Suitibility Investigation of the Different Classifiers in fNIRS Signal Classification

Brain-computer interface (BCI) can be a hope for the people who are not capable of interacting with the external environment. Classification accuracy is a vital issue for implementing BCI system. This study aims to compare the performance of four different machine learning models for fNIRS-based BCI and find out a suitable one. We also present a method of ensemble the models for slight enhancement of accuracy from the best performing model. We use fNIRS data acquired for two different mental tasks- right-hand and lefthand motor imagery movements. The signals were taken from the prefrontal cortex of six healthy subjects using the multichannel fNIRS system. Mean, median, variance, and slope were used as features for classification. Based on the features, four different models -Quadratic discriminant analysis (QDA), Naive Bayes approach, support vector machine (SVM), and Random forest, were used to figure out the best-fit algorithm. Eventually, the average test accuracies are found as 85%, 83%, 87.5%, and 92.5%, respectively. We present a 0.4% enhancement of accuracy through ensemble the models.