Radiology

LLM-assisted decision support in Kenyan primary care was safe but not clearly effective.

In 16 Kenyan primary care facilities, clinicians using an EMR with LLM assistance did not have fewer 14-day treatment failures. Treatment failure occurred in 2.2% of the intervention arm versus 2.0% of control (adjusted OR 0.77, 95% CI 0.55–1.08, P=0.13). No serious adverse events were attributed to the intervention and independent review found no safety signal. The trial supports short-term safety but suggests any clinical benefit of LLM assistance is small or context dependent. 1

Deep learning detects periapical radiolucent lesions with high pooled accuracy but heterogeneity is substantial.

Systematic review included 30 studies; six entered meta-analysis for PRL detection on panoramic radiographs. Pooled sensitivity was 0.80, specificity 0.98, and AUROC 0.93, but between-study heterogeneity was high. Over half of included studies had risk-of-bias or applicability concerns by QUADAS-2. Certainty of evidence was rated moderate, so external validation on diverse datasets is needed before clinical roll-out. 2

Right-ventricular imaging metrics have promise as trial end points but need standardization and validation.

RV structure and function measures from echo, CMR, CT, and catheterization can increase trial sensitivity across many diseases. Use is inconsistent and multicenter feasibility is limited by variable acquisition and incomplete data capture. Key gaps include lack of standardized protocols and insufficient validation linking short-term RV changes to long-term outcomes. Centralized analysis, agreed acquisition standards, and demonstration of clinically meaningful thresholds are essential next steps. 3