Ensemble Approach to Combining Episode Prediction Models Using Sequential Circadian Rhythm Sensor Data from Mental...

Area of Science:

• Digital Health
• Computational Psychiatry
• Machine Learning in Medicine

Background:

• Managing mood disorders presents challenges due to counseling and drug treatment limitations.
• Patient empowerment through self-monitoring and predictive tools is vital for managing mental health.
• Current methods lack continuous, real-time insights into mood disorder progression.

Purpose of the Study:

• To validate the prediction of future depressive episodes using lifelog sequence data from digital device sensors.
• To assess the efficacy of various machine learning models in predicting mood disorder episodes.
• To optimize model performance by exploring data parameters.

Main Methods:

• Utilized diverse machine learning models including random forest, hidden Markov model, and recurrent neural network.
• Analyzed time-series data from digital device sensors for mood disorder prediction.
• Developed and evaluated a hybrid model combining multiple predictive algorithms.

Main Results:

• The hybrid model achieved a prediction accuracy of 0.78 for depressive episodes.
• F1-score performance for rare episode prediction was approximately 1.88 times higher than a dummy model.
• Identified key parameters (data sequence size, train-to-test ratio, labeling time slots) optimizing model performance.

Conclusions:

• Lifelog sequence data from digital devices can effectively predict depressive episodes.
• Machine learning, particularly hybrid models, offers a promising approach for mental health self-management and clinical insights.
• This study provides experimental validation using large-scale, long-term participant data.