A new prompting framework improves how large language models spot mental health risk in everyday speech, moving AUC from 0.55 to 0.67 in a study published in npj Mental Health Research.
Plain LLMs struggle with spontaneous, non-declarative speech from people who are not seeking help and rarely self-disclose. The LEMon-AID system relies on multi-step prompts that pull in LLM emotion profiling. It then blends in predictions from conventional machine learning models.
Speech is first mapped for emotional cues. Those cues then connect to mental health labels through layered reasoning steps. The hybrid structure improves accuracy and, the authors argue, explainability, since clinicians can see why a person was flagged. It also protects privacy, because the pipeline can operate on local or de-identified audio-derived features rather than raw recordings.
The study targets an unmet need: most mental health screening happens after people seek help, while early risk detection in the general population lags. Speech tasks offer a scalable, low-cost signal. An AUC of 0.67 is still modest, and the authors say more work is needed before deployment. But the design points to a practical path, emotion-aware prompting plus cross-model integration, that could make LLM-based screening more reliable and more transparent at the same time.
