CBAM-DeepConvNet: Convolutional Block Attention Module-Deep Convolutional Neural Network for asymmetric visual evoked potentials recognition

Abstract

Purpose: This study aimed to improve the accuracy and the ITR in the stimulative paradigm of character spelling systems based on asymmetric Visual Evoked Potentials (aVEPs) by utilizing EEG signal and an improved Convolutional Block Attention Module-Deep Convolutional Neural Network. Methods: This study proposed a deep-learning analysis framework called Convolutional Block Attention Module-Deep Convolutional Neural Network (CBAM-DeepConvNet) to decode aVEPs-based characters. First, the spatial and temporal attention modules were utilized to acquire refined EEG signals, effectively capturing event-related information. Second, the EEG signals were fed into a DeepConvNet to extract discriminative features, enabling the recognition of VEPs-based characters. Third, the performance of CBAM-DeepConvNet was evaluated using two different datasets (i.e., Dataset A and Dataset B), which are available from the BCI Robot Contest at the World Robot Conference Contest 2020. Character Recognition Accuracy (CRA) and Information Transfer Rate (ITR) were utilized as metrics for the performance evaluation. Results: The experimental results show that CRAs are 65.57% ± 5.09% and 77.63% ± 3.41% for Dataset A and Dataset B, respectively. The highest ITR is 38.23 ± 6.00 bits/min for Dataset A and 66.74 ± 7.98 bits/min for Dataset B. The EEG topography learned from CBAM-DeepConvNet demonstrates higher energy in the right hemisphere for left-side stimulus and higher energy in the left hemisphere for right-side stimulus, matching the spatial asymmetry of aVEPs paradigm. Conclusion: The proposed CBAM-DeepConvNet model in this work is a general framework and is particularly suitable multi-channel EEG decoding, this general framework could have the potential application for other active/passive brain computer interface paradigm. This research establishes a robust foundation for future explorations into deep learning-based identification in aVEPs paradigms.