Traditional mental health management mainly relies on manual assessment and regular screening, which is characterized by inefficiency, limited coverage and lagging early warning. The rapid development of artificial intelligence technology provides new solutions for mental health management, and the early warning of students’ mental health risks based on intelligent algorithms is of great practical significance. In this study, the Trans-LSTM neural network model was used to integrate the SCL-90 psychological assessment data and students’ daily behavioral data to construct a multivariate time series classification mental health risk early warning system. The 8766 valid psychological assessment data were privacy-protected by k-anonymization technique, the SENet attention mechanism was applied to enhance the feature extraction capability, and the Transformer positional coding and multi-head attention mechanism were combined to optimize the time-series feature learning. The experimental results show that the proposed Trans-LSTM model performs well in the mental health prediction task, with an accuracy of 83.36%, a precision of 87.64%, a recall of 78.27%, and an F1 value of 80.14%, which are all significantly better than the comparative models such as GCN, SVM, and GraphSAGE. The study found that the detection rates of the three factors of anxiety, relationship sensitivity, and depression were 22.57%, 16.34%, and 14.69%, respectively, providing an important basis for mental health risk identification. The study shows that the model can effectively integrate heterogeneous data from multiple sources, realize accurate prediction and timely warning of students’ mental health status, and provide an intelligent solution for mental health management in colleges and universities.
