Related Experiment Video
Updated: Sep 6, 2025
12:51
Evaluation of Commercial-Off-The-Shelf Wrist Wearables to Estimate Stress on Students
Published on: June 16, 2018
7.6K
Predicting Depression in Adolescents Using Mobile and Wearable Sensors: Multimodal Machine Learning-Based Exploratory
Tahsin Mullick1, Ana Radovic2, Sam Shaaban3
1Department of Engineering Systems and Environment, University of Virginia, Charlottesville, VA, United States.
JMIR Formative Research
|June 24, 2022
Summary
Personalized machine learning models accurately predict adolescent depression using smartphone and wearable sensor data. This approach offers a promising avenue for early intervention and improved mental health outcomes in young individuals.
Area of Science:
- Digital health
- Machine learning in mental health
- Adolescent psychology
Background:
- Adolescent depression rates are rising, with a significant portion of affected individuals not receiving treatment.
- Early identification of worsening symptoms through mobile and wearable sensors could enable timely intervention.
- Existing research on sensor-based depression prediction primarily focuses on adults, with limited studies on adolescents.
Purpose of the Study:
- To investigate the predictive capabilities of machine learning models using passively sensed data for adolescent depression.
- To analyze key sensor-based features indicative of depressive symptom changes in adolescents.
- To assess the impact of data sample variations on model accuracy for adolescent depression prediction.
Main Methods:
- Passive monitoring of 55 adolescents (aged 12-17) with depression symptoms using smartphones and wearables for 24 weeks.
- Collection of daily data including calls, conversations, location, and heart rate.
- Application of regression-based approaches (linear and nonlinear) with universal and personalized modeling strategies to predict depression scores and changes, validated against weekly Patient Health Questionnaire-9 surveys.
Main Results:
- Personalized modeling strategies outperformed universal approaches in predicting adolescent depression.
- The Accumulated Weeks personalized strategy achieved root mean squared errors of 2.83 for depression score and 3.21 for weekly change.
- Screen time, call patterns, and location data were identified as significant predictive features for adolescent depression.
Conclusions:
- Passively sensed data holds feasibility for predicting adolescent depression scores and changes.
- Personalized models demonstrate superior performance compared to universal models for adolescent depression prediction.
- Feature importance analysis enhances understanding of depression indicators within sensor data, paving the way for advanced predictive tools.
