1 - (SYM 24) Development of a Multilevel Bayesian Model for Predicting Individual Daily Distress

Just-in-time (JIT) interventions are a promising, scalable intervention for reducing distress in real time that can be delivered remotely via smartphone. Determining the optimal time frame during which to deliver JITs is a major challenge for the implementation of these interventions. We often need to “get to know” the participant for days or weeks before developing an algorithm to determine when to send an intervention, which may be too long to wait for some high-risk individuals. This study presents a risk prediction algorithm that allows researchers to more quickly determine when to send pre-emptive interventions by integrating population-level and participant-level data, using morning survey responses to predict days when acute distress is most likely.

This proof of concept model uses secondary data analysis of a large ecological momentary assessment (EMA) study. A community sample of participants (N=546) completed up to 6 surveys daily for 8 weeks, including one delivered at the same time each morning and 5 momentary surveys delivered randomly throughout the day (Total # of surveys completed=86,626).  The overall prediction model will use both population and individual models to predict the highest rating of distress an individual will experience across a specific day based on their affect ratings reported in their morning survey. The first stage is an initial, multi-level Bayesian population model that uses data from the first 100 participants who completed the EMA protocol to determine population priors. Next, an initial individual risk model will be derived from weights from the population Bayes estimate for each participant in the subsequent cohort. After each week of the EMA protocol, the participant’s individual risk model will be updated using both the population priors and that participant’s data in a Bayes model to determine new risk prediction scores. The population level model will be updated to include new risk prediction data after every cohort of 10 participants, and this updated model will subsequently inform the population priors used for each new participant.

This proof of concept model will support the feasibility and utility of integrating population and individual real-time assessment data to predict days on which an individual is at elevated risk. An algorithm that predicts daily risk on the basis of the first response each day will bolster our ability to determine when pre-emptive interventions might be most effective in preventing the escalation of distress to acute levels.