AI in Healthcare: Research Roundup — October 9, 2026

Today's LLM additions cluster around oncology decision support and specialized clinical benchmarking. A multicenter study ran 1,063 thyroid ultrasound reports (306 histopathology-confirmed) through ChatGPT-4o and DeepSeek-R1, finding DeepSeek-R1 more accurate and consistent at benign-versus-malignant classification, while the two models matched on management recommendations. A new 511-question benchmark (MSIC-Bench) tested GPT-4o, Gemini 2.5 Pro, and Claude Opus 4 on microsatellite instability cancer tasks and concluded the gap is domain knowledge, not reasoning ability. In a prospective head-and-neck tumor board study, two locally run open-weight models (gemma-3-12b, gpt-oss-20b) generated treatment recommendations for 17 real cases, scored by board members on a 6-point adequacy scale after the human decision was already made. Separately, a retrieval-augmented generation framework from Mass General Brigham targeted a quieter but high-volume problem: identifying dementia cases buried in unstructured EHR notes, where rule-based and keyword methods are known to miss clinical nuance.

On the ML/DL side, several groups reported a familiar pattern — strong internal performance that erodes externally. A chronic critical illness model for ICU patients with acute pancreatitis (MIMIC-IV/eICU development, Chinese single-center external validation) dropped from AUROC 0.85 internally to 0.73 externally. A breast MRI deep learning model for BI-RADS assessment, trained on 12,241 exams, underperformed radiologists in external validation (76.7% vs 91.3% AUC), and a gastric cancer endoscopy meta-analysis saw pooled sensitivity/specificity fall from 0.91/0.93 internally to 0.82/0.83 across just four external-validation studies. Kidney-metabolic risk models fared better: a China Renal Data System-derived XGBoost model predicting 5-year heart failure in CKD patients held up across both Chinese (AUC 0.879) and UK Biobank (0.851) external cohorts, and a SHAP-interpreted ensemble reached AUC 0.979 for mortality risk in 512 maintenance hemodialysis patients.

Together, today's batch underscores external validation as the recurring fault line between promising internal results and real-world reliability.

Sources

1. ChatGPT-4o vs DeepSeek-R1 on thyroid nodule ultrasound reports
2. MSIC-Bench: benchmarking LLMs on microsatellite instability cancers
3. Locally run LLMs augmenting head and neck tumor board recommendations
4. RAG LLM framework for dementia identification from EHRs
5. Machine learning for chronic critical illness in acute pancreatitis (MIMIC-IV/eICU, external validation)
6. Deep learning BI-RADS model vs radiologists, breast MRI external validation
7. Deep learning for early gastric cancer detection, meta-analysis
8. XGBoost model for heart failure risk in chronic kidney disease (China Renal Data System, UK Biobank validation)
9. SHAP-explained machine learning for mortality risk in maintenance hemodialysis patients

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