Clinically Traceable Language Models for AI-Assisted Depression Screening in Digital Healthcare

Abstract

Digital mental health services increasingly depend on language interfaces that can interpret patient narratives, triage risk, and support timely clinical referral. Large language models are attractive for this task because they can read colloquial self-reports, infer symptom patterns, and produce natural explanations. However, depression screening is a high-stakes use case in which an unsupported statement, hallucinated diagnostic criterion, or opaque recommendation can cause clinical and ethical harm. This article develops a clinically traceable language-model framework for AI-assisted depression screening in digital healthcare. Drawing on the core logic of retrieval-augmented and agent-orchestrated diagnosis described in the source manuscript, we reframe the task as evidence-grounded screening rather than autonomous diagnosis. The proposed framework separates patient-text interpretation, symptom mapping, clinical-knowledge retrieval, evidence-ledger construction, model reasoning, and safety review into auditable modules. We introduce a mathematical traceability model linking a screening decision to retrieved evidence, define an evidence coverage index and a contradiction penalty, and evaluate a reconstructed public-data experiment comparing direct prompting with traceable prompting across four open language models. The analysis shows that traceable language models improve precision and overall reliability while making the basis of each recommendation inspectable. The paper argues that clinical traceability should be treated not as an optional explanation layer but as a core design requirement for mental-health language systems.

Keywords: Large Language Models; Depression Screening; Clinical Traceability; Retrieval-Augmented Generation; Digital Healthcare; AI Safety