A Spatio-Temporal Interactive Attention Network for Motor Imagery EEG Decoding

Abstract

Brain-computer interface (BCI) technology can link direct communication paths between human brain and external devices, where tasks of motor imagery (MI) electroencephalogram (EEG) decoding play important roles. Multi-channel electrode montage achieves EEG measurements with high spatial resolution. In previous studies of MI-EEG decoding, the extracted temporal features of multi-channel EEG measurement data were harnessed to recognize different MI-EEG patterns, while spatial features, especially those manifesting the intrinsic connectivity of EEG channels during different MI tasks, has often been overlooked. In this paper, we propose a spatio-temporal interactive attention network (STIA-Net), which exploits spatial features, temporal features, as well as their interaction, for MI-EEG decoding. Graph convolution is employed for spatial feature manipulation, where functional connectivity with phase locking value (PLV) is involved to establish a graph and hence exhibiting topological structural properties. The temporal features are extracted by dilated temporal convolutions, and spatio-temporal interaction is accomplished via attention mechanism. The STIA-Net utilizes the spatio-temporal feature fusion for ultimate MI-EEG classification. The experimental results demonstrate that the proposed STIA-Net performs well on the PhysioNet MI-EEG dataset, with a subject-independent classification accuracy of 83.9%, higher than state-of-the-art methods.