Predicting substance use behaviors among students taking university entrance test: a cross-sectional study with machine learning techniques

In Bangladesh, there is limited research examining substance use behaviors among university entrance test takers students. Therefore, this study aimed to investigate the current status of substance use behaviors within the Bangladeshi student community. A cross-sectional study was conducted among university entrance test takers from June 18 to 25, 2023, utilizing a convenience sampling technique. Data were collected on socio-demographic characteristics, admission-related variables, depression, and anxiety. Statistical analysis was conducted using the Chi-square test and logistic regression. Machine learning models, including CatBoost and XGBoost, were applied to predict substance use behaviors. The models were evaluated based on accuracy, precision, F1 score, and log loss, with CatBoost demonstrating superior performance across all metrics. During their lifetime, 15.1% reported smoking, while 3.6% acknowledged substance use, and 4.5% had consumed alcohol at some point in their lives. Adjusted logistic regression suggests that gender and depression were significantly associated with substance use behaviors. Within the machine learning framework, the CatBoost model performed better in every criterion than other models. CatBoost SHAP value detected depression and XGBoost gini importance detected gender as the main factor linked to risky substance use behaviors. The study sheds light on the concerning prevalence of risky substance use behaviors among university test takers, highlighting the need for preventive measures. By initiating further research and stimulating prevention programs, this study provides valuable insights for health professionals, students, guardians, policymakers, and educators.