Predicting student mental health through entropy-based features and interpretable cross-attention transformer networks

Mental health is becoming a major concern for students in today's fast-changing world. Mental health challenges have impact on every aspect of life including performance which pointed to early identification of risk levels. Recent reports show a rapid rise in anxiety, depression, and stress among students. This need urge for intelligent systems that can support mental health monitoring in educational environments. Existing methods often lack accuracy, and the ability to capture complex psychological patterns. This study aims to address these gaps by developing an interpretable deep learning model that predicts student mental health risk using FT-Transformer and LSTM architectures. The model integrates a Cross-Attention Attribution Layer (CAAL), which combines feature attention with temporal attention, making the ensemble intrinsically interpretable. The approach captures both global feature relationships and time-based emotional variations. Feature engineering based on entropy and uncertainty patterns further strengthens the model's ability to detect subtle risk signals. The proposed method is compared with several baseline models, including SVM, Logistic Regression, Random Forest, standalone LSTM, and FT-Transformer. Empirical analysis shows that the proposed model achieves the highest accuracy of 95%, outperforming all baselines. These findings are validated through explainable AI techniques, global feature-importance analysis, and multiple statistical tests for effective framework to support student mental health assessment.