Temporal and Type Correlation in Digital Phenotyping for Bipolar Disorder State Prediction Using Multitask Self-Supervised Learning

Bipolar disorder is a prevalent mental illness characterized by a high relapse rate. In this study, we propose an early warning system that utilizes digital phenotyping to collect various data points from bipolar patients, including location information, self-assessment scales, daily mood reports, sleep patterns, and multimedia records, through a mobile application. These collected data are utilized to develop a predictive model for assessing the risk state of bipolar disorder. Compared to traditional recurrence prediction methods, this study incorporates medical records, medication data, and emergency records, as suggested by medical professionals, to define the five states of bipolar disorder, leading to enhanced accuracy. To account for data type correlation and temporal correlation, we employ a multitask self-supervised learning mechanism. The proposed method is trained on a Gated recurrent unit and demonstrates an improved prediction accuracy of 88.2% on the collected test data, as compared to the baseline accuracy of 85.1%. These findings highlight the significant importance of considering data type and temporal correlations in digital phenotyping for predicting the state of bipolar disorder.