A Software-Oriented Deep Learning Framework for Multiclass Academic Stress and Mental Health Risk Prediction

Abstract

Student mental health and academic stress are significant concerns in higher education, creating a need for effective approaches to early risk identification. This study proposes a multiclass predictive framework for academic stress and mental health risk classification among students. The framework was developed using a Kaggle dataset containing 25,000 records with demographic, academic, behavioral, psychological, lifestyle, and medical attributes. A unified comparison was conducted across six conventional machine learning algorithms, namely Decision Tree, Random Forest, XGBoost, CatBoost, K-Nearest Neighbors (KNN), and Naïve Bayes, and three deep learning architectures, namely Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (BiLSTM). Class imbalance was addressed using the Synthetic Minority Over-sampling Technique (SMOTE), applied only to the training data, and model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrices. CatBoost and BiLSTM achieved the highest classification accuracy of 97%, with weighted F1-scores of 0.97 and macro F1-scores of 0.96. The results indicate that ensemble and deep learning approaches can effectively support multiclass mental health risk classification. Because the provenance and label construction of the secondary dataset are not documented, the results should be read as within-dataset performance. The findings provide a basis for future real-world validation and decision-support applications.