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

Background: Effective treatment of bipolar disorder (BD) requires prompt response to mood episodes. Preliminary studies suggest that predictions based on passive sensor data from personal digital devices can accurately detect mood episodes (e.g., between routine care appointments), but studies to date do not use methods designed for broad application. This study evaluated whether a novel, personalized machine learning approach, trained entirely on passive Fitbit data, with limited data filtering could accurately detect mood symptomatology in BD patients.

Methods: We analyzed data from 54 adults with BD, who wore Fitbits and completed bi-weekly self-report measures for 9 months. We applied machine learning (ML) models to Fitbit data aggregated over two-week observation windows to detect occurrences of depressive and (hypo)manic symptomatology, which were defined as two-week windows with scores above established clinical cutoffs for the Patient Health Questionnaire-8 (PHQ-8) and Altman Self-Rating Mania Scale (ASRM) respectively.

Results: As hypothesized, among several ML algorithms, Binary Mixed Model (BiMM) forest achieved the highest area under the receiver operating curve (ROC-AUC) in the validation process. In the testing set, the ROC-AUC was 86.0% for depression and 85.2% for (hypo)mania. Using optimized thresholds calculated with Youden's J statistic, predictive accuracy was 80.1% for depression (sensitivity of 71.2% and specificity of 85.6%) and 89.1% for (hypo)mania (sensitivity of 80.0% and specificity of 90.1%).

Conclusion: We achieved sound performance in detecting mood symptomatology in BD patients using methods designed for broad application. Findings expand upon evidence that Fitbit data can produce accurate mood symptomatology predictions. Additionally, to the best of our knowledge, this represents the first application of BiMM forest for mood symptomatology prediction. Overall, results move the field a step toward personalized algorithms suitable for the full population of patients, rather than only those with high compliance, access to specialized devices, or willingness to share invasive data.