Contactless Depression Screening via Facial Video-derived Heart Rate Variability

BackgroundDepression 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. MethodsWe 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. ResultsThe 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. LimitationsThe 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. ConclusionFacial 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. HighlightsO_LIFacial video-derived HRV enables non-invasive, contactless depression screening. C_LIO_LIStacking ensemble with SVM meta-learner optimized for MCC in depression screening. C_LIO_LICombining HRV with demographics improved depression classification vs. HRV alone. C_LIO_LIModerate yet consistent performance achieved with minimal, non-invasive inputs. C_LI