Contrastive self-supervised and causal inference-based contextual predictive model for international student mental health education

Drawing on 4.8 million unlabeled behavioural events from 17 universities on five continents, this study proposes the Contrastive Self-Supervised and Causal Inference-Based Contextual Predictive Model (CSCI-CPM) to forecast depression risk and quantify the value of counseling outreach for 12 438 internationally mobile students. Twin momentum-updated Transformers learn 128-dimensional, domain-invariant embeddings via an InfoNCE objective, sharply reducing label dependence and campus drift. A doubly robust head jointly models treatment propensity and counterfactual PHQ-9 outcomes, yielding unbiased individualized treatment-effect estimates. Leave-one-continent-out tests lift ROC-AUC from 0.882 to 0.931, cut root-mean-squared PHQ-9 error by 0.41, and trim PEHE to 0.027, surpassing five baselines at p < 0.01. A 16-week Thompson-sampling simulation with 25 weekly counseling slots lowers unmet-need days by 41.6 %, raises outreach to low-SES learners from 21 % to 34 %, and shrinks equal-opportunity gaps to 0.019. Real-time inference executes in 18 ms at < 0.001 kg CO-eq per student-day, enabling sustainable on-premise deployment. Clinician review validates 87 % of alerts, while integrated-gradients explanations highlight language-switch entropy, night-screen bursts, and weekend immobility as salient risk signals. CSCI-CPM thus offers a scalable, culturally responsive, and privacy-preserving framework for proactive mental-health governance in global higher education.