Mental health concerns among university students in Bangladesh are rising, with depression presenting as one of the most prevalent challenges. This study employs machine learning techniques to identify key factors associated with self-reported depression, used as a proxy for bipolar disorder in the absence of clinical diagnostic data. Data were collected through structured surveys covering academic stress, lifestyle habits, emotional well-being, and social relationships. To address class imbalance in the target variable, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. A comparative analysis of six supervised learning modelsRandom Forest, Support Vector Machine, Logistic Regression, Decision Tree, K-Nearest Neighbors, and Naïve Bayes-was conducted to determine the most effective approach for classification. The models were evaluated using accuracy, precision, recall, F1-score, and AUC-ROC metrics. In addition to classification performance, feature importance scores and Local Interpretable Model-Agnostic Explanations (LIME) were used to interpret individual predictions and identify influential features such as suicidal ideation, social stress, and academic burden. The findings underscore the value of interpretable, data-driven methods in enhancing mental health awareness and guiding early interventions. Although limited by the use of selfreported diagnoses and non-clinical instruments, this study provides a scalable analytical framework for supporting student mental health in low-resource academic settings.