Principal Components-Artificial Neural Network in Functional Near-Infrared Spectroscopy (fNIRS) for Brain Control Interface

Abstract

Functional near-infrared spectroscopy (fNIRS) is a non-invasive brain imaging technology that is widely utilized in Brain Control Interface (BCI) applications. Feature extraction is crucial to remove unwanted signals and improve the accuracy of a machine learning algorithm in BCI. Despite principal component analysis (PCA) is a popular feature extraction method in near-infrared spectroscopy, PCA is rarely studied in fNIRS. Thus, this study compared fNIRS-based BCI models that used PCA and that used statistical features in BCI for four mental activities classification. First, PCA was applied to transform pre-processed fNIRS signals into few principal components that were the inputs of artificial neural network (ANN) to form PCs-ANN. Three different combinations of fNIRS signals were used to study the performance of PCs-ANN using 10-fold cross-validation. The best PCs-ANN was compared with ANN that used statistical-based features. The finding shows that PCs-ANN outperformed ANN that used statistical-based features in the BCI classification application.