AI in Healthcare: Research Roundup — September 23, 2026

On the LLM side, several studies this week focused on diagnostic reasoning. GPT-4 evaluated against cardiologists at a Sub-Saharan Africa hospital identified the correct diagnosis with certainty in 64% of 50 cases, rising to 82% when counted among the top three hypotheses. A separate study tracking six successive GPT models on sleep medicine cases found diagnostic accuracy climbing from 74.4% (GPT-3.5 Turbo) to 93.6% (GPT-o3), with board-exam knowledge scores rising from 56.9% to 93.0% for GPT-5, though gains appear to be plateauing.

On safety and deployment, a risk-stratified evaluation of Claude Sonnet 4.0 across 200 consultation scenarios reported 78.3% diagnostic accuracy and 88.7% clinical safety, with acute-care scenarios flagged as needing more oversight than routine care. Applied tools also moved forward: a LangChain/GPT-3.5 chatbot cut pediatric EHR search time by 40.6% and raised retrieval accuracy (F1) from 0.71 to 0.89, while CLIN-LLM, a BioBERT-based diagnosis-and-treatment framework, reached 98% classification accuracy and cut unsafe antibiotic suggestions by 67% versus GPT-5.

On the ML/DL prognostic and diagnostic side, a meta-analysis of 69 studies confirmed XGBoost-family gradient boosting as the most consistent performer for post-MI mortality prediction, and a similar XGBoost model reached AUC 0.954 for ICU sepsis mortality. An explainable LightGBM model for COVID-19 mortality in older adults identified SaO2 (85%) and CRP (100 mg/L) as clinically meaningful risk thresholds. In imaging, deep learning models reached AUC 0.986 for proptosis detection on orbital MRI and improved differentiation of liver cancer subtypes on contrast-enhanced ultrasound.

Across both threads, today's studies point toward the same pattern: LLM performance keeps improving on diagnostic and reasoning tasks, while ML/DL prognostic models continue to deliver strong, well-validated accuracy on narrower, single-disease prediction problems.

Sources

1. GPT-4 diagnostic accuracy in cardiology (Sub-Saharan Africa)
2. Successive GPT models on sleep medicine cases and board questions
3. Risk-stratified safety assessment of Claude Sonnet 4.0 in medical consultations
4. LangChain/GPT-3.5 chatbot for pediatric EHR accessibility
5. CLIN-LLM safety-constrained diagnosis and treatment framework
6. XGBoost meta-analysis for post-MI mortality prediction
7. XGBoost model for ICU sepsis mortality (AUC 0.954)
8. Explainable LightGBM model for COVID-19 mortality in older adults

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