AI-driven early warning of psychological crisis in college students using network theory and few-shot learning through survey-parameterized simulation

Psychological crises among college students are often detected too late because current practice relies on periodic screening that is passive and infrequent. This paper proposes an AI-driven early-warning framework that organizes multi-source campus behavioral traces into a three-layer student–behavior–space–time heterogeneous dynamic graph. A heterogeneous graph neural network with hierarchical and temporal attention detects behavioral-pattern drift preceding a crisis, while graph contrastive pre-training and few-shot fine-tuning yield a student-specific dynamic behavioral representation from scarce crisis-positive labels; the risk score is decomposed into behavioral-dimension attributions for counselors. Evaluation uses a survey-parameterized simulation: eight thousand synthetic student records generated from distributions calibrated to the desensitized behavioral records of 316 real students, with a field-survey-parameterized crisis-generating mechanism applied to sixty designated cases. The method achieves an AUPRC of 0.249, recall of 0.607, and average lead time of 2.79 weeks, ahead of all baselines; ablation attributes most gain to contrastive pre-training and heterogeneous representation. Against stronger competitors the advantage narrows, and against four additional crisis-generating mechanisms evaluated with the model frozen, it persists when the signature carries relational and directionally persistent structure but disappears otherwise. An information-flow audit shows discriminative ability persisting to roughly two weeks ahead and flooring beyond three, indicating detection of ongoing deterioration rather than forecasting of future decline. At recall 0.700 the alert load reaches 8.2 per counselor per week; the framework is accordingly a methodological contribution, not a validated system for real crises.