Reinforcement Learning for Preventable Hospitalization and Readmission Reduction: A Structured Narrative Review

Purpose: To critically assess reinforcement learning (RL) for sequential interventions designed to reduce preventable hospitalizations and readmissions, distinguishing computational feasibility from demonstrated clinical effectiveness. Methods: A structured narrative review synthesized direct utilization-focused policies, adjacent inpatient decision policies, and contextual prediction research through 9 October 2026. Outcomes, Markov decision formulations, off-policy estimators, and methodological limitations were analyzed without meta-analysis. Results: Directly relevant peer-reviewed evidence centers on Medicaid care management and post-discharge intervention planning. A retrospective state-action-reward-state-action (SARSA) study estimated a 12-percentage-point reduction in an emergency visit and hospitalization composite; however, this reflects a model-based counterfactual estimate rather than an observed randomized treatment effect. A larger offline study reported non-inferiority on a shaped reward scale alongside narrower intervention-allocation disparities. Data-pooling research contributed theoretical regret bounds and empirical support for learning across patient classes. Critical limitations across the literature include unmeasured confounding, limited action support, surrogate rewards, overlapping cohorts, and absent external validation. No direct study established a prospective randomized reduction in hospitalizations from an executed RL policy. Conclusion: RL provides a coherent framework for longitudinal care allocation, but current evidence supports algorithmic development rather than proven clinical efficacy. Translation demands explicit endpoint definitions, rigorous off-policy evaluation, prospective silent workflow evaluation, and controlled trials with safety, equity, and resource constraints.