<![CDATA[Abstract: Mental health issues such as depression are prevalent among students and significantly impact both academic performance and psychological well-being. While machine learning techniques have been widely employed to predict mental health conditions, single-model approaches often suffer from limited generalizability and interpretability. This study proposes a Stacked Ensemble Learning framework that integrates three heterogeneous base classifiers—Logistic Regression (LR), Support Vector Machine (SVM) with RBF kernel, and Random Forest (RF)—combined with a meta-learner to enhance the accuracy and robustness of depression prediction among students. Experiments were conducted on a large-scale student mental health dataset comprising 27,901 records, with preprocessing steps including feature standardization, class balancing using SMOTE, and stratified cross-validation. Performance evaluation utilized Confusion Matrix, F1-Score, Recall, Precision, and the Area Under the ROC Curve (AUC-ROC). The proposed ensemble model achieved a classification accuracy of 84%, an AUC of 0.911, and an average precision of 0.89, consistently outperforming individual baseline classifiers. These results validate that combining margin-based, non-linear, and tree-based models can yield more reliable and interpretable predictions. The proposed architecture presents a promising and explainable tool for early detection of mental health issues within educational settings. Keywords: depression detection; student mental health; ensemble learning;machine learning. Abstrak: Permasalahan kesehatan mental seperti depresi kerap dialami oleh siswa dan berdampak pada performa akademik serta kesejahteraan psikologis mereka. Meskipun pendekatan pembelajaran mesin telah banyak digunakan untuk prediksi kondisi ini, model tunggal kerap menghadapi keterbatasan dalam hal generalisasi dan interpretabilitas. Studi ini mengusulkan kerangka kerja Stacking Ensemble Learning yang mengintegrasikan tiga model dasar—Logistic Regression (LR), Support Vector Machine (SVM) dengan kernel RBF, dan Random Forest (RF)—yang dikombinasikan dengan meta-learner untuk meningkatkan akurasi dan stabilitas prediksi depresi pada siswa. Eksperimen dilakukan pada dataset berskala besar yang mencakup 27.901 entri, dengan penerapan preprocessing, standardisasi, penyeimbangan kelas menggunakan SMOTE, dan validasi silang stratifikasi. Evaluasi performa menggunakan metrik Confusion Matrix, F1-Score, Recall, Precision, serta AUC-ROC Curve. Hasil menunjukkan bahwa model ansambel yang diusulkan mencapai akurasi 84%, AUC 0,911, dan rata-rata precision 0,905, yang secara konsisten melampaui performa model individual. Temuan ini menegaskan bahwa kombinasi antara model berbasis margin, non-linear, dan pohon keputusan mampu menghasilkan prediksi yang lebih andal dan dapat dijelaskan, sehingga potensial untuk diimplementasikan dalam sistem pemantauan kesehatan mental berbasis institusi pendidikan. Kata kunci: deteksi depresi, kesehatan mental siswa, stacking ensemble, machine learning.]]>
