An Integrated Learning Approach for Detecting Depression in University Students

Abstract

Depression among college students has become a major mental health issue that needs accurate and reliable computer screening methods. Conventional grading methods often have problems with uneven distribution of classes, duplicate features, and limited interpretability, which makes them less useful in the real world. This framework uses two different sets of data that are linked to depression: a Kaggle dataset on student depression and a large-scale mental health survey that was conducted during the COVID period and included demographic, academic, behavioral, and psychological questions. A lot of work goes into the preprocessing, like getting rid of nulls and duplicates, labeling, balancing the data with SMOTE, getting rid of features with RFECV and Stratified K-Fold validation, and making sure everything is normal with MinMax and Standard scaling. Random Forest, Gradient Boosting, Decision Tree, Gaussian Naive Bayes, Logistic Regression, Extra Trees, Support Vector Machine, XGBoost, LightGBM, CatBoost, SGD, LASSO, and MLP are some of the machine learning classifiers that are used. There are also hybrid ensemble strategies that combine MLP, SGD, and CatBoost. We check how well the model works by looking at its accuracy, precision, recall, F1-score, ROC-AUC, Cohen's kappa, and log loss. Voting and Stacking classifiers worked best for the sadness student dataset, getting a maximum accuracy of 98.0%. A Voting classifier worked best for the mental health dataset, getting a 99.3% accuracy. SHAP and LIME analysis make sure that things can be explained. Real-time prediction from user-provided inputs is possible with a Flask-based interface that uses SQLite for authentication. This shows that the sadness detection capability is strong, understandable, and deployable.