Toward a Mobile Platform for Real-world Digital Measurement of Depression: User-Centered Design, Data Quality, and...

Area of Science:

• Digital Phenotyping in Mental Health
• Mobile Health (mHealth) Interventions
• Computational Psychiatry

Background:

• Effective mental health treatments exist, but matching patients to optimal therapies and assessing response is challenging due to a lack of objective psychiatric symptom measurement.
• Smartphones offer a scalable, low-cost method for collecting real-world patient data via sensors and active tasks, potentially enabling measurement-based care.

Purpose of the Study:

• To test the feasibility of a remote study using a smartphone app to collect subjective and objective depression measures.
• To develop an engaging user interface for high task adherence, assess data quality from passive sensors, and build clinically relevant behavioral features.
• To explore preliminary connections between these digital features and depression severity.

Main Methods:

• A 12-week, fully remote, observational study involving 600 participants downloading an Android app.
• Passive collection of 20 sensor data streams, daily surveys, weekly voice diaries, and the Patient Health Questionnaire (PHQ-9).
• Computation of correlations between behavioral features (e.g., time spent at home) and PHQ-9 scores; construction of a logistic regression model.

Main Results:

• 415 individuals logged in, completing 83.35% of daily PHQ-9 surveys over 12 weeks.
• 11 out of 34 behavioral features significantly correlated with weekly PHQ-9 scores (P<.001), including voice sentiment and ambient audio.
• A logistic regression model predicting depression status achieved a cross-validated mean area under the curve of 0.656.

Conclusions:

• The study provides a strong proof of concept for smartphone-based depression outcome assessment.
• Behavioral features from passive sensors and active tasks show promising correlations with the PHQ-9.
• Future research should focus on increasing study scale for more complex predictive models and improving data missingness handling.