Comparison of Machine Learning Algorithms for Yes/No Decoding Using Functional Near-Infrared Spectroscopy (FNIRS)

Abstract

Brain-computer interface (BCI) can be an alternative to speech production for people with disabilities. More recently, a non-invasive optical technique called functional near-infrared spectroscopy (fNIRS) has gained popularity in BCI studies due to several advantages such as high mobility, being inexpensive, and being tolerant to motion artifacts. In this study, we compared the performance of machine learning algorithms to decode fNIRS signals acquired during a binary decision paradigm for motor-independent communication. Twenty healthy participants were asked to perform mental arithmetic tasks for the "yes" decision and the rest for the "no" decision. Three trials for each decision were conducted and oxyhemoglobin concentration changes were used to classify the decision using machine learning algorithms: linear support vector machine (SVM), logistic regression (LR), naive Bayes, and k-nearest neighbors (KNN). We observed subject-wise average accuracies across twenty participants, with the logistic regression classifier achieving an average accuracy of 80.65% for training, 83.81% for validation, and 82.75% for testing. Similarly, the linear SVM classifier achieved an average accuracy of 83.89% for training, 82.34% for validation, and 81.67% for testing. Our findings suggest that both logistic regression and linear SVM classifiers, in combination with fNIRS, have the potential to be used in the binary classification with individuals having motor disabilities.