This study proposes a machine learning framework to predict student dropout in higher education using academic, demographic, and financial data. The original dataset was reformulated into a binary classification problem distinguishing dropout from not dropout and adapted to the Vietnamese educational context. Since dropout cases represent only a minority of the data, the dataset is inherently imbalanced and presents challenges for accurate prediction. To address this issue, the study evaluated three resampling strategies based on the synthetic minority over-sampling technique (SMOTE), including its standalone version and two hybrid approaches with Tomek Links and edited nearest neighbors (ENN), in combination with four classification algorithms including random forest, support vector machine, extreme gradient boosting, and categorical boosting. The best performance was achieved by categorical boosting with synthetic minority over-sampling technique and edited nearest neighbors, with an F1 score of 0.810, an F1 macro of 0.857, a recall of 0.845, and a precision-recall area under the curve of 0.863. SHapley Additive exPlanations further identified failed courses, tuition payment status, and debtor history as the most influential factors in predicting dropout. These findings demonstrate the effectiveness of interpretable machine learning and hybrid resampling techniques in enabling early identification and intervention for at-risk students.