Motor imagery classification using parallel convolution transformer based feature extraction and RSA optimized ridge ELM classifier

Being a time-series data, EEG data has strong temporal and spatial dependencies across the various time points and channels that possess substantial information. Convolutional neural networks (CNNs) are widely employed for motor imagery (MI) classification to extract these correlations and decode electroencephalography (EEG) signals. However, the inherent small perceptual field of CNNs limits their use in extracting global dependencies, leading to substantial information loss over time. Transformers can potentially mitigate this problem with their attention mechanism capable of extracting the correlation among features in the long-sequence time-series data. This study proposes a novel end-to-end framework for EEG decoding using improved feature extraction and classification methods to enhance the MI classification performance. Specifically, a parallel architecture of transformer (TF) and CNN (TF-CNN) is proposed to extract the global dependencies and local temporal features. The local temporal features from CNN and the global correlation features from the transformer are further fused together and classified using ridge regression extreme learning machine (rELM). Moreover, reptile search algorithm (RSA) is utilized to optimize the classification performance of rELM. The proposed framework is evaluated on four public datasets i.e., BCI Competition III dataset IVa, Open BMI dataset, BCI Competition IV dataset 2a and 2b, yielding average accuracies of 92.64%, 79.26 %, 84.90% and 86.77%, respectively. Such experimental results demonstrate superior performance compared to the existing methods. In addition, the proposed method, CNN-TF-rELM-RSA, hereby named as CTER also highlights the potential of meta-heuristic algorithms to improve the generalization capabilities of the ELM classifier for EEG decoding.