Identifying factors associated with vaping cessation in young adults: A machine learning and XAI approach.

The public health impact of vaping in the United States reflects a complex balance of potential benefits and emerging risks, as e‑cigarettes may reduce exposure to toxic combustion byproducts and support adult smoking cessation, yet growing evidence links vaping to respiratory and cardiovascular harm and youth uptake remains concerning, with 38.4% of adolescent users in 2024 reporting habitual use. To inform the optimal use of predictive technologies in cessation efforts, this study sought to characterize cessation‑related behaviors and attitudes among young adult vapers and evaluate machine learning and explainable AI methods for predicting quit attempts and cessation success. A social media-based survey captured behavioral, contextual, and demographic factors, and cessation was defined as self‑reported abstinence from all vaping products for at least 30 days. Predictors were identified using forward selection and backward elimination, and data were split into training and testing sets. Linear models (LASSO, ridge regression, elastic net) and nonlinear models (random forest, support vector machine) were trained and evaluated using AUC and Brier scores. Linear models demonstrated the strongest overall performance: LASSO achieved AUCs of 0.89 (training) and 0.91 (testing), ridge regression 0.88 and 0.93, and elastic net 0.91 for both sets. Nonlinear models showed signs of overfitting, with random forest achieving 0.99 in training but only 0.70 in testing, and SVM achieving 0.89 and 0.72. Key predictors included age, environmental triggers, vaping frequency, sex, and long‑term behavioral outlook. Individuals under 25 showed greater vulnerability to continued use, environmental cues, especially social exposure, were strongly associated with relapse, and erratic vaping patterns predicted lower cessation success. While these models highlight behavioral and contextual factors that may influence cessation, findings should be interpreted as exploratory given the cross‑sectional design and sample characteristics. Larger, longitudinal studies are needed to validate these insights and clarify the potential of predictive modeling to inform targeted public health interventions.