Post-Market Surveillance of Predictive Decision Support Tools in the Presence of Confounding Medical Interventions using Causal Inference.

Predictive AI models have been widely adopted in various clinical applications. While there has been tremendous focus on the pre-deployment performance assessment of these models, there has been little attention paid to their post-deployment evaluation. It has previously been shown that post-market surveillance of predictive AI models can be complicated by confounding medical interventions (CMI), actions that clinicians take in response to the model predictions to prevent the predicted adverse outcomes. This impacts the labels that the model will be evaluated on, leading to biased performance estimates. The more successful the model is in preventing adverse outcomes, the lower its performance is going to appear. To address this issue, we propose a novel approach to monitor the performance of deployed predictive models using causal inference techniques. We provide simulation results demonstrating that the proposed approach can provide unbiased estimates of all performance metrics for models with binary target variable.