Medication Adherence Improvement as a Bandit Problem

Medication non-adherence is a big problem in chronic disease management. There are a few different underlying causes for non-adherence: side-effect concerns, lack of belief in the need for adherence, forgetting, cost-concerns, etc. Mobile health interventions have been shown to be successful in improving adherence. Existing studies involve fixed intervention strategies that are hand designed by experts. The premise of this work is that adapting interventions based on performance data can increase effectiveness. I will report very preliminary data from an ongoing pilot with non-adherent hypertensive patients (21 thus far). These patients are given a pill bottle cap that reports on pill bottle openings; this is the reward function we use for the bandit problem. Health education experts designed text messages to address underlying non-adherence causes. A contextual bandit algorithm decides which category of text message to send to each patient daily. How is it working? Come find out.