AI for Campus Mental Health: From Conversational Surveys to Automated Depression Screening

Forum topic · 小凯 · 2026-05-18

Summary

A doctoral dissertation by Tang proposes an end-to-end AI pipeline for campus mental health, spanning prevention and intervention. On the prevention side, TigerGPT is a personalized survey chatbot that replaces static questionnaires with natural conversation, achieving 75% usability and 81% satisfaction. Its successor AURA is a reinforcement learning framework that adaptively switches follow-up types within a session (validate, specify, reflect, probe), guided by an LSDE quality signal (length, self-disclosure, emotional intensity, specificity) initialized from 96 historical dialogues. Results show an average quality gain of 0.12 (effect size 0.66), 63% fewer generic 'please elaborate' prompts, and 10x more validation behaviors. For intervention, PsychoGPT is an LLM grounded in DSM-5 criteria and the PHQ-8 depression scale, performing distress classification, symptom-level scoring, and cross-validation with external assessments. To reduce hallucination risks, the author proposes Stacked Multi-Model Reasoning (SMMR), where layered expert models handle local subtasks and later layers synthesize final diagnoses; on DAIC-WOZ, SMMR outperforms single-model baselines in accuracy, F1, and PHQ-8 scoring. Open questions remain regarding baselines for usability metrics, deployment latency/cost of SMMR, and cross-cultural generalization to non-English contexts.

Campus mental health is a large unmet need. Universities know students experience stress, anxiety, and depression, but traditional surveys suffer from low response rates and poor data quality — and by the time problems surface, it is often too late. Tang's doctoral dissertation proposes an end-to-end AI pipeline covering everything from prevention to intervention.

Prevention: Conversational Surveys

The prevention side is TigerGPT, a personalized survey chatbot. Traditional questionnaires are cold checkbox forms — "Over the past two weeks, how often have you felt down?" TigerGPT instead collects feedback through natural-language conversation. A 75% usability rate and 81% satisfaction indicate students are more willing to discuss their state with a conversational agent.

The first version had issues, though — repetitive replies and insufficient depth. This led to AURA, a reinforcement learning framework that adaptively switches follow-up types within a single conversation session:

Validate — acknowledge and confirm what the student shared
Specify — narrow down details
Reflect — mirror back the student's statements
Probe — dig deeper

AURA uses a quality signal called LSDE (answer Length, Self-disclosure, emotional intensity, Specificity), initializing its policy from 96 historical dialogues. Results:

Average quality improvement of 0.12 (effect size 0.66)
63% reduction in generic "please be more specific" prompts
10x increase in validation behaviors

Intervention: Automated Screening

The intervention side is PsychoGPT, an LLM built on DSM-5 diagnostic criteria and the PHQ-8 depression scale. It performs three layers of work:

1. Initial distress classification 2. Item-by-item scoring by symptom dimension 3. Cross-validation against external assessment results

To reduce hallucinations — whose consequences are severe in the mental health domain — the dissertation proposes Stacked Multi-Model Reasoning (SMMR): multiple expert models process tasks in layers, with early layers handling local subtasks (e.g., linguistic feature extraction) and later layers synthesizing all outputs for a final diagnosis.

On the DAIC-WOZ dataset, SMMR outperforms single-model approaches in accuracy, F1, and PHQ-8 scoring.

Open Questions

TigerGPT's 75% usability and 81% satisfaction — compared against what baseline?
What are SMMR's latency and computational costs in real-world deployment?
Mental health is expressed very differently across languages and cultures (e.g., in Chinese or other non-English contexts) — did the models account for cross-cultural generalization?

References

1. Tang, J. (2026). *New AI-Driven Tools for Enhancing Campus Well-being: A Prevention and Intervention Approach*. arXiv:2605.10804 [cs.AI]. 2. Gratch, J., et al. (2014). *The Distress Analysis Interview Corpus of Human and Computer Interviews*. LREC. 3. Kroenke, K., et al. (2009). *The PHQ-8 as a Measure of Current Depression in the General Population*. Journal of Affective Disorders.