Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning

Researchers introduce Dep-LLM, a training-free framework that diagnoses depression from clinical interviews by decomposing dialogue into structured themes and using large language models without fine-tuning. The system outperforms supervised approaches and commercial LLMs while requiring no additional training, addressing critical gaps in mental health AI deployment.

Dep-LLM represents a meaningful advancement in computational psychiatry by solving two persistent obstacles in automated depression detection: handling sparse diagnostic signals within lengthy, multi-topic interviews and eliminating expensive training requirements. The framework operates entirely on frozen foundation models, making it immediately deployable across organizations without computational overhead or privacy concerns from retraining. This training-free approach holds particular significance for healthcare systems with strict data governance requirements and limited ML infrastructure. The method employs clinical reasoning principles by decomposing interviews into five psychiatrist-aligned themes, then quantifying confidence at the token level to filter unreliable signals—mimicking how human clinicians weight diagnostic evidence. Performance gains across 21 foundation models and 9 evaluation metrics suggest robust generalization rather than optimization toward specific benchmarks. The work addresses a genuine healthcare bottleneck: depression detection from clinical conversations requires domain expertise that supervised systems struggle to learn from scarce labeled data. By leveraging the reasoning capabilities already embedded in large language models, Dep-LLM bypasses the traditional ML pipeline entirely. For healthcare providers and mental health technology companies, this reduces barriers to deploying AI-assisted diagnosis tools while maintaining clinical validity. The framework's success on both DAIC-WOZ and E-DAIC datasets indicates practical applicability across different interview settings. However, real-world deployment success depends on integration with clinical workflows and validation with practicing psychiatrists before scaling to production environments.

• →Dep-LLM achieves depression diagnosis without training by structurally decomposing clinical interviews into five clinical themes and applying confidence-weighted reasoning.
• →The framework surpasses supervised domain-specific models and commercial LLMs on 21 different foundation models without fine-tuning, suggesting broad applicability.
• →Confidence analysis at the token level filters unreliable diagnostic signals, improving accuracy while maintaining transparency in clinical decision-making.
• →Training-free deployment eliminates privacy risks from retraining on sensitive mental health data and reduces computational costs for healthcare organizations.
• →The approach mirrors psychiatric reasoning methodology, enhancing clinical acceptability compared to black-box machine learning systems.