Research note /
What trustworthy mental-health AI requires
Reliability, causal sensitivity, and interpretation matter alongside predictive performance.
Mental-health applications create a demanding test for machine learning. A useful system must do more than predict labels: it should behave consistently when irrelevant details change, respond appropriately when meaningful evidence changes, and make its reasoning inspectable.
That motivates two complementary research directions in this portfolio. AttentionDep studies interpretable severity assessment through domain-aware attention. The causality-aware LLM study examines attention and domain reasoning in mental-health classification, including the limitations of treating attention as evidence of causality.
Trustworthiness is not a single metric. It is an evaluation program covering robustness, transparency, and the limits of deployment. A plausible explanation is not a clinical validation, and a predictive association is not a causal finding.