Digital phenotyping in bipolar disorder: Using longitudinal Fitbit data and personalized machine learning to predict...

Area of Science:

• Digital Health
• Machine Learning in Medicine
• Psychiatry

Background:

• Bipolar disorder (BD) treatment necessitates prompt identification of mood episodes.
• Passive sensor data from personal devices shows potential for mood episode detection.
• Existing methods lack broad applicability for mood symptomatology prediction.

Purpose of the Study:

• To evaluate a novel, personalized machine learning approach for detecting mood symptomatology in BD patients.
• To assess the accuracy of models trained solely on passive Fitbit data with minimal filtering.
• To determine the feasibility of broad application for mood prediction in BD.

Main Methods:

• Analysis of data from 54 adults with BD over 9 months, including Fitbit data and bi-weekly self-reports.
• Application of machine learning (ML) models to two-week aggregated Fitbit data.
• Detection of depressive and (hypo)manic symptomatology using established clinical cutoffs (PHQ-8, ASRM).

Main Results:

• The Binary Mixed Model (BiMM) forest algorithm demonstrated the highest predictive performance (ROC-AUC).
• In the testing set, ROC-AUC was 86.0% for depression and 85.2% for (hypo)mania.
• Optimized thresholds yielded 80.1% accuracy for depression and 89.1% for (hypo)mania detection.

Conclusions:

• Accurate detection of mood symptomatology in BD patients was achieved using broadly applicable methods.
• Findings support the use of Fitbit data for precise mood symptomatology predictions.
• This study introduces the BiMM forest for mood prediction, advancing personalized algorithms for wider patient populations.