Week-Ahead Prediction of High-Risk Drinking Episodes Among Young Adults Using Wearable Biosignals and Psychological Vulnerabilities: Prospective Observational Machine Learning Study.

BACKGROUND: Although machine learning has increasingly been used to predict mental health symptoms and maladaptive behaviors, real-world prediction of addiction-related risk remains limited. Emotional and temperamental vulnerabilities are established correlates of alcohol-related problems, yet few studies have integrated these factors with wearable-derived biosignals in alcohol-risk prediction models. OBJECTIVE: This study evaluated whether machine learning models could predict weekly high-risk drinking episodes among young adults with elevated alcohol-use risk by integrating wearable-derived health data with baseline emotional and personality vulnerability indicators. METHODS: In this prospective observational study, adults in their 20s completed weekly self-report surveys and wore Fitbit devices for 4 weeks. Features from week t were used to predict the Alcohol Use Disorders Identification Test-Korean version (AUDIT-K)-based high-risk drinking label at week t+1. Positive labels were defined using AUDIT-K high-risk drinking cutoffs, with scores of ≥20 for men and ≥10 for women. Extreme gradient boosting (XGBoost) and random forest models were evaluated across self-report-only, wearable-only, and integrated feature sets using 5-fold participant-level grouped cross-validation. RESULTS: A total of 206 participants contributed 620 week-level observations, of which 85 (13.7%) were labeled as positive high-risk drinking episodes. In participant-level grouped cross-validation, the integrated random forest model showed the most favorable sensitivity-oriented performance, with a mean accuracy of 0.617 (SD 0.078), recall/sensitivity of 0.653 (SD 0.144), area under the receiver operating characteristic curve (ROC AUC) of 0.681 (SD 0.079), and area under the precision-recall curve (PR AUC) of 0.255 (SD 0.090). The integrated XGBoost model achieved an accuracy of 0.670 (SD 0.089), recall/sensitivity of 0.399 (SD 0.174), ROC AUC of 0.651 (SD 0.089), and PR AUC of 0.228 (SD 0.074). Shapley additive explanations analyses indicated that both baseline vulnerability indicators and wearable-derived weekly summaries contributed to model predictions. CONCLUSIONS: Integrating baseline emotional and personality vulnerability indicators with wearable-derived weekly health signals may provide useful information for week-ahead prediction of high-risk drinking episodes. These findings provide preliminary support for wearable-assisted alcohol-risk stratification, although the modest positive predictive performance indicates that external validation and more proximal within-person measures are needed before real-world early-warning or just-in-time adaptive intervention applications.