InfoFlowNet: A Multi-head Attention-based Self-supervised Learning Model with Surrogate Approach for Uncovering Brain Effective Connectivity

Abstract

Deciphering brain network topology can enhance the depth of neuroscientific knowledge and facilitate the development of neural engineering methods. Effective connectivity, a measure of brain network dynamics, is particularly useful for investigating the directional influences among different brain regions. In this study, we introduce a novel brain causal inference model named InfoFlowNet, which leverages the self-attention mechanism to capture associations among electroencephalogram (EEG) time series. The proposed method estimates the magnitude of directional information flow (dIF) among EEG processes by measuring the loss of model inference resulting from the shuffling of the time order of the original time series. To evaluate the feasibility of InfoFlowNet, we conducted experiments using a synthetic time series and two EEG datasets. The results demonstrate that InfoFlowNet can extract time-varying causal relationships among processes, reflected in the fluctuation of dIF values. Compared with the Granger causality model and temporal causal discovery framework, InfoFlowNet can identify more significant causal edges underlying EEG processes while maintaining an acceptable computation time. Our work demonstrates the potential of InfoFlowNet for analyzing effective connectivity in EEG data. The findings highlight the importance of effective connectivity in understanding the complex dynamics of the brain network.