Optimizing Substance Use Treatment Selection Using Reinforcement Learning

Substance use disorder (SUD) exacts a substantial economic and social cost in the United States, and it is crucial for SUD treatment providers to match patients with feasible, effective, and affordable treatment plans. The availability of large SUD patient datasets allows for machine learning techniques to predict patient-level SUD outcomes, yet there has been almost no research on whether machine learning can be used to optimize or personalize which treatment plans SUD patients receive. We use contextual bandits (a reinforcement learning technique) to optimally map patients to SUD treatment plans, based on dozens of patient-level and geographic covariates. We also use near-optimal policies to incorporate treatments’ time-intensiveness and cost into our recommendations, to aid treatment providers and policymakers in allocating treatment resources. Our personalized treatment recommendation policies are estimated to yield higher remission rates than observed in our original dataset, and they suggest clinical insights to inform future research on data-driven SUD treatment matching.