SE-MDTCN: A Spatially Enhanced Temporal Convolutional Network for Imagined Speech Classification from Electroencephalography Signals

Electroencephalography (EEG)-based imagined speech decoding is a promising approach for non-invasive neural speech prostheses and related brain-computer interface (BCI) applications. However, this task remains challenging because EEG signals have a low signal-to-noise ratio (SNR), limited spatial resolution, and significant individual variability. To address these challenges, this study proposes a Spatially Enhanced Multi-branch Depthwise Temporal Convolutional Network (SE-MDTCN), an end-to-end spatiotemporal network designed to enhance spatial representation learning and temporal feature extraction for imagined speech decoding. The proposed model integrates electrode normalization, a modified squeeze-and-excitation (SE)- based spatial filter, a multi-branch depthwise temporal convolutional network (MDTCN), and a multilayer perceptron (MLP) classifier. SE-MDTCN was evaluated on the 2020 International BCI Competition Track 3 dataset under a articipant-dependent 5-class setting, using both raw and independent component analysis (ICA)-cleaned EEG. The model achieved average 5-class accuracies of 74.4% on raw EEG and 69.1% on ICA-cleaned EEG, outperforming several representative deep-learning baselines for EEG-based BCI decoding. Ablation studies and feature analyses further demonstrated the consistent contributions of electrode normalization and spatial attention, while temporal features extracted with smaller kernels were found to be more robust than those extracted with larger kernels. These findings suggest that SE-MDTCN improves imagined-speech EEG decoding by jointly modelling spatial structure and temporal dynamics, while providing interpretable spatiotemporal representations. This approach may contribute to the development of practical non-invasive speech BCIs. The code is publicly available to facilitate reproducibility.