JMIR Publications

ABSTRACT

Background:

Depression is one of the most prevalent mental illnesses worldwide, negatively impacting numerous aspects of individuals’ lives. It is typically diagnosed by trained professionals through structured interviews and questionnaires, yet the condition often remains undetected due to stigma, limited access to clinical services, and the reliance on self-reported symptoms. Recent advances in digital health have positioned digital biomarkers as promising complementary tools that enable passive, continuous monitoring of behavioral markers to support clinical assessment. However, existing approaches frequently rely on specialized sensing hardware or controlled recording environments, limiting their applicability in real-world settings and often remaining restricted to single-modality data, preventing them from unfolding their full potential by not considering relevant complementary data streams.

Objective:

This work evaluates the feasibility of detecting depressive symptoms through data captured by earables, including voice-related features, gait parameters, and head motion to enable unobtrusive, real-world mental health monitoring.

Methods:

A study was conducted with 25 participants, of whom 21 were included in the analysis (12 healthy, 9 with clinically elevated depressive symptoms, defined as a PHQ-8 (Patient Health Questionnaire-8) cutoff score of 10). Voice and motion data were captured via hearing aid–embedded sensors during semi-structured interviews and a free walking task designed to elicit behavioral and emotional responses in naturalistic settings. Audio features included a minimalistic set of paralinguistic and emotion-related features. Motion analysis incorporated spatio-temporal gait features, generic time-series descriptors, and expert-designed features. Group differences were assessed using Mann-Whitney U tests and age- and BMI-corrected ANCOVAs. Classification was performed using logistic regression (LR), k-nearest neighbors (kNN), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost).

Results:

Multi-modal fusion of voice-, gait-, and motion-derived features yielded the best classification performance, with a balanced accuracy of 0.85 (SD 0.13) using a Random Forest classifier. An ablation study revealed that performance declined as feature sets were reduced, though configurations restricted to the interview condition or IMU-derived features alone retained competitive accuracies of up to 0.80. Unimodal configurations – voice or gait features only – achieved moderate balanced accuracies of 0.62, indicating that each modality independently carries relevant discriminative information, while their combination provides the most robust representation of depressive symptomatology.

Conclusions:

These findings demonstrate the feasibility of passive, unobtrusive depression detection using a single ear-worn device. Rather than relying on specialized hardware or single-modality pipelines, the proposed approach simultaneously captures rich multimodal behavioral information across voice, gait, and motion. The results support earables as viable platforms for multimodal mental health assessment, establishing a foundation for future large-scale validation and real-world deployment in continuous mental health monitoring.