Evaluating Cellular Communication Sensing for Lapse Risk Prediction During Early Recovery from Alcohol Use Disorder: A Longitudinal Observational Study

Background: Alcohol Use Disorder (AUD) is a chronic, relapsing disease. An automated recovery support system using personal sensing and machine learning may help identify when individuals are at elevated lapse risk. Cellular communication sensing may detect dynamic changes in lapse risk and can be contextualized with self-reported, risk-relevant information about contacts. Objective: We evaluated a machine learning model predicting next-day alcohol lapse among individuals in early recovery from AUD using contextualized cellular communication data and baseline demographic and AUD characteristics. Methods: A total of 144 participants (49% male; mean age=40; 87% non-Hispanic White) with a goal of abstinence provided cellular communication data and alcohol use reports via a 4x daily EMA for up to three months. Models were trained and evaluated using repeated k-fold cross-validation. Results: The best-performing model used an elastic net algorithm and retained 13 features (median posterior auROC=0.68, 95% Bayesian credible interval (CI; [0.64, 0.71]). A baseline comparison model including only baseline features retained five features and demonstrated nearly identical performance (median auROC=0.68, 95% CI [0.64, 0.71]). Conclusions: Cellular communication data capture some risk-relevant signal for alcohol lapse but do not provide incremental predictive value beyond baseline measures. Several communication features were retained in the final model with moderately sized coefficients, suggesting that aspects of social communication may be important for understanding lapse risk. Although, limitations inherent to cellular communication as a sensing method may outweigh their added value.