Introduction: Depression and diabetes are highly disabling diseases with high comorbidity. Though typically treated separately, they share a common risk factor: low physical activity. Mobile apps might be effective self-management tools to help patients increase physical activity. However, most digital interventions do not tailor their content to individual users, which might impact their effectiveness. In this study, we explain the development of a smartphone application that uses a reinforcement learning algorithm to personalize text-messages that encourage physical activity in low-income ethnic minority patients with comorbid depression and diabetes. 

Methods: Every day of the study, three algorithms are trained to decide which feedback and motivational messages will be sent to the user, with the training data consisting of all previously collected user data. We use linear regression models to assess 1. which feedback message 2. which motivational message and 3. which time period is predicted to maximize the number of steps walked the next day, given an individual's current contextual variables, including demographics and past/expected activity. We use Thompson sampling on the model outputs to choose the message to be sent. We will compare this adaptive intervention to a static messaging component with motivational and educational content; more typical of common digital interventions.

Conclusion: We expect that a smartphone application that uses adaptive learning is an effective intervention to increase physical activity in low-income ethnic-minority patients with depression and diabetes. If proven successful, this approach could potentially be applied in other populations to encourage physical activity and other healthy behaviors.