JMIR Publications

ABSTRACT

Background:

Although recent advances in machine learning have accelerated the development of algorithms capable of predicting mental health symptoms and maladaptive behaviors, research aiming to forecast the emergence of addiction-related problems in real-life contexts remains limited. Notably, despite longstanding evidence that emotional and temperamental vulnerabilities are core predictors of alcohol-related problems, few studies have incorporated these factors into evidence-based machine learning or AI models.

Objective:

This study evaluated the extent to which machine learning (ML) models can accurately predict weekly drinking episodes among high-risk drinkers by integrating real-time health data collected from wearable devices with self-reported emotional and temperamental vulnerability indicators across a 4-week follow-up period.

Methods:

Using a prospective observational design, we collected weekly self-report surveys (five time points) and wearable-derived data (Fitbit; heart rate, physical activity, sleep) from adults in their twenties over four weeks. All variables were aggregated at the weekly level. Positive labels were defined according to the AUDIT-K hazardous-use cutoffs (≥20 for men, ≥10 for women). XGBoost and Random Forest models were trained, and performance was evaluated using K-fold cross-validation, yielding Accuracy, Precision, Recall, F1-score, and ROC AUC for comparison among self-report-only, wearable-only, and integrated models.

Results:

For week-ahead prediction of drinking episodes, the integrated model showed superior performance. XGBoost achieved an Accuracy of 0.887, Recall of 0.824, and ROC AUC of 0.906, while Random Forest achieved an Accuracy of 0.903, Recall of 0.824, and ROC AUC of 0.937—outperforming both self-report-only and wearable-only models.

Conclusions:

The combination of self-report and wearable biosignal data demonstrated strong utility for predicting weekly high-risk drinking episodes. These findings suggest that digital mental-health indicators can serve as personalized early-warning signals and intervention triggers in clinical and public-health settings. To support real-world deployment, future work should include threshold optimization, external validation, and time-series split validation to assess model generalizability.