with few-shot prompting versus 60.9% zero-shot against 12 board-certified radiologists. Three systematic/scoping reviews — covering LLMs in cardiology (33 studies), psychiatry, and mental-health prediction (201 studies on bias and clinical-utility reporting) — converged on the same message: performance is promising but inconsistent in high-stakes, acute scenarios, and clinical validation remains thin.

On the ML/DL side, an XGBoost model for chronic kidney disease progression trained on Chinese retirement-survey data held up well across two independent Western cohorts (AUC 0.867–0.892), outperforming the standard Kidney Failure Risk Equation (AUC 0.745). A cardiac surgery complication model built at a Korean academic hospital maintained its discrimination across two external hospitals, though precision metrics dropped outside the development site — a reminder that external validation often exposes calibration gaps rather than confirming them away. A deep learning model combining dual-modality ultrasound video with clinical data classified benign versus malignant superficial lymphadenopathy with AUCs of 0.91–0.95 across five centers and two external test sets. Rounding out the day, a conference-reported random forest model for sepsis prediction in aneurysmal subarachnoid hemorrhage dropped from AUC 0.915 internally to 0.833 on external eICU data, and a five-algorithm ML comparison improved prognostic prediction for NSTEMI patients across a five-hospital Chinese cohort.

Taken together, today's crop reinforces a now-familiar pattern: internal performance numbers look strong across both LLM and classical ML approaches, but the external and cross-population validation step continues to be where real-world confidence is won or lost.

Sources

1. Screening for Missed Opportunities for Diagnosis in the ED Using eTriggers and Large Language Models
2. Error Detection in Emergency Radiology Reports Using a Large Language Model (DeepSeek-R1)
3. Large Language Models in Cardiology: A Systematic Review
4. A Cross-Ethnicity Validated Machine Learning Model for the Progression of Chronic Kidney Disease
5. Machine Learning Model for Predicting Severe Immediate Postoperative Complications After Cardiac Surgery
6. Deep Learning Analysis of Dual-Modality US Videos for Characterization of Superficial Lymphadenopathy
7. Machine Learning Models for Sepsis Prediction in Aneurysmal Subarachnoid Hemorrhage (AAN 2026)

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