A generative two-stage semantic intermediary framework for explainable mental health early warning in higher education.

The psychological well-being of university students is an important public health concern and a growing implementation challenge for digital health systems. Cross-sectional psychometric screening is limited by temporal lag, selective self-disclosure, and the difficulty of distinguishing transient contextual disruption from clinically meaningful deterioration. Although digital phenotyping and predictive artificial intelligence (AI) have advanced mental health monitoring, real-world deployment in universities remains constrained by intrusive data collection, limited auditability, automation bias, and the risk that routine behavioral variation will be prematurely medicalized. In response to these implementation and governance challenges, this article proposes the Generative Semantic Intermediary Framework (GSIF), a behavior-first framework for explainable mental health early warning in higher education. GSIF is organized around three layers: ecologically feasible multimodal observation, two-stage generative semantic translation, and constrained review prioritization. Large language models (LLMs) are used not as autonomous diagnostic agents but as bounded semantic intermediaries: first translating heterogeneous institutional signals into plain-language descriptions of observable behavioral change, and then mapping these descriptions to cautious, reviewable symptom-related descriptors within established psychopathological frameworks. The framework emphasizes data minimization, role-bounded access, human-in-the-loop (HITL) verification, and explicit escalation thresholds. By making the pathway from routine data to review recommendations more transparent, GSIF offers a testable digital health architecture for earlier, more proportionate, and more governable student support workflows. Future work should evaluate its feasibility, acceptability, reviewer calibration, false-positive burden, and incremental value over existing screening and monitoring approaches.