Predictors of past 30-day vaping abstinence among young e-cigarette users: Machine learning analysis of a longitudinal cohort.

INTRODUCTION: Our existing knowledge on factors influencing vaping abstinence are still limited. The objective of this study was to build a machine learning (ML)-based model to predict past 30-day vaping abstinence and identify predictors among young e-cigarette users. METHODS: Data was taken from a Canadian past 30-day e-cigarette users aged 16-25 (n = 1,659), who were followed-up from 2020 to 2023 across 9 waves. For each outcome, predictors were taken from the immediately preceding wave, resulting in a dataset of 6,435 observations. This dataset was split into a training and testing set in 4:1 ratio and three ML models- random forest, gradient boosting machine, extreme gradient boosting were built on the training set to predict past 30-day vaping abstinence. Model performance was evaluated on the testing set and the best performing model was selected for further Shapley Additive ExPlanations analysis. RESULTS: The random forest model achieved the highest performance (AUC 0.737), and sensitivity analysis showed the robustness of the model. The topmost key predictors of past 30-day vaping abstinence were past month frequency of vaping and different measures of e-cigarette dependence. In addition, product characteristics (i.e., nicotine strength, flavor), intention to quit, and harm perception of nicotine vaping emerged as important predictors across different models. The model was used to estimate individual probability of abstinence and identify the barriers of successful cessation for each individual user. CONCLUSION: While these findings can inform targeted vaping cessation strategies for young people, further research is needed to develop a more generalizable and higher-performing model.