DDNet: A Robust, and Reliable Hybrid Machine Learning Model for Effective Detection of Depression Among University Students

Abstract

In recent years, the detection of depression among university students has become an increasingly critical issue. This paper presents a depression detection network (DDNet), a novel approach utilizing a three-stage stacked ensemble model to address this challenge. The proposed model incorporates two different Multilayer Perceptron (MLP) and a Stochastic Gradient Descent (SGD) as the 1st stage base classifier, an MLP, and CatBoost (CB) as 2nd stage base classifiers, with a Lasso Regressor (LASSO) serving as the meta-classifier. The hyperparameter of the used models has been optimized using random search. The optimal configuration has been determined through extensive experimentation with various machine learning (ML) models and settings to ensure high performance. Two different datasets (Dataset-1, Dataset-2) of depression detection among university students have been used to evaluate the model. The model has achieved 98.98% and 99.16% accuracy in Dataset-1 and Dataset-2, respectively. Paired t-test has been performed to ensure the statistical significance of the proposed model. To guarantee the model’s transparency, SHapley Additive exPlanations (SHAP) were employed, providing interpretability of the predictive factors. Additionally, a variation of Monte Carlo Dropout (MCD) has been used to assess the uncertainty of the model’s predictions, ensuring reliability. The results indicate that the proposed model is a promising tool for mental health professionals seeking effective, interpretable, and reliable solutions for early depression detection in educational settings.