JMIR Publications

ABSTRACT

Background:

Levels of depression in adolescents have been trending upwards over the past several years. According to a 2019 survey by the National Survey on Drug Use and Health, 3.8 million U.S. adolescents have experienced at least one major depressive episode. This number constitutes about 15% of adolescents between the age of 12 to 17. However, 57% of those with major depressive episodes did not receive treatment. Most studies on depression are geared towards the adult population. There are very few studies that look into predicting depression in adolescents. Better studies are necessary to detect and predict depressive behavior in adolescents and prevent its escalation.

Objective:

The aim of our work is to study passively sensed data of adolescents with depression and investigate predictive capabilities of two machine learning approaches to predict depression scores and change in depression levels of adolescents. This work also provides an in-depth analysis of sensor features that serve as key indicators of change in depressive behavior and the effect of variation of data samples on model accuracy levels.

Methods:

The study consisted of 55 adolescents with symptoms of depression between the ages of 12 and 17 inclusive. Each participant was passively monitored through the AWARE application on their personal smartphones and a Fitbit Inspire wearable device for 24 weeks. Passive sensors used in the study collected calls, conversations, location, heart rate, steps, sleep, screen, and WiFi information on a daily basis. Following data preprocessing, 37 participants in the aggregated dataset were analyzed. Weekly Patient Health Questionnaire-9 (PHQ-9) surveys answered by participants served as ground truth. We applied regression-based approaches to predict the PHQ-9 depression score and change in depression severity category. These approaches were consolidated by universal and personalized modeling strategies. The universal strategies consisted of leave one participant out (LOPO) and leave week X out (LWXO). The personalized strategies models were based on accumulated weeks (ACCU) and leave one week one user instance out (LOWOU).

Results:

We observed that personalized approaches performed better on adolescent depression prediction compared to universal approaches. The best models were able to predict depression score and change in depression level with RMSEs of 2.83 and 3.21 respectively following the ACCU personalized modeling strategy. Our feature importance investigation showed that the contribution of screen, calls and location-based features influenced optimum models and were predictive of adolescent depression.

Conclusions:

This study provides insight into the feasibility of using passively sensed data for predicting adolescent depression. We demonstrated prediction capabilities in terms of depression score and change in depression level. The prediction results revealed that personalized models performed better on adolescents in comparison to universal approaches. Frequently chosen features by optimum performing machine learning models were investigated to form a better understanding of depression and sensor data. Furthermore we also share the impact of varying data on model performance.