Can a Single Model Deep Learning Approach Enhance Classification Accuracy of an EEG-based Brain-Computer Interface?

Convolutional Neural Network (CNN) has outperformed many traditional linear and polynomial classifiers in various domains. CNN and other deep learning methods have gained attention in the Brain-Computer Interface (BCI) domain also. Here, we investigate a CNN-based model with two different optimizers for reducing error of classification of brain states using EEG motor imagery data. Two different optimizers namely stochastic gradient descent (SGD) and adaptive momentum (Adam) are investigated for increasing classification accuracy using the well-known BCI competition IV 2b dataset. The study was conducted to investigate the feasibility of a single deep learning model for all subjects without compromising on information decoding rate for any of the BCI participants. Using two different models, mean cross-validation accuracy of 80.32% (±2.2) was achieved across participants, which is significantly higher (p<0.05) compared to a state-of-the-art deep learning approach.