From Simulation to Reality and Back Again: A Hybrid Approach to Estimate the Compliance of ESM Study Participants to Different ESM Protocols

Sustaining sufficient compliance in long-running Experience Sampling Method (ESM) studies has remained a challenge. Participants of such studies usually drop out after a few weeks due to response fatigue, technical difficulties, the intrusiveness of the prompts, and changes in their motivation. One common approach to ensure higher compliance is to tailor the timing of the prompts. Different tailoring approaches that take into account the personal context of participants have been proposed. Such as considering calendar events, ESM device usage patterns, or information derived from physiological sensors. Recently, the application of the reinforcement learning (RL) approach in this domain has shown promise in learning the right timing of the prompts. However, RL agents require repeated inquiries at the beginning of their learning process which is from the participants’ point of view intrusive and may result in early dropouts. To overcome this problem, agents can pre-train from prior knowledge and avoid "cold start". Although real-life data about compliance in ESM studies is insufficient, agents could instead train with generated data that imitates real-life events. Accordingly, psychological theories should be involved in the simulation process that generates ESM-related data. We present our hybrid approach that utilizes both historical ESM data and synthesized data backed by psychological theories to provide sufficient data to train ML models and RL agents to predict the opportune moments of ESM prompts.