Machine learning models for temporally precise lapse prediction in alcohol use disorder

We developed three machine learning models that predict hour-by-hour probabilities of a future lapse back to alcohol use with increasing temporal precision (i.e., lapses in the next week, next day, and next hour). Model features were based on raw scores and longitudinal change in theoretically implicated risk factors collected through ecological momentary assessment. Participants (N = 151, 51% male, M_age = 41, 87% White, 97% non-Hispanic) in early recovery (1-8 weeks of abstinence) from alcohol use disorder provided 4 × daily ecological momentary assessment for up to 3 months. We used grouped, nested cross-validation to select the best models and evaluate the performance of those best models. Models yielded median areas under the receiver operating curves of 0.89, 0.90, and 0.93 in the 30 held-out test sets for week-, day-, and hour-level models, respectively. Some feature categories consistently emerged as being globally important to lapse prediction across our week-, day-, and hour-level models (i.e., past use, future self-efficacy). However, most of the more punctate, time-varying constructs (e.g., craving, past stressful events, arousal) appear to have a greater impact within the next-hour prediction model. This research represents an important step toward the development of a smart (machine learning guided) sensing system that can both identify periods of peak lapse risk and recommend specific supports to address factors contributing to this risk. (PsycInfo Database Record (c) 2024 APA, all rights reserved).