DSTA-Net: dynamic spatio-temporal feature augmentation network for motor imagery classification

Accurate decoding and strong feature interpretability of Motor Imagery (MI) are expected to drive MI applications in stroke rehabilitation. However, the inherent nonstationarity and high intra-class variability of MI-EEG pose significant challenges in extracting reliable spatio-temporal features. We proposed the Dynamic Spatio-Temporal Feature Augmentation Network (DSTA-Net), which combines DSTA and the Spatio-Temporal Convolution (STC) modules. In DSTA module, multi-scale temporal convolutional kernels tailored to the α and β frequency bands of MI neurophysiological characteristics, while raw EEG serve as a baseline feature layer to retain original information. Next, Grouped Spatial Convolutions extract multi-level spatial features, combined with weight constraints to prevent overfitting. Spatial convolution kernels map EEG channel information into a new spatial domain, enabling further feature extraction through dimensional transformation. And STC module further extracts features and conducts classification. We evaluated DSTA-Net on three public datasets and applied it to a self-collected stroke dataset. In tenfold cross-validation, DSTA-Net achieved average accuracy improvements of 6.29% (p < 0.01), 3.05% (p < 0.01), 5.26% (p < 0.01), and 2.25% over the ShallowConvNet on the BCI-IV-2a, OpenBMI, CASIA, and stroke dataset, respectively. In hold-out validation, DSTA-Net achieved average accuracy improvements of 3.99% (p < 0.01) and 4.2% (p < 0.01) over the ShallowConvNet on the OpenBMI and CASIA datasets, respectively. Finally, we applied DeepLIFT, Common Spatial Pattern, and t-SNE to analyze the contributions of individual EEG channels, extract spatial patterns, and visualize features. The superiority of DSTA-Net offers new insights for further research and application in MI. The code is available in https://github.com/CL-Cloud-BCI/DSTANet-code.