A1306
Title: Adaptive frequency band learning of nonstationary functional time series: An application to high-dimensional EEG signals
Authors: Pramita Bagchi - George Washington University (United States) [presenting]
Abstract: Frequency-domain properties of high-dimensional EEG (HD-EEG) signals encode rich information about brain function, varying over time through their time-varying spectral density. Instead of analyzing this density directly, researchers often summarize power within pre-defined frequency bands. However, how frequencies are partitioned crucially influences how well these summaries capture underlying dynamics, an aspect often overlooked. Treating HD-EEGs as nonstationary functional time series, a method is proposed to learn frequency-band structures that optimally summarize time-varying spectral behavior. Scan statistics and a scalable search algorithm are introduced to detect frequency-domain changes for adaptive band estimation. Theoretical results establish key properties, and simulations confirm the accuracy of band recovery. Applied to HD-EEG data from alternating eyes-open and eyes-closed conditions, the method uncovers spatially localized nonstationary patterns and cross-frequency synchronization consistent with known physiological phenomena.
