Prospective and Interventional Clinical Evidence in Medical AI Research

Artificial intelligence (AI) is rapidly moving from retrospective benchmarking studies into clinical workflows. However, much of the literature still focuses on model accuracy as the primary endpoint. This Collection focuses on evaluating clinical AI beyond accuracy by examining the pathway from model output to clinical action and ultimately to patient and health-system outcomes. In particular, a model paired with different actions can actually produce different results. However, the importance of downstream action exploration is still not widely recognized. We particularly welcome studies built around AI–action bundles, the combination of an AI result and the predefined action expected from clinicians, patients, or care teams. Studies focused only on model development, benchmark performance, external validation, fairness, or implementation experience without a defined downstream action are outside the core scope of this Collection. Across all submissions, the causal pathway from model output to action to outcome should be explicit.

This collection seeks research that evaluates the full pathway from model output to clinical action and health outcomes in the following areas:

• Prospective pragmatic trials, cluster-randomized trials, and stepped-wedge designs that evaluate AI-enabled screening, diagnosis, prognosis, triage, monitoring, and treatment.
• Research on integrating generative AI into the clinical workflow, including referral pathways, and the design of human-AI interaction
• High-quality retrospective causal analyses and target trial emulations that justify specific downstream actions.
• Robust governance and surveillance strategies that monitor real-world effectiveness, model drift, safety, retraining, and recalibration of AI-action bundles.
• Health technology assessments, cost-effectiveness, and economic evaluations of AI-action bundles.
• Reporting, governance, and lifecycle frameworks that support clinically meaningful evaluation beyond accuracy alone.