Beyond performance metrics: modeling outcomes and cost for clinical machine learning

Abstract

Advances in medical machine learning are expected to help personalize care, improve outcomes, and reduce wasteful spending. In quantifying potential benefits, it is important to account for constraints arising from clinical workflows. Practice variation is known to influence the accuracy and generalizability of predictive models, but its effects on cost-effectiveness and utilization are less well-described. A simulation-based approach by Mišić and colleagues goes beyond simple performance metrics to evaluate how process variables may influence the impact and financial feasibility of clinical prediction algorithms.