Abstract

Abstract In brain-computer interface (BCI) systems, the non-stationarity of brain signals is known to be a challenge for training robust classifiers as other brain processes produce signals that coincide with those resulting from the desired brain activity. One source of interference is the user's cognitive response to the provided BCI feedback. In the case of motor imagery paradigms, this feedback can for instance be a cursor moving on the screen. The response to such feedback has been shown in general to be a source of noise that can add to the non-stationarity of the brain signal; however, in this work, we show that the user’s brain response to this feedback can be used to improve the BCI performance. We first show in a motor imagery task that the user’s brain responds to the direction of cursor movement, which is different for the cursor moving towards or away from the target (i.e. BCI feedback), and this feedback-related information is present in frequency bands similar to those used in motor imagery. Next, we propose a classifier that combines the user response to feedback together with the motor imagery signal itself, and show that this combined classifier can significantly outperform a conventional motor imagery classifier. Our results show an average of 11% and up to 22% improvement in classification accuracy across 10 participants.