Min Jhon
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Min Jhon has a total of 84 co-authors. Here are the top co-authors by number of shared publications:
- Sung-Wan Kim
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Korea
30 shared publications - Jae-Min Kim
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Republic of Korea
27 shared publications - Ju-Yeon Lee
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Republic of Korea
17 shared publications - Il-Seon Shin
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Korea
12 shared publications - Seunghyong Ryu
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Korea
11 shared publications - Hee-Ju Kang
Department of Psychiatry, Chonnam National University Medical School, Gwangju, Korea
10 shared publications
Latest Publications
- Open Access
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26  (22) , 34521-34533
A Frequency-Band Siamese Network With Dual Functional Connectivity Metrics for Subject-Independent EEG Depression Detection
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26  (22) , 34521-34533
Open AccessA Frequency-Band Siamese Network With Dual Functional Connectivity Metrics for Subject-Independent EEG Depression Detection
- Open Access
- 15 October 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e102768
Evaluating Large Language ModelâDerived Outputs as Psychological Measures: A Five-Domain Psychometric Framework With Application to Daily Stress
Background: Large language models (LLMs) are increasingly used to quantify psychological constructs from natural language, yet their outputs are typically evaluated only as predictions of an external label. When that label is itself imperfect, optimizing prediction conflates label agreement with construct measurement. No study has applied a multidomain measurement-oriented evaluation to LLM-derived psychological outputs. Objective: We applied 5 psychometric evaluation domains to LLM-derived outputs to test whether measurement-oriented evaluation surfaces properties that prediction-oriented evaluation does not: extraction repeatability, convergent validity with a lexicon-based tool, ecological concurrent validity decomposed into within-person and between-person components, incremental value in a 4-arm comparison with that tool, and method-blinded expert evaluation with structured error attribution. Methods: Thirteen linguistic features were extracted by GPT-5.2 (OpenAI) from 3556 person-days of daily voice narratives collected from 279 call center workers over 15 days via ecological momentary assessment. Features were evaluated across the 5 domains, with Linguistic Inquiry and Word Count, version 22, as the lexicon-based comparator. Multimodel robustness was assessed by re-extracting all features with Google Gemini 2.5 Flash. Results: Features showed high repeatability (mean intraclass correlation coefficient=0.987), construct-dependent convergent validity (Spearman Ď=0.08 to 0.79), and within-person tracking of same-day stress (negative emotion β=0.68). In a 4-arm comparison (baseline, baseline plus lexicon features, baseline plus LLM-derived features, and baseline plus both), LLM features increased explained variance in daily stress from 13.2% to 30.4%, versus 16.9% for the lexicon-based comparator. Among transcripts sampled for maximal method discordance, experts preferred LLM-derived profiles in 83.7% (261/312) of blinded decisions. Error analysis showed that 53% (35/66) of false-positive attributions reflected reference-label limitations, not model failures. Re-extraction with Google Gemini 2.5 Flash replicated key properties (person-level r =0.94). Three phenomena were visible under measurement-oriented evaluation that prediction-oriented evaluation does not typically examine: divergence between within-person and between-person effects, reference-label limitations as a dominant error source, and moderation of feature meaning by individual differences. Conclusions: Prediction metrics alone are insufficient for evaluating LLM-derived psychological measures. Measurement-oriented evaluation revealed properties, including level-dependent effects and asymmetric error attribution, that have direct implications for how LLM-based assessment studies interpret their results. For clinical use, the measure is a complement to existing screening instruments for longitudinal monitoring, not a stand-alone diagnostic.
- 15 October 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e102768
Open AccessEvaluating Large Language ModelâDerived Outputs as Psychological Measures: A Five-Domain Psychometric Framework With Application to Daily Stress
Background: Large language models (LLMs) are increasingly used to quantify psychological constructs from natural language, yet their outputs are typically evaluated only as predictions of an external label. When that label is itself imperfect, optimizing prediction conflates label agreement with construct measurement. No study has applied a multidomain measurement-oriented evaluation to LLM-derived psychological outputs. Objective: We applied 5 psychometric evaluation domains to LLM-derived outputs to test whether measurement-oriented evaluation surfaces properties that prediction-oriented evaluation does not: extraction repeatability, convergent validity with a lexicon-based tool, ecological concurrent validity decomposed into within-person and between-person components, incremental value in a 4-arm comparison with that tool, and method-blinded expert evaluation with structured error attribution. Methods: Thirteen linguistic features were extracted by GPT-5.2 (OpenAI) from 3556 person-days of daily voice narratives collected from 279 call center workers over 15 days via ecological momentary assessment. Features were evaluated across the 5 domains, with Linguistic Inquiry and Word Count, version 22, as the lexicon-based comparator. Multimodel robustness was assessed by re-extracting all features with Google Gemini 2.5 Flash. Results: Features showed high repeatability (mean intraclass correlation coefficient=0.987), construct-dependent convergent validity (Spearman Ď=0.08 to 0.79), and within-person tracking of same-day stress (negative emotion β=0.68). In a 4-arm comparison (baseline, baseline plus lexicon features, baseline plus LLM-derived features, and baseline plus both), LLM features increased explained variance in daily stress from 13.2% to 30.4%, versus 16.9% for the lexicon-based comparator. Among transcripts sampled for maximal method discordance, experts preferred LLM-derived profiles in 83.7% (261/312) of blinded decisions. Error analysis showed that 53% (35/66) of false-positive attributions reflected reference-label limitations, not model failures. Re-extraction with Google Gemini 2.5 Flash replicated key properties (person-level r =0.94). Three phenomena were visible under measurement-oriented evaluation that prediction-oriented evaluation does not typically examine: divergence between within-person and between-person effects, reference-label limitations as a dominant error source, and moderation of feature meaning by individual differences. Conclusions: Prediction metrics alone are insufficient for evaluating LLM-derived psychological measures. Measurement-oriented evaluation revealed properties, including level-dependent effects and asymmetric error attribution, that have direct implications for how LLM-based assessment studies interpret their results. For clinical use, the measure is a complement to existing screening instruments for longitudinal monitoring, not a stand-alone diagnostic.
- Open Access
- 28 September 2026
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
Inflammation-based Precision Medicine for Depression: From Biomarkers to Clinical Application
Hee-Ju Kang, Ju-Wan Kim, Min Jhon, Seon-Young Kim, Ju-Yeon Lee, Sung-Wan Kim, Il-Seon Shin, Jae-Min Kim. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.26.1491
- 28 September 2026
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
Open AccessInflammation-based Precision Medicine for Depression: From Biomarkers to Clinical Application
Hee-Ju Kang, Ju-Wan Kim, Min Jhon, Seon-Young Kim, Ju-Yeon Lee, Sung-Wan Kim, Il-Seon Shin, Jae-Min Kim. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.26.1491
- Open Access
- 25 September 2026
- Published by Korean Neuropsychiatric Association in Psychiatry Investigation
- Vol. 23  (9) , 1096-1106
Daily Language as an Objective Indicator of Depressive Affective States: A Two-Week Ecological Momentary Assessment Study in Emotional Labor Workers
Objective Workers performing emotional labor are at increased risk for depression, yet conventional self-report assessments often lack objectivity and temporal sensitivity. This study aimed to examine whether longitudinal analysis of daily natural language collected through ecological momentary assessment (EMA) can serve as an indicator of depressive affective states and to compare its temporal sensitivity with traditional self-report measures. Methods A total of 400 call center employees completed three voice-recorded free-text entries per day for two weeks using an EMA application. Transcriptions were analyzed using a lexicon-based sentiment approach (Linguistic Inquiry and Word Count [LIWC]) and three large language models (LLMs; GPT-4o-mini, Qwen, and Mistral) under zero-shot prompting. Depressive symptoms were assessed using the Patient Health Questionnaire-9, and neuroticism was measured using both self-reported questionnaires and language-inferred scores. Results LLM-derived sentiment scores significantly differentiated groups across levels of depressive symptom burden and consistently outperformed LIWC. Longitudinal analyses demonstrated clear group-level separation, particularly in morning entries, whereas self-reported mood ratings failed to distinguish groups and showed lower adherence over the two-week period. Language-inferred neuroticism exhibited stronger associations with depressive symptoms than self-reported neuroticism. A cumulative model based on morning sentiment scores showed progressively improved discrimination over time, reaching its highest performance on Day 10 (area under the curve=0.75). Conclusion Daily natural language captures meaningful longitudinal affective dynamics associated with depressive symptoms and may complement conventional self-report assessments as a low-burden and scalable indicator of depressive affective states in real-world occupational settings.
- 25 September 2026
- Published by Korean Neuropsychiatric Association in Psychiatry Investigation
- Vol. 23  (9) , 1096-1106
Open AccessDaily Language as an Objective Indicator of Depressive Affective States: A Two-Week Ecological Momentary Assessment Study in Emotional Labor Workers
Objective Workers performing emotional labor are at increased risk for depression, yet conventional self-report assessments often lack objectivity and temporal sensitivity. This study aimed to examine whether longitudinal analysis of daily natural language collected through ecological momentary assessment (EMA) can serve as an indicator of depressive affective states and to compare its temporal sensitivity with traditional self-report measures. Methods A total of 400 call center employees completed three voice-recorded free-text entries per day for two weeks using an EMA application. Transcriptions were analyzed using a lexicon-based sentiment approach (Linguistic Inquiry and Word Count [LIWC]) and three large language models (LLMs; GPT-4o-mini, Qwen, and Mistral) under zero-shot prompting. Depressive symptoms were assessed using the Patient Health Questionnaire-9, and neuroticism was measured using both self-reported questionnaires and language-inferred scores. Results LLM-derived sentiment scores significantly differentiated groups across levels of depressive symptom burden and consistently outperformed LIWC. Longitudinal analyses demonstrated clear group-level separation, particularly in morning entries, whereas self-reported mood ratings failed to distinguish groups and showed lower adherence over the two-week period. Language-inferred neuroticism exhibited stronger associations with depressive symptoms than self-reported neuroticism. A cumulative model based on morning sentiment scores showed progressively improved discrimination over time, reaching its highest performance on Day 10 (area under the curve=0.75). Conclusion Daily natural language captures meaningful longitudinal affective dynamics associated with depressive symptoms and may complement conventional self-report assessments as a low-burden and scalable indicator of depressive affective states in real-world occupational settings.
- Open Access
- 29 August 2026
- Published by Elsevier in SSRN Electronic Journal
Robust Temporal Edge-aware Heterogeneous Graph for Social Media Depression Detection under Visual Modality Incompleteness
Depression is a complex and temporally evolving mental health disorder whose assessment benefits from modeling users' longitudinal behavioral patterns. Social media platforms provide rich textual and visual records that enable data-driven depression-risk prediction. However, existing approaches often rely on text-only analysis or simplistic multimodal fusion, and struggle with the inherently incomplete and irregular nature of social media data, where visual information may be sparse or missing over time. To address these challenges, we propose TE-HGN, a Temporal Edge-aware Heterogeneous Graph Network for multimodal depression detection. The proposed model represents user activity as a heterogeneous graph with text and image nodes, and models temporal dynamics through ordinal, forward, and backward edges, along with synchronous cross-modal relations. All temporal relations are constructed exclusively among posts within a predefined observation window, enabling bidirectional contextual reasoning without accessing information beyond the observed user history. Furthermore, an edge-aware message passing mechanism explicitly incorporates temporal direction and relation types, while an adaptive gated fusion module dynamically adjusts modality contributions according to modality availability. Missing visual observations are handled without constructing placeholder image nodes or imputing missing visual information during graph message passing. Experiments on the Twitter and Reddit benchmark datasets demonstrate that TE-HGN consistently achieves superior or competitive performance and improved robustness under visual modality incompleteness, outperforming or matching strong state-of-the-art baselines on key evaluation metrics. The source code is available on Github.
- 29 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessRobust Temporal Edge-aware Heterogeneous Graph for Social Media Depression Detection under Visual Modality Incompleteness
Depression is a complex and temporally evolving mental health disorder whose assessment benefits from modeling users' longitudinal behavioral patterns. Social media platforms provide rich textual and visual records that enable data-driven depression-risk prediction. However, existing approaches often rely on text-only analysis or simplistic multimodal fusion, and struggle with the inherently incomplete and irregular nature of social media data, where visual information may be sparse or missing over time. To address these challenges, we propose TE-HGN, a Temporal Edge-aware Heterogeneous Graph Network for multimodal depression detection. The proposed model represents user activity as a heterogeneous graph with text and image nodes, and models temporal dynamics through ordinal, forward, and backward edges, along with synchronous cross-modal relations. All temporal relations are constructed exclusively among posts within a predefined observation window, enabling bidirectional contextual reasoning without accessing information beyond the observed user history. Furthermore, an edge-aware message passing mechanism explicitly incorporates temporal direction and relation types, while an adaptive gated fusion module dynamically adjusts modality contributions according to modality availability. Missing visual observations are handled without constructing placeholder image nodes or imputing missing visual information during graph message passing. Experiments on the Twitter and Reddit benchmark datasets demonstrate that TE-HGN consistently achieves superior or competitive performance and improved robustness under visual modality incompleteness, outperforming or matching strong state-of-the-art baselines on key evaluation metrics. The source code is available on Github.
- Open Access
- 25 February 2026
- Published by Korean Neuropsychiatric Association in Psychiatry Investigation
- Vol. 23  (2) , 227-235
The Impact of 5-HTTLPR Polymorphism on Depression and Lymphocyte Changes in Colorectal Cancer Patients Undergoing FOLFOX Chemotherapy
Objective Depression could be related to immune function among cancer patients. The serotonin transporter gene has been reported for its associations with both depression and immune regulation. This study investigates the interaction between depression, immunity, and 5-HT transporter gene-linked polymorphism (5-HTTLPR) among colorectal cancer (CRC) patients undergoing chemotherapy.Methods This prospective longitudinal study collected information on depression and lymphocyte percentages at two time points: the first cycle and the final cycles of 5-fluorouracil, leucovorin, and oxaliplatin (FOLFOX) chemotherapy. Clinical depression was assessed using the Hospital Anxiety and Depression Scale-depression subscale (HADS-D) score. Genotyping identified 5-HTTLPR alleles. The dependent variables were changes in the percentages of CD4+, CD8+, CD19+, and CD16/56+ lymphocytes between the two time points. Moderated regression analysis was used to find interactions.Results Among 104 patients, no significant direct associations were observed between changes in lymphocyte percentages and HADS-D scores. However, with the interaction of 5-HTTLPR polymorphism, the moderated regression analysis revealed two significant associations between HADS-D scores and changes in the percentages of CD4+ and CD16/56+ lymphocytes. Specifically, the percentage of CD4+ cells decreased, and the percentage of CD16/56+ cells increased, in relation to the s allele as depression worsened. These findings were consistent in a sensitivity analysis.Conclusion Changes in the percentage of CD4+ cells and CD16/56+ cells under depression were moderated by 5-HTTLPR alleles among CRC patients undergoing FOLFOX chemotherapy, suggesting a gene-environment interaction. Further research on the role of 5-HTTLPR in the immune system under depression among CRC patients is warranted.
- 25 February 2026
- Published by Korean Neuropsychiatric Association in Psychiatry Investigation
- Vol. 23  (2) , 227-235
Open AccessThe Impact of 5-HTTLPR Polymorphism on Depression and Lymphocyte Changes in Colorectal Cancer Patients Undergoing FOLFOX Chemotherapy
Objective Depression could be related to immune function among cancer patients. The serotonin transporter gene has been reported for its associations with both depression and immune regulation. This study investigates the interaction between depression, immunity, and 5-HT transporter gene-linked polymorphism (5-HTTLPR) among colorectal cancer (CRC) patients undergoing chemotherapy.Methods This prospective longitudinal study collected information on depression and lymphocyte percentages at two time points: the first cycle and the final cycles of 5-fluorouracil, leucovorin, and oxaliplatin (FOLFOX) chemotherapy. Clinical depression was assessed using the Hospital Anxiety and Depression Scale-depression subscale (HADS-D) score. Genotyping identified 5-HTTLPR alleles. The dependent variables were changes in the percentages of CD4+, CD8+, CD19+, and CD16/56+ lymphocytes between the two time points. Moderated regression analysis was used to find interactions.Results Among 104 patients, no significant direct associations were observed between changes in lymphocyte percentages and HADS-D scores. However, with the interaction of 5-HTTLPR polymorphism, the moderated regression analysis revealed two significant associations between HADS-D scores and changes in the percentages of CD4+ and CD16/56+ lymphocytes. Specifically, the percentage of CD4+ cells decreased, and the percentage of CD16/56+ cells increased, in relation to the s allele as depression worsened. These findings were consistent in a sensitivity analysis.Conclusion Changes in the percentage of CD4+ cells and CD16/56+ cells under depression were moderated by 5-HTTLPR alleles among CRC patients undergoing FOLFOX chemotherapy, suggesting a gene-environment interaction. Further research on the role of 5-HTTLPR in the immune system under depression among CRC patients is warranted.
- Open Access
- 28 January 2026
- Published by Springer Nature in Translational Psychiatry
- Vol. 16  (1)
Contactless depression screening via facial video-derived heart rate variability
Depression is a prevalent mental health condition that frequently remains undiagnosed, highlighting the need for objective and scalable screening tools. Heart rate variability (HRV) has emerged as a potential physiological marker of depression, and facial video-based HRV measurement offers a novel, contactless approach that could facilitate widespread, non-invasive depression screening. We analyzed data from 1453 individuals who completed facial video recordings and the Patient Health Questionnaire-9 (PHQ-9). A stacking ensemble classifier was developed using HRV features and basic demographic information to classify individuals with depressive symptoms. The ensemble incorporated four base learners (logistic regression, gradient boosting, XGBoost, and SVM) with an SVM meta-learner. Model performance was evaluated using 5-fold cross-validation. The stacking model achieved its best discrimination of AUROC 0.64 (AUPRC 0.45 and MCC 0.21). Incorporating demographic features alongside HRV improved performance over HRV alone. Feature importance analysis revealed that smoking status, sex, and medical comorbidities were the strongest contributors to the predictions. Facial video-derived HRV, combined with simple demographic factors, can moderately distinguish individuals with depressive symptoms in a contactless manner. Although predictive performance was modest, this non-invasive approach shows promise for accessible, large-scale depression screening.
- 28 January 2026
- Published by Springer Nature in Translational Psychiatry
- Vol. 16  (1)
Open AccessContactless depression screening via facial video-derived heart rate variability
Depression is a prevalent mental health condition that frequently remains undiagnosed, highlighting the need for objective and scalable screening tools. Heart rate variability (HRV) has emerged as a potential physiological marker of depression, and facial video-based HRV measurement offers a novel, contactless approach that could facilitate widespread, non-invasive depression screening. We analyzed data from 1453 individuals who completed facial video recordings and the Patient Health Questionnaire-9 (PHQ-9). A stacking ensemble classifier was developed using HRV features and basic demographic information to classify individuals with depressive symptoms. The ensemble incorporated four base learners (logistic regression, gradient boosting, XGBoost, and SVM) with an SVM meta-learner. Model performance was evaluated using 5-fold cross-validation. The stacking model achieved its best discrimination of AUROC 0.64 (AUPRC 0.45 and MCC 0.21). Incorporating demographic features alongside HRV improved performance over HRV alone. Feature importance analysis revealed that smoking status, sex, and medical comorbidities were the strongest contributors to the predictions. Facial video-derived HRV, combined with simple demographic factors, can moderately distinguish individuals with depressive symptoms in a contactless manner. Although predictive performance was modest, this non-invasive approach shows promise for accessible, large-scale depression screening.
- Open Access
- 19 January 2026
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
- Vol. 24  (4) , 732-745
Predicting Polysomnographyâderived Sleep Efficiency and Total Sleep Time Using Clinical Questionnaires: A Machine Learning Approach
Ju-Wan Kim, Yeongmin Kim, Inyong Jeong, Taewon Jung, Geesoo Hong, Dae-Kwang Kim, Hee-Ju Kang, Seunghyong Ryu, Il-Seon Shin, Sung-Wan Kim, Min Jhon, Hwamin Lee, Jae-Min Kim. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.25.1354
- 19 January 2026
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
- Vol. 24  (4) , 732-745
Open AccessPredicting Polysomnographyâderived Sleep Efficiency and Total Sleep Time Using Clinical Questionnaires: A Machine Learning Approach
Ju-Wan Kim, Yeongmin Kim, Inyong Jeong, Taewon Jung, Geesoo Hong, Dae-Kwang Kim, Hee-Ju Kang, Seunghyong Ryu, Il-Seon Shin, Sung-Wan Kim, Min Jhon, Hwamin Lee, Jae-Min Kim. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.25.1354
- Open Access
- 13 November 2025
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
- Vol. 24  (1) , 106-117
Sex-specific Serum Biomarker Associations with Antidepressant Treatment Outcomes in Depressive Disorders
Jae-Min Kim, Hee-Ju Kang, Ju-Wan Kim, Min Jhon, Min-Gon Kim, Ju-Yeon Lee, Sung-Wan Kim, Il-Seon Shin. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.25.1331
- 13 November 2025
- Published by Korean College of Neuropsychopharmacology in Clinical Psychopharmacology and Neuroscience
- Vol. 24  (1) , 106-117
Open AccessSex-specific Serum Biomarker Associations with Antidepressant Treatment Outcomes in Depressive Disorders
Jae-Min Kim, Hee-Ju Kang, Ju-Wan Kim, Min Jhon, Min-Gon Kim, Ju-Yeon Lee, Sung-Wan Kim, Il-Seon Shin. Clin Psychopharmacol Neurosci -0001;0:. https://doi.org/10.9758/cpn.25.1331
- Open Access
Contactless Depression Screening via Facial Video-derived Heart Rate Variability
Background: Depression is a prevalent mental health condition that frequently goes undiagnosed. Heart rate variability (HRV) has emerged as a potential objective marker of depression. Facial video-based HRV measurement offers a novel, contactless approach that could facilitate widespread, non-invasive depression screening. Methods: We analyzed data from 1,453 individuals who completed facial video recordings for HRV analysis and the Patient Health Questionnaire-9 (PHQ-9). A stacking ensemble classifier was developed using HRV features and basic demographic information to classify individuals with depressive symptoms. The ensemble incorporated four base learners (logistic regression, gradient boosting, XGBoost, and SVM) with an SVM meta-learner. Model performance was evaluated using 5-fold cross-validation. Results: The stacking model achieved its best discrimination of AUROC 0.64 (AUPRC 0.45 and MCC 0.21). Incorporating demographic features alongside HRV improved performance over HRV alone. Feature importance analysis revealed that smoking status, sex, and medical comorbidities were the strongest contributors to the predictions. Limitations: The predictive performance was modest, and HRV alone showed limited discrimination. Additionally, the findings are based on a single cohort and require validation in more diverse populations. Conclusion: Facial video-derived HRV, combined with simple demographic factors, can moderately distinguish individuals with depressive symptoms in a contactless manner. Although predictive performance was modest, this non-invasive approach shows promise for accessible, large-scale depression screening. Highlights: Facial video-derived HRV enables non-invasive, contactless depression screening. Stacking ensemble with SVM meta-learner optimized for MCC in depression screening. Combining HRV with demographics improved depression classification vs. HRV alone. Moderate yet consistent performance achieved with minimal, non-invasive inputs.
Open AccessContactless Depression Screening via Facial Video-derived Heart Rate Variability
Background: Depression is a prevalent mental health condition that frequently goes undiagnosed. Heart rate variability (HRV) has emerged as a potential objective marker of depression. Facial video-based HRV measurement offers a novel, contactless approach that could facilitate widespread, non-invasive depression screening. Methods: We analyzed data from 1,453 individuals who completed facial video recordings for HRV analysis and the Patient Health Questionnaire-9 (PHQ-9). A stacking ensemble classifier was developed using HRV features and basic demographic information to classify individuals with depressive symptoms. The ensemble incorporated four base learners (logistic regression, gradient boosting, XGBoost, and SVM) with an SVM meta-learner. Model performance was evaluated using 5-fold cross-validation. Results: The stacking model achieved its best discrimination of AUROC 0.64 (AUPRC 0.45 and MCC 0.21). Incorporating demographic features alongside HRV improved performance over HRV alone. Feature importance analysis revealed that smoking status, sex, and medical comorbidities were the strongest contributors to the predictions. Limitations: The predictive performance was modest, and HRV alone showed limited discrimination. Additionally, the findings are based on a single cohort and require validation in more diverse populations. Conclusion: Facial video-derived HRV, combined with simple demographic factors, can moderately distinguish individuals with depressive symptoms in a contactless manner. Although predictive performance was modest, this non-invasive approach shows promise for accessible, large-scale depression screening. Highlights: Facial video-derived HRV enables non-invasive, contactless depression screening. Stacking ensemble with SVM meta-learner optimized for MCC in depression screening. Combining HRV with demographics improved depression classification vs. HRV alone. Moderate yet consistent performance achieved with minimal, non-invasive inputs.