A P300-Based Speller Design Using a MINMAX Riemannian Geometry Scheme and Convolutional Neural Network

A P300-based speller BCI with large target symbols was designed to improve the detection accuracy and information transfer rate (ITR) as well as to overcome adjacency, crowding, and fatigue problems. The speller paradigm consists of a
$8\times 7$
speller matrix which uses only 8 flashing characters for selection. Another critical issue in the design of the BCI speller is the feature extraction and classification algorithms. The symmetric positive definite (SPD) matrix of data has emerged as a compact and informative representation of brain signals conveying spatial information. A geometric way to analyze SPD matrices is through the Riemannian geometry (RG) which has been shown to overcome the limitations of Euclidean geometry and provides robustness with respect to outlier and possess good generalization abilities. However, for large data sets, when the number of channels is large, we encounter the curse of dimensionality problem in computing the covariance matrix and constructing the Riemannian graph. To overcome this problem and be able to construct a low-dimensional SPD, we propose a hybrid method based on a convolutional neural network and RG. A mapping that relies on the RG, is proposed to extract the maximum discriminative features. The experimental results from 6 subjects showed that the proposed BCI speller achieved promising accuracy (ITR) of 95±3.6% (26.73±1.8 bits/min) in offline analyses and 94.3±3.8% (54.6±3.8 bits/min) in online experiments. Additionally, the proposed feature extraction method was applied to the BCI competition datasets and compared with most of the state-of-the-art algorithms. The results demonstrated that the accuracy (ITR) of the proposed method was 100% (8.9 bits/min) and 97% (8.5 bits/min) on dataset II from BCI-competition II and dataset II from BCI-competition III, respectively.