Heinz MV, Price GD, Singh A, et al. A longitudinal observational study with ecological momentary assessment and deep learning to predict non-prescribed opioid use, treatment retention, and medication nonadherence among persons receiving medication treatment for opioid use disorder. J Subst Use Addict Treat. 2025;173:209685. doi:10.1016/j.josat.2025.209685
In this analysis, previously recorded data from app-based ecological momentary assessments (EMAs) was assessed in combination with deep learning (recurrent neural networks) to predict four outcomes in people receiving medication for opioid use disorder (MOUD): (1) self-reported non-prescribed opioid use (NPOU), (2) self-reported medication nonadherence, (3) objective medication nonadherence (from electronic health records or EHR), and (4) treatment retention (also from EHR data). The study included 62 participants from the NIDA DTECT study. Each participant received EMA prompts three times per day, covering 12 topics such as mood, cravings, withdrawal symptoms, substance use, sleep, stress, and self-regulation. A separate model was trained for each EMA subtype and outcome. The goal was to see if EMA data could predict outcomes with at least moderate accuracy (AUC > 0.70). SHAP analysis was used to understand which EMA features were most influential. Results showed that all EMA types could predict the four outcomes better than chance, up to seven days in advance. The most predictive EMA subtype was recent substance use (HR_USED), which ranked in the top three across all outcomes. Self-regulation (SR) was also highly predictive for NPOU, EMA-based nonadherence, and retention. The least predictive EMAs varied by outcome but commonly included SLEEP, WITHDWL, and STRESS. Overall, the study suggests EMAs, especially those assessing behavior, mood, and context, can be powerful tools for predicting key clinical outcomes in people undergoing MOUD.