Predictors of vaping relapse among young e-cigarette users: Machine learning-based survival analysis of a smartphone app-derived longitudinal data

Introduction Although many young people attempt to quit vaping nicotine, limited evidence exists on factors associated with relapse. We aimed to predict time to vaping relapse and identify key predictors among young e-cigarette users attempting to quit. Methods Data was sourced from the “Stop Vaping Challenge” smartphone app users aged 15-35 years (n = 311), who initiated 387 quit challenges. We used 10x5-fold nested cross-validation to train and test three machine learning-based survival models, including random survival forest (RSF), GBM survival, and XGBoost survival to predict time to vaping relapse (i.e., duration of a single quit challenge). The best performing model was carried forward for Shapely Additive ExPlanations analysis, providing individual-level interpretations of predictive factors. Results More than two-thirds of the 387 quit challenges resulted in relapse. Approximately 12.8% of abstinent challenges and 9.71% of relapsed challenges lasted at least 7 days. The GBM survival model achieved the highest average C-index (0.629 ± 0.063), followed by the RSF model (0.624 ± 0.073). The final GBM survival model (C-index 0.694) demonstrated comparable performance in sensitivity analysis. Lower self-confidence in quitting and weaker intention to quit were the most important predictors of faster relapse. Additionally, pod depletion time, time to first vape, past 30-day alcohol drinking, past month frequency of vaping, mood trend during challenge, and reasons for quitting consistently emerged as important predictors across different models. Conclusions We identified key predictors of relapse among young e-cigarette users using smartphone app-based longitudinal data. Future research is needed to assess reliability and generalizability of the model across diverse populations.