Investigating Smartphone-Based Sensing Features for Depression Severity Prediction: Observation Study

Unobtrusively collected objective sensor data from everyday devices like smartphones provide a novel paradigm to infer mental health symptoms. This process, called smart sensing, allows a fine-grained assessment of various features (eg, time spent at home based on the GPS sensor). Based on its prevalence and impact, depression is a promising target for smart sensing. However, currently, it is unclear which sensor-based features should be used in depression severity prediction and if they hold an incremental benefit over established fine-grained assessments like the ecological momentary assessment (EMA).

The aim of this study was to investigate various features based on the smartphone screen, app usage, and call sensor alongside EMA to infer depression severity. Bivariate, cluster-wise, and cluster-combined analyses were conducted to determine the incremental benefit of smart sensing features compared to each other and EMA in parsimonious regression models for depression severity.

In this exploratory observational study, participants were recruited from the general population. Participants needed to be 18 years of age, provide written informed consent, and own an Android-based smartphone. Sensor data and EMA were collected via the INSIGHTS app. Depression severity was assessed using the 8-item Patient Health Questionnaire. Missing data were handled by multiple imputations. Correlation analyses were conducted for bivariate associations; stepwise linear regression analyses were used to find the best prediction models for depression severity. Models were compared by adjusted R². All analyses were pooled across the imputed datasets according to Rubin's rule.

A total of 107 participants were included in the study. Ages ranged from 18 to 56 (mean 22.81, SD 7.32) years, and 78% of the participants identified as female. Depression severity was subclinical on average (mean 5.82, SD 4.44; Patient Health Questionnaire score ≥10: 18.7%). Small to medium correlations were found for depression severity and EMA (eg, valence: r=-0.55, 95% CI -0.67 to -0.41), and there were small correlations with sensing features (eg, screen duration: r=0.37, 95% CI 0.20 to 0.53). EMA features could explain 35.28% (95% CI 20.73% to 49.64%) of variance and sensing features (adjusted R²=20.45%, 95% CI 7.81% to 35.59%). The best regression model contained EMA and sensing features (R²=45.15%, 95% CI 30.39% to 58.53%).

Our findings underline the potential of smart sensing and EMA to infer depression severity as isolated paradigms and when combined. Although these could become important parts of clinical decision support systems for depression diagnostics and treatment in the future, confirmatory studies are needed before they can be applied to routine care. Furthermore, privacy, ethical, and acceptance issues need to be addressed.