Reinforcement Learning to Prevent Acute Care Events Among Medicaid Populations: Mixed Methods Study.

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Multidisciplinary teams require efficient prioritization for patients with complex medical and social needs.
• Current care management relies on experience, missing opportunities for data-driven intervention recommendations.
• Recommender systems can potentially prevent adverse outcomes by learning from patient trajectories.

Purpose of the Study:

• To evaluate if a reinforcement learning (RL) approach can surpass standard care management in recommending optimal interventions.
• To compare a SARSA RL model against experience-based methods for complex patient care.

Main Methods:

• Compared a SARSA RL model with the standard experience-based approach using data from 3175 Medicaid beneficiaries.
• Utilized counterfactual causal inference to estimate reductions in acute care events.
• Assessed model fairness across demographic subgroups and conducted qualitative chart reviews.

Main Results:

• SARSA-guided care management reduced acute care events by 12 percentage points (20.7% relative reduction) compared to the status quo.
• The RL approach demonstrated improved fairness across gender and race/ethnicity subgroups.
• Qualitative reviews showed SARSA identifying interventions for medical-social interactions like housing quality and food insecurity.

Conclusions:

• SARSA-guided care management shows promise in reducing acute care utilization compared to standard practices.
• Reinforcement learning can enhance complex decision-making for patients with concurrent clinical and social factors.
• The RL approach maintains safety and fairness across diverse patient populations.