Data-driven Interpretable Policy Construction for Personalized Mobile Health.

Area of Science:

• Digital Health
• Behavioral Science
• Machine Learning

Background:

• Mobile health apps use message-based interventions to promote healthy behaviors.
• Current methods, like HeartSteps, lack optimized policies for message delivery, potentially overwhelming users.
• There's a need for effective and interpretable policies to maximize intervention benefits.

Purpose of the Study:

• To develop an effective and interpretable policy for optimizing message delivery in mobile health interventions.
• To reduce sedentary behavior by improving intervention strategies.
• To provide insights for future behavioral science research.

Main Methods:

• Proposed Optimal Policy Trees + (OPT+), a batch off-policy learning method.
• Combined personalized threshold learning with an extension of Optimal Policy Trees.
• Tested the method using data from the HeartSteps V2/V3 physical-activity intervention.

Main Results:

• Demonstrated a significant reduction in sedentary behavior.
• Achieved these reductions with a lower message delivery budget.
• OPT+ generated a highly interpretable and stable decision tree.

Conclusions:

• OPT+ offers an effective and interpretable approach to optimizing mobile health interventions.
• The method can significantly reduce sedentary behavior while managing message frequency.
• The interpretable policy facilitates theoretical insights for future research in digital health and behavior change.