Identifying risk factors for drug use recurrence with ecological momentary assessment, wearable technologies, and machine learning: a feasibility trial of peer recovery support specialist intervention

James J. Mahoney III, Victor S. Finomore, Jennifer L. Marton, Lucinda J. England, Sara McFoy, Danielle Romanoff, Jad Ramadan, Anahita Zarei, Amer Mahyoub, Jessie Crooks, James H. Berry, Steven D. Shirk, Manish Ranjan, Ali R. Rezai

Background Identifying predictors of relapse/drug use recurrence (DUR) in real-time could allow the rapid implementation of overdose prevention interventions for those with substance use disorders (SUD). Wearable devices and phone-based applications for self-reported assessments in the patient's natural environment [e.g., ecological momentary assessment (EMA)] have potential for predicting DUR. Peer recovery support specialists (PRSS) play a critical role in reducing DUR risk. The benefits of utilizing PRSS resources in response to alerts derived from wearable technology and EMA data are unknown. This feasibility study investigates a) the use of wearable technologies/EMA to predict physiological/behavioral biomarkers of DUR and b) opportunities for PRSS-based interventions based on predictions. Methods Participants were recruited from various settings (e.g., SUD treatment, sober living), were provided a commercial wearable device (Oura ring), and were prompted daily to complete an EMA application assessing mood and substance cravings. Participants were monitored for 90 days (Baseline Phase), then randomized to the Standard of Care (SoC) or PRSS Intervention arm and followed for two 90-day phases. When machine learning algorithms detected an anomaly, an alert was sent to the participant's phone. The PRSS was sent an alert to contact participants in the PRSS Intervention arm. Results Of 229 participants enrolled, 108 provided EMA and Oura data for ≥1 of 90 days during Baseline, Phase 1, and Phase 2; 63 provided ≥30% of the data across the 3 phases. Twenty-six were randomized to the PRSS Intervention arm, and 37 were randomized to the SoC arm. The PRSS made 483 call attempts for unique alerts; the average per participant across the study was ∼20 (median = 15; range = 1–53). The PRSS intervention arm had modest but significant decreases in anxiety, stress, depression, angst composite (all p 's < 0.001), and maximum craving ( p = 0.011) relative to the SoC arm. Discussion This feasibility study highlights the potential benefits and barriers for using models to predict risk factors for DUR via wearable devices and EMA and supports the utility of a PRSS intervention once the model detects an elevated risk for DUR. Patient compliance and attrition must be improved to optimize this approach as a clinical tool. We discuss challenges and recommend strategies for future studies.