Enhancing motor imagery classification: a novel CNN with self-attention using local and global features of filtered EEG data

Abstract

Listen

Translate article

Motor imagery (MI)-based brain computer interfaces (BCIs) frequently use convolutional neural networks (CNNs) to analyse electroencephalography (EEG) signals. In this study, we proposed a novel methodology that includes an innovative preprocessing step and a new model for MI EEG classification. In the preprocessing, we use common average reference (CAR) filtering and Laplace filtering of EEG signals. The CAR filter eliminates the overall noise and Laplace filter removes the neighbouring electrode noise. Additionally, a sliding window method is used to increases the number of small-time segments which prevents overfitting. Next, the time segments are converted into spectrograms using the short-time Fourier transform (STFT). Further, the concatenated spectrogram images of mu and beta bands are processed using a CNN model with self-attention. The proposed model uses both local and global information to effectively extract features The EEG signals obtained from BCI competition IV dataset-2a are divided into 80:20 ratio for training and testing. Moreover, the ablation study highlights the importance of the combination of CAR and Laplace filters. The classification results obtained using proposed methodology shows advancement as compared to state-of-the-art methods. Finally, the proposed CNN model learning and feature distribution are visualized with the gradient weight class activation map.