Improved Decoding of EEG-Based Motor Imagery Using Convolutional Neural Network and Data Space Adaptation

One challenge in decoding EEG-based motor imagery for brain computer interface (BCI) is the inter-session non-stationarities due to factors like subject fatigue or electrode placements. This non stationarity leads to lower decoding accuracy, especially if data from the training and testing sessions are collected on separate days. An existing technique termed data space adaptation (DSA) transforms EEG data such that the distribution of test and training data are aligned. Decoding accuracy has been shown to improve when transformed data was used for classification instead of the original data. In this study, space transformed data from six subjects were fed into a Convolutional Neural Network (CNN) and a Filter Bank Common Spatial Pattern (FBCSP) model. The results show that when no adaptation was applied, there was on average a 2.8% improvement in decoding accuracy when CNN was used compared to FBCSP. When DSA was used in conjunction with FBCSP, the majority of the subjects had improved decoding accuracy. However, DSA together with CNN on average yielded a lower decoding accuracy compared to CNN alone. Hence the results suggest that when decoding non-stationary EEG data for motor imagery, DSA should be used only in conjunction with FBCSP, but not when CNN is used as a decoding model.