YISROEL MIRSKY

The deployment of clinical artificial intelligence (AI) systems lies at the core of their purpose—only then can they be used to reduce clinical labor and improve patient care. However, despite a growth in both research on and approved clinical AI systems, there still exists a big gap between the research and application of clinical AI systems. In this review, we discuss robust clinical AI from the perspective of AI model development as the intersection of four pillars—generalizability, explainability, uncertainty, and adversarial resistance. We discuss how their neglect has affected or can affect the success and performance of clinical AI systems, positing that the success of clinical AI systems cannot be expected without addressing these four key issues.