Machine Learning-Based Depression Detection Using Physiological Signals and Smartphone-Based Behavioral Interventions

Sadeghi M, Rahimi F, Singh S, Richer R, Schindler-Gmelch L, Rupp L, Keinert M, Hager M, Berking M, Eskofier BM (2025)

Publication Language: English

Publication Type: Conference contribution

Publication year: 2025

Publisher: SPRINGER

City/Town: INTERNATIONAL JOURNAL OF BEHAVIORAL MEDICINE

Book Volume: 32

Conference Proceedings Title: International Congress of Behavioral Medicine

Event location: Vienna

Abstract

1 Background

Depression impacts mental health worldwide, with symptoms of persistent depressed mood, cognitive impairment, loss of interest, and suicidal ideation. Traditional diagnosis relies on time-intensive assessments by qualified clinicians, who may not always be available. Advances in machine learning (ML) enable objective, non-invasive depression detection using physiological signals such as electrocardiograms (ECG), respiratory signals (RSP), and electromyograms (EMG).

2 Purpose

This study aims to enhance depression detection and symptom severity prediction using ML on physiological signals collected during user interaction with a cognitive-behavioral-therapy-based mobile app. Analyzing biosignals, we aim to develop objective methods for identifying depressive symptoms and evaluating mobile interventions.

3 Method

In the Collaborative Research Centre “EmpkinS,” 256 participants underwent a randomized controlled trial, 128 diagnosed with depression via SCID-5-CV. Physiological data (ECG, RSP, EMG from facial muscles) were collected during app interaction. Data preprocessing included noise filtering, window segmentation, and feature extraction via NeuroKit Toolbox. Feature selection employed Recursive Feature Elimination. ML models (Decision Trees, Random Forest, XGBoost) classified depressive symptoms and predicted PHQ-8 severity scores. Early fusion combined modalities for multimodal analyses.

4 Results

Multimodal classification using XGBoost achieved an F1 score of 85.45%, outperforming single-modality ECG (83.16%). For severity prediction, the RSP-based XGBoost model reached a mean absolute error (MAE) of 3.49. Multi-modal regression yielded a slightly higher MAE of 3.70.

5 Conclusion(s)

ML applied to physiological data collected during mobile-based interventions offers a promising, objective method for depression detection. While multimodal

approaches improve classification, severity prediction remains modality-specific. Future work should focus on improving multimodal fusion strategies and advancing feature engineering.

Authors with CRIS profile

Involved external institutions

Helmholtz Zentrum München - Deutsches Forschungszentrum für Gesundheit und Umwelt (HMGU) / Helmholtz Munich Germany (DE)

How to cite

APA:

Sadeghi, M., Rahimi, F., Singh, S., Richer, R., Schindler-Gmelch, L., Rupp, L.,... Eskofier, B.M. (2025). Machine Learning-Based Depression Detection Using Physiological Signals and Smartphone-Based Behavioral Interventions. In International Congress of Behavioral Medicine. Vienna, AT: INTERNATIONAL JOURNAL OF BEHAVIORAL MEDICINE: SPRINGER.

MLA:

Sadeghi, Misha, et al. "Machine Learning-Based Depression Detection Using Physiological Signals and Smartphone-Based Behavioral Interventions." Proceedings of the 18th International Congress of Behavioral Medicine (ICBM 2025), Vienna INTERNATIONAL JOURNAL OF BEHAVIORAL MEDICINE: SPRINGER, 2025.

BibTeX: Download