AN EXPLAINABLE AND ROBUST AI FRAMEWORK FOR MENTAL HEALTH SCREENING IN HIGHER EDUCATION

Abstract

University student depression is a serious mental health problem, which requires accurate, interpretable and practical computational screening methods. The paper presents a full machine learning system to identify depression, with two different datasets of depression: a student depression dataset on Kaggle and a massive mental health survey in the COVID-19 pandemic. The files contain details on the demographics, academics, behavior, and mental health of people. Quite a bit of preprocessing is performed, including removing null and duplicate data, encoding labels, correcting class asymmetry using SMOTE, selecting features with RFECV and Stratified K-Fold validation, and normalizing with MinMax and Standard scaling. We consider a variety of ML models, including RF, Gradient Boosting, DT, Naive Bayes, LR, Extra Trees, SVM, XGBoost, LightGBM, CatBoost, SGD, LASSO, and MLP. Ensemble Voting and Stacking classifiers are also investigated by us. Performance is measured by standard evaluation measures and the findings indicate that the accuracy is 98.0% and 99.3% in the two sets. SHAP and LIME ensure that models are comprehensible. It can predict real-time sadness with a web application created in Flask and authenticated with SQLite as it proves to be powerful and effective.