MI-CES: An explainable weak labelling approach to example selection for Motor Imagery BCI classification

Motor Imagery (MI) Brain Computer Interfaces (BCI) can be used to control assistive devices such as wheelchairs. These systems require a training period to get both the user and the machine to learn and adapt to each other, achieving an acceptable control accuracy. Previous systems have discovered that providing a form of feedback to the user about what the system thinks the user is thinking can increase the effect of training and increase both the control accuracy of the user and the classification accuracy of the BCI system. However, if this feedback is 'incorrect' due to the classifier behind the BCI system having a poor accuracy, this may cause the user to 'incorrectly' adapt to the feedback, providing the system with further poor examples of MI. In this paper, we propose MI-CES, an explainable 'example selection' approach based on the neuro-physiological principle of MI. We found that while using 2 classification techniques, we achieved a statistically significant increase in classification accuracy across 3 datasets that were comprised of both multi-participant and multi-session recordings.