Investigation of different approaches for noise reduction in functional near-infrared spectroscopy signals for brain–computer interface applications

• Published in: Neural computing & applications 2017-10, Vol.28 (10), p.2889-2903
• Main Authors: Janani, A., Sasikala, M.
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Brain; Cerebral cortex; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Datasets; Filtration; Image Processing and Computer Vision; Independent component analysis; Infrared spectra; Infrared spectroscopy; Motion perception; Near infrared radiation; New Trends in data pre-processing methods for signal and image classification; Noise reduction; Oscillations; Physiology; Preprocessing; Probability and Statistics in Computer Science; Signal processing; Spectrum analysis; Wave attenuation; Wavelet analysis
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-017-2961-4

Functional near-infrared spectroscopy (fNIRS) is a noninvasive technique to measure the hemodynamic response from the cerebral cortex. The acquired fNIRS signal usually contains influences generated from physiological processes, also called “global” oscillations, in addition to motion artifacts that impede detection of the localized hemodynamic response due to cortical activation. Preprocessing is the fundamental step to enhance the quality of fNIRS signals corresponding to movement tasks for efficient classification of brain–computer interface (BCI) application. Various signal preprocessing approaches such as band-pass filtering, correlation-based signal improvement, median filtering, Savitzky–Golay filtering, wavelet denoising and independent component analysis (ICA) have been investigated on experimental datasets acquired during hand movement tasks and are compared to one another using artifact power attenuation and contrast-to-noise ratio (CNR) metrics. The results showed that wavelet denoising method attenuated the artifact energy of the datasets belonging to Subjects 1 and 2 as well as enhanced the CNR. In the case of Subject 1, before denoising the values of ΔHbR and ΔHbO were 0.6392 and 0.8710, respectively. Wavelet method improved these values to 0.8085 and 0.9790. In the case of Subject 2, the CNR values of ΔHbR and ΔHbO signals were improved from 0.0221 and 0.0638 to 1.1242 and 0.3460, respectively. In this study, ICA was also demonstrated to suppress noises related to physiological oscillations including Mayer wave influence and other unknown artifacts. It greatly reduced the sharp spikes present in the Subject 2 dataset. On the basis of the results obtained, it can be shown that application of such filtering algorithms for fNIRS signal could effectively classify motor tasks to develop BCI applications.