Estimating brain effective connectivity from time series using recurrent neural networks

Effective Connectivity (EC) reflects the causal influence between brain regions. Identifying Effective Connectivity Networks (ECN) in the brain can enhance our understanding of brain functions and reveal the impact of mental illnesses on these functions. However, existing EC estimation methods face challenges in extracting deep features from functional magnetic resonance imaging (fMRI) data. In this study, we propose a novel Time Series to Effective Connectivity (TS2EC) prediction model based on recurrent neural networks, which directly extracts deep features from fMRI time series without relying on a fixed model order. Specifically, we introduce a method for generating EC labels from electrocortical stimulation fMRI (es-fMRI) data, representing the first attempt to use es-fMRI for EC estimation. We evaluated TS2EC on three datasets: an es-fMRI dataset with 23 subjects (augmented to 7,082 samples), a multivariate autoregressive simulated dataset, and a Smith simulated dataset. On the es-fMRI dataset, TS2EC achieved a mean squared error of 0.0057, significantly outperforming existing methods. Experiments on the simulated datasets demonstrated that TS2EC attained superior performance in accuracy, recall, structural Hamming distance, and F1-score. Experimental results demonstrate that the EC prediction performance of TS2EC is significantly higher than current EC analysis methods. TS2EC holds promise as a novel tool for the analysis of ECN in the brain.