Depression Diagnosis and Forecast based on Mobile Phone Sensor Data.

Area of Science:

• Digital Phenotyping
• Computational Psychiatry
• Machine Learning in Mental Health

Background:

• Mobile phone sensor data correlates with human depression states.
• Passive data collection offers a less time-consuming alternative to traditional self-assessment questionnaires.
• Previous research primarily focused on depression diagnosis, with limited attention to forecasting.

Purpose of the Study:

• To investigate the potential of passive mobile phone sensor data for depression forecasting.
• To compare the performance of depression forecasting against depression diagnosis using machine learning models.
• To identify key passive features from mobile phone usage for mental health monitoring.

Main Methods:

• Extraction of four types of passive features: phone call, phone usage, user activity, and GPS data.
• Implementation of a long short-term memory (LSTM) network.
• A subject-independent 10-fold cross-validation setup for both diagnostic and forecasting tasks.

Main Results:

• The forecasting task achieved comparable performance to the diagnostic task.
• Achieved 77.0% accuracy for major depression forecasting (binary classification).
• Achieved 53.7% accuracy for depression severity forecasting (5 classes) and an RMSE of 4.094 for PHQ-9 scores.

Conclusions:

• Passive mobile phone sensor data holds significant potential for forecasting depression.
• The findings suggest the feasibility of early intervention through continuous mental state monitoring.
• LSTM networks are effective for modeling and predicting depression states from digital footprints.