A comparative study of motor imagery based BCI classifiers on EEG and iEEG data

There are many state-of-the-art Brain Computer Interface (BCI) classification algorithms designed to perform well when applied to signals acquired using electroencephalography (EEG). EEG has the advantage of being non-invasive in nature, easy to use, and effective in capturing signals in the mu (7-13 Hz) and beta (13-30 Hz) bands during motor imagery tasks. However, EEG recordings are more susceptible to movement artifacts and capture a lower frequency of neural activity when compared with invasive techniques such as electrocorticography (ECoG) or intracranial EEG (iEEG). In this paper, we analyze the performance of four different EEG motor imagery classification algorithms (both classical machine learning methods and deep learning-based methods) on a two-hand motor imagery task using both EEG and iEEG data sets. Using various feature visualization techniques, we provide insight into why deep learning-based classifiers designed to learn features end-to-end may perform better than the classical machine learning-based models. We also showed on average iEEG-based motor imagery BCIs, using our iEEG data set, do not perform as well as EEG-based BCIs. This work provides a starting point for the implementation of BCI applications using iEEG data.