Machine learning for passive mental health symptom prediction: Generalization across different longitudinal mobile...

Area of Science:

• Digital health
• Machine learning in mental health

Background:

• Mobile sensing and machine learning offer passive, remote mental health assessment.
• Previous models lacked generalizability due to homogeneous data and single platforms.

Purpose of the Study:

• To assess the generalizability of machine learning models trained on combined longitudinal study data for predicting mental health symptoms.
• To explore the impact of model personalization and oversampling on performance.

Main Methods:

• Combined data from CrossCheck and StudentLife studies.
• Used leave-one-subject-out cross-validation (LOSO-CV) for sleep quality and stress prediction.
• Investigated personalization and oversampling techniques.

Main Results:

• Models trained on combined data showed improved generalizability and predictive power compared to single-study models.
• Personalization offered limited improvements; oversampling enhanced severe symptom detection but reduced specificity.
• Decreased feature distribution distance correlated with improved model performance.

Conclusions:

• Machine learning models trained on aggregated, diverse mobile sensing data can generalize across heterogeneous datasets for mental health monitoring.
• Data sharing and standardization are crucial for advancing generalizability assessments in digital mental health research.