Automated speech content analysis to detect depression with large language models: towards multilingual and few-shot capabilities

Automated speech content analysis to detect depression with large language models: towards multilingual and few-shot capabilities | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Automated speech content analysis to detect depression with large language models: towards multilingual and few-shot capabilities Rachid Riad, Alexandre Ducorroy, Sélim Benjamin GUESSOUM, Filomène ROQUEFORT, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6594999/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Large Language Models (LLMs) offer potential solutions for scalable depression detection across diverse populations. This study evaluates LLM-based speech content analysis for multilingual depression detection in clinical and general populations. We analyzed speech transcripts from three distinct cohorts: Chinese clinical (n = 52), Italian clinical (n = 116), and French general population (n = 1,347). Our LLM-based system, using state-of-the-art open source LLM-model with few-shot prompting, was compared against traditional audio embedding and text embedding approaches for detecting depression and secondary symptoms (anxiety, insomnia, fatigue). The LLM system achieved excellent depression detection with F1-scores of 0.96 (Chinese), 0.85 (Italian), and 0.40 (French), consistently outperforming baseline methods. Depression sensitivity reached 1.00 (Chinese) and 0.93 (French), with high specificity in clinical populations (0.93 Chinese, 0.88 Italian). For secondary symptoms, anxiety detection performed well with high sensitivity (0.85 Chinese, 0.97 French) and F1-scores of 0.78 (Chinese) and 0.31 (French), while performance varied for other symptoms with fatigue detection performing at near-random levels. Statistical analysis revealed language-dependent benefits from few-shot learning, with Chinese datasets particularly benefiting from additional examples when using larger models. Our findings demonstrated that LLM-based speech analysis provides robust multilingual capabilities for depression detection without requiring language-specific training data, offering a scalable solution for mental health screening across diverse populations. Health sciences/Diseases/Psychiatric disorders/Depression Health sciences/Biomarkers speech large language models depression multilingual anxiety insomnia fatigue machine learning few-shot learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Major Depressive Disorder (MDD) affects 30 to 40% of individuals at some point in their lifetime, across genders and regions worldwide 1 . Depression is one of the leading contributors to global disability, associated with decreased functioning, and premature mortality. Despite its prevalence and profound impact, depression remains underdiagnosed and undertreated worldwide, highlighting the need for greater recognition and innovative approaches to care 2 . In the general population, there is a need to identify depression to guide individuals towards mental healthcare. In clinical populations, there is a need for objective assessment, follow up and prediction of patient outcomes 3 . The current standard of care relies on subjective clinical assessments of mental health, constrained by the limited number of trained clinicians, limited time, and economic resources. In most countries, there is a lack of trained clinicians to address population needs for mental healthcare 4 , 5 . In clinical practice, complementary exams are mostly unhelpful for depression management, despite the important efforts of research in imagery, biology, and genetics. Yet, measurement-based care with validated clinical scales has been proven useful for better and faster response and remission of depressive patients 6 , 7 . These challenges necessitate acceptable, cheap, easy-to-implement, scalable, and non-invasive methods to identify, assess, and monitor depressive and other mental health symptoms. Among these, automated analysis of speech, a widely available and minimally intrusive medium, has emerged as a key modality for the evaluation of mental health conditions 8 – 10 . Psychiatry practice is based on patient-doctor social interaction. Until now it has been based on clinicians' analysis of patient's behavior and thoughts, through their speech, expressivity, and communication. Developing speech analysis tools is consistent with the observation that speech analysis in psychiatry has been irreplaceable with biological or imagery exams. Speech provides a rich source of information about the speaker’s psychological and cognitive state, offering both acoustic and linguistic markers for assessment 11 . Historical observations, such as those by Kraepelin in 1921 12 , described speech abnormalities in mood disorders as “slow, hesitating, monotonous, sometimes stuttering,” laying the foundation for modern computational methods. Clinicians assess depression by focusing on some symptoms that include: monotonous speech, decrease in speech speed, and content of speech encompassing negative perception of self, past, future, and the others. Acoustic features such as prosody, pitch, intensity, and temporal dynamics have proven valuable for identifying emotional and cognitive disturbances 13 , 14 . For instance, the Geneva Minimalistic Acoustic Parameter Set (GeMAPS) has been widely adopted for detecting depression or neurological conditions 15 . The increase in dataset size and the improvements of techniques allowed the use of deep learning to further enhance audio-based assessments, enabling the automatic discovery of complex acoustic patterns and improving diagnostic accuracy 16 , 17 . These speech excerpts can be elicited from simple, semi-structured tasks, such as narrating daily events or reflecting on recent challenges, which are neither taxing for patients nor difficult for practitioners to administer. Linguistic analysis complements these acoustic approaches by capturing self-referential language, emotional valence, and syntactic complexity, all of which are known to correlate with depressive state 18 – 20 . Predictive modeling studies have identified specific linguistic markers of depression 21 , 22 , which can be combined with audio inputs for more comprehensive assessment 23 . These approaches have also been successfully implemented in longitudinal trial designs to monitor changes over time 24 . Multilingual studies are particularly important to ensure that speech-based tools can be widely disseminated and adapted to diverse cultural and clinical contexts 8 , 20 . Depression is a universal condition, yet its manifestations vary significantly across languages and cultural settings 2 . While these acoustic and linguistic approaches have shown promise, they face significant limitations. Each new application, clinical scale, or language requires substantial parallel data between validated clinical scales and speech samples in the target language 25 , 26 . This necessity for extensive language-specific datasets creates barriers to global clinical implementation, particularly in regions with limited research resources. The development process typically requires collecting, annotating, and analyzing large corpora of clinical speech data – a process that is time-consuming, expensive, and difficult to scale across multiple languages and cultural contexts 27 . Large Language Models (LLMs) have emerged as a potential solution to these challenges. As state-of-the-art approaches in natural language processing, LLMs are trained on a vast corpora of multilingual text data and demonstrate remarkable zero-shot and few-shot capabilities 28 , 29 , even in health-care 30 . These capabilities enable LLMs to perform tasks with minimal or no task-specific training examples, potentially addressing the data scarcity issues that plague traditional acoustic and linguistic analysis methods 31 . Such models significantly reduce the effort required to train and deploy systems, particularly in multilingual applications where large-scale clinical trials are often impractical. Recent research has begun exploring LLMs for mental health applications with promising results. Wu et al. 32 introduced the CALLM model, which leverages LLMs to generate and extend datasets for mental health analysis. Similarly, the MentaLLaMA approach 33 has demonstrated the capability to tackle diverse mental health tasks based on social media posts. Other researchers have used LLMs to generate synthetic training examples to improve traditional machine learning models for mental health applications 32 , 34 These approaches suggest that LLMs can augment existing methods by addressing data scarcity issues while maintaining clinical validity. Despite these promising developments, significant challenges remain in applying LLMs to mental health care. As highlighted in a recent viewpoint in Lancet Digital Health 35 , while LLMs offer potential solutions to treatment gaps and research limitations, issues concerning technical costs, literacy barriers, algorithmic biases, and inequitable data representation must be addressed. Our work directly contributes to this framework by addressing two key priorities: culturally-informed development through multilingual applications and refined diagnostic approaches through specific LLM outputs for each target condition. Building on these developments and addressing the identified challenges, our study aims to evaluate the potential of LLM-based speech content analysis for mental health assessment across languages and populations. Specifically, the main objective is to assess LLM-based speech content analysis models in detecting depression in diverse linguistic contexts (French, Chinese, and Italian), in both clinical and general populations. The secondary objective is to identify anxiety, fatigue, and insomnia in the same populations. Methods This study employed a novel methodological framework to assess the effectiveness of Large Language Models (LLMs) for detecting mental health conditions through speech analysis across multiple languages. Our approach combines state-of-the-art automatic speech recognition with prompt-engineered LLMs that function as direct classifiers, reducing the need for language-specific training data that traditional methods require. We first explain our study design and corpora, then data preprocessing pipeline, we also detail both traditional baseline approaches and our innovative LLM-based system, followed by our implementation of prompt engineering and statistical analysis methods. This methodology was applied consistently across three diverse linguistic and clinical contexts. Study Design and Participants The study is based on three distinct cohorts: a large general population sample collected in French (n = 1347), and two specialized clinical populations collected in Italian (n = 116) and Chinese (n = 52) languages. The Italian and the Chinese corpus are publicly available corpuses for research uses. In Fig. 1 .A, Fig. 1 .B, and Fig. 1 .C, we present the study populations, symptoms assessed through questionnaires, and co-occurrence of symptoms in patients. Chinese Clinical Population (MODMA Corpus) The MODMA corpus 36 includes 52 Chinese-speaking participants with a mean age of 31.3 years (SD = 9.2, range 18–52). The sample comprised 36 males and 16 females, with education levels distributed as: no diploma (n = 7), secondary education (n = 8), and long-cycle higher education (n = 37), with no participants reporting short-cycle higher education. The dataset included 23 outpatients diagnosed with Major Depressive Disorder (MDD) (16 males and 7 females, aged 16–56) and 29 healthy controls (20 males and 9 females, aged 18–55). MDD diagnoses were confirmed by clinical psychiatrists at Lanzhou University Second Hospital based on the Mini-International Neuropsychiatric Interview (MINI) 37 and DSM-IV diagnostic criteria 38 . Inclusion criteria for patients included a PHQ-9 score above 10 and no psychotropic drug treatment in the two weeks prior to participation. Healthy controls were recruited through public advertisements, with exclusion criteria that ruled out personal or family histories of mental disorders. Speech samples consisted of responses to 18 questions derived from DSM-IV and extracted from depression scales 39 , including questions such as: "What is the best gift you have ever received, and how did you feel?", "Please describe one of your friends, including age, job, characters, and hobbies," "What would you like to do when you are unable to fall asleep?", and "What makes you desperate?" 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 $ 16 compensation. Italian Clinical Population (Androids Corpus) The Androids Corpus was specifically designed for automatic depression detection from speech 40 . It includes 228 recordings from 116 native Italian speakers with a mean age of 37.4 years (SD = 12, range 19–71). The gender distribution was 32 males and 84 females. Educational levels included: no diploma (n = 11), secondary education (n = 37), short-cycle higher education (n = 52), and long-cycle higher education (n = 16). Of the 116 participants, 64 were clinically diagnosed with depression by professional psychiatrists using the Montgomery-Asberg Depression Rating Scale (MADRS) 41 , providing a reliable characterization of depressed individuals compared to self-reported assessments. The interview task involved answering questions about everyday life (e.g., "What did you do last weekend"). The corpus contains both elicited and spontaneous speech samples, with approximately 9 hours of total recording time. For this study, we focused exclusively on spontaneous speech productions. French General Population Dataset This dataset comprises data from 1,347 French-speaking participants recruited from the general population in France. Participants had a mean age of 37.8 years (SD = 18.2, range 18.4–91.4), with 479 males and 860 females (8 participants did not disclose gender). The education level distribution was: no diploma (n = 47), secondary education (n = 218), short-cycle higher education (n = 166), and long-cycle higher education (n = 916). Participants completed a series of speech tasks and self-administered questionnaires through a mobile research application specifically designed for clinical studies. Two types of speech tasks were included: These included an elicited production task (constrained-vocabulary) where participants counted from 1 to 20, and a spontaneous speech task (open-vocabulary) where participants responded to open-ended questions such as: "Describe how you are feeling at the moment and how your sleep has been lately," "Describe your last 24 hours," "Describe a negative event or situation you think might happen in the future (week, month, or year)," and "Describe a positive event or situation you think might happen in the future (week, month, or year)." The total duration of recorded speech was 10.12 hours, sampled at 16kHz. Self-administered questionnaires assessed participants' mental health status, including depressive symptoms using the Patient Health Questionnaire-9 (PHQ-9), with scores ≥ 10 classified as moderate depression 42 ; anxiety using the Generalized Anxiety Disorder questionnaire (GAD-7), with scores ≥ 10 classified as moderate anxiety 43 ; and insomnia using the Athens Insomnia Scale (AIS), with scores ≥ 6 classified as mild insomnia 44 . All participants provided informed consent in accordance with the Declaration of Helsinki, Good Clinical Practice guidelines, and local regulations. The 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 €15 compensation for their time (See 17 for more details). The study includes clinical cohorts of Italian- and Chinese-speaking participants and a French-speaking general population cohort. Psychiatric symptoms were assessed using self-report questionnaires — including the Patient Health Questionnaire (PHQ-9), the Generalized Anxiety Disorder scale (GAD-7), the Athens Insomnia Scale (AIS), the Beck Depression Inventory (BDI), the Multidimensional Fatigue Inventory (MFI), the Mini International Neuropsychiatric Interview (MINI), and the Pittsburgh Sleep Quality Index (PSQI) — as well as clinician-administered measures such as the Montgomery-Åsberg Depression Rating Scale (MADRS). Data preprocessing, transcription and predictions of symptoms Only audio recordings were available and are used for any mental health symptoms predictions. These recordings were transcribed using the Whisper multilingual automatic speech recognition model 45 , the large-v3-turbo version, and by specifying the language. To ensure compatibility with the Whisper model, all audio files were resampled to 16 kHz before processing. For the Chinese clinical population (n = 52), individuals were classified based on depression status determined through structured clinical interviews using the Mini International Neuropsychiatric Interview [MINI] and supplemented by the Patient Health Questionnaire-9(measured via PHQ-9). Additionally, anxiety severity (Generalized Anxiety Disorder-7 (GAD-7)) and sleep disturbances (Pittsburgh Sleep Quality Index [PSQI]) were assessed. The use of both diagnostic interviews and psychometric scales provided comprehensive clinical assessment in this cohort. For the Italian clinical population (n = 116), classification focused exclusively on depression status as assessed by psychiatrists using the Montgomery–Åsberg Depression Rating Scale (MADRS). This provided a clinician-validated ground truth rather than relying solely on self-reported measures. For the French general population cohort (n = 1347), classification encompassed multiple self-reported mental health dimensions: depression severity (measured via PHQ-9), anxiety symptoms (GAD-7), fatigue levels (Multidimensional Fatigue Inventory (MFI)), and insomnia severity (Athens Insomnia Scale (AIS)). A positive classification was assigned if the participant's score exceeded clinically validated thresholds for each scale (PHQ-9 ≥ 10 for moderate depression, GAD-7 ≥ 10 for moderate anxiety, and AIS ≥ 6 for mild insomnia). Large Language Model as classifier Our system introduces a novel approach using a LLM as a classifier for mental health assessment, as illustrated in Fig. 2 . This approach differs from conventional methods that have dominated the field of automatic speech-based mental health assessment 26 , 46 . The first approach to tackle speech-based mental health assessment is to model acoustic realization from raw audio, using established feature sets such as eGeMAPS (extended Geneva Minimalistic Acoustic Parameter Set) 13 (See Fig. 2 A). These audio features extract statistics from articulatory and prosodic elements including pitch, intensity, spectral, and voice quality features. These acoustic features form a fixed-size vector representation (embeddings) that serves as input to conventional machine learning algorithms, which are trained on parallel datasets of audio features and clinical labels to predict mental health status 14 , 47 . We refer to this approach as Audio embedding with ML in Fig. 2 A and in the Results section. The second approach is to model the content of the speech production using an automatic speech recognition (ASR) system. This transcribed speech content (words) are then transformed into linguistic embeddings, fixed-size vectors, using large pre-trained transformer models such as BERT 48 (See Fig. 2 B). These embeddings capture semantic and syntactic information in fixed-size embeddings, which are then fed into conventional machine learning classifiers trained on parallel corpora of text embeddings and clinical labels 21 , 23 . We refer to this approach as Text embedding with ML in Fig. 2 B and in the Results section. We used the multilingual text embedding mpnet-base-v2 49 . For audio embedding, we used a 20-second sliding window; and for text embedding, we used a 70-word sliding window. We used as final classifier, a logistic regression and used a mean pooling to aggregate predictions. These two traditional approaches require substantial parallel data between speech samples and clinical assessments for each target language and clinical scale, creating significant barriers for multilingual and cross-cultural implementation (Low et al., 2020). Our proposed system fundamentally differs in its approach (See Fig. 2 C). We also first transcribed speech to convert audio to text across all three languages (French, Italian, and Chinese). Instead of generating fixed-size embeddings followed by a separate classifier, we employed a few-shot prompting technique with the LLM. This involves providing the model with selected examples of transcribed speech paired with their corresponding clinical classifications (e.g., "depressed" or "not depressed"). The LLM is then prompted with a new patient's transcribed speech and directly asked to classify the mental health status (e.g., "Is this patient depressed?"), returning a binary classification (Yes/No). This approach leverages the LLM's pre-trained knowledge of linguistic patterns associated with psychological states across multiple languages, potentially reducing the need for extensive language-specific training data 29 , 31 . Prompt engineering Unlike traditional machine learning classifiers, LLMs generate open-ended text rather than discrete labels, introducing variability in outputs. To mitigate this, structured prompt engineering was applied, utilizing soft constraints rather than enforcing strict response formats, as it yielded the best results. Implementation relied on the DSPy prompting framework for structured prompt generation 50 and Ollama for efficient LLM inference. We leveraged DSPy's prompt management capabilities to maintain consistent prompting patterns across languages while allowing for dataset-specific adaptations. The framework enabled us to programmatically construct complex few-shot learning contexts that included both examples and target tasks within a unified prompt structure. This systematic approach ensured reproducibility and minimized variability in prompt formatting across experimental conditions, which is particularly important when working with open-ended LLM responses in clinical applications. We used the following prompting strategy to obtain binary responses and minimize model hallucinations and biases: we sampled examples such that we respected the balance between the number of positive and negative examples in the prompt, and we did not enforce a grammar on the output. For the LLM weights, we selected the state-of-the-art Gemma open source version 3 model with 27 billions of parameters 51 , due to its performance on multilingual tasks and its large 128K token context window, which is critical for our few-shot learning approach that requires integrating multiple examples alongside patient transcripts. For comparative in-context learning analysis, we also evaluated a smaller 12 billion parameter version of the same model family. We specifically chose these two model sizes because smaller models (< 12B parameters) did not reliably support the extensive context windows. In contrast with our LLM system, the classic ML models learned on the full training set for each task. The LLM sees only the examples given in the prompts, which limits the number of examples available to perform each task. This number to be seen is limited by the maximum context window of the LLM, in our case 128K tokens. Based on our preliminary experiments and different speech transcripts across datasets, our upper limit, found empirically, for the number of examples to provide the LLM was 32. Given the heterogeneous nature of the datasets, prompts were tailored to maximize context inclusion. For the French general population corpus, a structured question-answer format was employed, compiling all responses into a single text input. Participants described their past 24 hours, significant negative and positive events from the past year, anticipated events in the coming year, their emotional state, and sleep quality. For the Italian Androids dataset, since questions were unavailable, transcripts were constructed from concatenated responses. For the Chinese Lanzhou corpus, prompts explicitly specified four question categories derived from DSM-IV, Hamilton Depression Rating Scale, and other validated questionnaires. All prompts were written in English, while participants' responses were retained in their original language, with no demographic information included in the prompts. We implemented two distinct prompting techniques: zero-shot prompting, where the LLM was prompted to classify responses without prior labeled examples; and few-shot prompting, where the model was provided with pre-labeled examples structured as a dialogue format (prompt → LLM answer) to contextualize the task to solve for the LLM. These examples were aggregated into a single prompt, including the participant's transcript for classification, along the speech transcript to classify. This approach is illustrated in Fig. 2 C. Model Training and Evaluation For the general population dataset, models were trained on 80% of the data and tested on the remaining 20%. For the Chinese and Italian clinical datasets, a five-fold cross-validation approach was employed to stay consistent with original studies, partitioning the dataset into five subsets and training on four while evaluating on the remaining one. For our LLM system all the prior examples are only extracted from the training set. Stratified sampling ensured that each subset maintained the same distribution as the full dataset, based on depression class (Chinese and Italian populations) or PHQ-9 scores (general population). Performance was assessed using the F1-score, which integrates Positive Predictive Value (PPV) and Sensitivity (Se): where an F1-score of 1 represents perfect classification, and 0 indicates complete misclassification ( F1 = 2*(PPV*Se)/(PPV + Se)). For our LLM-based system, we also reported the detailed Se, Specificity (Spe), Positive Predictive Value (PPV) and Negative Predictive Value (NPV). Statistical analysis of in-context learning To analyze the relationship between depression detection performance and the number of examples provided to the large language model (LLM), we performed extensive experiments using two different versions of the models that vary in size. We computed the F1 score as a function of the number of examples provided (0, 1, 2, 4, 8, 16, 32). To focus on comparable conditions, we utilized two subtasks involving depression detection where labels were provided by psychiatrists in two different languages (Italian and Chinese). This allowed us to examine cross-cultural consistency while maintaining clinical validity of the classifications. We ran a multiple linear regression for each language with interaction terms to quantify the relationships between the model size and the example count. The model was specified as: $$\:\:{F}_{1}\:score\:\sim\:\:example\:count+model\:size+example\:count\times\:model\:size\:$$ Results The main results of our study are presented in Fig. 3 . Overall, our LLM-based approach outperformed traditional machine learning and acoustic methods in most mental health detection tasks across languages, with particularly strong performance in depression detection. The F1-scores ranged from 0.85–0.96 for depression in clinical populations (Chinese and Italian) and 0.40 in the general population (French), compared to substantially lower scores for baseline methods. Our LLM-based system demonstrated excellent sensitivity for depression detection in both Chinese (Se = 1.00) and French (Se = 0.93) populations, with good specificity in Chinese (Sp = 0.93) and Italian (Sp = 0.88) clinical populations. Anxiety detection was also successful with high sensitivity in both Chinese (Se = 0.85) and French (Se = 0.97) populations, though with moderate specificity (Sp = 0.39 and Sp = 0.46, respectively). Detection of secondary symptoms like sleep difficulties, fatigue, and insomnia showed more variable results across populations. Depression Across all three studied populations, our LLM-based system outperformed alternative methods in detecting depression in both clinical and general populations across different languages. F1-scores for depression reached 0.96 in the Chinese-speaking dataset with perfect sensitivity (Se = 1.00) and excellent specificity (Sp = 0.93), resulting in strong predictive values (PPV = 0.92, NPV = 1.00). Similarly, in the Italian-speaking dataset, the F1-score was 0.85 with good sensitivity (Se = 0.81) and specificity (Sp = 0.88), yielding high predictive values (PPV = 0.90, NPV = 0.79). The F1-score in the French-speaking general population was lower at 0.40, despite high sensitivity (Se = 0.93) but limited specificity (Sp = 0.47), resulting in low positive predictive value (PPV = 0.26) but high negative predictive value (NPV = 0.97). In contrast, baseline methods achieved scores at random level in the French dataset, with null F1-scores. Anxiety, Insomnia, and Fatigue For anxiety detection, our system achieved an F1-score of 0.78 in the Chinese dataset with good sensitivity (Se = 0.85) but limited specificity (Sp = 0.39), resulting in moderate predictive values (PPV = 0.72, NPV = 0.58). In the French dataset, anxiety detection reached an F1-score of 0.31 with excellent sensitivity (Se = 0.97) but poor specificity (Sp = 0.46), yielding low positive predictive value (PPV = 0.18) but very high negative predictive value (NPV = 0.99). For sleep/insomnia problems, our system performed below random level in the Chinese dataset with low sensitivity (Se = 0.53) and specificity (Sp = 0.39), while achieving better results in the French dataset (F1 = 0.64) with high sensitivity (Se = 0.94) but poor specificity (Sp = 0.34). For fatigue detection in the French general population, all systems performed similarly at random level. Statistical analysis of in-context learning capabilities In Fig. 4 , we presented the F1-scores for each depression detection task according to the number of examples included in the prompt. For the sub-analysis concerning the number of examples, we focused on the severe depression tasks. We found out that the relationship between few-shot learning (example count) and depression detection performance varies substantially between languages. For Chinese, the relationship depends strongly on model size (R² = 0.781, F = 11.91, p = 0.00122), with larger models benefiting more from additional examples (interaction coefficient = 0.0007, p = 0.002). Specifically, while additional examples showed a negative effect for smaller models (coefficient = -0.0080, p = 0.049), the significant positive interaction indicates this effect reverses for larger models. For Italian, performance remains relatively stable regardless of model size and example count, with the regression model explaining only 16.4% of variance and showing no significant effects (F = 0.6561, p = 0.597). The zero-shot prompting method is already performing at a high level in Italian. This suggests that few-shot learning strategies may need to be language-specific, and that scaling to larger models may be more beneficial for some languages than others. Discussion In this study, we demonstrated that automated speech content analysis with a LLM-based system exhibits a good to excellent ability to identify depression, and good results for anxiety, insomnia and fatigue, across diverse languages and populations. Our LLM-based system consistently outperformed traditional machine learning techniques and acoustic feature analysis methods, achieving impressive F1 scores of 0.96 and 0.85 for depression detection in Chinese and Italian clinical populations, respectively. In the French general population, while performance was more moderate, our approach still surpassed baseline methods that performed at chance level. Multilingual capabilities of LLMs in mental health assessment Our work contributes to the emerging literature on LLMs for mental health support by providing additional empirical evidence and more extensive evaluations across diverse linguistic and clinical contexts 35 . Multilingual capacity and transcultural validity are essential for mental health monitoring across diverse populations. Our study demonstrates that depression can be identified reliably across multiple languages, including French, Italian, and Chinese, addressing a critical need for equitable mental healthcare in various cultural settings 52 , 53 . A model trained and evaluated on non-representative data, for example, exclusively on English-speaker Caucasian adults without comorbidities, will likely underperform in minority groups, thereby exacerbating healthcare disparities 54 . It is therefore imperative to evaluate model performance across different demographics and different languages 17 . While recent studies have shown promising results, our approach demonstrates extends to more languages and more mental health symptoms assessments. Compared to 55 who achieved an F1 score of 0.73 for depression detection in a German-speaking general population cohort, our system delivered substantially higher performance (F1 = 0.96 in Chinese and F1 = 0.85 in Italian clinical populations). Similarly, 56 explored multilingual applications, obtaining F1 scores of 84.62 for depression detection in English-speaking PTSD populations and 75.31 for depression among Chinese university students. While these results demonstrate the multilingual capabilities of LLMs, our study extends this work by showing even stronger performance across clinically validated populations in multiple non-English languages, suggesting that our approach may better capture culture-specific manifestations of depression. Unlike 57 who found no benefit from in-context learning despite strong zero-shot performance (> 90% F1-score) in analyzing diabetic patients' text messages, our results revealed language-dependent benefits from few-shot learning. The Chinese dataset particularly benefited from additional examples with larger models, suggesting that optimal prompting strategies may vary across languages (See Fig. 4 ). Additionally, our use of open-source models (< 30B parameters), contrary to Kim et al 2025, addresses practical implementation concerns, enabling single-Graphical Processing Unit (GPU) deployment and on-premise processing compatible with regional data protection regulations like GDPR. A notable important aspect of our study is the use of open-ended interview transcripts with different psychological themes and questions asked to patients. This approach parallels Cummins et al.'s ecological methods in their RADAR-MD cohort without requiring psychiatrist-administered MADRS interviews. While 8 identified acoustic correlates of depression severity across European populations and later, 20 used automatic topic modeling to extract content patterns longitudinally correlated with PHQ-8 scores. Our LLM-based approach leverages pre-trained linguistic knowledge without requiring explicit feature engineering or topic extraction. This complementarity suggests promising integration opportunities, combining LLM content analysis with acoustic features and topic modeling could enhance detection accuracy while maintaining ecological validity. Such multimodal assessment would capture both linguistic patterns and paralinguistic information, creating an easy-to-use automatic assessment tool for patients when mental health professionals may not be available at specific times. Complexity of patients and measures of clinical signs not diagnosis through LLM The complex symptom patterns usually reported in psychiatry 58 are clearly reflected in our datasets (See Fig. 1 ). In the Chinese clinical population, we observed significant overlap between depression, anxiety, and sleep difficulties, with 18 participants simultaneously experiencing all three symptom clusters. Similarly, in the French general population, nearly all possible symptom combinations occurred, including isolated presentations of insomnia without fatigue or fatigue without insomnia. This heterogeneity underscores the challenges clinicians face in accurately assessing and treating psychiatric conditions, highlighting the need for nuanced assessment tools that can detect specific symptom profiles beyond broad diagnostic categories. The variability in ground truth methodology across our datasets further emphasizes this complexity. There are different uses for rating scales for psychiatrists and researchers: as screening instruments to detect the possible presence of a disorder, as measures to establish a symptom profile, as indicators of illness severity, and as measures of drug effect in controlled trials 59 . For the French general population, we relied on self-assessment scales, which introduce potential bias due to their dependence on participant insight. In the Chinese population, we even observed discrepancies between clinician-administered MINI interviews and PHQ-9 self-evaluations, reflecting the known challenges in achieving diagnostic consistency across assessment methods. These methodological differences align with findings from 60 who found out that patient characteristics and symptoms significantly affect the correlation between observer- and self-rating scales. Younger age, higher educational attainment, and depressive subtype (atypical, non-melancholic) predicted higher self-reported scores relative to clinician ratings, while personality traits such as high neuroticism, low extraversion, and low agreeableness were associated with higher endorsement of depressive symptoms on self-reports. These discrepancies were more pronounced for psychological symptoms than somatic symptoms, reaffirming the value of multi-modal assessment in depression research. Despite these methodological variations, our LLM-based system still demonstrated robust performance across datasets for depressive aspects, suggesting potential utility in diverse clinical contexts. Importantly, our results reveal that not all symptoms are equally expressed in language (or adequately captured by our LLM-based system). While depression detection was consistently strong across languages, performance varied for secondary symptoms. Sleep difficulties in Chinese speakers showed poorer results (Se = 0.53, Sp = 0.39), and fatigue detection in the French population performed at near-random levels with very low specificity (Se = 0.94, Sp = 0.14). Yet, insomnia was detected more effectively with our LLM approach than with other methods. This discrepancy between Chinese clinical detection of sleep difficulties and better performance on insomnia in the French population suggests that symptom reporting patterns may vary between clinical and general populations, with the latter potentially providing more concrete descriptions of sleep patterns. As underlined by Martin et al. (2024), it is of prime importance of assessing specific symptom severity beyond diagnosis for effective therapeutic relationships and personalized treatment planning. While precision medicine aims to adapt interventions to individual characteristics 61 , 62 , its implementation in psychiatry has proven challenging despite expensive neuroimaging and genetic approaches yielding limited clinical utility 63 . Our findings suggest that speech-based assessments using LLMs offer a complementary, practical path toward precision psychiatry—not as a comprehensive solution for all symptoms, but as an accessible tool for specific clinical challenges. By effectively detecting depression across diverse languages and capturing certain secondary symptoms like insomnia, these methods provide clinically useful insight without the resource-intensive requirements of traditional biomarker approaches. However, the limitations in detecting fatigue and certain sleep disturbances highlight that speech-content analysis should be viewed as one component in a broader assessment framework rather than a standalone diagnostic solution. Limitations Medication use. Medication use represents another critical variable in mental health assessment that our study did not specifically address. Antidepressants and other psychotropic medications can potentially impact language patterns and disease presentation, as noted by 64 , 65 . In our study, medication data was not available, which could influence both language production and symptom manifestation. Future work should systematically account for medication type and dosage to better understand their effects on speech-based assessments. Stratification. In this study, we focus on a binary classification of the presence or absence of at least moderate depression. However, future research should aim to develop methods that allow for more nuanced stratification across multiple severity levels—such as none, mild, moderate, and severe depression. Such graduated assessments could enhance clinical applicability, particularly in monitoring patients’ partial or complete responses to treatment over time 66 . Complete speech analysis. Future developments should also consider incorporating a broader range of parameters—including linguistic content, acoustic features, and sociodemographic variables—to improve the accuracy, robustness, and contextual relevance of assessments. Conclusion Our LLM-based system demonstrates good to excellent performance in identifying depression across multiple languages in both clinical and general populations, with more variable results for anxiety, sleep problems, and fatigue. This approach leverages pre-trained linguistic knowledge without requiring language-specific feature engineering, showing particular promise for cross-cultural applications. The consistent performance across French, Italian, and Chinese populations—despite variations in assessment methodologies—highlights the potential for supporting equitable mental healthcare globally. Further developments should focus on providing more precise classifications of symptom intensity beyond binary detection and integrating multiple data streams including acoustic features and sociodemographic variables to improve robustness. Future applications could address unmet challenges in psychiatry, such as predicting medication response and detecting suicide risk through linguistic markers. While requiring careful validation, our findings establish a foundation for speech-based assessment tools that complement clinical expertise, potentially expanding access to mental healthcare while supporting more personalized treatment approaches. Declarations Conflict of interest RR, XNC, AL, MD, and AB are shareholders of Callyope, and AD, FR are former employees of Callyope. SBG is a consultant for Callyope. Acknowledgements The authors are thankful to all the participants who volunteered for this research study. Without their active involvement, this study would not have been possible. The authors also would like to thank each of the speech pathology interns who helped with the participant recruitment and made sure that the protocol was completed successfully. References Bromet, E. et al. Cross-national epidemiology of DSM-IV major depressive episode. BMC Med. 9, 90 (2011). Herrman, H. et al. Time for united action on depression: a Lancet-World Psychiatric Association Commission. Lancet Lond. Engl. 399, 957–1022 (2022). Fernández, A. et al. Is major depression adequately diagnosed and treated by general practitioners? Results from an epidemiological study. Gen. Hosp. Psychiatry 32, 201–209 (2010). R. M. Scheffler et al. Human Resources for Mental Health in Low and Middle Income Countries: Evidence from Bangladesh. in ResearchGate vol. 8 56 (2011). Kakuma, R. et al. Human resources for mental health care: current situation and strategies for action. Lancet Lond. Engl. 378, 1654–1663 (2011). Xiao, L. et al. The effectiveness of enhanced evidence-based care for depressive disorders: a meta-analysis of randomized controlled trials. Transl. Psychiatry 11, 1–8 (2021). Guo, T. et al. Measurement-Based Care Versus Standard Care for Major Depression: A Randomized Controlled Trial With Blind Raters. Am. J. Psychiatry 172, 1004–1013 (2015). Cummins, N. et al. Multilingual markers of depression in remotely collected speech samples: A preliminary analysis. J. Affect. Disord. 341, 128–136 (2023). Williamson, J. R. et al. Vocal biomarkers of depression based on motor incoordination. in Proceedings of the 3rd ACM international workshop on Audio/visual emotion challenge 41–48 (Association for Computing Machinery, New York, NY, USA, 2013). doi: 10.1145/2512530.2512531 . Cohn, J. F. et al. Detecting depression from facial actions and vocal prosody. in 2009 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops 1–7 (2009). doi: 10.1109/ACII.2009.5349358 . Weintraub, M. J., Posta, F., Ichinose, M. C., Arevian, A. C. & Miklowitz, D. J. Word usage in spontaneous speech as a predictor of depressive symptoms among youth at high risk for mood disorders. J. Affect. Disord. 323, 675–678 (2023). Kraepelin, E. Manic Depressive Insanity and Paranoia. J. Nerv. Ment. Dis. 53, 350 (1921). Eyben, F. et al. The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing. IEEE Trans. Affect. Comput. 7, 190–202 (2016). Ringeval, F. et al. AVEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition. Preprint at https://doi.org/10.48550/arXiv.1907.11510 (2019). Pérez-Toro, P. A. et al. Depression assessment in people with Parkinson’s disease: The combination of acoustic features and natural language processing. Speech Commun. 145, 10–20 (2022). Tasnim, M., Ehghaghi, M., Diep, B. & Novikova, J. DEPAC: a Corpus for Depression and Anxiety Detection from Speech. in Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology (eds. Zirikly, A. et al.) 1–16 (Association for Computational Linguistics, Seattle, USA, 2022). doi: 10.18653/v1/2022.clpsych-1.1 . Riad, R. et al. Automated Speech Analysis for Risk Detection of Depression, Anxiety, Insomnia, and Fatigue: Algorithm Development and Validation Study. J. Med. Internet Res. 26, e58572 (2024). Rude, S., Gortner,Eva-Maria & and Pennebaker, J. Language use of depressed and depression-vulnerable college students. Cogn. Emot. 18, 1121–1133 (2004). Tackman, A. M. et al. Depression, negative emotionality, and self-referential language: A multi-lab, multi-measure, and multi-language-task research synthesis. J. Pers. Soc. Psychol. 116, 817–834 (2019). Zhang, Y. et al. Identifying depression-related topics in smartphone-collected free-response speech recordings using an automatic speech recognition system and a deep learning topic model. J. Affect. Disord. 355, 40–49 (2024). Burdisso, S., Villatoro-Tello, E., Madikeri, S. & Motlicek, P. Node-weighted Graph Convolutional Network for Depression Detection in Transcribed Clinical Interviews. in INTERSPEECH 2023 3617–3621 (2023). doi: 10.21437/Interspeech.2023-1923 . Villatoro-Tello, E., Ramírez-de-la-Rosa, G., Gática-Pérez, D., Magimai.-Doss, M. & Jiménez-Salazar, H. Approximating the Mental Lexicon from Clinical Interviews as a Support Tool for Depression Detection. in Proceedings of the 2021 International Conference on Multimodal Interaction 557–566 (ACM, Montréal QC Canada, 2021). doi: 10.1145/3462244.3479896 . Al Hanai, T., Ghassemi, M. & Glass, J. Detecting Depression with Audio/Text Sequence Modeling of Interviews. in 1716–1720 (2018). doi: 10.21437/Interspeech.2018-2522 . Arevian, A. C. et al. Clinical state tracking in serious mental illness through computational analysis of speech. PloS One 15, e0225695 (2020). Cummins, N., Baird, A. & Schuller, B. W. Speech analysis for health: Current state-of-the-art and the increasing impact of deep learning. Methods 151, 41–54 (2018). Low, D. M., Bentley, K. H. & Ghosh, S. S. Automated assessment of psychiatric disorders using speech: A systematic review. Laryngoscope Investig. Otolaryngol. 5, 96–116 (2020). MacWhinney, B. & Wagner, J. Transcribing, searching and data sharing: The CLAN software and the TalkBank data repository. Gesprachsforschung Online-Z. Zur Verbalen Interakt. 11, 154–173 (2010). Bommasani, R. et al. On the Opportunities and Risks of Foundation Models. Preprint at https://doi.org/10.48550/arXiv.2108.07258 (2022). Kojima, T., Gu, S. S., Reid, M., Matsuo, Y. & Iwasawa, Y. Large language models are zero-shot reasoners. in Proceedings of the 36th International Conference on Neural Information Processing Systems 22199–22213 (Curran Associates Inc., Red Hook, NY, USA, 2022). Agrawal, M., Hegselmann, S., Lang, H., Kim, Y. & Sontag, D. Large language models are few-shot clinical information extractors. in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (eds. Goldberg, Y., Kozareva, Z. & Zhang, Y.) 1998–2022 (Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022). doi: 10.18653/v1/2022.emnlp-main.130 . Pourpanah, F. et al. A Review of Generalized Zero-Shot Learning Methods. IEEE Trans. Pattern Anal. Mach. Intell. 45, 4051–4070 (2023). Wu, Y., Mao, K., Zhang, Y. & Chen, J. CALLM: Enhancing Clinical Interview Analysis Through Data Augmentation With Large Language Models. IEEE J. Biomed. Health Inform. 28, 7531–7542 (2024). Yang, K. et al. MentaLLaMA: Interpretable Mental Health Analysis on Social Media with Large Language Models. in Proceedings of the ACM Web Conference 2024 4489–4500 (Association for Computing Machinery, New York, NY, USA, 2024). doi: 10.1145/3589334.3648137 . Liyanage, C., Garg, M., Mago, V. & Sohn, S. Augmenting Reddit Posts to Determine Wellness Dimensions impacting Mental Health. Proc. Conf. Assoc. Comput. Linguist. Meet. 2023, 306–312 (2023). Malgaroli, M. et al. Large language models for the mental health community: framework for translating code to care. Lancet Digit. Health 7, e282–e285 (2025). Cai, H. et al. A multi-modal open dataset for mental-disorder analysis. Sci. Data 9, 178 (2022). Lecrubier, Y. et al. The Mini International Neuropsychiatric Interview (MINI). A short diagnostic structured interview: reliability and validity according to the CIDI. Eur. Psychiatry 12, 224–231 (1997). American Psychiatric Association, A. & Association, A. P. Diagnostic and Statistical Manual of Mental Disorders: DSM-IV. vol. 4 (American psychiatric association Washington, DC, 1994). Hamilton, M. A rating scale for depression. J. Neurol. Neurosurg. Psychiatry 23, 56–62 (1960). Tao, F., Esposito, A. & Vinciarelli, A. The Androids Corpus: A New Publicly Available Benchmark for Speech Based Depression Detection. in 4149–4153 (2023). doi: 10.21437/Interspeech.2023-894 . Montgomery, S. A. & Asberg, M. A new depression scale designed to be sensitive to change. Br. J. Psychiatry J. Ment. Sci. 134, 382–389 (1979). Levis, B., Benedetti, A., Thombs, B. D., & DEPRESsion Screening Data (DEPRESSD) Collaboration. Accuracy of Patient Health Questionnaire-9 (PHQ-9) for screening to detect major depression: individual participant data meta-analysis. BMJ 365, l1476 (2019). Plummer, F., Manea, L., Trepel, D. & McMillan, D. Screening for anxiety disorders with the GAD-7 and GAD-2: a systematic review and diagnostic metaanalysis. Gen. Hosp. Psychiatry 39, 24–31 (2016). Okajima, I. et al. Evaluation of Severity Levels of the Athens Insomnia Scale Based on the Criterion of Insomnia Severity Index. Int. J. Environ. Res. Public. Health 17, 8789 (2020). Radford, A. et al. Robust Speech Recognition via Large-Scale Weak Supervision. in Proceedings of the 40th International Conference on Machine Learning 28492–28518 (PMLR, 2023). Malgaroli, M., Hull, T. D., Zech, J. M. & Althoff, T. Natural language processing for mental health interventions: a systematic review and research framework. Transl. Psychiatry 13, 1–17 (2023). Cummins, N. et al. A review of depression and suicide risk assessment using speech analysis. Speech Commun. 71, 10–49 (2015). Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (eds. Burstein, J., Doran, C. & Solorio, T.) 4171–4186 (Association for Computational Linguistics, Minneapolis, Minnesota, 2019). doi: 10.18653/v1/N19-1423 . Reimers, N. & Gurevych, I. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (eds. Inui, K., Jiang, J., Ng, V. & Wan, X.) 3982–3992 (Association for Computational Linguistics, Hong Kong, China, 2019). doi: 10.18653/v1/D19-1410 . Khattab, O. et al. DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines. in (2023). Team, G. et al. Gemma 3 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2503.19786 (2025). Guessoum, S. B. et al. Articulating biological and social approaches in child and adolescent psychiatry. Front. Child Adolesc. Psychiatry 1, (2022). Guinart, D., Kane, J. M. & Correll, C. U. Is Transcultural Psychiatry Possible? JAMA 322, 2167–2168 (2019). Kordzadeh, N. & Ghasemaghaei, M. Algorithmic bias: review, synthesis, and future research directions. Eur. J. Inf. Syst. 31, 388–409 (2022). Ohse, J. et al. Zero-Shot Strike: Testing the generalisation capabilities of out-of-the-box LLM models for depression detection. Comput. Speech Lang. 88, 101663 (2024). Chen, Z. et al. Depression Detection in Clinical Interviews with LLM-Empowered Structural Element Graph. in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) (eds. Duh, K., Gomez, H. & Bethard, S.) 8181–8194 (Association for Computational Linguistics, Mexico City, Mexico, 2024). doi: 10.18653/v1/2024.naacl-long.452 . Kim, J. et al. Optimizing Large Language Models for Detecting Symptoms of Comorbid Depression or Anxiety in Chronic Diseases: Insights from Patient Messages. Preprint at https://doi.org/10.48550/arXiv.2503.11384 (2025). Öngür, D. & Paulus, M. P. Embracing complexity in psychiatry—from reductionistic to systems approaches. Lancet Psychiatry 12, 220–227 (2025). Möller, H. J. Rating depressed patients: observer- vs self-assessment. Eur. Psychiatry 15, 160–172 (2000). Enns, M. W., Larsen, D. K. & Cox, B. J. Discrepancies between self and observer ratings of depression: The relationship to demographic, clinical and personality variables. J. Affect. Disord. 60, 33–41 (2000). Bzdok, D. & Meyer-Lindenberg, A. Machine Learning for Precision Psychiatry: Opportunities and Challenges. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 3, 223–230 (2018). König, I. R., Fuchs, O., Hansen, G., Mutius, E. von & Kopp, M. V. What is precision medicine? Eur. Respir. J. 50, (2017). Winter, N. R. et al. A Systematic Evaluation of Machine Learning-Based Biomarkers for Major Depressive Disorder. JAMA Psychiatry 81, 386–395 (2024). Berardi, M. et al. Relative importance of speech and voice features in the classification of schizophrenia and depression. Transl. Psychiatry 13, 1–8 (2023). Silva, W. J., Lopes, L., Galdino, M. K. C. & Almeida, A. A. Voice Acoustic Parameters as Predictors of Depression. J. Voice 38, 77–85 (2024). Martin, V. P. & Rouas, J.-L. Estimating Symptoms and Clinical Signs Instead of Disorders: The Path Toward The Clinical Use of Voice and Speech Biomarkers In Psychiatry. in ICASSP 2024–2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 10606–10610 (2024). doi: 10.1109/ICASSP48485.2024.10445888 . Additional Declarations Yes RR, XNC, AL, MD, and AB are shareholders of Callyope, and AD, FR are former employees of Callyope. SBG is a consultant for Callyope. This is reported in the manuscript. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6594999","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":453262160,"identity":"c0c7fb30-7c35-400c-8e1a-52c7629566e4","order_by":0,"name":"Rachid Riad","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzUlEQVRIiWNgGAWjYHACNoYENhsGBnbmBpK0pDEwMDOSooWB7TAJWszbDz978KDsvDx/M2ObxMcdNgz87d0JeLXInEkzN0g4d9twxmHGNsmZZ9IYJM6c3YBXiwRDDptEYtvtBAagFmnetsMMBhK5BLTwvwFpOZcgD9HynwgtEmBbDiQYQLQcIEbLMzOJhHPJhhsPMzZbzjyTzEPYL/zJzyR/lNnJyx1vPnjj4w47Of72XvxakAGLBDBqeIhWDgLMH0hIAKNgFIyCUTCCAABCv0Odl+XohQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-7753-1219","institution":"Callyope","correspondingAuthor":true,"prefix":"","firstName":"Rachid","middleName":"","lastName":"Riad","suffix":""},{"id":453262161,"identity":"18bbec80-848b-45a3-a983-5d979ccc9fae","order_by":1,"name":"Alexandre Ducorroy","email":"","orcid":"","institution":"Callyope","correspondingAuthor":false,"prefix":"","firstName":"Alexandre","middleName":"","lastName":"Ducorroy","suffix":""},{"id":453262162,"identity":"bc17e730-da03-47ef-8036-54d606754957","order_by":2,"name":"Sélim Benjamin GUESSOUM","email":"","orcid":"","institution":"Callyope","correspondingAuthor":false,"prefix":"","firstName":"Sélim","middleName":"Benjamin","lastName":"GUESSOUM","suffix":""},{"id":453262163,"identity":"9baac0ed-75d7-4be1-b52a-fba979835d50","order_by":3,"name":"Filomène ROQUEFORT","email":"","orcid":"","institution":"Callyope","correspondingAuthor":false,"prefix":"","firstName":"Filomène","middleName":"","lastName":"ROQUEFORT","suffix":""},{"id":453262164,"identity":"b60f66ad-72bf-4ebc-b2d2-6ef4868ff929","order_by":4,"name":"Adrien Lesage","email":"","orcid":"","institution":"Callyope","correspondingAuthor":false,"prefix":"","firstName":"Adrien","middleName":"","lastName":"Lesage","suffix":""},{"id":453262165,"identity":"8bd25ba3-6293-4598-ae9a-9cad69634993","order_by":5,"name":"Xuan-Nga Cao","email":"","orcid":"","institution":"Callyope","correspondingAuthor":false,"prefix":"","firstName":"Xuan-Nga","middleName":"","lastName":"Cao","suffix":""},{"id":453262166,"identity":"1708281b-f0da-4a45-97c6-6461b55a6f6a","order_by":6,"name":"Alexis Bourla","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Alexis","middleName":"","lastName":"Bourla","suffix":""}],"badges":[],"createdAt":"2025-05-05 13:51:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6594999/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6594999/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82571599,"identity":"97ce916a-9790-489f-a742-05938eb95542","added_by":"auto","created_at":"2025-05-13 04:25:25","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1152457,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCo-occurrences of symptoms among the 3 populations under study\u003c/strong\u003e\u003cbr\u003e\nThe study includes clinical cohorts of Italian- and Chinese-speaking participants and a French-speaking general population cohort. Psychiatric symptoms were assessed using self-report questionnaires — including the \u003cem\u003ePatient Health Questionnaire\u003c/em\u003e (PHQ-9), the \u003cem\u003eGeneralized Anxiety Disorder scale\u003c/em\u003e(GAD-7), the \u003cem\u003eAthens Insomnia Scale\u003c/em\u003e(AIS), the \u003cem\u003eBeck Depression Inventory\u003c/em\u003e(BDI), the \u003cem\u003eMultidimensional Fatigue Inventory\u003c/em\u003e (MFI), the \u003cem\u003eMini International Neuropsychiatric Interview\u003c/em\u003e (MINI), and the \u003cem\u003ePittsburgh Sleep Quality Index\u003c/em\u003e (PSQI) — as well as clinician-administered measures such as the \u003cem\u003eMontgomery-Åsberg Depression Rating Scale\u003c/em\u003e (MADRS).\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6594999/v1/96aff63d42141d708ccb1c5a.jpg"},{"id":82570828,"identity":"92da1235-809b-45f1-9f6c-fb5770baa25c","added_by":"auto","created_at":"2025-05-13 04:17:25","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":508370,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGeneral Overview of our LLM-based system versus traditional machine learning approaches\u003c/strong\u003e (A) Audio embedding with ML: Speech signals are processed to extract acoustic features using eGeMAPS before classification with ML algorithms. (B) Text embedding with ML: Speech is transcribed and transformed into linguistic embeddings before ML classification. Both traditional methods require substantial parallel datasets for each language. (C) Our LLM-based system: Transcribed speech is analyzed by a LLM using few-shot prompting with N examples that contextualize each classification task.\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6594999/v1/5e28bb5f0402b6267d963ffb.jpg"},{"id":82570835,"identity":"50c086c5-77b7-47a3-969b-75c7c221512c","added_by":"auto","created_at":"2025-05-13 04:17:25","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3297776,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOur LLM-based system outperformed other methods in identifying depression across languages, in both clinical and general population samples\u003c/strong\u003e\u003cbr\u003e\n(A) Bar chart comparing F1-scores across different mental health classification tasks in three languages (Chinese, Italian, and French). Our LLM-based system consistently outperformed traditional methods including Text embedding with machine learning, Audio embedding with machine learning, and a baseline random prediction model (Dummy) across most classification tasks. (B) Detailed performance metrics for our LLM-based approach across all classification tasks, showing Sensitivity (recall), Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV). Color coding indicates performance levels, with darker green representing higher values.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6594999/v1/9e6fdf774f7006f3832cd455.jpg"},{"id":82570834,"identity":"cd7ff3e0-6e86-4313-ae6c-49f720e56817","added_by":"auto","created_at":"2025-05-13 04:17:25","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":433643,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eF1-score for depression detection tasks according to the number of examples provided in the prompt\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"fullfigure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6594999/v1/b42db27bf24bc82cf5bfc17f.jpg"},{"id":94092916,"identity":"6987c470-ffb2-4b30-b6eb-e7f6bdb179e9","added_by":"auto","created_at":"2025-10-22 09:15:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6350084,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6594999/v1/80e02c06-ef50-4184-8f85-067296e7413e.pdf"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e\nRR, XNC, AL, MD, and AB are shareholders of Callyope, and AD, FR are former employees of Callyope. SBG is a consultant for Callyope. This is reported in the manuscript.","formattedTitle":"Automated speech content analysis to detect depression with large language models: towards multilingual and few-shot capabilities","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMajor Depressive Disorder (MDD) affects 30 to 40% of individuals at some point in their lifetime, across genders and regions worldwide \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Depression is one of the leading contributors to global disability, associated with decreased functioning, and premature mortality. Despite its prevalence and profound impact, depression remains underdiagnosed and undertreated worldwide, highlighting the need for greater recognition and innovative approaches to care \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. In the general population, there is a need to identify depression to guide individuals towards mental healthcare. In clinical populations, there is a need for objective assessment, follow up and prediction of patient outcomes \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. The current standard of care relies on subjective clinical assessments of mental health, constrained by the limited number of trained clinicians, limited time, and economic resources. In most countries, there is a lack of trained clinicians to address population needs for mental healthcare \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. In clinical practice, complementary exams are mostly unhelpful for depression management, despite the important efforts of research in imagery, biology, and genetics. Yet, measurement-based care with validated clinical scales has been proven useful for better and faster response and remission of depressive patients \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThese challenges necessitate acceptable, cheap, easy-to-implement, scalable, and non-invasive methods to identify, assess, and monitor depressive and other mental health symptoms. Among these, automated analysis of speech, a widely available and minimally intrusive medium, has emerged as a key modality for the evaluation of mental health conditions \u003csup\u003e\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Psychiatry practice is based on patient-doctor social interaction. Until now it has been based on clinicians' analysis of patient's behavior and thoughts, through their speech, expressivity, and communication. Developing speech analysis tools is consistent with the observation that speech analysis in psychiatry has been irreplaceable with biological or imagery exams.\u003c/p\u003e \u003cp\u003eSpeech provides a rich source of information about the speaker\u0026rsquo;s psychological and cognitive state, offering both acoustic and linguistic markers for assessment \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Historical observations, such as those by Kraepelin in 1921 \u003csup\u003e12\u003c/sup\u003e, described speech abnormalities in mood disorders as \u0026ldquo;slow, hesitating, monotonous, sometimes stuttering,\u0026rdquo; laying the foundation for modern computational methods. Clinicians assess depression by focusing on some symptoms that include: monotonous speech, decrease in speech speed, and content of speech encompassing negative perception of self, past, future, and the others. Acoustic features such as prosody, pitch, intensity, and temporal dynamics have proven valuable for identifying emotional and cognitive disturbances \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. For instance, the Geneva Minimalistic Acoustic Parameter Set (GeMAPS) has been widely adopted for detecting depression or neurological conditions \u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. The increase in dataset size and the improvements of techniques allowed the use of deep learning to further enhance audio-based assessments, enabling the automatic discovery of complex acoustic patterns and improving diagnostic accuracy \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. These speech excerpts can be elicited from simple, semi-structured tasks, such as narrating daily events or reflecting on recent challenges, which are neither taxing for patients nor difficult for practitioners to administer.\u003c/p\u003e \u003cp\u003eLinguistic analysis complements these acoustic approaches by capturing self-referential language, emotional valence, and syntactic complexity, all of which are known to correlate with depressive state \u003csup\u003e\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Predictive modeling studies have identified specific linguistic markers of depression \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, which can be combined with audio inputs for more comprehensive assessment \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. These approaches have also been successfully implemented in longitudinal trial designs to monitor changes over time \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Multilingual studies are particularly important to ensure that speech-based tools can be widely disseminated and adapted to diverse cultural and clinical contexts \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Depression is a universal condition, yet its manifestations vary significantly across languages and cultural settings \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eWhile these acoustic and linguistic approaches have shown promise, they face significant limitations. Each new application, clinical scale, or language requires substantial parallel data between validated clinical scales and speech samples in the target language \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. This necessity for extensive language-specific datasets creates barriers to global clinical implementation, particularly in regions with limited research resources. The development process typically requires collecting, annotating, and analyzing large corpora of clinical speech data \u0026ndash; a process that is time-consuming, expensive, and difficult to scale across multiple languages and cultural contexts \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eLarge Language Models (LLMs) have emerged as a potential solution to these challenges. As state-of-the-art approaches in natural language processing, LLMs are trained on a vast corpora of multilingual text data and demonstrate remarkable zero-shot and few-shot capabilities \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, even in health-care \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. These capabilities enable LLMs to perform tasks with minimal or no task-specific training examples, potentially addressing the data scarcity issues that plague traditional acoustic and linguistic analysis methods \u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. Such models significantly reduce the effort required to train and deploy systems, particularly in multilingual applications where large-scale clinical trials are often impractical.\u003c/p\u003e \u003cp\u003eRecent research has begun exploring LLMs for mental health applications with promising results. Wu et al.\u003c/p\u003e \u003cp\u003e \u003csup\u003e \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e \u003c/sup\u003e introduced the CALLM model, which leverages LLMs to generate and extend datasets for mental health analysis. Similarly, the MentaLLaMA approach \u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e has demonstrated the capability to tackle diverse mental health tasks based on social media posts. Other researchers have used LLMs to generate synthetic training examples to improve traditional machine learning models for mental health applications \u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e These approaches suggest that LLMs can augment existing methods by addressing data scarcity issues while maintaining clinical validity.\u003c/p\u003e \u003cp\u003eDespite these promising developments, significant challenges remain in applying LLMs to mental health care. As highlighted in a recent viewpoint in Lancet Digital Health \u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e, while LLMs offer potential solutions to treatment gaps and research limitations, issues concerning technical costs, literacy barriers, algorithmic biases, and inequitable data representation must be addressed. Our work directly contributes to this framework by addressing two key priorities: culturally-informed development through multilingual applications and refined diagnostic approaches through specific LLM outputs for each target condition.\u003c/p\u003e \u003cp\u003eBuilding on these developments and addressing the identified challenges, our study aims to evaluate the potential of LLM-based speech content analysis for mental health assessment across languages and populations. Specifically, the main objective is to assess LLM-based speech content analysis models in detecting depression in diverse linguistic contexts (French, Chinese, and Italian), in both clinical and general populations. The secondary objective is to identify anxiety, fatigue, and insomnia in the same populations.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis study employed a novel methodological framework to assess the effectiveness of Large Language Models (LLMs) for detecting mental health conditions through speech analysis across multiple languages. Our approach combines state-of-the-art automatic speech recognition with prompt-engineered LLMs that function as direct classifiers, reducing the need for language-specific training data that traditional methods require. We first explain our study design and corpora, then data preprocessing pipeline, we also detail both traditional baseline approaches and our innovative LLM-based system, followed by our implementation of prompt engineering and statistical analysis methods. This methodology was applied consistently across three diverse linguistic and clinical contexts.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy Design and Participants\u003c/h2\u003e \u003cp\u003eThe study is based on three distinct cohorts: a large general population sample collected in French (n\u0026thinsp;=\u0026thinsp;1347), and two specialized clinical populations collected in Italian (n\u0026thinsp;=\u0026thinsp;116) and Chinese (n\u0026thinsp;=\u0026thinsp;52) languages. The Italian and the Chinese corpus are publicly available corpuses for research uses. In Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.A, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.B, and Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.C, we present the study populations, symptoms assessed through questionnaires, and co-occurrence of symptoms in patients.\u003c/p\u003e \u003cp\u003eChinese Clinical Population (MODMA Corpus)\u003c/p\u003e \u003cp\u003eThe MODMA corpus \u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e includes 52 Chinese-speaking participants with a mean age of 31.3 years (SD\u0026thinsp;=\u0026thinsp;9.2, range 18\u0026ndash;52). The sample comprised 36 males and 16 females, with education levels distributed as: no diploma (n\u0026thinsp;=\u0026thinsp;7), secondary education (n\u0026thinsp;=\u0026thinsp;8), and long-cycle higher education (n\u0026thinsp;=\u0026thinsp;37), with no participants reporting short-cycle higher education.\u003c/p\u003e \u003cp\u003eThe dataset included 23 outpatients diagnosed with Major Depressive Disorder (MDD) (16 males and 7 females, aged 16\u0026ndash;56) and 29 healthy controls (20 males and 9 females, aged 18\u0026ndash;55). MDD diagnoses were confirmed by clinical psychiatrists at Lanzhou University Second Hospital based on the Mini-International Neuropsychiatric Interview (MINI)\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e and DSM-IV diagnostic criteria \u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. Inclusion criteria for patients included a PHQ-9 score above 10 and no psychotropic drug treatment in the two weeks prior to participation. Healthy controls were recruited through public advertisements, with exclusion criteria that ruled out personal or family histories of mental disorders.\u003c/p\u003e \u003cp\u003eSpeech samples consisted of responses to 18 questions derived from DSM-IV and extracted from depression scales \u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e, including questions such as: \"What is the best gift you have ever received, and how did you feel?\", \"Please describe one of your friends, including age, job, characters, and hobbies,\" \"What would you like to do when you are unable to fall asleep?\", and \"What makes you desperate?\"\u003c/p\u003e \u003cp\u003e 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 \u003cspan\u003e$\u003c/span\u003e16 compensation.\u003c/p\u003e \u003cp\u003eItalian Clinical Population (Androids Corpus)\u003c/p\u003e \u003cp\u003eThe Androids Corpus was specifically designed for automatic depression detection from speech \u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. It includes 228 recordings from 116 native Italian speakers with a mean age of 37.4 years (SD\u0026thinsp;=\u0026thinsp;12, range 19\u0026ndash;71). The gender distribution was 32 males and 84 females. Educational levels included: no diploma (n\u0026thinsp;=\u0026thinsp;11), secondary education (n\u0026thinsp;=\u0026thinsp;37), short-cycle higher education (n\u0026thinsp;=\u0026thinsp;52), and long-cycle higher education (n\u0026thinsp;=\u0026thinsp;16).\u003c/p\u003e \u003cp\u003eOf the 116 participants, 64 were clinically diagnosed with depression by professional psychiatrists using the Montgomery-Asberg Depression Rating Scale (MADRS) \u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e, providing a reliable characterization of depressed individuals compared to self-reported assessments. The interview task involved answering questions about everyday life (e.g., \"What did you do last weekend\"). The corpus contains both elicited and spontaneous speech samples, with approximately 9 hours of total recording time. For this study, we focused exclusively on spontaneous speech productions.\u003c/p\u003e \u003cp\u003eFrench General Population Dataset\u003c/p\u003e \u003cp\u003eThis dataset comprises data from 1,347 French-speaking participants recruited from the general population in France. Participants had a mean age of 37.8 years (SD\u0026thinsp;=\u0026thinsp;18.2, range 18.4\u0026ndash;91.4), with 479 males and 860 females (8 participants did not disclose gender). The education level distribution was: no diploma (n\u0026thinsp;=\u0026thinsp;47), secondary education (n\u0026thinsp;=\u0026thinsp;218), short-cycle higher education (n\u0026thinsp;=\u0026thinsp;166), and long-cycle higher education (n\u0026thinsp;=\u0026thinsp;916). Participants completed a series of speech tasks and self-administered questionnaires through a mobile research application specifically designed for clinical studies.\u003c/p\u003e \u003cp\u003e Two types of speech tasks were included: These included an elicited production task (constrained-vocabulary) where participants counted from 1 to 20, and a spontaneous speech task (open-vocabulary) where participants responded to open-ended questions such as: \"Describe how you are feeling at the moment and how your sleep has been lately,\" \"Describe your last 24 hours,\" \"Describe a negative event or situation you think might happen in the future (week, month, or year),\" and \"Describe a positive event or situation you think might happen in the future (week, month, or year).\"\u003c/p\u003e \u003cp\u003eThe total duration of recorded speech was 10.12 hours, sampled at 16kHz. Self-administered questionnaires assessed participants' mental health status, including depressive symptoms using the Patient Health Questionnaire-9 (PHQ-9), with scores\u0026thinsp;\u0026ge;\u0026thinsp;10 classified as moderate depression \u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e; anxiety using the Generalized Anxiety Disorder questionnaire (GAD-7), with scores\u0026thinsp;\u0026ge;\u0026thinsp;10 classified as moderate anxiety \u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e; and insomnia using the Athens Insomnia Scale (AIS), with scores\u0026thinsp;\u0026ge;\u0026thinsp;6 classified as mild insomnia \u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e All participants provided informed consent in accordance with the Declaration of Helsinki, Good Clinical Practice guidelines, and local regulations. The 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 \u0026euro;15 compensation for their time (See \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e for more details).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe study includes clinical cohorts of Italian- and Chinese-speaking participants and a French-speaking general population cohort. Psychiatric symptoms were assessed using self-report questionnaires \u0026mdash; including the \u003cem\u003ePatient Health Questionnaire\u003c/em\u003e (PHQ-9), the \u003cem\u003eGeneralized Anxiety Disorder scale\u003c/em\u003e (GAD-7), the \u003cem\u003eAthens Insomnia Scale\u003c/em\u003e (AIS), the \u003cem\u003eBeck Depression Inventory\u003c/em\u003e (BDI), the \u003cem\u003eMultidimensional Fatigue Inventory\u003c/em\u003e (MFI), the \u003cem\u003eMini International Neuropsychiatric Interview\u003c/em\u003e (MINI), and the \u003cem\u003ePittsburgh Sleep Quality Index\u003c/em\u003e (PSQI) \u0026mdash; as well as clinician-administered measures such as the \u003cem\u003eMontgomery-\u0026Aring;sberg Depression Rating Scale\u003c/em\u003e (MADRS).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData preprocessing, transcription and predictions of symptoms\u003c/h3\u003e\n\u003cp\u003eOnly audio recordings were available and are used for any mental health symptoms predictions. These recordings were transcribed using the Whisper multilingual automatic speech recognition model \u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e, the large-v3-turbo version, and by specifying the language. To ensure compatibility with the Whisper model, all audio files were resampled to 16 kHz before processing.\u003c/p\u003e \u003cp\u003eFor the Chinese clinical population (n\u0026thinsp;=\u0026thinsp;52), individuals were classified based on depression status determined through structured clinical interviews using the Mini International Neuropsychiatric Interview [MINI] and supplemented by the Patient Health Questionnaire-9(measured via PHQ-9). Additionally, anxiety severity (Generalized Anxiety Disorder-7 (GAD-7)) and sleep disturbances (Pittsburgh Sleep Quality Index [PSQI]) were assessed. The use of both diagnostic interviews and psychometric scales provided comprehensive clinical assessment in this cohort.\u003c/p\u003e \u003cp\u003eFor the Italian clinical population (n\u0026thinsp;=\u0026thinsp;116), classification focused exclusively on depression status as assessed by psychiatrists using the Montgomery\u0026ndash;\u0026Aring;sberg Depression Rating Scale (MADRS). This provided a clinician-validated ground truth rather than relying solely on self-reported measures.\u003c/p\u003e \u003cp\u003eFor the French general population cohort (n\u0026thinsp;=\u0026thinsp;1347), classification encompassed multiple self-reported mental health dimensions: depression severity (measured via PHQ-9), anxiety symptoms (GAD-7), fatigue levels (Multidimensional Fatigue Inventory (MFI)), and insomnia severity (Athens Insomnia Scale (AIS)). A positive classification was assigned if the participant's score exceeded clinically validated thresholds for each scale (PHQ-9\u0026thinsp;\u0026ge;\u0026thinsp;10 for moderate depression, GAD-7\u0026thinsp;\u0026ge;\u0026thinsp;10 for moderate anxiety, and AIS\u0026thinsp;\u0026ge;\u0026thinsp;6 for mild insomnia).\u003c/p\u003e\n\u003ch3\u003eLarge Language Model as classifier\u003c/h3\u003e\n\u003cp\u003eOur system introduces a novel approach using a LLM as a classifier for mental health assessment, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. This approach differs from conventional methods that have dominated the field of automatic speech-based mental health assessment \u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe first approach to tackle speech-based mental health assessment is to model acoustic realization from raw audio, using established feature sets such as eGeMAPS (extended Geneva Minimalistic Acoustic Parameter Set) \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e (See Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). These audio features extract statistics from articulatory and prosodic elements including pitch, intensity, spectral, and voice quality features. These acoustic features form a fixed-size vector representation (embeddings) that serves as input to conventional machine learning algorithms, which are trained on parallel datasets of audio features and clinical labels to predict mental health status \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. We refer to this approach as Audio embedding with ML in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA and in the Results section.\u003c/p\u003e \u003cp\u003eThe second approach is to model the content of the speech production using an automatic speech recognition (ASR) system. This transcribed speech content (words) are then transformed into linguistic embeddings, fixed-size vectors, using large pre-trained transformer models such as BERT \u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e (See Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). These embeddings capture semantic and syntactic information in fixed-size embeddings, which are then fed into conventional machine learning classifiers trained on parallel corpora of text embeddings and clinical labels \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. We refer to this approach as Text embedding with ML in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB and in the Results section. We used the multilingual text embedding mpnet-base-v2 \u003csup\u003e49\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFor audio embedding, we used a 20-second sliding window; and for text embedding, we used a 70-word sliding window. We used as final classifier, a logistic regression and used a mean pooling to aggregate predictions.\u003c/p\u003e \u003cp\u003eThese two traditional approaches require substantial parallel data between speech samples and clinical assessments for each target language and clinical scale, creating significant barriers for multilingual and cross-cultural implementation (Low et al., 2020).\u003c/p\u003e \u003cp\u003eOur proposed system fundamentally differs in its approach (See Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC). We also first transcribed speech to convert audio to text across all three languages (French, Italian, and Chinese). Instead of generating fixed-size embeddings followed by a separate classifier, we employed a few-shot prompting technique with the LLM. This involves providing the model with selected examples of transcribed speech paired with their corresponding clinical classifications (e.g., \"depressed\" or \"not depressed\"). The LLM is then prompted with a new patient's transcribed speech and directly asked to classify the mental health status (e.g., \"Is this patient depressed?\"), returning a binary classification (Yes/No). This approach leverages the LLM's pre-trained knowledge of linguistic patterns associated with psychological states across multiple languages, potentially reducing the need for extensive language-specific training data \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003ePrompt engineering\u003c/h3\u003e\n\u003cp\u003eUnlike traditional machine learning classifiers, LLMs generate open-ended text rather than discrete labels, introducing variability in outputs. To mitigate this, structured prompt engineering was applied, utilizing soft constraints rather than enforcing strict response formats, as it yielded the best results.\u003c/p\u003e \u003cp\u003eImplementation relied on the DSPy prompting framework for structured prompt generation \u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e and Ollama for efficient LLM inference. We leveraged DSPy's prompt management capabilities to maintain consistent prompting patterns across languages while allowing for dataset-specific adaptations. The framework enabled us to programmatically construct complex few-shot learning contexts that included both examples and target tasks within a unified prompt structure. This systematic approach ensured reproducibility and minimized variability in prompt formatting across experimental conditions, which is particularly important when working with open-ended LLM responses in clinical applications. We used the following prompting strategy to obtain binary responses and minimize model hallucinations and biases: we sampled examples such that we respected the balance between the number of positive and negative examples in the prompt, and we did not enforce a grammar on the output.\u003c/p\u003e \u003cp\u003eFor the LLM weights, we selected the state-of-the-art Gemma open source version 3 model with 27 billions of parameters \u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e, due to its performance on multilingual tasks and its large 128K token context window, which is critical for our few-shot learning approach that requires integrating multiple examples alongside patient transcripts. For comparative in-context learning analysis, we also evaluated a smaller 12\u0026nbsp;billion parameter version of the same model family. We specifically chose these two model sizes because smaller models (\u0026lt;\u0026thinsp;12B parameters) did not reliably support the extensive context windows. In contrast with our LLM system, the classic ML models learned on the full training set for each task. The LLM sees only the examples given in the prompts, which limits the number of examples available to perform each task. This number to be seen is limited by the maximum context window of the LLM, in our case 128K tokens. Based on our preliminary experiments and different speech transcripts across datasets, our upper limit, found empirically, for the number of examples to provide the LLM was 32.\u003c/p\u003e \u003cp\u003eGiven the heterogeneous nature of the datasets, prompts were tailored to maximize context inclusion. For the French general population corpus, a structured question-answer format was employed, compiling all responses into a single text input. Participants described their past 24 hours, significant negative and positive events from the past year, anticipated events in the coming year, their emotional state, and sleep quality. For the Italian Androids dataset, since questions were unavailable, transcripts were constructed from concatenated responses. For the Chinese Lanzhou corpus, prompts explicitly specified four question categories derived from DSM-IV, Hamilton Depression Rating Scale, and other validated questionnaires. All prompts were written in English, while participants' responses were retained in their original language, with no demographic information included in the prompts.\u003c/p\u003e \u003cp\u003eWe implemented two distinct prompting techniques: zero-shot prompting, where the LLM was prompted to classify responses without prior labeled examples; and few-shot prompting, where the model was provided with pre-labeled examples structured as a dialogue format (prompt \u0026rarr; LLM answer) to contextualize the task to solve for the LLM. These examples were aggregated into a single prompt, including the participant's transcript for classification, along the speech transcript to classify. This approach is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC.\u003c/p\u003e\n\u003ch3\u003eModel Training and Evaluation\u003c/h3\u003e\n\u003cp\u003eFor the general population dataset, models were trained on 80% of the data and tested on the remaining 20%. For the Chinese and Italian clinical datasets, a five-fold cross-validation approach was employed to stay consistent with original studies, partitioning the dataset into five subsets and training on four while evaluating on the remaining one. For our LLM system all the prior examples are only extracted from the training set. Stratified sampling ensured that each subset maintained the same distribution as the full dataset, based on depression class (Chinese and Italian populations) or PHQ-9 scores (general population). Performance was assessed using the F1-score, which integrates Positive Predictive Value (PPV) and Sensitivity (Se): where an F1-score of 1 represents perfect classification, and 0 indicates complete misclassification (\u003cem\u003eF1\u0026thinsp;=\u0026thinsp;2*(PPV*Se)/(PPV\u0026thinsp;+\u0026thinsp;Se)).\u003c/em\u003e For our LLM-based system, we also reported the detailed Se, Specificity (Spe), Positive Predictive Value (PPV) and Negative Predictive Value (NPV).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis of in-context learning\u003c/h2\u003e \u003cp\u003eTo analyze the relationship between depression detection performance and the number of examples provided to the large language model (LLM), we performed extensive experiments using two different versions of the models that vary in size. We computed the F1 score as a function of the number of examples provided (0, 1, 2, 4, 8, 16, 32).\u003c/p\u003e \u003cp\u003eTo focus on comparable conditions, we utilized two subtasks involving depression detection where labels were provided by psychiatrists in two different languages (Italian and Chinese). This allowed us to examine cross-cultural consistency while maintaining clinical validity of the classifications.\u003c/p\u003e \u003cp\u003eWe ran a multiple linear regression for each language with interaction terms to quantify the relationships between the model size and the example count. The model was specified as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\:{F}_{1}\\:score\\:\\sim\\:\\:example\\:count+model\\:size+example\\:count\\times\\:model\\:size\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eThe main results of our study are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Overall, our LLM-based approach outperformed traditional machine learning and acoustic methods in most mental health detection tasks across languages, with particularly strong performance in depression detection. The F1-scores ranged from 0.85\u0026ndash;0.96 for depression in clinical populations (Chinese and Italian) and 0.40 in the general population (French), compared to substantially lower scores for baseline methods. Our LLM-based system demonstrated excellent sensitivity for depression detection in both Chinese (Se\u0026thinsp;=\u0026thinsp;1.00) and French (Se\u0026thinsp;=\u0026thinsp;0.93) populations, with good specificity in Chinese (Sp\u0026thinsp;=\u0026thinsp;0.93) and Italian (Sp\u0026thinsp;=\u0026thinsp;0.88) clinical populations. Anxiety detection was also successful with high sensitivity in both Chinese (Se\u0026thinsp;=\u0026thinsp;0.85) and French (Se\u0026thinsp;=\u0026thinsp;0.97) populations, though with moderate specificity (Sp\u0026thinsp;=\u0026thinsp;0.39 and Sp\u0026thinsp;=\u0026thinsp;0.46, respectively). Detection of secondary symptoms like sleep difficulties, fatigue, and insomnia showed more variable results across populations.\u003c/p\u003e\n\u003ch3\u003eDepression\u003c/h3\u003e\n\u003cp\u003eAcross all three studied populations, our LLM-based system outperformed alternative methods in detecting depression in both clinical and general populations across different languages. F1-scores for depression reached 0.96 in the Chinese-speaking dataset with perfect sensitivity (Se\u0026thinsp;=\u0026thinsp;1.00) and excellent specificity (Sp\u0026thinsp;=\u0026thinsp;0.93), resulting in strong predictive values (PPV\u0026thinsp;=\u0026thinsp;0.92, NPV\u0026thinsp;=\u0026thinsp;1.00). Similarly, in the Italian-speaking dataset, the F1-score was 0.85 with good sensitivity (Se\u0026thinsp;=\u0026thinsp;0.81) and specificity (Sp\u0026thinsp;=\u0026thinsp;0.88), yielding high predictive values (PPV\u0026thinsp;=\u0026thinsp;0.90, NPV\u0026thinsp;=\u0026thinsp;0.79). The F1-score in the French-speaking general population was lower at 0.40, despite high sensitivity (Se\u0026thinsp;=\u0026thinsp;0.93) but limited specificity (Sp\u0026thinsp;=\u0026thinsp;0.47), resulting in low positive predictive value (PPV\u0026thinsp;=\u0026thinsp;0.26) but high negative predictive value (NPV\u0026thinsp;=\u0026thinsp;0.97). In contrast, baseline methods achieved scores at random level in the French dataset, with null F1-scores.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eAnxiety, Insomnia, and Fatigue\u003c/h2\u003e \u003cp\u003eFor anxiety detection, our system achieved an F1-score of 0.78 in the Chinese dataset with good sensitivity (Se\u0026thinsp;=\u0026thinsp;0.85) but limited specificity (Sp\u0026thinsp;=\u0026thinsp;0.39), resulting in moderate predictive values (PPV\u0026thinsp;=\u0026thinsp;0.72, NPV\u0026thinsp;=\u0026thinsp;0.58). In the French dataset, anxiety detection reached an F1-score of 0.31 with excellent sensitivity (Se\u0026thinsp;=\u0026thinsp;0.97) but poor specificity (Sp\u0026thinsp;=\u0026thinsp;0.46), yielding low positive predictive value (PPV\u0026thinsp;=\u0026thinsp;0.18) but very high negative predictive value (NPV\u0026thinsp;=\u0026thinsp;0.99). For sleep/insomnia problems, our system performed below random level in the Chinese dataset with low sensitivity (Se\u0026thinsp;=\u0026thinsp;0.53) and specificity (Sp\u0026thinsp;=\u0026thinsp;0.39), while achieving better results in the French dataset (F1\u0026thinsp;=\u0026thinsp;0.64) with high sensitivity (Se\u0026thinsp;=\u0026thinsp;0.94) but poor specificity (Sp\u0026thinsp;=\u0026thinsp;0.34). For fatigue detection in the French general population, all systems performed similarly at random level.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis of in-context learning capabilities\u003c/h2\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, we presented the F1-scores for each depression detection task according to the number of examples included in the prompt. For the sub-analysis concerning the number of examples, we focused on the severe depression tasks. We found out that the relationship between few-shot learning (example count) and depression detection performance varies substantially between languages. For Chinese, the relationship depends strongly on model size (R\u0026sup2; = 0.781, F\u0026thinsp;=\u0026thinsp;11.91, p\u0026thinsp;=\u0026thinsp;0.00122), with larger models benefiting more from additional examples (interaction coefficient\u0026thinsp;=\u0026thinsp;0.0007, p\u0026thinsp;=\u0026thinsp;0.002). Specifically, while additional examples showed a negative effect for smaller models (coefficient = -0.0080, p\u0026thinsp;=\u0026thinsp;0.049), the significant positive interaction indicates this effect reverses for larger models. For Italian, performance remains relatively stable regardless of model size and example count, with the regression model explaining only 16.4% of variance and showing no significant effects (F\u0026thinsp;=\u0026thinsp;0.6561, p\u0026thinsp;=\u0026thinsp;0.597). The zero-shot prompting method is already performing at a high level in Italian.\u003c/p\u003e \u003cp\u003eThis suggests that few-shot learning strategies may need to be language-specific, and that scaling to larger models may be more beneficial for some languages than others.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we demonstrated that automated speech content analysis with a LLM-based system exhibits a good to excellent ability to identify depression, and good results for anxiety, insomnia and fatigue, across diverse languages and populations. Our LLM-based system consistently outperformed traditional machine learning techniques and acoustic feature analysis methods, achieving impressive F1 scores of 0.96 and 0.85 for depression detection in Chinese and Italian clinical populations, respectively. In the French general population, while performance was more moderate, our approach still surpassed baseline methods that performed at chance level.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eMultilingual capabilities of LLMs in mental health assessment\u003c/h2\u003e \u003cp\u003eOur work contributes to the emerging literature on LLMs for mental health support by providing additional empirical evidence and more extensive evaluations across diverse linguistic and clinical contexts \u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Multilingual capacity and transcultural validity are essential for mental health monitoring across diverse populations. Our study demonstrates that depression can be identified reliably across multiple languages, including French, Italian, and Chinese, addressing a critical need for equitable mental healthcare in various cultural settings \u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. A model trained and evaluated on non-representative data, for example, exclusively on English-speaker Caucasian adults without comorbidities, will likely underperform in minority groups, thereby exacerbating healthcare disparities \u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. It is therefore imperative to evaluate model performance across different demographics and different languages \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eWhile recent studies have shown promising results, our approach demonstrates extends to more languages and more mental health symptoms assessments. Compared to \u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e who achieved an F1 score of 0.73 for depression detection in a German-speaking general population cohort, our system delivered substantially higher performance (F1\u0026thinsp;=\u0026thinsp;0.96 in Chinese and F1\u0026thinsp;=\u0026thinsp;0.85 in Italian clinical populations). Similarly, \u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e explored multilingual applications, obtaining F1 scores of 84.62 for depression detection in English-speaking PTSD populations and 75.31 for depression among Chinese university students. While these results demonstrate the multilingual capabilities of LLMs, our study extends this work by showing even stronger performance across clinically validated populations in multiple non-English languages, suggesting that our approach may better capture culture-specific manifestations of depression.\u003c/p\u003e \u003cp\u003eUnlike \u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e who found no benefit from in-context learning despite strong zero-shot performance (\u0026gt;\u0026thinsp;90% F1-score) in analyzing diabetic patients' text messages, our results revealed language-dependent benefits from few-shot learning. The Chinese dataset particularly benefited from additional examples with larger models, suggesting that optimal prompting strategies may vary across languages (See Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Additionally, our use of open-source models (\u0026lt;\u0026thinsp;30B parameters), contrary to Kim et al 2025, addresses practical implementation concerns, enabling single-Graphical Processing Unit (GPU) deployment and on-premise processing compatible with regional data protection regulations like GDPR.\u003c/p\u003e \u003cp\u003eA notable important aspect of our study is the use of open-ended interview transcripts with different psychological themes and questions asked to patients. This approach parallels Cummins et al.'s ecological methods in their RADAR-MD cohort without requiring psychiatrist-administered MADRS interviews. While \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e identified acoustic correlates of depression severity across European populations and later, \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e used automatic topic modeling to extract content patterns longitudinally correlated with PHQ-8 scores. Our LLM-based approach leverages pre-trained linguistic knowledge without requiring explicit feature engineering or topic extraction. This complementarity suggests promising integration opportunities, combining LLM content analysis with acoustic features and topic modeling could enhance detection accuracy while maintaining ecological validity. Such multimodal assessment would capture both linguistic patterns and paralinguistic information, creating an easy-to-use automatic assessment tool for patients when mental health professionals may not be available at specific times.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eComplexity of patients and measures of clinical signs not diagnosis through LLM\u003c/h2\u003e \u003cp\u003eThe complex symptom patterns usually reported in psychiatry \u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e are clearly reflected in our datasets (See Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In the Chinese clinical population, we observed significant overlap between depression, anxiety, and sleep difficulties, with 18 participants simultaneously experiencing all three symptom clusters. Similarly, in the French general population, nearly all possible symptom combinations occurred, including isolated presentations of insomnia without fatigue or fatigue without insomnia. This heterogeneity underscores the challenges clinicians face in accurately assessing and treating psychiatric conditions, highlighting the need for nuanced assessment tools that can detect specific symptom profiles beyond broad diagnostic categories.\u003c/p\u003e \u003cp\u003eThe variability in ground truth methodology across our datasets further emphasizes this complexity. There are different uses for rating scales for psychiatrists and researchers: as screening instruments to detect the possible presence of a disorder, as measures to establish a symptom profile, as indicators of illness severity, and as measures of drug effect in controlled trials \u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e. For the French general population, we relied on self-assessment scales, which introduce potential bias due to their dependence on participant insight. In the Chinese population, we even observed discrepancies between clinician-administered MINI interviews and PHQ-9 self-evaluations, reflecting the known challenges in achieving diagnostic consistency across assessment methods. These methodological differences align with findings from \u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e who found out that patient characteristics and symptoms significantly affect the correlation between observer- and self-rating scales. Younger age, higher educational attainment, and depressive subtype (atypical, non-melancholic) predicted higher self-reported scores relative to clinician ratings, while personality traits such as high neuroticism, low extraversion, and low agreeableness were associated with higher endorsement of depressive symptoms on self-reports. These discrepancies were more pronounced for psychological symptoms than somatic symptoms, reaffirming the value of multi-modal assessment in depression research. Despite these methodological variations, our LLM-based system still demonstrated robust performance across datasets for depressive aspects, suggesting potential utility in diverse clinical contexts.\u003c/p\u003e \u003cp\u003eImportantly, our results reveal that not all symptoms are equally expressed in language (or adequately captured by our LLM-based system). While depression detection was consistently strong across languages, performance varied for secondary symptoms. Sleep difficulties in Chinese speakers showed poorer results (Se\u0026thinsp;=\u0026thinsp;0.53, Sp\u0026thinsp;=\u0026thinsp;0.39), and fatigue detection in the French population performed at near-random levels with very low specificity (Se\u0026thinsp;=\u0026thinsp;0.94, Sp\u0026thinsp;=\u0026thinsp;0.14). Yet, insomnia was detected more effectively with our LLM approach than with other methods. This discrepancy between Chinese clinical detection of sleep difficulties and better performance on insomnia in the French population suggests that symptom reporting patterns may vary between clinical and general populations, with the latter potentially providing more concrete descriptions of sleep patterns. As underlined by Martin et al. (2024), it is of prime importance of assessing specific symptom severity beyond diagnosis for effective therapeutic relationships and personalized treatment planning. While precision medicine aims to adapt interventions to individual characteristics \u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e,\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e, its implementation in psychiatry has proven challenging despite expensive neuroimaging and genetic approaches yielding limited clinical utility \u003csup\u003e\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e. Our findings suggest that speech-based assessments using LLMs offer a complementary, practical path toward precision psychiatry\u0026mdash;not as a comprehensive solution for all symptoms, but as an accessible tool for specific clinical challenges. By effectively detecting depression across diverse languages and capturing certain secondary symptoms like insomnia, these methods provide clinically useful insight without the resource-intensive requirements of traditional biomarker approaches. However, the limitations in detecting fatigue and certain sleep disturbances highlight that speech-content analysis should be viewed as one component in a broader assessment framework rather than a standalone diagnostic solution.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003e \u003cb\u003eMedication use.\u003c/b\u003e Medication use represents another critical variable in mental health assessment that our study did not specifically address. Antidepressants and other psychotropic medications can potentially impact language patterns and disease presentation, as noted by \u003csup\u003e\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e,\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e. In our study, medication data was not available, which could influence both language production and symptom manifestation. Future work should systematically account for medication type and dosage to better understand their effects on speech-based assessments.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStratification.\u003c/b\u003e In this study, we focus on a binary classification of the presence or absence of at least moderate depression. However, future research should aim to develop methods that allow for more nuanced stratification across multiple severity levels\u0026mdash;such as none, mild, moderate, and severe depression. Such graduated assessments could enhance clinical applicability, particularly in monitoring patients\u0026rsquo; partial or complete responses to treatment over time \u003csup\u003e\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eComplete speech analysis.\u003c/b\u003e Future developments should also consider incorporating a broader range of parameters\u0026mdash;including linguistic content, acoustic features, and sociodemographic variables\u0026mdash;to improve the accuracy, robustness, and contextual relevance of assessments.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eOur LLM-based system demonstrates good to excellent performance in identifying depression across multiple languages in both clinical and general populations, with more variable results for anxiety, sleep problems, and fatigue. This approach leverages pre-trained linguistic knowledge without requiring language-specific feature engineering, showing particular promise for cross-cultural applications. The consistent performance across French, Italian, and Chinese populations\u0026mdash;despite variations in assessment methodologies\u0026mdash;highlights the potential for supporting equitable mental healthcare globally.\u003c/p\u003e \u003cp\u003eFurther developments should focus on providing more precise classifications of symptom intensity beyond binary detection and integrating multiple data streams including acoustic features and sociodemographic variables to improve robustness. Future applications could address unmet challenges in psychiatry, such as predicting medication response and detecting suicide risk through linguistic markers. While requiring careful validation, our findings establish a foundation for speech-based assessment tools that complement clinical expertise, potentially expanding access to mental healthcare while supporting more personalized treatment approaches.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eConflict of interest\u003c/h2\u003e \u003cp\u003eRR, XNC, AL, MD, and AB are shareholders of Callyope, and AD, FR are former employees of Callyope. SBG is a consultant for Callyope.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThe authors are thankful to all the participants who volunteered for this research study. Without their active involvement, this study would not have been possible. The authors also would like to thank each of the speech pathology interns who helped with the participant recruitment and made sure that the protocol was completed successfully.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBromet, E. \u003cem\u003eet al.\u003c/em\u003e Cross-national epidemiology of DSM-IV major depressive episode. BMC Med. 9, 90 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHerrman, H. \u003cem\u003eet al.\u003c/em\u003e Time for united action on depression: a Lancet-World Psychiatric Association Commission. Lancet Lond. Engl. 399, 957\u0026ndash;1022 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFern\u0026aacute;ndez, A. \u003cem\u003eet al.\u003c/em\u003e Is major depression adequately diagnosed and treated by general practitioners? Results from an epidemiological study. Gen. Hosp. Psychiatry 32, 201\u0026ndash;209 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR. M. Scheffler et al. Human Resources for Mental Health in Low and Middle Income Countries: Evidence from Bangladesh. in ResearchGate vol. 8 56 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKakuma, R. \u003cem\u003eet al.\u003c/em\u003e Human resources for mental health care: current situation and strategies for action. Lancet Lond. Engl. 378, 1654\u0026ndash;1663 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiao, L. \u003cem\u003eet al.\u003c/em\u003e The effectiveness of enhanced evidence-based care for depressive disorders: a meta-analysis of randomized controlled trials. Transl. Psychiatry 11, 1\u0026ndash;8 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo, T. \u003cem\u003eet al.\u003c/em\u003e Measurement-Based Care Versus Standard Care for Major Depression: A Randomized Controlled Trial With Blind Raters. Am. J. Psychiatry 172, 1004\u0026ndash;1013 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCummins, N. \u003cem\u003eet al.\u003c/em\u003e Multilingual markers of depression in remotely collected speech samples: A preliminary analysis. J. Affect. Disord. 341, 128\u0026ndash;136 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilliamson, J. R. et al. Vocal biomarkers of depression based on motor incoordination. in Proceedings of the 3rd ACM international workshop on Audio/visual emotion challenge 41\u0026ndash;48 (Association for Computing Machinery, New York, NY, USA, 2013). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/2512530.2512531\u003c/span\u003e\u003cspan address=\"10.1145/2512530.2512531\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohn, J. F. et al. Detecting depression from facial actions and vocal prosody. in 2009 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops 1\u0026ndash;7 (2009). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACII.2009.5349358\u003c/span\u003e\u003cspan address=\"10.1109/ACII.2009.5349358\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeintraub, M. J., Posta, F., Ichinose, M. C., Arevian, A. C. \u0026amp; Miklowitz, D. J. Word usage in spontaneous speech as a predictor of depressive symptoms among youth at high risk for mood disorders. J. Affect. Disord. 323, 675\u0026ndash;678 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKraepelin, E. Manic Depressive Insanity and Paranoia. J. Nerv. Ment. Dis. 53, 350 (1921).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEyben, F. \u003cem\u003eet al.\u003c/em\u003e The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing. IEEE Trans. Affect. Comput. 7, 190\u0026ndash;202 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRingeval, F. et al. AVEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1907.11510\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1907.11510\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eP\u0026eacute;rez-Toro, P. A. \u003cem\u003eet al.\u003c/em\u003e Depression assessment in people with Parkinson\u0026rsquo;s disease: The combination of acoustic features and natural language processing. Speech Commun. 145, 10\u0026ndash;20 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTasnim, M., Ehghaghi, M., Diep, B. \u0026amp; Novikova, J. DEPAC: a Corpus for Depression and Anxiety Detection from Speech. in Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology (eds. Zirikly, A. et al.) 1\u0026ndash;16 (Association for Computational Linguistics, Seattle, USA, 2022). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2022.clpsych-1.1\u003c/span\u003e\u003cspan address=\"10.18653/v1/2022.clpsych-1.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRiad, R. \u003cem\u003eet al.\u003c/em\u003e Automated Speech Analysis for Risk Detection of Depression, Anxiety, Insomnia, and Fatigue: Algorithm Development and Validation Study. J. Med. Internet Res. 26, e58572 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRude, S., Gortner,Eva-Maria \u0026amp; and Pennebaker, J. Language use of depressed and depression-vulnerable college students. Cogn. Emot. 18, 1121\u0026ndash;1133 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTackman, A. M. \u003cem\u003eet al.\u003c/em\u003e Depression, negative emotionality, and self-referential language: A multi-lab, multi-measure, and multi-language-task research synthesis. J. Pers. Soc. Psychol. 116, 817\u0026ndash;834 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Y. \u003cem\u003eet al.\u003c/em\u003e Identifying depression-related topics in smartphone-collected free-response speech recordings using an automatic speech recognition system and a deep learning topic model. J. Affect. Disord. 355, 40\u0026ndash;49 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBurdisso, S., Villatoro-Tello, E., Madikeri, S. \u0026amp; Motlicek, P. Node-weighted Graph Convolutional Network for Depression Detection in Transcribed Clinical Interviews. in INTERSPEECH 2023 3617\u0026ndash;3621 (2023). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2023-1923\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2023-1923\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVillatoro-Tello, E., Ram\u0026iacute;rez-de-la-Rosa, G., G\u0026aacute;tica-P\u0026eacute;rez, D., Magimai.-Doss, M. \u0026amp; Jim\u0026eacute;nez-Salazar, H. Approximating the Mental Lexicon from Clinical Interviews as a Support Tool for Depression Detection. in Proceedings of the 2021 International Conference on Multimodal Interaction 557\u0026ndash;566 (ACM, Montr\u0026eacute;al QC Canada, 2021). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3462244.3479896\u003c/span\u003e\u003cspan address=\"10.1145/3462244.3479896\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl Hanai, T., Ghassemi, M. \u0026amp; Glass, J. Detecting Depression with Audio/Text Sequence Modeling of Interviews. in 1716\u0026ndash;1720 (2018). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2018-2522\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2018-2522\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArevian, A. C. \u003cem\u003eet al.\u003c/em\u003e Clinical state tracking in serious mental illness through computational analysis of speech. PloS One 15, e0225695 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCummins, N., Baird, A. \u0026amp; Schuller, B. W. Speech analysis for health: Current state-of-the-art and the increasing impact of deep learning. Methods 151, 41\u0026ndash;54 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLow, D. M., Bentley, K. H. \u0026amp; Ghosh, S. S. Automated assessment of psychiatric disorders using speech: A systematic review. Laryngoscope Investig. Otolaryngol. 5, 96\u0026ndash;116 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacWhinney, B. \u0026amp; Wagner, J. Transcribing, searching and data sharing: The CLAN software and the TalkBank data repository. Gesprachsforschung Online-Z. Zur Verbalen Interakt. 11, 154\u0026ndash;173 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBommasani, R. et al. On the Opportunities and Risks of Foundation Models. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2108.07258\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2108.07258\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKojima, T., Gu, S. S., Reid, M., Matsuo, Y. \u0026amp; Iwasawa, Y. Large language models are zero-shot reasoners. in Proceedings of the 36th International Conference on Neural Information Processing Systems 22199\u0026ndash;22213 (Curran Associates Inc., Red Hook, NY, USA, 2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAgrawal, M., Hegselmann, S., Lang, H., Kim, Y. \u0026amp; Sontag, D. Large language models are few-shot clinical information extractors. in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (eds. Goldberg, Y., Kozareva, Z. \u0026amp; Zhang, Y.) 1998\u0026ndash;2022 (Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2022.emnlp-main.130\u003c/span\u003e\u003cspan address=\"10.18653/v1/2022.emnlp-main.130\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePourpanah, F. \u003cem\u003eet al.\u003c/em\u003e A Review of Generalized Zero-Shot Learning Methods. IEEE Trans. Pattern Anal. Mach. Intell. 45, 4051\u0026ndash;4070 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, Y., Mao, K., Zhang, Y. \u0026amp; Chen, J. CALLM: Enhancing Clinical Interview Analysis Through Data Augmentation With Large Language Models. IEEE J. Biomed. Health Inform. 28, 7531\u0026ndash;7542 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, K. et al. MentaLLaMA: Interpretable Mental Health Analysis on Social Media with Large Language Models. in Proceedings of the ACM Web Conference 2024 4489\u0026ndash;4500 (Association for Computing Machinery, New York, NY, USA, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3589334.3648137\u003c/span\u003e\u003cspan address=\"10.1145/3589334.3648137\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiyanage, C., Garg, M., Mago, V. \u0026amp; Sohn, S. Augmenting Reddit Posts to Determine Wellness Dimensions impacting Mental Health. \u003cem\u003eProc. Conf. Assoc. Comput. Linguist. Meet.\u003c/em\u003e 2023, 306\u0026ndash;312 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalgaroli, M. \u003cem\u003eet al.\u003c/em\u003e Large language models for the mental health community: framework for translating code to care. Lancet Digit. Health 7, e282\u0026ndash;e285 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai, H. \u003cem\u003eet al.\u003c/em\u003e A multi-modal open dataset for mental-disorder analysis. Sci. Data 9, 178 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLecrubier, Y. \u003cem\u003eet al.\u003c/em\u003e The Mini International Neuropsychiatric Interview (MINI). A short diagnostic structured interview: reliability and validity according to the CIDI. Eur. Psychiatry 12, 224\u0026ndash;231 (1997).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmerican Psychiatric Association, A. \u0026amp; Association, A. P. Diagnostic and Statistical Manual of Mental Disorders: DSM-IV. vol. 4 (American psychiatric association Washington, DC, 1994).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHamilton, M. A rating scale for depression. J. Neurol. Neurosurg. Psychiatry 23, 56\u0026ndash;62 (1960).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTao, F., Esposito, A. \u0026amp; Vinciarelli, A. The Androids Corpus: A New Publicly Available Benchmark for Speech Based Depression Detection. in 4149\u0026ndash;4153 (2023). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21437/Interspeech.2023-894\u003c/span\u003e\u003cspan address=\"10.21437/Interspeech.2023-894\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMontgomery, S. A. \u0026amp; Asberg, M. A new depression scale designed to be sensitive to change. Br. J. Psychiatry J. Ment. Sci. 134, 382\u0026ndash;389 (1979).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevis, B., Benedetti, A., Thombs, B. D., \u0026amp; DEPRESsion Screening Data (DEPRESSD) Collaboration. Accuracy of Patient Health Questionnaire-9 (PHQ-9) for screening to detect major depression: individual participant data meta-analysis. BMJ 365, l1476 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlummer, F., Manea, L., Trepel, D. \u0026amp; McMillan, D. Screening for anxiety disorders with the GAD-7 and GAD-2: a systematic review and diagnostic metaanalysis. Gen. Hosp. Psychiatry 39, 24\u0026ndash;31 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOkajima, I. \u003cem\u003eet al.\u003c/em\u003e Evaluation of Severity Levels of the Athens Insomnia Scale Based on the Criterion of Insomnia Severity Index. Int. J. Environ. Res. Public. Health 17, 8789 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRadford, A. et al. Robust Speech Recognition via Large-Scale Weak Supervision. in Proceedings of the 40th International Conference on Machine Learning 28492\u0026ndash;28518 (PMLR, 2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalgaroli, M., Hull, T. D., Zech, J. M. \u0026amp; Althoff, T. Natural language processing for mental health interventions: a systematic review and research framework. Transl. Psychiatry 13, 1\u0026ndash;17 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCummins, N. \u003cem\u003eet al.\u003c/em\u003e A review of depression and suicide risk assessment using speech analysis. Speech Commun. 71, 10\u0026ndash;49 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDevlin, J., Chang, M.-W., Lee, K. \u0026amp; Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (eds. Burstein, J., Doran, C. \u0026amp; Solorio, T.) 4171\u0026ndash;4186 (Association for Computational Linguistics, Minneapolis, Minnesota, 2019). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/N19-1423\u003c/span\u003e\u003cspan address=\"10.18653/v1/N19-1423\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReimers, N. \u0026amp; Gurevych, I. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (eds. Inui, K., Jiang, J., Ng, V. \u0026amp; Wan, X.) 3982\u0026ndash;3992 (Association for Computational Linguistics, Hong Kong, China, 2019). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/D19-1410\u003c/span\u003e\u003cspan address=\"10.18653/v1/D19-1410\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhattab, O. et al. DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines. in (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeam, G. et al. Gemma 3 Technical Report. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2503.19786\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2503.19786\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuessoum, S. B. \u003cem\u003eet al.\u003c/em\u003e Articulating biological and social approaches in child and adolescent psychiatry. Front. Child Adolesc. Psychiatry 1, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuinart, D., Kane, J. M. \u0026amp; Correll, C. U. Is Transcultural Psychiatry Possible? JAMA 322, 2167\u0026ndash;2168 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKordzadeh, N. \u0026amp; Ghasemaghaei, M. Algorithmic bias: review, synthesis, and future research directions. Eur. J. Inf. Syst. 31, 388\u0026ndash;409 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOhse, J. \u003cem\u003eet al.\u003c/em\u003e Zero-Shot Strike: Testing the generalisation capabilities of out-of-the-box LLM models for depression detection. Comput. Speech Lang. 88, 101663 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, Z. et al. Depression Detection in Clinical Interviews with LLM-Empowered Structural Element Graph. in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) (eds. Duh, K., Gomez, H. \u0026amp; Bethard, S.) 8181\u0026ndash;8194 (Association for Computational Linguistics, Mexico City, Mexico, 2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18653/v1/2024.naacl-long.452\u003c/span\u003e\u003cspan address=\"10.18653/v1/2024.naacl-long.452\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, J. et al. Optimizing Large Language Models for Detecting Symptoms of Comorbid Depression or Anxiety in Chronic Diseases: Insights from Patient Messages. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2503.11384\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2503.11384\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026Ouml;ng\u0026uuml;r, D. \u0026amp; Paulus, M. P. Embracing complexity in psychiatry\u0026mdash;from reductionistic to systems approaches. Lancet Psychiatry 12, 220\u0026ndash;227 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM\u0026ouml;ller, H. J. Rating depressed patients: observer- vs self-assessment. Eur. Psychiatry 15, 160\u0026ndash;172 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEnns, M. W., Larsen, D. K. \u0026amp; Cox, B. J. Discrepancies between self and observer ratings of depression: The relationship to demographic, clinical and personality variables. J. Affect. Disord. 60, 33\u0026ndash;41 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBzdok, D. \u0026amp; Meyer-Lindenberg, A. Machine Learning for Precision Psychiatry: Opportunities and Challenges. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 3, 223\u0026ndash;230 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK\u0026ouml;nig, I. R., Fuchs, O., Hansen, G., Mutius, E. von \u0026amp; Kopp, M. V. What is precision medicine? Eur. Respir. J. 50, (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWinter, N. R. \u003cem\u003eet al.\u003c/em\u003e A Systematic Evaluation of Machine Learning-Based Biomarkers for Major Depressive Disorder. JAMA Psychiatry 81, 386\u0026ndash;395 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerardi, M. \u003cem\u003eet al.\u003c/em\u003e Relative importance of speech and voice features in the classification of schizophrenia and depression. Transl. Psychiatry 13, 1\u0026ndash;8 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSilva, W. J., Lopes, L., Galdino, M. K. C. \u0026amp; Almeida, A. A. Voice Acoustic Parameters as Predictors of Depression. J. Voice 38, 77\u0026ndash;85 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartin, V. P. \u0026amp; Rouas, J.-L. Estimating Symptoms and Clinical Signs Instead of Disorders: The Path Toward The Clinical Use of Voice and Speech Biomarkers In Psychiatry. in ICASSP 2024\u0026ndash;2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 10606\u0026ndash;10610 (2024). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICASSP48485.2024.10445888\u003c/span\u003e\u003cspan address=\"10.1109/ICASSP48485.2024.10445888\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"speech, large language models, depression, multilingual, anxiety, insomnia, fatigue, machine learning, few-shot learning","lastPublishedDoi":"10.21203/rs.3.rs-6594999/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6594999/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge Language Models (LLMs) offer potential solutions for scalable depression detection across diverse populations. This study evaluates LLM-based speech content analysis for multilingual depression detection in clinical and general populations. We analyzed speech transcripts from three distinct cohorts: Chinese clinical (n\u0026thinsp;=\u0026thinsp;52), Italian clinical (n\u0026thinsp;=\u0026thinsp;116), and French general population (n\u0026thinsp;=\u0026thinsp;1,347). Our LLM-based system, using state-of-the-art open source LLM-model with few-shot prompting, was compared against traditional audio embedding and text embedding approaches for detecting depression and secondary symptoms (anxiety, insomnia, fatigue). The LLM system achieved excellent depression detection with F1-scores of 0.96 (Chinese), 0.85 (Italian), and 0.40 (French), consistently outperforming baseline methods. Depression sensitivity reached 1.00 (Chinese) and 0.93 (French), with high specificity in clinical populations (0.93 Chinese, 0.88 Italian). For secondary symptoms, anxiety detection performed well with high sensitivity (0.85 Chinese, 0.97 French) and F1-scores of 0.78 (Chinese) and 0.31 (French), while performance varied for other symptoms with fatigue detection performing at near-random levels. Statistical analysis revealed language-dependent benefits from few-shot learning, with Chinese datasets particularly benefiting from additional examples when using larger models. Our findings demonstrated that LLM-based speech analysis provides robust multilingual capabilities for depression detection without requiring language-specific training data, offering a scalable solution for mental health screening across diverse populations.\u003c/p\u003e","manuscriptTitle":"Automated speech content analysis to detect depression with large language models: towards multilingual and few-shot capabilities","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-13 04:17:20","doi":"10.21203/rs.3.rs-6594999/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"15227150-b0e3-4cd8-abcd-a248e89fb971","owner":[],"postedDate":"May 13th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":48201637,"name":"Health sciences/Diseases/Psychiatric disorders/Depression"},{"id":48201638,"name":"Health sciences/Biomarkers"}],"tags":[],"updatedAt":"2025-10-22T09:15:13+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-13 04:17:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6594999","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6594999","identity":"rs-6594999","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}