Leveraging Temporal Causal Discovery Framework to Explore Event-Related EEG Connectivity

Abstract

Deciphering the cause-and-effect relationships between brain regions not only can provide insights into the mechanism of brain networking but also facilitate the development of the brain-computer interface. Numerous studies have adopted effective connectivity measurements such as transfer entropy or multivariate autoregressive model to investigate the synchronous neuronal coupling across brain regions. Recent successes with representation learning in deep neural networks deserve further investigation on their capability for causal discovery. To this end, this study modified the Temporal Causal Discovery Framework (TCDF) with a sliding-window approach to explore the event-related intra-brain electroencephalogram (EEG) dynamics. The TCDF used a convolution neural network with an attention mechanism to predict causal direction among multi-channel time series data. The resultant array of attention scores was used to gauge the causality magnitude. The TCDF was first validated through a simulation study, followed by real EEG data collected in a virtual reality-based multitasking experiment. Results showed that the time-varying causality magnitudes were positively correlated with the predefined regression coefficients of simulated data, and the TCDF could successfully capture the coupling between brain regions, particularly for adjacent channels. However, some issues were raised in modeling as the number of channels increased or the causal order became larger. The current findings might shed some light on the development of deep learning-based causality analysis.