Using wearable data to detect depression severity across clinical and non-clinical samples

Abstract

Early detection of depression is crucial, yet current assessment methods depend on self-report questionnaires and clinical interviews, which are resource-intensive. Wearable devices provide a scalable way to assess physiological and behavioral features, but their predictive value across clinical and non-clinical populations remains insufficiently established. Wearable-derived features were collected from a student sample (n = 187) and an outpatient sample (n = 95). Depressive symptoms were assessed using the Patient Health Questionnaire-9 (PHQ-9), and participants were categorized as screen-positive for depressed (≥ 10) or non-depressed (< 10). An elastic net regularized logistic regression model was used for classification, with performance evaluated in held-out test data. Sensitivity analyses controlled for age and bedtime, tested alternative PHQ-9 cutoffs, and comparisons to baseline models with and without wearable features. Across the combined sample (n = 282), the model achieved good discriminative performance (area under the curve = 0.82; accuracy = 79%). Sensitivity analyses revealed that sample was a strong predictor, but wearable-derived features still added incremental value. Minimum awake heart rate, variability in sleep duration, and maximum step count emerged as the strongest predictors. Wearable-derived features show promise for detecting depressive symptoms across clinical and non-clinical populations. Sample-specific factors should be considered in future research.