Drug addiction intent requires proper prediction to be able to intervene early and make effective choices concerning the issue of population health. Nevertheless, most current machine learning models are characterized by class imbalance, optimistic performance estimates because of data leakage, and low interpretability. In order to solve these difficulties, this paper suggests a strictly leakage-resistant stacking ensemble algorithm for multi-class drug addiction intent identification. The model combines four linear base learners, that is, the Logistic Regression, Linear Discriminant Analysis, Linear Support Vector Machine, and L2-regularized Logistic Regression, and a Logistic Regression meta-learner. The sample size is 1,139 survey responses that are described in 35 demographic, social, and psychological features. Stratified 5×5 nested cross-validation is used as the model analysis tool, as all preprocessing steps (cleaning, encoding, scaling, and handling missing values) are trained only on training data, and imbalance in classes is solved with the help of SMOTE that is applied only to training folds. The meta-learner is trained in infold predictions created in an inner validation scheme to avoid information leakage during the stacking. Moreover, modelagnostic interpretability methods, such as SHAP and LIME, are only used on untouched testing data to give both faithful global and class-specific explanations. The findings demonstrate strong and consistent results in a rigid evaluation protocol, and the interpretability analysis shows Drug Occasions and Current Drug Use to be the most important predictors, along with Mental Health, Addicted Friends, Age at First Use, and Relationship Status. In general, the suggested framework is a valid and explainable method of analyzing the intent of drug addiction that may be used to inform research and clinical practice.