Detecting Depression using Machine Learning Approaches through Digital Biomarkers

Sadeghi M, Rahimi F, Biradar PU, Richer R, Rupp LH, Schindler-Gmelch L, Keinert M, Hager M, Capito K, Berking M, Eskofier B (2025)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2025

Conference Proceedings Title: Bavarian Conference on AI in Medicine 2025

Event location: Munich

Abstract

Depression is a prevalent mental health disorder characterized by persistent sadness and loss of interest in enjoyable activities. Traditional diagnostic methods are subjective, highlighting a need for more objective and accurate approaches. This study investigates the application of classical machine learning (ML) algorithms to detect depression and predict its severity using facial digital biomarkers extracted from video data within the EmpkinS D02 dataset. 

Facial features, including Action Units (AUs), emotional cues, head orientation, and 2D landmarks, were extracted frame-by-frame using FaceReader 9. Data preprocessing involved handling missing values, feature standardization, and statistical aggregation. Feature selection was performed with SelectKBest using ANOVA F-statistics to identify the most informative features. 

Binary classification (depressed vs. healthy) and regression models (predicting PHQ-8 severity scores) were evaluated across three distinct experimental conditions: a Depressed Mood Induction phase using negative self-statements, a Coping phase (where participants actively invalidated depressogenic self-statements with disapproving facial expressions or validated antidepressive self-statements with affirming facial expressions), and a Comprehensive phase (including data from all experimental conditions). XGBoost achieved the highest classification performance in the Comprehensive phase with an accuracy of 81.82% and an F1-score of 80.36%, surpassing the targeted benchmark of 80%. For severity prediction, the LightGBM regression model performed best during the Comprehensive phase, attaining a Mean Absolute Error (MAE) of 2.02 and Root Mean Squared Error (RMSE) of 2.52. 

Statistical analyses identified significant facial biomarkers differentiating depressed and healthy participants, notably eyebrow movements (AU 04) and lip movements (AU 25 and AU 20). Performance variability across experimental conditions emphasized the sensitivity of ML models to data context and distribution.

The results demonstrate the feasibility and effectiveness of using ML-based facial feature analysis as a robust tool for enhancing depression diagnostics, suggesting promising integration potential in remote and resource-limited clinical settings.

Authors with CRIS profile

Related research project(s)

Involved external institutions

Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI) GmbH Germany (DE)

How to cite

APA:

Sadeghi, M., Rahimi, F., Biradar, P.U., Richer, R., Rupp, L.H., Schindler-Gmelch, L.,... Eskofier, B. (2025). Detecting Depression using Machine Learning Approaches through Digital Biomarkers. Poster presentation at Bavarian Conference on AI in Medicine, Munich.

MLA:

Sadeghi, Misha, et al. "Detecting Depression using Machine Learning Approaches through Digital Biomarkers." Presented at Bavarian Conference on AI in Medicine, Munich 2025.

BibTeX: Download