Abstract

Depression is a prevalent mental disorder. Current clinical and self-reported assessment methods of depression are laborious and incur recall bias. Their sporadic nature often misses severity fluctuations. Previous research highlights the potential of in-situ quantification of human behaviour using mobile sensors to augment traditional methods of depression management. In this paper, we study whether self-reported mood scores and passive smartphone and wearable sensor data could be used to classify people as depressed or non-depressed. In a longitudinal study, our participants provided daily mood (valence and arousal) scores and collected data using their smartphones and Oura Rings. We computed daily aggregations of mood, sleep, physical activity, phone usage, and GPS mobility from raw data to study the differences between the depressed and non-depressed groups and created population-level Machine Learning classification models of depression. We found statistically significant differences in GPS mobility, phone usage, sleep, physical activity and mood between depressed and non-depressed groups. An XGBoost model with daily aggregations of mood and sensor data as predictors classified participants with an accuracy of 81.43% and an Area Under the Curve of 82.31%. A Support Vector Machine using only sensor-based predictors had an accuracy of 77.06% and an Area Under the Curve of 74.25%. Our results suggest that digital biomarkers are promising in differentiating people with and without depression symptoms. This study contributes to the body of evidence supporting the role of unobtrusive mobile sensor data in understanding depression and its potential to augment depression diagnosis and monitoring.