INTRODUCTION
Depression, a prevalent mental disorder, is projected to become the leading contributor to global disease burden by 2030 [1]. In particular, workers performing emotional labor face extreme stress and a high burden of mental health issues. A recent survey found that 26% of female call-center employees in Seoul had clinically significant depressive symptoms, exceeding the 9%-21% prevalence observed in the general Korean population [2]. Research on the detection and management of stress-related depression in occupations characterized by emotional labor, such as call center work, is becoming increasingly topical. Most previous studies on this population have relied on self-report surveys or interviews to identify mental health problems [2,3]. Survey-based assessments provide valuable snapshots of prevalence and risk factors; however, their cross-sectional designs are limited. Workers typically report their distress retrospectively, which can introduce recall bias and overlook day-to-day fluctuations. Recent efforts have therefore explored more objective and temporally sensitive approaches to capture stress and affective burden in real time.
Language use is widely recognized as a reflection of a person’s emotional state. Depressed individuals tend to use more negative emotion words and fewer positive ones when describing their lives [4-6]. For example, one study found that the emotional tone of brief written language responses predicted changes in depressive symptoms three weeks later, even after controlling for current mood and depression severity [5]. Similarly, Kaźmierczak et al. [7], who analyzed longer spoken narratives collected at three time points over a year, found an association between frequent use of negatively valenced words and more severe depression and anxiety symptoms. In addition, a recent ambulatory assessment study by Hartnagel et al. [8] highlighted the link between moment-to-moment language use and real-time affective states. By collecting brief spoken responses multiple times a day during a rapid mood-shifting intervention (sleep-deprivation therapy), this study showed that the use of positive and negative emotion words closely tracked concurrent depression severity. Similarly, Shin et al. [9] applied large language model (LLM) to user-generated diary texts collected over a two-week period and found that LLMderived sentiment features from free-text narratives could accurately detect depression severity.
Previous studies have employed a range of methods to analyze linguistic features associated with depression. Traditional lexicon-based approaches, such as Linguistic Inquiry and Word Count (LIWC), quantify the frequencies of psychologically relevant word categories (e.g., positive and negative emotion words and personal pronouns) and offer a rule-based framework for sentiment analysis [10]. However, these methods often fail to account for word order, negation, and contextual nuances. Recent advances in natural language processing, particularly in the use of LLMs, have enabled sentiment and symptom analyses that incorporate semantic depth and contextual meaning. LLM-based classification studies have generally focused on social media data, such as that from Twitter or Reddit, to identify depression-related language patterns [11,12]. For example, Bao et al. [11] applied GPT-3.5 in a few-shot setting on labeled datasets such as PsySym and BDI-Sen, achieving an F1 score of 0.91 in binary depression classification. Shin et al. [9] constructed a novel dataset by collecting user-generated emotional diary entries through a mobile app and tagging them with depression labels based on the Patient Health Questionnaire-9 (PHQ-9), Beck Scale for Suicide Ideation, and expert ratings. GPT-3.5 was fine-tuned and various prompt formats were tested, with the best F1 score of 0.685 achieved using chain-of-thought prompting [9]. In the clinical context, Ohse et al. [13] applied GPT-3.5 to transcribed interview text, classifying individual sentences into depression, anxiety, somatic, and insomnia symptom clusters. They achieved an F1 score of 0.82 for depression detection using PHQ-8 scores as the reference.
Neuroticism, a core personality dimension, is a well-established risk factor for major depressive disorder [14]. Individuals high in neuroticism tend to rely on maladaptive cognitive emotion-regulation strategies, which may impair their ability to effectively cope with stressful situations [15]. Neuroticism is closely associated with increased sensitivity to negative emotional stimuli [16]. This heightened affective reactivity is often reflected in language. For example, neurotic individuals have been found to use more negative emotion words [17]. Identifying linguistic markers of neuroticism in naturalistic data may enhance the predictive performance of language-based depression detection systems. Recent advances in natural language processing, particularly in LLM development, offer new opportunities for inferring personality-related traits from text. By leveraging the contextual and semantic understanding of LLMs, trait-level estimates of neuroticism could be derived directly from selfgenerated language samples.
While previous studies have demonstrated the utility of language-based approaches in mental-health assessments, many have relied on single-time-point or retrospective data and therefore lack temporal resolution. Moreover, recent studies using LLMs have focused on social media or aggregated diary data, thereby limiting their applicability in structured real-world settings. To address these limitations, this study implemented a high-frequency ecological momentary assessment (EMA) design. Prompted free-text responses and momentary mood ratings were collected three times daily over a two-week period from call-center workers. Sentiment scores were computed from each text using both LIWC- and LLM-based approaches. In addition, we evaluated whether neuroticism scores inferred from language using LLMs contributed to group differentiation indexed by depression severity. Using a repeated-measures approach, we examined whether temporal patterns in emotional language reflect differences in affective profiles indexed by depression severity and assessed the discriminative utility of these time-series sentiment features.
METHODS
Study design and participants
Participants were recruited from three call centers and one public agency. These workers performed emotional labor by providing civil consultations, customer service, sales, or administrative duties. Data were collected on-site at the participants’ workplaces between August 3, 2023 and January 25, 2024.
Figure 1A outlines the study design and participant flow. In total, 480 customer service agents participated. Participants completed in-person sociodemographic, psychiatric, and personality assessments through baseline surveys. They were asked to install an EMA app specifically developed for this study to facilitate real-time data collection. During a two-week period, participants completed three EMA sessions per day using the app, in which they reported their moods and stress levels and submitted voice recordings. Follow-up assessments, including psychiatric scales and in-person surveys, were conducted at the end of the two-week period. All participants provided written informed consent. The study was approved by the Institutional Review Boards (IRB) of CNUH and CNUHH (approval numbers: CNUH-2023-156 and CNUHH-2023-118).
EMA protocol
Figure 1B shows the structure and timing of the EMA protocol. Participants completed EMA tasks three times per day over the two-week study period. At 08:00 on each study day, participants rated their mood on a 10-point scale and recorded corresponding voice descriptions. At 13:00 and 18:00, they answered two stress-related questions and recorded voice descriptions of stressful experiences. The 18:00 session also included mood ratings, PHQ-2 items modified for daily assessment purposes within this study, and an anger-related item. All audio recordings were collected at 8 kHz.
Depressive symptoms and personality measures
Depression severity was measured using the PHQ-9 [18], a validated screening tool for depressive symptoms. The nine items are scored from 0-3, yielding a total score ranging from 0-27. Severity levels were categorized as follows: 0-4 (none), 5-9 (mild), 10-14 (moderate), 15-19 (moderately severe), and 20-27 (severe).
Personality traits were assessed using the Korean version of the Big Five Inventory-10 (BFI-K-10). This measure evaluates extraversion, agreeableness, conscientiousness, neuroti-cism, and openness, with two items for each trait rated on a 5-point Likert scale [19].
Voice transcription and text preprocessing
All voice recordings were transcribed using Vrew software [20]. To protect participants’ personal information, transcriptions were anonymized by trained annotators through the removal of identifying information (e.g., names and birthdates). Utterances by other individuals, as well as irrelevant content, were removed to ensure data quality. Text data were cleaned and standardized to enhance consistency, readability, and reliability for downstream linguistic and computational analyses. Korean spacing rules were applied, sentence boundaries were marked with periods, and grammatical and typographical errors were corrected using a spell-check tool. Nonsemantic expressions, such as filler words (“um,” “eh”), technical remarks (“Is it working?”), and nonverbal sounds (e.g., laughter), were excluded.
Final dataset
The initial dataset comprised 14,733 entries from 480 individuals. After excluding 467 entries because of unprocessable audio, 13 participants with missing PHQ-9 scores, and 67 participants whose PHQ-9 categories changed between baseline and follow-up, the final dataset included 12,418 entries from 400 participants (Supplementary Figure 1).
Analysis
Sentiment analysis approaches
Linguistic features were extracted using two complementary methods: lexical analysis using the LIWC and LLM based analysis. LIWC quantifies psychological constructs in texts based on word frequencies within predefined categories. This study used the KNU Korean Sentiment Lexicon [21], which includes general emotion words, emoticons, and abbreviations, with valence scores ranging from -2 to +2. This approach facilitated structured sentiment scoring based on established psycholinguistic theory. LLMs utilize transformer-based architectures capable of understanding contextual and semantic information in free text. This study selected Mistral Small (24 B; Mistral AI), Qwen 2.5 (32 B; Alibaba Cloud), and GPT-4o-mini (OpenAI) for LLM-based analysis in light of their performance on Korean language tasks and their instructionfollowing capabilities. Furthermore, this study aimed to incorporate a representative range of publicly available LLMs with diverse parameter scales, pre-training corpora, and training contexts. This selection was intended to mitigate potential model-specific biases and to observe varied response behaviors across different LLMs, rather than to identify a superior model (Supplementary Table 1). Zero-shot inference was performed to extract sentiment-relevant features directly from diary entries using these models, without the need for finetuning.
Sentiment and neuroticism scoring procedures
Using LIWC, the frequencies of positive and negative emotion words were calculated. In the LLM pipeline, zero-shot prompts were used to identify positive or negative emotional expressions (Supplementary Table 2), and sentiment scores were assigned on a -5 to +5 scale (Supplementary Table 3). To facilitate interpretation of the LLM-derived sentiment scores, representative diary excerpts with corresponding LLM-derived mood scores and self-rated mood scores are provided in Supplementary Table 4. These examples are included solely for illustrative purposes, highlighting cases in which numeric self-ratings appear less aligned with the emotional content of the text, and are not intended to imply ground-truth superiority of either measure.
To estimate neuroticism, each participant’s daily diary text was segmented by time of day and processed using LLMs. Scores were mapped to a scale of 1-5 for comparison with the BFI-K-10. The prompts used for this task are presented in Supplementary Table 5.
Statistical analysis of linguistic features
Several statistical approaches were used to examine the relationship between linguistic features and depression severity. First, t-tests and analysis of variance were applied to compare the frequencies of positive and negative words across depression severity groups. Second, group-level sentiment trajectories across different times of the day (morning, afternoon, and evening) were visualized to examine variations in temporal patterns in sentiment scores. Third, LLMs were used to generate neuroticism scores based on the linguistic content of participants’ diary entries, and logistic regression was applied to examine whether these scores could predict depression (PHQ-9 <5 vs. ≥5), adjusting for sex and education level.
To assess the psychological validity of LLM-derived features, two validation analyses were conducted. First, the consistency of LIWC- and LLM-derived sentiment scores with self-reported mood ratings was assessed by calculating their correlations with morning and evening entries. Separate calculations were performed as mood-related questions differed between the time periods. Second, the consistency of LLM-predicted neuroticism scores with BFI-K-10 scores was assessed by computing their correlations using both the original dataset and a version adjusted through imputation and normalization.
Modeling sentiment trajectories by depression severity
We constructed linear mixed-effects models using GPT-derived sentiment scores from Day 1-11, the period with sufficient participant coverage (n >200), to examine whether daily sentiment trajectories differed according to depression severity. This analysis collapsed the PHQ-9 depression severity classification into three categories (normal, mild, and moderate- to-severe) to enhance group-level contrasts and facilitate interpretation of trajectory differences. GPT-derived sentiment scores were selected for modeling because they demonstrated the most robust differentiation between depression severity groups in the preliminary analyses. Depression severity was modeled as a fixed effect, and a participant random intercept was included. The models were fitted across all time points as well as separately for each segment (morning, afternoon, and evening). Time was not included as a fixed or random effect.
Comparing model-derived sentiment and self-reported mood
Separate mixed-effects analyses were conducted to compare the respective abilities of LLM-derived sentiment scores and self-reported mood ratings to reflect depression severity. Only morning data were used because both self-reported mood and mood-related diary texts were available. Each score type was modeled independently to assess its ability to differentiate between the depression severity groups.
Cumulative sentiment-based approach for depression classification
A Long Short-Term Memory (LSTM) model was employed to evaluate whether morning sentiment scores can serve as distinctive features for depression classification. These scores were treated as sequential daily data, and the classification performance was assessed across cumulative monitoring durations from Days 1 to 11. A three-step preprocessing pipeline was applied to address missing data: participants with more than 50% missing values were excluded; gaps within the time series were filled using forward linear interpolation; and any remaining missing values were imputed using each participant’s average sentiment score. The data were then split into 80% training and 20% test sets using stratified sampling to preserve class distributions. Model performance was evaluated using area under the curve (AUC) scores.
RESULTS
Participant characteristics
Table 1 presents the participants’ demographic and psychological characteristics. Among the 400 call center employees analyzed, 73.2% (n=293) were female and 26.8% (n=107) were male. The participants had a mean age of 42.5 years (standard deviation=8.4). Smoking status (p=0.023), cohabitation status (p=0.006), and the four dimensions of the BFI-K-10, excluding openness (p=0.220), differed significantly between the normal and depressive groups. Notably, the depressive group had a higher proportion of smokers (29.6%) and individuals living alone (25.6%) than the normal group (19.3% and 14.2%, respectively). The depressive group also showed lower scores on extraversion, agreeableness, and conscientiousness, and higher scores on neuroticism. Regarding these group-level differences, an adjusted logistic regression confirmed that neuroticism was a significant independent risk factor for depression (odds ratio [OR]=1.105, p<0.001) (Supplementary Table 6).
Word frequency patterns by depression severity
Using LIWC and three LLMs (Mistral, Qwen, and GPT), the frequency of positive and negative words was analyzed across the depression groups, both as a binary variable (depressive group vs. normal group) and as a categorical variable based on PHQ-9 severity levels.
As shown in Table 2, individuals in the depressive group consistently used more negative words and fewer positive ones than those in the normal group. This pattern was most pronounced with GPT, which produced the most polarized distributions of emotional word usage. In contrast, LIWC showed weaker group differentiation with no significant difference in positive word frequency (p=0.294).
While the LLMs generally reflected a severity gradient across PHQ levels (i.e., higher severity correlated with more negative word use), this trend was most pronounced for GPT and Qwen but less evident for Mistral.
Temporal patterns in sentiment scores
We compared the sentiment scores generated by LIWC and LLMs across the time segments (morning, afternoon, and evening) and depression severity groups. Supplementary Figure 2 illustrates that sentiment scores were consistently lower in the depressive group than in the normal group, with LLMs showing greater separation between groups than LIWC. Among the LLMs, GPT demonstrated the clearest distinction between the depression severity groups. Temporal trends were also observed; in all groups, sentiment scores were highest in the morning and decreased in the evening. Notably, the LIWC sentiment scores remained positive throughout the day, whereas the LLM-based sentiment scores varied by group and time. Among the LLMs, GPT and Qwen produced negative sentiment scores for the depressive group, particularly in the morning, whereas Mistral showed consistently negative scores for both groups. These findings suggest that differences in sentiment according to depression severity are most detectable in the morning.
Neuroticism and depression severity: LLM-derived vs. self-reported measures
To determine whether language-derived neuroticism scores could predict depression, self-reported BFI-K-10 neuroticism scores were compared with scores predicted by LLMs based on two-week diary texts. All measures of neuroticism were significantly associated with depression. Notably, most LLMderived scores yielded larger ORs than self-reported measures, indicating stronger associations with depression. GPT produced the highest OR (1.193) among the LLMs, followed closely by Mistral (OR=1.191) and Qwen (OR=1.069). In contrast, the self-reported neuroticism score showed a slightly lower but still significant association (OR=1.119, p<0.001) (Supplementary Table 7).
Convergent validity of sentiment and personality features
Convergent validity was assessed by examining the correlations between LIWC- and LLM-derived sentiment scores and self-reported mood ratings using morning and evening entries. As shown in Supplementary Table 8, the correlations were generally low, with the strongest correlation observed between GPT sentiment and evening mood (r=0.535, p< 0.0001). The morning correlations were weaker across all LLMs. Furthermore, the LIWC sentiment scores showed minimal associations at both time points. We then assessed the correspondence between the self-reported and LLM-predicted neuroticism scores (Supplementary Table 9). The correlation coefficients were moderately low across the models. Once again, GPT showed the highest alignment after adjustment (r=0.264).
Sentiment score trajectory and depression group classification
The GPT-derived sentiment scores differed significantly across depression severity groups over the 11-day EMA period (Figure 2). The number of participant entries varied across days because of missing data (n=235-330 per day). The normal group had consistently higher sentiment scores than the mild and moderate-to-severe groups. Supplementary Figure 3 illustrates the sentiment trajectories in the morning, afternoon, and evening. Across all time segments, the Normal group had the highest sentiment scores, whereas the moderate-to-severe group scored the lowest. The group separation was most pronounced in the morning.
Consistent with these visual patterns, mixed-effects model results confirmed significant between-group differences across all time points (Normal vs. Mild: β=-0.516, p<0.001; Mild vs. Moderate-to-Severe: β=-0.736, p<0.001). Stratified analyses revealed that the morning sentiment scores provided the greatest discriminatory power (Normal vs. Mild: β=-0.551, p=0.005; Mild vs. Moderate-to-Severe: β=-0.933, p<0.001). The afternoon and evening scores also demonstrated smaller but significant effects (Table 3).
Comparing LLM-derived sentiment and self-reported mood across depression severity levels
Supplementary Figure 4 shows the self-reported morning mood score trajectories over the 11-day study period. Although the Normal group generally reported a higher mood than the Moderate-to-Severe group, there was substantial overlap with no consistent separation of groups over time. While the GPTderived sentiment scores identified significant pairwise group differences, the self-reported mood scores did not have comparable discriminatory power. The differences between the normal and mild groups (β=-0.117, p=0.494) and the mild and moderate-to-severe groups (β=-0.133, p=0.5080) were not significant (Table 3).
Notably, on average, fewer participants (mean, 256 vs. 289 per day) provided daily self-reported mood scores than voicebased sentiment responses, potentially reflecting lower adherence to numeric self-reporting.
Prediction performance of cumulative morning sentiment scores across monitoring durations
Figure 3 shows the cumulative prediction performance of the LSTM model using morning sentiment scores for up to 11 days. The AUC scores declined until Day 6 and then increased from Day 7 onward. The LSTM model achieved its highest performance on Day 10 with an AUC score of 75.1% (71.5% on Day 11). Confusion matrices are presented in Supplementary Figure 5.
DISCUSSION
This study evaluated whether LLMs, which incorporate contextual understanding, can extract sentiments and personality traits from daily free-text entries to characterize longitudinal affective patterns associated with depression severity. Compared with traditional lexicon-based approaches, LLM-derived features demonstrated superior discriminative power in differentiating affective profiles across levels of depression severity. By conducting a longitudinal analysis of two weeks of EMA data, we found that LLM-derived sentiment scores showed clearer separation between depression severity groups than self-reported mood scores. The cumulative data yielded a discrimination performance exceeding 70%, demonstrating the feasibility of this approach. These findings highlight the potential utility of LLMs for deriving objective, language-based indicators of longitudinal affective burden in real-world occupational settings. Further methodological enhancements, such as model fine-tuning and personalization, may extend this approach to support scalable monitoring of affective states in high-stress occupational groups.
The LLM-based approach outperformed LIWC in both word-frequency analysis and sentiment scoring. This may be attributed to LLMs’ superior ability to capture contextual nuances within diary entries [22]. This contextual sensitivity may also have enhanced the estimation of personality traits. The LLM-predicted neuroticism scores demonstrated stronger associations with depression severity than the self-reported BFI scores. These findings suggest that LLMs may complement conventional self-reporting instruments when analyzing naturalistic language data. Notably, we observed LLM-specific differences in emotional sensitivity, temporal trends, and alignment with psychological constructs such as depression and neuroticism. The study did not aim to compare model performance directly, owing to substantial differences in their training data, fine-tuning scope, and parameter sizes. Nevertheless, differences were observed in how each model captured affective patterns. GPT demonstrated the most distinct separation between groups with differing levels of depressive symptoms, whereas Mistral showed flattened affective distinctions and Qwen showed intermediate performance. Given its relatively stronger discriminative capacity across key outcomes, subsequent analyses to further explore model-driven insights were conducted using GPT alone.
The LLM-derived sentiment trajectories revealed clear distinctions among groups defined by depression severity across the two-week EMA period. The value of a time-series perspective was both visually and statistically evident. When examined cross-sectionally, the mild group frequently overlapped with either the normal or moderate-to-severe groups. When considering a single time point, accurate differentiation between depression severity groups was difficult. By contrast, sentiment trajectories over time showed more consistent and separable patterns across groups. This supports prior research using speech and self-reported mood scores, which has emphasized the advantages of temporal modeling. For instance, daily monitoring of speech or mood captures subtle changes in depressive symptoms more effectively than infrequent as-sessments, as demonstrated by Gumus et al. [23] and Horwitz et al. [24] Furthermore, van de Leemput et al. [25] showed that early warning patterns, such as increasing mood instability, frequently precede depressive-episode onset and that these patterns can only be detected through a time-series approach. Shin et al. [9] also showed that LLM-based linguistic analysis of aggregated diary texts can reflect depression severity. Our study extends this work by demonstrating how longitudinal daily language trajectories enhance group-level differentiation when analyzed within a time-series framework.
In the stratified analysis, the data collected during the morning session yielded the most pronounced group differences in the LLM-derived sentiment scores. This likely reflects the EMA structure in which morning questions explicitly ask about the respondent’s current mood, whereas afternoon and evening questions focus on stress-related events. These findings suggest that the question content directly influences the strength of emotional signals captured by LLMs, with moodrelated questions providing more diagnostic information. Question design and timing shape the data quality, as has been emphasized in prior research [26]. This effect is clearly demonstrated by the stronger sentiment-based group separation observed in the morning responses.
Despite the low correlations between LLM-derived sentiment scores and self-rated mood, as well as between LLMpredicted and self-reported neuroticism, both LLM approaches showed comparable or superior discriminative capacities in distinguishing between depression severity groups. This discrepancy may reflect limitations of subjective self-reporting, particularly among individuals with reduced emotional awareness or introspective accuracy. This phenomenon is also commonly observed in depression [27]. Self-reporting can additionally be affected by recall bias, situational mood fluctuations, and social desirability, especially in occupational settings with emotional display rules [28,29]. In contrast, languagebased approaches may capture latent affective patterns that are less accessible to conscious self-report. At the same time, LLM-derived signals may reflect linguistic or cultural patterns embedded in natural language rather than affective states per se, and the possibility of model hallucinations cannot be fully excluded. Future studies should validate these findings across diverse linguistic and cultural contexts. Taken together, these findings indicate that language-derived affective indicators may serve as a complementary signal to traditional self-report measures, although broader cross-linguistic and cross-cultural validation will be required.
We observed a discrepancy between the number of completed morning voice recordings and mood ratings. The number of completed morning voice recordings per day ranged from 196-328, whereas the corresponding mood ratings ranged from 191-318, indicating that the participants completed the speech task more frequently than the self-rating task. Direct mood reporting may therefore be perceived as effortful or intrusive, particularly when repeated over time. This may reflect a general tendency to avoid repetitive or high-effort self-reporting tasks. These tendencies may be amplified in individuals with depressive symptoms who are more likely to experience reduced motivation and fatigue [30]. Unlike self-reporting, LLM-derived sentiment scores are passively extracted from transcribed speech, allowing more complete and consistent data to be captured with minimal participant effort.
Cumulative data yielded an AUC of 0.75 and a macro F1 of 0.683 for depression classification by Day 10. This performance may be considered relatively modest. A recent study that applied chain-of-thought prompting to two weeks of diary text from 91 participants to predict depression achieved a comparable F1 score of 0.685 and a balanced accuracy of 0.799 using a fine-tuned GPT-3.5 model [9]. In another EMA-based study, Lei et al. [31] analyzed 25 days of twice-daily mood and stress ratings in 103 participants to predict suicidal ideation. Their AUC remained below 0.7 until Day 11 and only exceeded 0.8 after the full 25 days were included. Although neither of these studies incorporated time-series modeling, we achieved a comparable performance using a single feature type without fine-tuning. Furthermore, the marked improvement observed after Day 8 suggests that a minimum duration of approximately one week is required for group-level differentiation. This is consistent with a prior finding that at least eight days of smartphone and wearable data are needed to achieve stable prediction of depression severity [32]. Given the larger sample size and use of directly collected natural-language data in our design, there remains potential to further improve classification performance through future model optimization.
This study has several limitations. First, relatively simple prompting strategies were used, and further validation of LLMderived sentiment and neuroticism scores against established psychometric instruments is needed. Second, the study population consisted of workers in high-stress occupational environments, which may limit the generalizability of the findings to other populations. Third, although transcription and preprocessing procedures included anonymization, naturalistic language data may still contain contextual details that could be recognizable to the participants themselves or to individuals familiar with them, reflecting an inherent limitation of real-world data. Emerging on-device language models may help by enabling local processing of sensitive data and reducing manual review, thereby supporting scalable deployment. Finally, while LLM-derived features demonstrated meaningful associations with depression severity, further work is needed to evaluate their robustness across different linguistic, cultural, and modeling contexts. As vision-language models continue to advance, this framework may be extended to incorporate behavioral observations, enabling a more comprehensive multimodal assessment of affective states.
The LLM-derived sentiment and personality-related features in this study carry important clinical implications, as they enable the inference of affective patterns without directly querying individuals. This approach may facilitate more efficient and low-burden monitoring in real-world settings, particularly in contexts where repeated self-reporting is challenging. Furthermore, the ability to leverage large volumes of naturally occurring language data in daily life suggests a pathway toward scalable mental health assessment and may contribute to the development of personalized and precision psychiatry. Future work could strengthen and extend these findings through validation against clinician-administered assessments, fine-tuning to enhance sensitivity, and integration with behavioral or physiological signals for multimodal assessment across diverse populations. In this study, LLMs successfully extracted affective signals from daily free-text entries and differentiated patterns associated with depression severity more effectively than self-reported measures, with performance improving over time. To our knowledge, this is one of the first large-scale studies to apply LLMs to time-series free-text data, highlighting their potential to advance workplace mental health assessment and population-level monitoring.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement1
Print
