Estimation of Temporally Correlated Directed Graphs to Describe Neural Connectivity and Its Dynamics

• Activity Number: 569
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Catherine Stamoulis*+ and Bernard Chang
• Companies: Harvard Medical School and Harvard Medical School
• Keywords: Directed graphs ; Time series ; High-dimensional data ; Mutual information

Abstract

Multivariate neural signals are typically measured simultaneously at multiple spatial locations in the brain and sometimes over long periods of time. The dimension of these non-stationary signals is often reduced prior to analysis, either through simple averaging or estimation of parameters of interest, e.g., pairwise correlation measures, in discrete time segments. Dimensionality-reduced signals are then used to estimate properties of the whole brain spatial network, such as connectivity and directionality, i.e., conditional dependences. Weighted directed graphs are appropriate for describing these network properties but need to be estimated dynamically, i.e., the associated covariance matrix needs to be updated as correlations between nodes vary in a time-dependent manner. This study describes a novel approach for estimating these graphs using information theoretic measures to reduce the dimensionality of continuously-recorded electroencephalograms (EEG), and estimate the directed edges of the associated graphs. This framework is used to characterize network dynamics (and their time scales) encoded in continous EEG from epilepsy patients who are monitored for diagnostic purposes.