In prognostic, diagnostic, and risk-prediction models, a meta-analysis of 40 studies (17,246 patients) found deep learning and radiomics detect extracranial carotid plaques with pooled sensitivity 0.88 and specificity 0.89, while a separate meta-analysis of axillary lymph node metastasis detection in breast cancer reported sensitivity/specificity of 0.79/0.86 for ultrasound-based models and 0.78/0.82 for MRI-based models. In a cohort of 11,361 tuberculosis patients, random forest and XGBoost reached validation AUCs of 0.960 and 0.954 for predicting antituberculosis drug-induced liver injury, with population-stratified models outperforming a single universal one. Elsewhere, a CT-CLIP foundation model improved lung cancer prognosis prediction over a clinical-variable-only baseline, and a GRU-based model generated time-updated risk scores for metastatic prostate cancer from routine visit data alone.
In large language models in healthcare, PrIME-LLM, tested across 21 LLMs including GPT-5 and Grok 4 on 29 standardized clinical vignettes, found models reached correct final diagnoses over 90% of the time with full case information but performed far worse at the earlier, open-ended stages of a case. GPT-5 reached 82.1% accuracy on TNM cancer staging from 552 PET-CT reports, edging out junior radiologists (77.0%). A randomized controlled trial found AI-generated previsit document summaries shortened neurosurgical consultations from a median of 33 to 25 minutes and left patients feeling more informed, though physicians rated their own autonomy somewhat lower.
In agentic AI for health research, a hierarchical multi-agent system trained with process-level rewards raised gene-disease validity curation accuracy to 0.750 and process-fidelity F1 to 0.520, both above outcome-only reward training. EmulatRx, a five-agent LLM framework for clinical trial design using real-world data, reproduced many previously reported treatment effects across historical trials spanning heart failure, septic shock, and other conditions. A separate study found an orchestrated multi-agent architecture held accuracy far better than a single agent as clinical workload scaled up (90.6% to 65.3% versus 73.1% to 16.6% accuracy), while cutting token usage up to 65-fold.
Across today's studies, stratifying by subgroup, supervising intermediate reasoning steps, and testing under realistic workload or case-complexity conditions prove more informative than single aggregate accuracy figures.
Sources
- Diagnostic performance of deep learning and radiomics in extracranial carotid plaque detection
- Accuracy of deep learning for detecting axillary lymph node metastasis in breast cancer
- Stratified machine learning models for predicting antituberculosis drug-induced liver injury
- Large Language Model Performance and Clinical Reasoning Tasks (PrIME-LLM)
- Evaluating LLMs for automated TNM staging from PET-CT reports
- Effect of previsit AI-generated document summaries on consultation time and communication (AI-INFOCARE RCT)
- Process-supervised multi-agent reinforcement learning for reliable clinical reasoning
- Empowering clinical trial design with agentic intelligence and real-world data (EmulatRx)
- Orchestrated multi-agents sustain accuracy under clinical-scale workloads
Prepared with AI assistance from newly published research. See the sources for the original studies.



