Improved Classification Accuracy of Four Class FNIRS-BCI

Abstract

Experimentation and analysis in brain-computer interface (BCI) has increasingly been receiving quite some consideration as a substitute communication possibility for patients who are severely paralyzed in the last few years. To measure brain activities using optical signals a fairly new and non-invasive brain imaging tool can be put to test know as Functional near-infrared spectroscopy (fNIRS). Comparability low cost, safety, portability and wear ability are some of the main advantages of imaging of brain using this non-invasive modality. We have applied this relatively new non-invasive fNIRS technique to make an image of brain activities during four different mental tasks. These tasks include Mental Arithmetic (MA), Motor Imagery (i.e. Left-Hand and Right-Hand Motor Imagery) and Rest. fNIRS data used is an open access dataset of 29 individuals which was collected by Continuous-wave imaging system (NIR Scout NIRx GmbH, Berlin, Germany) with the sampling frequency of 10 Hz. In this research Data synchronization is performed before the data is preprocessed. Usual preprocessing is done using Butterworth filter of 4th order to minimize or eliminate any unwanted signal distortion. After that an extensive signal analysis is done in which Six different statistical features (Signal Mean (SM), Skewness (SK), Kurtosis (KR), Standard Deviation (SD), Signal Peak (SP), and Signal Variance (SV)) are obtained in the time domain and 13 MFCC (Mel Frequency Cepstral Coefficients) features are obtained from the frequency domain. After the preprocessing and signal analysis of data our results shows hemodynamic behavior of multiple patterns during the tasks performed. These unique patterns of hemodynamic behavior can be used to differentiate and distinguish different task. Separate Classification analysis is performed on each domain features. We were able to compare, differentiate and distinguish the brain signal activities captured while performing 4 different tasks using 3 different classifiers i.e. Linear Discriminant Analysis (LDA), Support Vector Machine (SVM) and K Nearest Neighbor (KNN). The average classification accuracy of above 90% is achieved from K Nearest Neighbors (KNN) using the time domain features and same accuracy is achieved from Support Vector Machine (SVM) using the frequency domain features.