Identifying Predictors of Smoking Switching Behaviours Among Adult Smokers in the United States: A Machine Learning Approach.

Completely abstaining from cigarette smoking or fully switching to e-cigarette (EC) use may be beneficial for reducing the global burden of smoking-related diseases. This study aimed to identify and compare the top 10 prospective predictors of smokers switching away from smoking in the United States. Data from adult exclusive cigarette smokers at Wave 4 of the Population Assessment of Tobacco and Health (PATH) study, who were followed up at Wave 6, were analysed. An Xgboost-based machine learning (ML) approach with a nested cross-validation scheme was utilised to develop a multiclass predictive model to classify smokers' behavioural changes from W4 to W6, including smoking cessation, full and partial switching to EC, and cigarette non-switching. The SHapley Additive exPlanations (SHAP) algorithm was deployed to interpret the top 10 predictors of each switching behaviour. A total of 396 variables were selected to generate the four-class prediction model, which demonstrated a micro- and macro-average area under the receiver operating characteristics curve (ROC-AUC) of 0.91 and 0.81, respectively. The top three predictors of smoking cessation were prior regular EC use, age, and household rules about non-combusted tobacco. For full switching to EC use, the leading predictors were age, type of living space, and frequency of social media visits. For partial switching to EC use, the key predictors were daily cigarette consumption, the time from waking up to smoking the first cigarette, and living with tobacco users. ML is a promising technique for providing comprehensive insights into predicting smokers' behavioural changes. Public health interventions aimed at helping adults switch away from smoking should consider the predictors identified in this study.