Dynamics of data-driven microstates in bipolar disorder.

Many of the existing models of mood in bipolar disorder can largely be divided into two camps, tracking mood as either a discrete or continuous variable. Both groups rely upon certain assumptions, with most considering only aggregate scores on clinical instruments. In this study, we propose a novel framework that combines elements from both discrete and continuous mood models, using a machine learning pipeline to detect subtle patterns across individuals. Latent factors are constructed from assessments at the item level, then clustered into groups referred to as microstates. Transitions between microstates are captured via a discrete-time Markov chain, allowing for characterization of mood's dynamic nature. Key findings include a factor mapping heavily onto irritability and aggression, as well as a hierarchical pattern of microstates within depression and mania. Validity of these results is confirmed by reproduction in an unseen data set from a separate subject cohort.