Comparison of machine learning approaches for motor imagery based optical brain computer interface

Abstract

A Brain-computer Interface (BCI) is a system that interprets specific patterns in human brain activity, such as the intention to perform motor functions, in order to generate a signal which can be used for communication or control. Functional near infrared spectroscopy (fNIRS) is an emerging optical neuroimaging technique which is a relatively new modality for BCI systems. As such, the optimal paradigms and classification techniques for the interpretation of fNIRS-BCI systems is an area of active investigation. Presently, most fNIRS BCIs have adopted Linear Discriminant Analysis (LDA) algorithm as the primary classification approach, however other alternative methods may offer increased performance. In order to compare different algorithms, a dataset from a four-class motor imagery-based fNIRS-BCI study was re-analyzed, and we systematically compared the performance of different machine learning algorithms: Naïve Bayes (NB), LDA, Logistic Regression (LR), Support Vector Machines (SVM) and Multi-layer Perception (MLP). Our findings suggest that the LR classifier slightly outperformed other classifiers, unlike most fNIRS-BCI studies which reported LDA or SVM as the best classifier. The results presented here suggest that an LR classifier could be a potential replacement for LDA classifiers in motor imagery tasks.