Subject-Independent Multi-Class Motor Imagery EEG Classification Using Hybrid CNN–BiLSTM with Channel Attention

Abstract

The classification of MI-EEG is complex due to the noise, non-stationarity and inter-subject variation of these signals.  Some of these challenges are magnified in subject-agnostic settings, where a model must generalise to an unseen user without subject-specific tuning. This study presents a hybrid deep-learning architecture that combines convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM), and squeeze-and-excitation feature-channel attention for subject-independent four-class MI-EEG classification. The CNN extracts local spatiotemporal representations, the BiLSTM models forward and backward temporal dependencies, and the attention module recalibrates informative feature maps before classification. The framework was evaluated on BCI Competition IV Dataset 2a using leave-one-subject-out cross-validation across nine subjects. It achieved a mean accuracy of 69.75% and Cohen’s kappa of 0.5967, exceeding EEGNet by 6.94 percentage points and ShallowConvNet by 9.72 percentage points. Ablation analysis showed that removing attention reduced accuracy by 2.47 percentage points. Although the proposed model has higher computational cost and latency than EEGNet, it provides a favorable accuracy–complexity trade-off for subject-independent multi-class BCI applications in which decoding performance is prioritized over minimal latency.