AI in healthcare research — a few things from today’s reading — 17 September 2026

Today's LLM papers focused mainly on diagnostic accuracy in imaging and specialist domains, plus documentation support. On orofacial pain, a comparison of four chatbots (ChatGPT 5.5, Claude Opus 4.1, Gemini 3 Flash, Grok 4) across 30 ICOP-based scenarios found Grok 4 scoring highest overall, though the other three did not differ significantly, and subclassification accuracy trailed primary diagnosis accuracy across the board. In neurosurgery, ChatGPT-4 led a five-model comparison on 148 clinical vignettes with 74% diagnostic accuracy, ahead of AtlasGPT (63%), ChatGPT-3.5 (53%), Copilot (48%), and Gemini (36%). Two brain tumor studies showed a similar pattern: GPT-4o reached 76% final-diagnosis and 83% differential-diagnosis accuracy when given structured findings instead of images alone, and LLM assistance lifted human readers' top-three differential accuracy (radiologists 80.7% to 90.2%, trainees 62.5% to 81.1%). On the workflow side, an LLM-based inpatient documentation tool scored 4.68/5 across quality domains from clinicians at three hospitals, and a simulated LLM-CDS platform improved tumor-board summary efficiency for 26 oncologists working synthetic breast cancer cases.

On the ML/DL side, several prognostic and diagnostic models targeted specific cancers and acute conditions. An XGBoost model for hepatocellular carcinoma showed robust cross-center generalizability for predicting early recurrence after hepatectomy, while random forest-based ensembles were identified as the strongest approach for uncovering breast cancer prognostic factors, and a SEER-database model added population-level risk stratification for liver cancer. In acute care, an interpretable CatBoost model built on 85,886 pediatric pneumonia cases flagged chloride and glucose as novel biomarkers for simultaneous severe-pneumonia diagnosis and progression prediction, and a deep learning ensemble reached AUC 0.9388 (87.29% accuracy) classifying mpox from skin lesion images. A separate comparison of five lightweight deep learning architectures (CNN, LSTM, MLP, CNN-MLP, ConvLSTM) benchmarked ECG-based cardiovascular disease classification for accuracy against computational cost, and a new online tool targets one-year mortality risk in 601 maintenance hemodialysis patients.

Across both threads, today's work leaned toward models built for deployment — web tools, calibrated risk scores, and documentation aids — rather than isolated benchmark exercises.

Sources

1. Four LLMs compared on ICOP-based orofacial pain diagnosis
2. ChatGPT-4 leads five-model comparison on neurosurgery vignettes
3. GPT-4o feasibility study for brain tumor diagnosis
4. LLM assistance improves radiologist/trainee brain tumor diagnostic accuracy
5. LLM-based inpatient documentation summarization quality assessment
6. Simulated LLM clinical decision support platform in oncology
7. XGBoost explainable model for HCC recurrence after hepatectomy
8. Interpretable CatBoost model for pediatric severe pneumonia risk
9. Deep learning ensemble for mpox image classification (AUC 0.94)
10. Deep learning architecture comparison for ECG-based cardiovascular diagnosis

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