Prediction of self-reported depression scores using person-generated health data from a virtual 1-year mental health observational study

Background: Undiagnosed mental illnesses represent one of the biggest challenges in our society. Due to stigma surrounding mental health, many people experience symptoms years before diagnosis and often never receive active management.Objectives:We use person-generated health data, consisting of self-reports and data from consumer wearable devices to predict an individual's depression severity level. Methods: Reference labels and input feature sets were derived from a 1-year long longitudinal cohort study consisting of 10,036 individuals. Participant-reported PHQ-9 scores were used as reference labels for depression severity, and input feature sets consisted of self-reported socio-demographic information, lifestyle and medication change surveys, and objective behavioral data collected using consumer wearables. Results: Our best performing model achieved an adjacent accuracy of 0.889 (CI±0.006) and a Kappa of 0.655 (CI±0.015). We observe that socio-demographic features contribute strongly to model performance, and that although good performance can be achieved with self-reported features, the addition of a small number of threshold-based features, derived from objective wearable data, improves model robustness. Conclusions: To our knowledge, the presented classification model is developed using the largest longitudinal cohort study ever considered for depression diagnosis, and one of the first attempts to predict granular depression severity, beyond bi-nary classification of depressed individuals versus healthy controls.We demonstrate the feasibility of our approach for this non-trivial problem. Future work will focus on combining the output labels of this model with self-reports in order to attempt to predict changes in individual, longitudinal mental health status.