Statelets: A Novel Multi-Dimensional State-Shape Representation Of Brain Functional Connectivity Dynamics

Time series motifs discovery and summarization is an established powerful tool for modeling and analyzing dynamical systems. In a similar spirit, we propose a state-space data mining approach called “statelets.” It is a probabilistic pattern-summarization framework operating on time series correlations and capturing their dynamics. Statelets rely on earth mover distance (EMD): a simple yet effective similarity metric that provides a scale-independent, variable-length comparison between substructures and accounts for partial matching. The efficiency of computing EMD supports its application in a kernel density estimator to approximate local motifs’ probability density across the dataset. We validate statelets’ utility on the dynamic functional network connectivity (dFNC) of patients with schizophrenia and healthy controls. The typical approach to the dFNC analysis, sliding window plus clustering (SWC), models the system continuously through a fixed set of connectivity states. Here, rather than using all the time points and searching for the patterns that span throughout the brain, we adopt time series motifs to capture the most recurring and transient patterns of the signals. Statelets produce the synopsis of dynamic connectivity shapes from both dynamics that provide valuable insights into the human brain’s dynamic features. We observe these state-shapes are highly informative about the disorder. Also, they reveal significant group differences in connectivity strength across various regions of the brain.