Letter: Spillover Effects in Randomized Evaluations of Translational AI

The Perspective “Evaluating Translational AI: A Two-Way Moving Target Problem” aptly describes two fundamental challenges that occur when health systems evaluate predictive AI models that inform clinical decision-making. Leuchter et al. highlight how traditional evaluation approaches are rendered inadequate due to a two-way moving target problem caused by (1) concurrent-intervention confounding prior to implementation and (2) action-induced outcome bias that occurs after implementation. To address these challenges, they recommend short-term randomized controlled evaluations to assess model effectiveness prior to complete adoption. Although such randomized evaluations are warranted, they face a third challenge that could lead to misestimation of AI model effectiveness: spillover due to constraints on care delivery.