A Modified Motor Imagery Classification Method Based on EEGNet

Brain-computer interfaces (BCIs) are a growing field of scientific research aiming to study how computers interpret human neural signals and take corresponding actions for communication purposes. Motor imagery-related EEG (MI-EEG) signals, generated by humans mentally simulating a given action, can be analyzed and learned by machine learning algorithms and deep learning models. EEGNet is one of the convolutional neural networks (CNNs) robust for different BCI paradigms because of its two-dimensional filters: temporal and spatial filters. In this study, I modified the EEGNet configurations into three new models with double convolutional layers for temporal, spatial, and both filters, respectively. I compared four models' binary classification performances on downsampling and normalized MI-EEG datasets without prior feature extraction methods. The results show that: 1. All four EEGNet models with max-pooling layers perform better than using their original average-pooling layers. 2. The model with double two-dimensional filters' layers received the highest mean F1 score (71.71%) and the lowest sample standard deviation (0.0168) among all four models after repeated five runs. Compared to the original EEGNet model, only the best-performed model has a statistically significant difference (p < 0.05) in their mean F1 scores. The results suggest doubling layers for both temporal and spatial filters can increase EEGNet performance on MI-EEG classification. Therefore, it can be considered an approach to improve EEGNet in future studies.