Authors
al., & C.Y.E.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6079469Abstract
Fast and reliable electroencephalogram (EEG) directed network estimation is urgently needed to comprehend brain organization with direction and causality in real-time scenarios such as Brain computer interface, closed-loop intervention, etc. This study proposes a high-efficiency method, Deep Neural Network Directed Network Estimation (DeepNNetDNE), to estimate EEG brain networks straightforwardly. In DeepNNetDNE, an end-to-end Spatio-Temporal Neural Relation Extraction (STNRE) model constructs the direct map between the brain signals and network interactions. The trained model assimilates the inherent rules of relation expression from a large volume of data, serving as a generalized neural relation extractor that is responsible for end-to-end estimating the directed networks of new instances of brain signal across scenarios (e.g., different numbers of channels and data lengths). The proposed DeepNNetDNE exhibited superior capacity to mine network patterns from noisy EEG signals while significantly reducing computational time compared to benchmark methods. Besides, DeepNNetDNE better captures default-mode network patterns during the resting state, i.e., showing network hubs that are more consistent with empirical knowledge; and further revealed network reorganizations during motor imagery execution in post-stroke hemiplegic patients, i.e., the lateralization in the uninjured hemisphere and regions serves as compensation, which may serve as potential indicators for the assessment of patients’ motor function. DeepNNetDNE distinguishes itself by obviating the requirements for individual instance-level prior assumptions, modeling, and hyperparameter searching that are usually encountered in existing network analysis; thus, it holds strong potential in real-time directed network analysis scenarios.