Forecasting alcohol lapse risk up to two weeks in advance using time-lagged machine learning models.

Alcohol use disorder is a chronic, relapsing condition. Individuals must monitor risk and take proactive steps to manage and reduce it indefinitely to prevent lapse. Smartphone sensing and machine learning offer promise as scalable tools for automating risk monitoring and delivering personalized support during recovery. Research has demonstrated that machine learning models using ecological momentary assessment can generate highly accurate predictions of immediate lapse risk (e.g., within the next hour or day). However, many risks require supports that are not immediately available. These supports may need to be planned or involve coordination with others. In such cases, individuals may benefit from advance warning about changes in their lapse risk and the contributing factors. To meet this need, we developed machine learning models to predict future alcohol lapses within 24-hour prediction windows, lagged by 1 day, 3 days, 1 week, and 2 weeks from the prediction timepoint. We engineered features from 4x daily ecological momentary assessments from individuals (N = 151; 51% male; mean age = 41; 87% non-Hispanic White) in early recovery (≤ 8 weeks of abstinence) from alcohol use disorder over a three-month period. We trained and evaluated models using nested cross-validation. Median posterior auROC values were high (0.85-0.89) across models, though performance decreased modestly with longer lag time. Models performed worse for non-advantaged groups (non-White and/or Hispanic, income below federal poverty line, female) compared to advantaged groups (non-Hispanic White, income above federal poverty line, male). Past alcohol use, abstinence self-efficacy, and craving were the most important features, with the magnitude of their importance varying meaningfully by lag time. These findings demonstrate the feasibility of predicting alcohol lapses up to two weeks in advance. Embedding these models within a recovery monitoring and support system could enable adaptive, personalized care with enough warning to implement recovery supports not immediately available. Improving model fairness and optimizing the delivery of model feedback to sustain engagement remain critical next steps.