Predicting Mental Health Challenges Among University Students Using a Federated CNN-BiLSTM-Attention Framework

Early detection and effective management of rising mental health issues among university students is a major challenge. This work proposes Fed-CBA, a novel deep learning framework combining CNN, BiLSTM, and Attention to predict future mental health risks. Fed-CBA integrates multiple data sources (LMS, EMA surveys, wearables, lifestyle data) and operates in a decentralized, privacy-preserving manner using FedAvg and differential privacy. It also provides explainable predictions through SHAP and LIME. Experimental results show that Fed-CBA outperforms DNN, CNN, and LSTM models, while also improving fairness (EOR) and interpretability. Overall, the framework enables early detection and supports personalized, equitable student mental health services.