Comparison of classification performance of handpicked, handcrafted and automated-features for fNIRS-BCI system

Abstract

In this paper, we have assessed and investigated the classification accuracies of three different techniques to classify functional near-infrared spectroscopy (fNIRS) signals. Signals were extracted from the motor cortex of the brain using a continuous wave multichannel imaging system. The acquired signals were filtered for noise and converted to oxygenated- and deoxygenated- hemoglobin using modified Beer-Lambert law. From the hemodynamic responses statistical features like slope, mean, skewness, kurtosis, peak and variance were extracted, this was trained on a machine learning classifier giving a classification accuracy of 60.66% for support vector machine (SVM) and 57.22% for k nearest neighbor (KNN), likewise from the hemodynamic response we extracted principal component analysis (PCA) vectors and independent component analysis (ICA) vectors, this along with statistical features were trained on the same SVM and KNN classifier yielding a classification accuracy of 71.4% and 71.8% respectively. Instead of handpicking or handcrafting features, if we let deep learning models, in our case, convolutional neural network (CNN) and long short-term memory (LSTM), choose the features and classify them, they gave a jump of 25% accuracy to over 95% for both architectures.