Title: Decomposition of multivariate brain signals in multi-subject replicated setting
Authors: Guillermo Cuauhtemoctzin Granados Garcia - King Abdullah Universe of Science and Technology (Saudi Arabia) [presenting]
Raquel Prado - UCSC (United States)
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Abstract: A Bayesian nonparametric model is proposed to understand (change to investigate) the dynamic dependence between multivariate brain signals in experiments involving several trials and subjects. The proposed method models a multiple-channel signal as a mixture of latent second-order autoregressive processes (AR(2)). Each latent AR(2) represents a unique quasi-periodic wave shared across channels. The channel dependence structure is inferred via a Hierarchical Dirichlet process model allowing us to model the strength of each latent wave by borrowing information across trials and subjects. A Metropolis within Gibbs algorithm is implemented for posterior distribution computation, mixed with optimization strategies to improve computational performance. The model's effectiveness is demonstrated in a simulation study of two groups with smooth and abrupt changes in the channel dependency structure and signals oscillatory behavior. Lastly, the novel method is used to compare the EEG recordings of a group of alcoholics and a control group during a visual recognition experiment.