B1496
Title: Spectral Granger causality using neural networks for biological signals
Authors: Samuel Horvath - Muhammad Bin Zayed University of Artificial Intelligence (United Arab Emirates)
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Malik Shahid Sultan - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Abstract: Granger Causality (GC) between channels of electroencephalograms (EEG) will be investigated. Since brain signals are complex, we expect a non-linear dependence structure; therefore, vector autoregressive (VAR) models may not completely characterize GC in the brain networks. To address this limitation, we shall apply deep learning (DL) tools which can learn the non-linear dependence structure in the data. However, these models are inherently black boxes and difficult to interpret. We shall use the learned kernel vector autoregressive (LeKVAR) model, component-wise multi-layer perceptron (cMLPwF), and component-wise long short-term memory (cLSTMwF) proposed in Horvath et al. We demonstrate that these models can learn the non-linear frequency band-specific dependence structure in the time series data and give an estimate of the GC through the filter layer and decoupling lags and time series. We identify GC between and across different signals based on the decomposed spectrum, which gives an insight into the GC between oscillations. We estimate GC using LeKVAR, cLSTMwF, and cMLPwF for the spectral decomposed data of an epilepsy patient and study the evolution of the GC between EEG electrodes pre, during, and post-seizure.
