Learning to safely approve updates to machine learning algorithms

Machine learning algorithms in healthcare have the potential to continually learn from real-world data generated during healthcare delivery and adapt to dataset shifts. As such, regulatory bodies like the US FDA have begun discussions on how to autonomously approve modifications to algorithms. Current proposals evaluate algorithmic modifications via hypothesis testing and control a definition of online approval error that only applies if the data is stationary over time, which is unlikely in practice. To this end, we investigate designing approval policies for modifications to ML algorithms in the presence of distributional shifts. Our key observation is that the approval policy most efficient at identifying and approving beneficial modifications varies across problem settings. So, rather than selecting a fixed approval policy a priori, we propose learning the best approval policy by searching over a family of approval strategies. We define a family of strategies that range in their level of optimism when approving modifications. To protect against settings where no version of the ML algorithm performs well, this family includes a pessimistic strategy that rescinds approval. We use the exponentially weighted averaging forecaster (EWAF) to learn the most appropriate strategy and derive tighter regret bounds assuming the distributional shifts are bounded. In simulation studies and empirical analyses, we find that wrapping approval strategies within EWAF is a simple yet effective approach to protect against distributional shifts without significantly slowing down approval of beneficial modifications.