Motor Imagery EEG Classification using Convolutional Neural Networks with Data Augmentation via Multiple Channel Sets

Abstract

Brain-Computer Interfaces (BCIs) provide a direct communication link between the human brain and external devices, with Electroencephalography (EEG) being a widely adopted non-invasive recording modality. Motor Imagery (MI), the mental simulation of movement, is a common BCI paradigm, though its accurate classification from EEG signals is challenging. This research proposes a Convolutional Neural Network (CNN) model for MI-EEG classification, incorporating a channel-set-based data augmentation strategy using three 6-channel EEG sets. For each original MI trial from the EEG Motor Movement/Imagery Dataset (103 subjects, 11,503 unique trials), three augmented samples were generated from three distinct 6-channel sets (fronto-central, central, and centro-parietal), each sample having dimensions of 6 channels × 640 time points.

After band-pass filtering (8–40 Hz) and subject-wise channel z-score normalization, the augmented dataset (34,509 samples) was used to train and evaluate the CNN. Under a mixed-subject/sample-wise evaluation protocol, the model achieved a validation accuracy of 91.55% for five classes. In addition, a stricter subject-wise 5-fold cross-validation protocol was added to evaluate generalization to unseen subjects; this protocol yielded a trial-level accuracy of 50.52% ± 5.82% and a trial macro-F1 of 50.65% ± 5.83%. These results clarify the difference between within-dataset mixed evaluation and cross-subject generalization, and demonstrate the usefulness and limitations of channel-set-based augmentation for MI-EEG classification.