PsyCrisCare: A Multimodal BERT–DNN Fusion with Hierarchical Classification for Student Psychological crisis risk Risk Prediction

Abstract

In education, stress, anxiety, and depression are developing issues that impair student performance and well-being. Slow and unscalable surveys and manual counseling dominate existing methods. This research introduces PsyCrisCare, a machine learning-based paradigm for student psychological crisis risk risk prediction and hierarchical intervention. The goal is to identify dangers early, classify kids by well-being, and propose support.GPA, sleep hours, daily steps, emotional ratings (stress, anxiety, depression), and self-reported text inputs (daily reflections) are integrated by PsyCrisCare. A hybrid pipeline uses feature engineering, sentiment analysis, and hierarchical classification to classify kids as Healthy, At-risk, or Struggling. In experiments, PsyCrisCare outperformed baseline classifiers by 8–12% with an accuracy of 91.3%, F1-score of 0.89, and AUROC of 0.92. The classifier initially divides all the students into two groups: those who are healthy and those who are not. On a public dataset of 500 students, PsyCrisCare fares better than baseline classifiers like Random Forest and XGBoost. The approach improves the Struggling group recall from 0.71 to 0.86, detecting at-risk students early. Analysis of 5-fold cross-validation reveals stable performance with limited volatility (±1.2%), demonstrating fairness and robustness across subgroups. In conclusion, PsyCrisCare has excellent potential as an AI-powered platform for proactive mental health monitoring, early crisis risk risk prediction, and targeted educational intervention.