Toward trustworthy healthcare AI: designing academic research for translation readiness.

Abstract

In controlled proof-of-concept studies, healthcare artificial intelligence (AI) systems now routinely match or exceed specialist-level performance. Despite these impressive technical achievements, most systems remain confined to research settings, highlighting an opportunity to better prepare academic AI research for subsequent clinical translation. Successful translation requires attention to technical, regulatory, organizational, and infrastructural factors, ideally from the earliest stages of research. This Perspective focuses on trustworthiness as a key factor that academic researchers can address proactively. First, we outline how trustworthiness requirements differ across three stakeholder groups-patients, clinicians, and regulatory bodies-and how academic research can better align with these expectations. We then discuss key dimensions of trustworthy AI-explainability, uncertainty quantification, and evaluation-and how researchers can address them during study design. We also consider foundation models, which require adapted validation approaches. Based on this analysis, we propose a stakeholder-centered framework that emphasizes proactive integration of trust requirements during early development phases rather than post hoc considerations.