Explainable Machine Learning Models for Prediction of Smoking Cessation Outcome in New Zealand

Isidro C. Medina, Mahsa Mohaghegh

Abstract: In many countries including New Zealand, tobacco use has been considered as the leading cause of preventable diseases and death. Among the adult smokers, more than one fourth makes quit attempts every year. Therefore, selecting the most suitable smoking cessation program and support system is crucial to help smokers achieve the desirable outcomes. A total of 14,443 enrolments were analyzed in this study using machine learning techniques based on tree ensemble and gradient boosted trees. The CatBoost model was adopted which performed better in all data set analyzed compared to Decision Trees, Random Forest or XGBoost. Predictive models were constructed for different priority populations and SHAP TreeExplainer was used to determine the most important predictors associated with smoking cessation outcome. These interpretable machine learning models could help support healthcare providers in making reliable decisions in assisting smokers throughout their quit smoking journey.