A benchmark of three motor-imagery EEG classifiers, with a custom 1D CNN reaching 83.8% accuracy on held-out data.
Implemented, trained and benchmarked three binary motor-imagery classifiers.
Used BCI Competition 2005 Dataset I with 64-channel EEG sampled at 250 Hz.
Implemented a custom 1D CNN, EEGNet and CSP+SVM.
Used PyTorch and 5-fold stratified cross-validation.
Reported 1D CNN (83.8%), EEGNet (83.4%), CSP+SVM (81.2%).
The custom 1D CNN reached 83.8% accuracy on held-out data.