Classification Accuracy Enhancement of fNIR based Imagery Movement by Modified Common Spatial Pattern

Abstract

Motor imagery event classification from functional near-infrared spectroscopy (fNIRS) is one of the most interesting problems of current brain-computer interfaces (BCIs) challenges because it needs no additional data guiding visual or listening protocol. A vital step of the fNIRS signal classification by machine learning approach is feature extraction. The feature extraction from multiple channel fNIRS signal is always challenging due to its high dimensionality. There exist several conventional feature extraction procedures like principal component analysis, nonlinear principal component analysis, independent component analysis, norm analysis, spectral norm analysis, etc. This research work studies such existing feature extraction method to classify the hand movement events of fNIRS signals. The accuracies of these methods have been found less than the expectation. Therefore, some more accurate method is needed. In this regard, usually common spatial pattern (CSP) is used to reduce the dimensional reduction and improving the classification accuracy. The conventional CSP method can be proven also ineffective for the motor imagery fNIRS signal due to its high level of trial to trial variations. The present research work proposes an algorithm named by standardized common spatial pattern (SCSP) based feature extraction method for fNIRS based motor imagery classification which can perform well in the context of the trial to trial significant variation. The classification results corresponding to the proposed feature extraction method reveal that the proposed SCSP algorithm outperformed the conventional CSP method and channel-wise method for classifying the two motor imagery event classifications. For classification accuracy measurement, four well-known classifiers: artificial neural network (ANN), k-nearest neighbor (kNN), support vector machine (SVM), and linear discriminant analysis (LDA) have been used. We have utilized both the fNIRS data of oxidized hemoglobin (HbO) and deoxidized hemoglobin (HbR) for classifying the motor imagery fNIRS data with the conventional and proposed methods. From the comparisons, we have found that in both cases, the proposed method outperforms the conventional methods in the context of the classification accuracies. To validate the classification accuracies, the sensitivities and specificities of the classifier are also calculated in this work. We believe that the proposed SCSP method will contribute in the field of feature extraction method for other types of fNIRS based BCI system, effectively. In addition, this method may be applied for the feature extraction to the other multidimensional signals so far.