The French general population cohort contained 1809 assessment visits from 1338 French-speaking participants (1161 contributed a single visit and the remainder 2 to 6 visits, with 1 participant assessed 16 times). The unit of analysis is the per-question transcript. As some participants contributed repeated visits, a one-visit-per-participant sensitivity analysis (Multimedia Appendix 2) leaves the associations between cluster membership and clinical scores essentially unchanged. They responded to open-ended questions, including the following: “describe your last 24 hours,” “describe a negative event that happened to you in the past,” “describe a positive event that happened to you in the past,” “describe a negative event or situation you think might happen in the future (wk, mo, or y),” “describe a positive event or situation you think might happen in the future (wk, mo, or y),” and “describe how you are feeling at the moment and how your sleep has been lately.”

Self-administered questionnaires assessed participants’ mental health status, including depressive symptoms using the Patient Health Questionnaire-9 (PHQ-9) [38] and the Beck Depression Inventory (BDI) [39]; anxiety using the Generalized Anxiety Disorder 7-item scale (GAD-7); insomnia using the Athens Insomnia Scale (AIS) [40]; and fatigue using the Multidimensional Fatigue Inventory (MFI) [41]. Using the PHQ-9 score, we obtained a control group (PHQ-9 ≤9) and a probable-depression group (PHQ-9 ≥10). PHQ-9 was at or above 10, the conventional threshold for probable depression, in 290 of 1809 (16%; 221 of 1338, 16.5% participants at first visit) assessments.

Participants completed a series of speech tasks and self-administered questionnaires through a mobile research app specifically designed for clinical studies.

The Androids cohort was specifically designed for automatic depression detection from speech [42], and includes 228 recordings from 116 native Italian speakers, collected across 2 tasks (a spontaneous interview task and a read-speech task). We analyzed the spontaneous interview task only. After removing duplicates and transcripts shorter than 20 characters, 110 transcripts entered the clustering analysis (Multimedia Appendix 1). Of the 116 participants, 64 participants were clinically diagnosed with depression by professional psychiatrists, providing a reliable characterization of individuals with depression compared to self-reported assessments. The interview task involved answering questions about everyday life (eg, “What did you do last weekend?”).

Participants were native Italian speakers recruited across 5 Mental Health Centers in Italy. Depression was diagnosed by psychiatrists according to DSM-5 (Diagnostic and Statistical Manual of Mental Disorders [Fifth Edition]) criteria, spanning major depressive disorder (MDD), bipolar disorder in a depressive phase, and related depressive conditions [42]. The Androids corpus provides this binary clinical diagnosis (depressed vs control). Controls reported no prior mental health history and were selected to match the patient group on age, gender, and education.

The multi-modal open dataset for mental-disorder analysis (MODMA) cohort [43] included 52 Chinese-speaking participants, with 23 outpatients diagnosed with MDD (16 males and 7 females, aged 18‐52 years) and 29 healthy controls (20 males and 9 females, aged 19‐52 years). MDD diagnoses were confirmed by clinical psychiatrists at Lanzhou University Second Hospital based on the Mini-International Neuropsychiatric Interview [44] and DSM-IV (Diagnostic and Statistical Manual of Mental Disorders [Fourth Edition]) diagnostic criteria [45].

Speech samples consisted of responses to 18 interview questions derived from DSM-IV and the Hamilton Depression Rating Scale [46]. Following the dataset design [43], these questions are grouped by emotional valence into 6 positive, 6 neutral, and 6 negative questions. The analysis reported here and shown in Figure 1 uses the neutral-valence questions, everyday topics such as “describe one of your friends, including age, job, character, and hobbies.” For each participant, the answers to the questions within a valence group were concatenated into a single transcript before clustering. The positive-valence questions (such as “What is the best gift you have ever received, and how did you feel?”) and the negative-valence questions (such as “What makes you desperate?”) were analyzed separately (Multimedia Appendix 3).

Inclusion criteria included a PHQ-9 score above 10 and no psychotropic drug treatment in the 2 weeks before participation. Healthy controls were recruited through public advertisements, with exclusion criteria that ruled out personal or family histories of mental disorders.

The Spanish clinical cohort for suicide-risk research (VOCES) cohort [47] comprised 90 Spanish-speaking adults who responded to open-ended questions such as: “Can you recall some recent good news you had and how did that make you feel?” “Do you get a characteristic feeling when you’re sad or down, and what do you normally do to cheer yourself up?” and “Describe your average day.”

Suicidal risk was assessed by professional psychiatrists using the validated Spanish version of the Columbia Suicide Severity Rating Scale (C-SSRS) [48,49]. Each participant’s summed C-SSRS severity score ranged from 0 to 21 in this cohort (mean 9.2, SD 7.5). For analysis, participants were dichotomized into a suicidal-risk group, defined as any positive C-SSRS finding (a score of 1 or more, indicating suicidal ideation or behavior), and a no-risk group (a score of 0). This binary risk class is the outcome tested in Figure 2. A total of 60 of the 90 (67%) participants were in the suicidal-risk group.

Inclusion criteria required participants to be at least 18 years of age, fluent in Spanish, and capable of understanding and signing the informed-consent form. Psychiatric outpatients were recruited consecutively during routine consultations, and healthy volunteers were enrolled through snowball sampling among mental-health staff and university students.

We adopted a data-driven pipeline that clusters transcripts based on their semantic similarity using multilingual language models (Figure 3). No machine translation was performed as transcripts were analyzed in the original language. Hyperparameters for each step are available in Multimedia Appendix 5.

We transcribed all speech recordings with Whisper large-v3-turbo, a state-of-the-art neural model for automatic speech recognition [50]. We removed duplicates and small transcripts (less than 20 characters).

To obtain multilingual semantic representations, each transcript was mapped to a 4096-dimensional vector (also called an embedding) using Qwen3-Embedding-8B, a language model for text embedding [51]. We selected this model because, as of June 2025, it ranked first on the multilingual track of the Massive Text Embedding Benchmark [52], covering all 4 languages of our cohorts, and because it is released under a permissive open-weight (Apache-2.0) license, which allowed us to run the model ourselves, a requirement for processing sensitive clinical speech without transmitting data to any third-party or commercial service. The model turned every transcript into a point in a vector space: transcripts dealing with similar issues have similar vectors; hence, the clustering step can identify clusters that share the same topics.

Dimension reduction transforms high-dimensional embeddings into a lower-dimensional representation that approximately preserves their geometric structure, thereby simplifying clustering and visualization. We projected the embeddings to a 5D space using UMAP (Uniform Manifold Approximation and Projection) [53]. Reducing embeddings with UMAP before density-based clustering is standard practice in modern topic modeling and is the default pipeline in BERTopic [21]. This step facilitates clustering, as we reduced the search space from 4096 to 5 dimensions. The number of components is not a sensitive design choice: the clinical associations are essentially unchanged when it is varied from 2 to 10. The choice of reduction method does matter. Replacing UMAP with a linear (principal component analysis) or alternative nonlinear (t-distributed stochastic neighbor embedding) reduction recovers weaker associations, because neither preserves the local neighborhood density that density-based clustering relies on (Multimedia Appendix 4). A neighborhood-preserving nonlinear reduction such as UMAP is therefore an important part of the method, whereas its exact dimensionality is not.

Clustering consisted of finding groups of patients with semantically similar transcripts. These transcripts may contain common topics that are clinically relevant. We extracted clusters using HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications With Noise) [54], an algorithm that (1) requires no a priori specification of the number of clusters, (2) adapts to variable-density clusters, and (3) labels ambiguous points as noise.

We automatically labeled each cluster by randomly sampling 30 transcripts and asking a language model to summarize them, focusing on what is common between the transcripts. This sample size was chosen because (1) descriptions generated from 30 transcripts are reproducible across resampling, repeated generation, and prompt wording (Multimedia Appendix 6), and (2) it approaches the maximum input that fits within the model’s context window (40,960 tokens for Qwen3-14B). As clustered transcripts are semantically similar by construction, 30 samples typically capture the cluster’s core themes without redundancy. The model received only the transcripts, so the generated descriptions are produced independently of any clinical score or demographic variable. As with the embedding model, we required an open-weight model under a permissive license, so that sensitive clinical speech was never sent to a third-party or commercial service. We used Qwen3-14B [55], an instruction-tuned 14-billion-parameter model from a leading open-weight family, with strong multilingual coverage and a context window large enough to take the 30-transcript input in a single pass. We automatically obtained textual descriptions of clusters that contain several topics. This generative approach yields fine-grained topic descriptions automatically and at scale, going beyond keyword-based descriptions, which stay coarse, and manual coding, which does not scale. The prompt is provided in Multimedia Appendix 5.

A board-certified psychiatrist (one of the coauthors) read the generated cluster descriptions as a coherence check, a qualitative reading rather than a clinical validation.

Each transcript is associated with a clinical score or diagnosis, such as a PHQ-9 score or a clinician-assigned depression diagnosis. Each cluster is therefore associated with a distribution of clinical scores. We performed statistical tests to determine whether there are statistical differences between the clusters based on their clinical scores. Transcripts that HDBSCAN labeled as noise were excluded from these cluster-level tests.

For continuous scores such as PHQ-9, we performed a Kruskal-Wallis H test to determine whether the clusters follow the same distribution or not. If we rejected the null hypothesis, we performed a pairwise Dunn test (with Benjamini-Hochberg corrections) to find significant differences between clusters. We reported effect sizes using eta-squared (η²=(H−k+1)/(n−k), where H is the Kruskal-Wallis statistic, k the number of clusters, and n the number of transcripts), interpreted following Cohen conventions: 0.01=small, 0.06=medium, and 0.14=large [56]. We also report adjusted P values.

For categorical outcomes such as a clinician-assigned depression diagnosis, we tested whether the distribution of the outcome differed across clusters using a chi-square test of independence. If the omnibus test was significant, we performed pairwise chi-square tests with Benjamini-Hochberg corrections. We report effect sizes using Cramér V and adjusted P values. For the small clinical cohorts, where sparse expected cell counts can violate the chi-square large-sample assumption, we also report the expected counts and confirm each principal association with an assumption-free Monte-Carlo permutation chi-square (Multimedia Appendix 7).

We applied 2 Benjamini-Hochberg false-discovery-rate corrections. Within each significant omnibus test, the pairwise comparisons (pairwise Dunn or pairwise chi-square) were corrected together as one family. Across the question-by-clinical-score grid in Multimedia Appendix 3, the omnibus P values were corrected jointly across all cells within a cohort. These omnibus tests and effect sizes are exploratory rather than prespecified: this study was designed to identify which questions and clinical constructs yield clusters associated with clinical status, not to test a small set of hypotheses fixed in advance.

We report effect sizes from the primary clustering, the single pipeline run on which all figures and cluster descriptions are based. Where reclustering is unstable, as for the Spanish suicide-risk association, we instead report the seed-averaged value and label the finding exploratory. Multimedia Appendix 4 shows these associations are robust to the clustering seed and reports a bootstrap 95% CI for every effect size stated in this paper. These CIs are conditional on the reported clustering: they quantify the sampling uncertainty of an effect size for a fixed partition and do not capture the additional uncertainty from rerunning the clustering or from choosing which interview questions to report. Uncertainty from reclustering is characterized separately by the seed-stability analysis (Multimedia Appendix 4), and the choice of questions and scores is exploratory rather than prespecified. The CIs are therefore widest, and read most cautiously, in the small clinical cohorts. Throughout the Results section, a parenthetical range accompanying an effect size is its 95% CI (for example, η²=0.19, 0.17 to 0.24).

All 4 cohorts used in this work had previously obtained informed consent from participants and ethical approval for speech data collection and secondary analysis.

All computational processing used open-weight models (Whisper large-v3-turbo, Qwen3-Embedding-8B, and Qwen3-14B) that the authors ran themselves on controlled computing infrastructure. No transcripts, audio recordings, or derived data were transmitted to any third-party or external commercial API service at any stage of the pipeline. Data were stored without directly identifying information.

All participants provided informed consent in accordance with the Declaration of Helsinki, Good Clinical Practice guidelines, and local regulations. This study received approval from the French National Institutional Review Board (identifier 23.00748.OOO2L7#I). Data was securely stored without identifying information, and participants received a gift card for their time.

All participants volunteered and provided written informed consent in accordance with Italian and European Union data-protection law. The protocol was approved by the Ethics Committee of the Department of Psychology, Università degli Studi della Campania “Luigi Vanvitelli” (protocol 09/2016).

All recordings were collected under ethical guidelines approved by the Ethics Committee of the Second Affiliated Hospital of Lanzhou University. Participants provided written informed consent and received approximately US $16 as compensation.

All procedures were approved by the Ethics Committee of the University Hospital Fundación Jiménez Díaz (PIC128-21_FJD) and complied with the Declaration of Helsinki. Participants provided written informed consent, and this study involved no costs or monetary compensation.

We set out the boundaries of this study and how they shape its interpretation.

The clinical cohorts are modest in size (n=52‐116). We therefore tested cluster stability directly: a stability analysis across random seeds and minimum cluster sizes, covering every interview question in all 4 cohorts, together with internal cluster-validity indices for each cohort’s principal clustering, is reported in Multimedia Appendix 4. The French general population cohort, while larger (n=1809), over-represents highly educated participants. Larger-scale investigations with more balanced recruitment are needed to establish population-level generalizability.

We performed the covariate-adjusted confound analysis only on the French general population cohort, as the clinical cohorts lacked sufficient sample sizes for it. In the French cohort, the association between cluster membership and clinical scores survived adjustment for age, sex, and education (Results section and Multimedia Appendix 2). Whether the same holds in the clinical samples remains open, because their size did not support the adjustment there. Extending covariate-adjusted or age-matched designs to the clinical cohorts is the next step. Future work should also incorporate demographic covariates directly into clustering.

Our cohorts used different clinical instruments: PHQ-9 (self-reported depression severity), a clinician-assigned depression diagnosis, and the C-SSRS (suicide risk). These measure related but distinct constructs, limiting direct cross-cohort comparisons. This heterogeneity reflects the reality of working with independently collected clinical datasets, and it bounds the cross-cohort comparisons to the exploratory level.

Although we used state-of-the-art speech-to-text models, transcript noise remains a potential source of bias. Mental disorders can alter speech production (eg, reduced energy and dysfluencies) [58], which may increase automatic speech recognition error rates that propagate to embeddings and clusters. We could not measure word or character error rates on our recordings, because no human reference transcripts were collected for any of the 4 cohorts. Published evaluations of the Whisper family report strong recognition accuracy for French, Italian, and Spanish and a higher error rate for Mandarin Chinese [50]. A further concern is differential error. If transcription quality varies with clinical status, transcription error could itself contribute to the associations between cluster membership and the clinical scales. We have no human reference transcripts to test this directly, and because recognition is weakest for Chinese, the Chinese cohort warrants the most caution here. Creating reference transcripts for a small subset would permit direct error-rate measurement and is a worthwhile direction for future work. Additionally, the language models used to embed transcripts and generate cluster descriptions may encode cultural or demographic biases [59]. LLMs also differ systematically in the values they prioritize in the text they generate [60]. As all cluster descriptions in this study were generated by a single model, Qwen3-14B, they may reflect that model’s value priorities rather than entirely neutral summaries of the transcripts.

The cluster descriptions are an interpretability layer, and this study makes no claim about their clinical accuracy, completeness, or relevance. Establishing that would require independent, blinded evaluation by multiple clinicians with formal interrater reliability (Cohen κ or the intraclass correlation coefficient), a distinct line of work. A board-certified psychiatrist read the descriptions as a coherence check. As all statistical associations are computed from cluster membership rather than from the descriptions, the findings do not depend on such clinical evaluation. The clusters that the pipeline recovers separate validated clinical scales in the French, Italian, and Chinese cohorts, and they are formed from the transcripts alone, without the clinical scores. The grounding of the descriptions in their transcripts (faithfulness), a comparison of each summary against its source text rather than a clinical judgment, is reported in Multimedia Appendix 9, where 39 of 40 descriptions assert no content that their transcripts do not support and their stability across transcript sampling, repeated generation, and prompt wording in Multimedia Appendix 6, where they prove highly reproducible.

Across all 4 cohorts, clusters with higher depression or suicide risk scores consistently contained descriptions of negative emotional states, functional impairment, and disrupted relationships, while lower-risk clusters emphasized positive coping strategies, supportive relationships, and engagement in meaningful activities. We report this recurrence as a preliminary cross-cohort observation, not as evidence of cross-cultural generalizability, because the 4 datasets differ in language, population, clinical instruments, and interview questions; whether these markers generalize across linguistic and cultural contexts is a hypothesis for future studies with parallel data-collection protocols.

The descriptions also surfaced content specific to individual cohorts. In the Italian cohort, participants expressed strong regional identity and affiliation with Naples. In the Chinese cohort, family communication challenges and feelings of inadequacy featured prominently in higher-risk clusters. These cases show the pipeline captures cohort-specific themes a generic analysis might miss. Whether such content reflects broader cultural patterns or the cohorts’ particular recruitment and interview design cannot be determined from these nonparallel datasets, and would require dedicated cross-cultural studies with parallel protocols [18].

We compared our generated topic descriptions with known risk and protective topics for depression studied in psychiatry research. A psychiatrist and a researcher in natural language processing reviewed the generated descriptions to relate them to risk and protective topics established in the literature.

Several clusters recapitulate established risk topics. Clusters with sleep disturbance or fatigue align with meta-analyses showing a bidirectional link between insomnia and depressive morbidity [61]. Stressor-laden clusters (stress due to unemployment or financial difficulties) track established longitudinal associations between adverse psychosocial work environments or job loss and increased depression risk [62]. Student-specific clusters may contain topics such as examination pressure that show associations between academic stressors and depressive symptoms [63].

Other clusters map onto protective topics. Clusters emphasizing arts participation and creative activity correspond to the WHO scoping review concluding that arts engagement contributes to prevention and symptom reduction in mental health [64]. Gardening and nature exposure clusters show consistency with evidence of small-to-moderate improvements in depressive symptoms from horticultural and green-space activities [65]. Mentions of holidays or travel appear predominantly in low-severity clusters and are congruent with longitudinal work showing short-lived but reliable gains in mood and well-being during vacations [66]. Finally, physical activity topics align with prospective meta-analyses, indicating that even subguideline levels of activity are associated with lower incident depression [67].

Taken together, the fine-grained topics generated by the language model are consistent with core psychiatric constructs established in the literature, a qualitative correspondence rather than a statistical one. We found risk topics (eg, related to sleep disruption, workload, and financial stressors) and protective topics (eg, related to arts, nature, holidays as short-term relief, and physical activity) in a multilingual, data-driven fashion that yields human-readable descriptions while scaling beyond manual thematic analysis.

Prior work on topic modeling in depression has relied on keyword-based approaches that surface topics as word lists. For instance, the study by Zhang et al [28] applied BERTopic to free-response speech in the RADAR-MDD dataset and identified 29 topics represented by keywords such as "daughter," "son," "parents," "hospital," "operation," "pain," "sleep," "tired," and "rest." While such keywords indicate broad thematic areas, they lack the semantic context and narrative coherence that clinicians use when interpreting patient discourse. That study is the closest precedent to our work. It analyzes a single English-language dataset, represents each topic as a keyword list, and relates the topics to Patient Health Questionnaire-8 depression scores. Our study builds on this line of work and broadens it to 4 cohorts in 4 languages, including a large general-population sample and 3 clinical cohorts, and to several clinical constructs (depression, anxiety, insomnia, fatigue, and suicide risk).

Our approach produces qualitatively different outputs and requires no manual post hoc categorization: the language model directly produces interpretable summaries, whereas keyword-based approaches typically require researchers to manually group and label topics. To make this concrete on our own data, we labeled the same clusters in 2 ways, with keyword labels from a standard keyword-based topic model (BERTopic, which ranks the words most distinctive to each cluster after removing common stopwords in each language) and with our language-model descriptions. We hold the clustering fixed, so every statistical association is unchanged and only what a reader can take from the label differs. The descriptions state the clinical character in a single readable sentence and disambiguate context a bare keyword cannot, whereas keyword lists mix the theme with idiosyncratic or generic words that carry no clinical content. For example, the most depressed cluster in the Chinese cohort carries the keyword label "feeling," especially, "think," "before," "now," "nothing," "seem," "do not want," "life," which conveys little clinical content, whereas its language-model description names physical health issues and emotional struggles such as "depression," "anxiety," and "loss of motivation." In the Spanish sadness-and-coping cohort, the description reports self-inflicted physical pain as a maladaptive coping strategy that the cluster’s keywords ("sad," "bad," "to feel," "sensation," "going out," "music," "week," "news," and "Madrid") do not surface. This contrast is not specific to our pipeline. Evaluations of topic models report that keyword-coherence metrics capture surface word statistics rather than the semantic quality a human reader needs, whereas language-model readings of topics align with human judgment [68,69]. The full side-by-side examples for every cohort, with the descriptions reproduced verbatim, are provided in Multimedia Appendix 8.