Forecasting Risk of Alcohol Lapse up to Two Weeks in Advance using Time-lagged Machine Learning Models

We developed machine learning models to predict future alcohol lapses within 24-hour windows lagged 1 day, 3 days, 1 week, and 2 weeks. We engineered features from 4x daily ecological momentary assessment from individuals (N=151; 51% male; mean age=41; 87% non-Hispanic White) in early recovery over three months. We trained and evaluated models using nested cross-validation. Median posterior auROC was high (0.85–0.91) for all models but decreased modestly with increasing lag. Models performed worse for non-advantaged groups (non-White and/or Hispanic, below poverty, female). Past alcohol use, abstinence self-efficacy, and craving were the most important features, with the magnitude of importance varying meaningfully by lag. These findings demonstrate feasibility of predicting next-day lapses up to two weeks in advance. Embedding these models in a recovery monitoring support system could enable adaptive, personalized care. Improving model fairness and optimizing the delivery of model feedback to sustain engagement remain critical next steps.