The Paradoxical Challenge of High-Value Medical Artificial Intelligence

AbstractMedical artificial intelligence (AI) is rapidly transitioning from proof of concept to clinical implementation. Although many medical AI advances represent incredible technological feats and break ground on new applications, feeding tremendous intellectual excitement, the greatest potential lies in those AI tools that provide high value to multiple medical stakeholders, including patients, providers, and payers. Indeed, it is possible that even high-performing tools could have little or even negative value to stakeholders. Paradoxically, many high-value applications may be dependent on training data that are expensive, difficult to obtain, or limited by substantial embedded biases. In this perspective article, the authors propose a framework for identifying potential high-value applications and the approaches to the attendant challenges of training data quality.