Active Feedback using Riemannian Features for Motor Imagery Classification

Brain-Computer Interface (BCI) technologies employing electroencephalography (EEG) signals heavily depend on effective and accurate signal classification strategies. Researchers have extensively developed various machine learning (ML) algorithms. However, very little has been done to improve the user’s ability to elicit better brain patterns easily distinguishable across different stimuli. This paper proposes four feedback mechanisms through which the user trains himself for a motor imagery (MI) task via active feedback. The feedback strategies involve t-Distributed Stochastic Neighbor Embedding (t-SNE)-transformed Riemannian covariance matrices, mean correlations between C3 and C4 channels, tangent space transformed mean correlations between C3 and C4 channels and the power spectral density difference between C3 and C4 channels. Using a standard SVM classifier, the subjects showed significant improvement in MI accuracy post the training session. An increase in accuracy of more than 9% is achieved for two feedback mechanisms on an in-house dataset of 24 subjects indicating the effectiveness of proper feedback for eliciting better MI ability from users.

I. Introduction

Brain-computer interfaces (BCIs), also known as neural interfaces, allow for direct brain-to-external device connection [1]–[3]. Electroencephalography (EEG) based non-invasive BCIs [4], [5] have shown promise for a variety of uses [3], including rehabilitation [6]. Assistive technologies (such as communication or smart wheelchair control) [7]–[9]. Mental rotation, mental calculation [10], or motor imagery [11–13] are some of the tasks performed in order to control BCIs. We shall refer to BCIs that are motor imagery (MI) based in this article. In MI-based BCIs, users are encouraged to visualize the limb movements rather than performing them [14].