Background: Although many approaches have been developed to promote medication adherence, most are resource intensive. Text messaging is a scalable approach but its effectiveness is modest, as content has largely been generic. Reinforcement learning is a machine learning method that can adjust a messaging strategy through an iterative and systematic feedback loop. This approach has been widely used to personalize content in e-commerce but only to a limited extent in health.
Objectives: To test the impact of applying reinforcement learning on text messages for adherence to diabetes medications
Methods: We randomized patients with suboptimally controlled type 2 diabetes to receive a reinforcement learning intervention or control. Both arms received electronic pillbottles to measure adherence. The intervention arm received daily texts individually adapted using a reinforcement learning algorithm based on feedback from daily latent pillbottle data. The algorithm chose from messages that differed with respect to: 1) message framing (positive, negative, or neutral), 2) inclusion of observed pillbottle feedback (yes/no), 3) inclusion of social reinforcement (yes/no), 4) informational or reminder content, and 5) inclusion of a question prompting reflection (yes/no). The primary outcome was average adherence over 6 months. We used generalized estimating equations with an identity link function and normally-distributed errors, with adjustment for baseline patient characteristics. We conducted pre-specified analyses in key subgroups. We also evaluated patient clusters of responsiveness to different message types using k-means clustering.
Results: Among 60 randomized patients, the reinforcement learning intervention personalized daily texts and resulted in a 13.6% greater average absolute adherence (95%CI: 1.7%-27.1%) versus control. In those non-adherent at baseline, the intervention improved adjusted absolute adherence by 33.0% (95%CI: 13.1%-52.8%). And, in those with moderately poor glycemic control (glycated hemoglobin A1c [HbA1c]: 7.5-9%), the intervention improved adherence by 36.6% (95%CI: 25.1-48.2%). We also identified three clusters of patients, including (1) responding best to observed feedback, (2) responding to social reinforcement or observed feedback, and (3) responding equally to all message types.
Conclusions: Compared with control, the reinforcement learning intervention significantly improved adherence to medication when adjusting for baseline characteristics. Patients who were non-adherent at baseline or moderately poor disease control at baseline may especially derive benefit. We also observed some differences in patient responsiveness to types of messages. Reinforcement learning may be a promising approach to personalizing communication.
