A0295
Title: Robust multivariate time series connectivity clustering through Ken-Coh
Authors: Mara Sherlin Talento - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Abstract: Brain functions are primarily organized at the regional level; hence, analyzing connectivity on a larger scale, such as regional connectivity, offers greater interpretability. A novel, interpretable approach is introduced for identifying brain states using a robust global dependence measure between brain regions (i.e., aggregating signals from multiple voxels or channels). The objective is to classify frequency-based connectivity structures at the regional level. Canonical coherence analysis is first discussed, a method that can be used for assessing dependence between two groups of signals. This framework is extended by proposing a generalized version that accounts for heavy-tailed brain signals and the nonlinear dynamics of brain functional connectivity. A robust clustering method is developed for time series, specifically designed to learn connectivity structures among neuronal signals. The approach effectively identifies signals with homogeneous connectivity patterns and uncovers meaningful brain activity under different conditions. Simulation results demonstrate that the method is highly accurate, particularly in the presence of outliers. Additionally, findings reveal that the Beta-band (12 to 30 Hz) in the frontal-parietal region exhibits strong discriminatory properties for alert brain states. Specifically, at this frequency range, alert-state signals display more homogeneous and distinct features, making them a key marker for classification.
