Predicting Student Depression Using Machine Learning

Abstract

The rising prevalence of depression among students has drawn significant attention. As data mining and artificial intelligence technologies continue to evolve, leveraging behavioral and textual data for early depression prediction offers new opportunities for timely mental health interventions. This study uses 17 depression-related features, including gender, financial stress, academic pressure, and working hours. To evaluate model performance, this study focused on three representative classification algorithms. The first is logistic regression, known for its interpretability; the second is random forest, which leverages ensemble learning; and the third is XGBoost, a powerful gradient boosting framework. To assess model performance, this study considered multiple quantitative indicators. These include the proportion of correct predictions (accuracy), the model’s ability to identify true positives (precision and recall), the harmonic means of those two (F1 score), and the area under the ROC curve (AUC), which reflects the overall classification capability. A 50-fold robustness test was also conducted to validate model stability. SHAP plots were utilized to interpret model predictions and to identify the most influential features contributing to depression risk at the end of this paper. The experimental data showed that Logistic Regression had the highest AUC score, which was 0.913, while the AUC scores of Random Forest and XGBoost were 0.906 and 0.903, respectively. Calibration curve analysis verified that Logistic Regression had the best calibration performance. The result of this study supports the feasibility and value of Logistic Regression in predicting student depression.