Emotion-informed associative learning in large language models for mental health detection from speech

Early detection of mental health risk in non-clinical, non-help-seeking populations is a persistent challenge. We investigate large language models (LLMs) for identifying individuals at high risk of mental illness using standardized speech tasks. LLMs perform poorly on spontaneous, non-declarative speech due to limited self-disclosure and the lack of suitable training data. To address this, we propose the Latent Emotion Proxy for Mental Health Detection with AI-integrated Decision-making (LEMon-AID), a multi-step prompting framework that leverages LLMs’ emotion profiling capabilities, integrated with output from machine learning models. This enables LLMs to link emotional features with mental health labels, significantly improving AUC from 0.55 to 0.67. While more improvement is needed, this novel hybrid design introduces layered reasoning that enhances performance while improving explainability and protecting data privacy. Our results demonstrate that emotion-aware prompting, augmented with cross-model integration, improves early risk detection and LLM interpretability.