Dual-Branch Temporal Attention Architecture for Improved EEG Motor Imagery Classification

AbstractFor high-performance brain-computer interface(BCI) systems, it is important to be able to accurately classifymotor imagery (MI) from electroencephalogram (EEG) signals.We present a Dual-Branch Temporal Attention Network (DB-TAN) that simultaneously utilizes temporal and spectral EEGinformation to improve MI decoding. The temporal branch usesconvolutional and LSTM layers with temporal attention to workwith raw EEG signals. The spectral branch uses a parallelattention-augmented pipeline to work with features producedfrom FFT. A gated fusion method combines the outputs from bothbranches to create a feature representation that may be used forclassification. The BCI Competition IV-2A dataset shows that DB-TAN has a mean accuracy of 94.29%, which is better than LDA(60.72%), CNN-only (72.50%), and LSTM-only (93.65%). Theseresults highlight the effectiveness of integrating multi-domainfeatures with temporal attention mechanisms, offering a robustframework for reliable MI classification in BCI systems.Index Terms—EEG, Motor Imagery, Brain-Computer Inter-face, Temporal Attention, Dual-Branch Temporal Attention Net-work (DBTAN)