LLM-clustered spontaneous speech was linked to depression and related symptom burden across 3 multilingual cohorts—French (n=1338), Italian (n=116), and Chinese (n=52)—with the strongest signal in the French sample, where PHQ-9 scores differed across clusters with a large effect (η²=0.19, 95% CI 0.17-0.24). In this study of 4 cohorts, the “feelings and sleep” prompt produced the strongest clinical associations, while past- or future-focused questions had small effects.

Why It Matters To Your Practice

• Depression remains underdiagnosed, and clinicians often depend on subjective interpretation of patient speech.

• This approach suggests AI may help scale speech-based phenotyping by turning open-ended responses into human-readable clusters tied to validated symptom scales.

• Because the signal appeared across multiple languages, the method may be relevant to diverse practice settings.

Clinical Implications

• Not all interview questions are equally informative: prompts about feelings and sleep showed the strongest associations with clinical status.

• In the Italian cohort, cluster membership aligned with clinician-assigned depression diagnosis (Cramér V=0.55, 95% CI 0.36-0.78; P=.001).

• Potential near-term uses include screening support and structured intake augmentation, but prospective validation for treatment planning or decision support is still needed.

Insights

• The model used unsupervised clustering of multilingual speech transcripts, then generated natural-language summaries of each cluster with an LLM.

• Associations extended beyond depression to anxiety, insomnia, fatigue, and in one smaller cohort an exploratory suicide-risk signal.

• Sociodemographic factors also shaped cluster membership—especially age (η²=0.27) and sex (Cramér V=0.22) for the “last 24 hours” question—highlighting a key confound for clinical AI tools.

The Bottom Line

• AI-based clustering of spontaneous speech may help clinicians detect depression-related patterns at scale, but question design matters and demographic confounding is real.

• For now, this looks more like a promising workflow aid than a practice-ready diagnostic tool.