2 - (SYM 133) Mood Episode Prediction in Patients with Bipolar Disorder Using Passively Collected Data from a Commercially-available Smartwatch

Research objective: Prompt identification of mood episodes is essential for effective treatment of bipolar disorder (BD). Passive sensor data from personal digital devices could allow detection of mood episodes between routine care appointments, but existing studies have not tested clinically feasible methods. This study investigated whether machine learning models trained on data from Fitbit smartwatches could detect depression and mania in patients with BD using clinically feasible methods. 

Methods: 57 adult patients, with a BD I or II diagnosis were recruited to complete 9-months of digital monitoring. During the monitoring period, patients were asked to wear a Fitbit device and complete a two biweekly, digital, symptom self-report measures, the Patient Health Questionnaire-8 (PHQ-8) and Altman Self-Rating Mania Scale (ASRM). Machine learning-based prediction models to detect depression and mania utilized Fitbit data aggregated over two-week observation windows. The presence of depression and mania was defined as having a score over established clinical cutoffs for the PHQ-8 or ASRM during the corresponding observation window. In the validation process, predictive accuracy of five machine learning-based prediction models was evaluated with the area under the receiver operating curve (ROC-AUC). Prediction accuracy was then characterized in the best performing model.

Results: 50 patients had sufficient data for prediction modeling (78.0% women; mean [SD] age, 39.6 [14.3]; 78.0% BD I]. In the validation process, among several machine learning algorithms, Binary Mixed Model (BiMM) forest achieved the highest ROC-AUC for depression (85%) and mania (89%). In the testing set, the ROC-AUC was 84% for depression and 87% for mania. Using optimized thresholds calculated with Youden’s J statistic, predictive accuracy was 79% for depression (sensitivity of 70% and specificity of 86%) and 79% for mania (sensitivity of 89% and a specificity of 78%).  

Implications: Findings suggest that it is feasible to predict mood episodes in BD from Fitbit-derived data features alone with clinically-useful levels of accuracy. In contrast to prior studies, prediction methods utilized did not require clinician input or perfect Fitbit compliance. Findings suggest that these methods may be feasible to use in clinical care, where they could provide objective and seamless progress monitoring and drive between-appointment outreach to treat mood episodes as soon as they emerge. Steps required for clinical application of findings are discussed.