Digital phenotyping using wearable-determined physical behaviors and machine learning to detect depression and anxiety in a general population
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
Depression and anxiety are widespread mental health disorders, yet their diagnosis remains challenging. Digital phenotyping with wearable devices provides a promising approach for detecting depression and anxiety in the general population. This study aims to explore the extent to which wearable accelerometer-determined physical behavior metrics can be used as digital phenotypes for identifying individuals with and without depression and anxiety symptoms using machine learning (ML) algorithms.
Methods
At age 46 years old, participants (N = 2,810) from the Northern Finland Birth Cohort 1966 carried wrist- and waist-worn accelerometers for 14 consecutive days. Physical activity and sedentary behaviors were measured using data from the waist-worn device, while sleep behavior was identified based on data from the wrist-worn accelerometer. A total of 54 physical behavior metrics were extracted for each participant. Severity of the depression and anxiety symptoms were assessed using three validated instruments: the Beck Depression Inventory-II, Generalized Anxiety Disorder-7, and the Hopkins Symptom Checklist-25. Five ML algorithms were applied to identify individuals with and without depression and anxiety symptoms. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP) to assess the contribution of individual features.
Results
Among ML models, random forest achieved the best performance with accuracy (66%–72%) and AUC (66%–70%) for all three instruments. Physical behavior metrics extracted from accelerometers emerged as potential predictors of depression and anxiety. In SHAP analysis wake up time, time in bed, bed time, physical activity intensity proportions and prolonged sedentary bouts emerged as most important features.
Conclusions
Wearable-derived metrics of physical behaviors combined with ML models can be utilized with reasonably good accuracy to differentiate between participants with and without depression and anxiety symptoms. Our findings support the utility of wearable-derived physical behavior digital phenotypes for differentiating between participants with and without depression and anxiety symptoms in a general population.
