Brain-computer interface (BCI) is one technology that allows a user to communicate with external devices through detecting brain activity. As a promising noninvasive technique, functional near-infrared spectroscopy (fNIRS) has recently earned increasing attention in BCI studies. However, in practice fNIRS measurements can suffer from significant physiological interference, for example, arising from cardiac contraction, breathing, and blood pressure fluctuations, thereby severely limiting the utility of the method. Here, we apply the multidistance fNIRS method, with short-distance and long-distance optode pairs, and we propose the combination of independent component analysis (ICA) and least squares (LS) with the fNIRS recordings to reduce the interference. The short-distance fNIRS measurement is treated as the virtual channel and the long-distance fNIRS measurement is treated as the measurement channel. Least squares is used to optimize the reconstruction value for brain activity signal. Monte Carlo simulations of photon propagation through a five-layered slab model of a human adult head were implemented to evaluate our methodology. The results demonstrate that the ICA method can separate the brain signal and interference; the further application of least squares can significantly recover haemodynamic signals contaminated by physiological interference from the fNIRS-evoked brain activity data.
Near-infrared spectroscopy (NIRS), employing continuous wave (CW) instrumentation with several wavelengths and a variety of source-detector configurations, has attracted growing interest recently [
Near-infrared, spectroscopy uses stimuli to evoke physiological responses, and it has developed into a commonly used method called functional near-infrared spectroscopy (fNIRS). The concentration of HbO2 and HHb changes in brain activities can be measured simultaneously and this allows a user to interact with the outside world through the measurement of correlates of neural activity associated with mental processes. Brain-computer interfaces (BCIs) can be characterized in a number of ways, such as functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG), positron emission tomography (PET), and electroencephalography (EEG) [
In an attempt to decorrelate the physiological fluctuations from the evoked haemodynamic response, various groups have made extensive research on fNIRS data. Bandpass filtering, low-pass filtering, moving averaging, and Wiener filtering [
Recently, independent component analysis (ICA) has been shown to be able to separate and identify interference arising from some cardiac events and breathing and has been widely used in the EEG and fMRI research community [
In our study presented here, we propose a new method of isolating useful information about hemodynamics in the cerebral layer. We used a multidistance measurement method and a theoretical analysis of global interference reduction based on ICA and the least squares (LS) criterion. The short-distance fNIRS measurement is treated as the reference channel comprising superficial haemodynamic changes induced by physiological fluctuations and the long-distance fNIRS measurement is treated as the measurement channel containing both the functional haemodynamic response and global interference. We aim to remove global interference that is correlated with superficial haemodynamic fluctuations, evoked by cardiac contractions, breathing, blood pressure, and so forth. ICA is a powerful method to separate the mixed multichannel recordings and recover them into their constituent latent sources. By decomposing the long-distance measurement with the ICA algorithm, we separated the signal into different components based on the virtual channel. The least squares criterion was then used to adjust the corresponding weighting coefficients and the direction to estimate the real signal with the obtained component. Monte Carlo simulations of a five-layer human model were used to investigate the performance of the ICA-LS for removing global interference in brain activity measurement.
The use of near-infrared (NIR) light (700~1000 nm) for functional measurement in biological tissues depends, firstly, on the fact that biological tissues are relatively transparent at these wavelengths and, secondly, on the oxygen dependence of haemoglobin and cytochrome aa3 optical absorption properties [
Schematic illustration of five-layered slab human head model and multidistance optode configuration. S is the photon source, and D1 and D2 are detectors.
When we obtained the change of optical density with the source and the detectors, the concentration changes of HbO2 and HHb at a given time could be estimated using the modified Lambert-Beer law (MLBL) as follows:
ICA is a multivariate statistical technique which is used to estimate a set of signals of which only a mixture is available. The sources and the mixing process are both unknown, and the sources are estimated on the assumption that they are statistically independent from one another [
For a one-dimensional signal with additive noise, the mathematical model can be represented by the following equation:
A layered model of the adult human head was used to investigate the light propagation and attenuation. The model was composed of five layers including scalp, skull, cerebrospinal fluid (CSF), gray matter, and white matter [
The Monte Carlo code used here is an extension of the general multilayer, three-dimensional, weighted photon Monte Carlo codes developed by Wang et al. [
Haemodynamic parameters, thickness, and optical properties for each layer of an adult head model.
Tissue type | Baseline concentration | Absorption coefficient |
Transport scattering coefficient |
Thickness | |
---|---|---|---|---|---|
HbO2 ( |
HHb ( |
750/830 nm (mm−1) | 750/830 nm (mm−1) | (mm) | |
Scalp | 64 | 27 | 0.0177/0.0203 | 2.11/1.84 | 3 |
Skull | 57 | 24 | 0.0158/0.0181 | 1.79/1.47 | 7 |
CSF | 14 | 6 | 0.0039/0.0045 | 0.27/0.22 | 2 |
Gray matter | 128 | 55 | 0.0358/0.0409 | 2.39/2.10 | 4 |
White matter | 50 | 21 | 0.0138/0.0158 | 9.62/8.82 | 30 |
To further quantify the utility of ICA-LS in the removal of physiological noise, and therefore the improved recovery of a functional response, we introduced a series of simulated haemodynamic functional responses function. The functional haemodynamic responses in gray matter were defined as the convolution of the stimulation
In order to make the simulation as realistic as possible, the haemodynamic changes were simulated as a combination of the functional haemodynamic responses and the physiological interference. The physiological interference is generated by a combination of cardiac fluctuation
Simulation parameters for amplitude and frequency of interference oscillations and haemodynamic changes.
Head layers | Blood contents |
|
|
|
|
|
---|---|---|---|---|---|---|
Scalp | HbO2 | 0.2 | 0.6 | 0.9 | 1.0 | 0 |
HHb | 0.013 | 0.04 | 0.058 | 0.06 | 0 | |
Skull | HbO2 | 0.2 | 0.63 | 0.96 | 1.1 | 0 |
HHb | 0.012 | 0.045 | 0.06 | 0.07 | 0 | |
CSF | HbO2 | 0.02 | 0.06 | 0.08 | 0.11 | 0 |
HHb | 0.001 | 0.004 | 0.005 | 0.007 | 0 | |
Gray matter | HbO2 | 0.2 | 0.65 | 0.92 | 1.1 | 15 |
HHb | 0.014 | 0.043 | 0.07 | 0.072 | −4 | |
White matter | HbO2 | 0.2 | 0.6 | 0.9 | 0.9 | 0 |
HHb | 0.012 | 0.04 | 0.06 | 0.06 | 0 |
The simulated haemodynamic changes were used to calculate the optical measurement by Monte Carlo method. We derived the simulated optical measurements by launching 108 photon packets and the running time of Monte Carlo simulation is about 10 hours on the desktop (Intel Core i5-2320 CPU). The sampling rate was set to 10 Hz and the whole time series for the changes of optical density were acquired under the assumption that the scattering properties of the head do not vary with time. The experiment is designed as a 5-epoch block and each individual epoch consisted of a series of 400 points, 200 points of rest, and 200 points of stimulation. The short source-detector spacing was set to 5 mm, which only considers penetration of light into the extra cerebral tissue. The long source-detector spacing was set to 45 mm, which is long enough for penetration into the cerebral cortex. Therefore, the short-distance optode pair was used as the virtual channel and the long-distance optode pair was used as the measurement channel in the ICA model. Using (
The simulated optical measurements here were obtained by employing S-D1 with a separation of 5 mm and S-D2 with a separation of 45 mm. The concentration changes of oxyhaemoglobin,
ICA method to separate physiological interference. (a) The time series of
ICA method to separate physiological interference. (a) The time series of
ICA is a signal discrimination method that extracts independent components from multiple signals without knowledge of the obtained signal by utilizing the statistical independence of the source components [
In Figure
In Figure
The results of hemodynamic changes and the associated processed results are shown in Figures
However, the magnitudes of
Concentration changes of HbO2 during the simulated measurement are shown in Figure
The haemodynamic changes calculated with ICA-LS method. (a) The time series of
Several methods have been explored in the literature to attempt to remove interference from fNIRS recordings. If a reference signal for physiological interference is available, such as the pulse oximeter, electrocardiogram (ECG), it can be subtracted from the fNIRS data after scaling it by an appropriate factor determined by regression either in the time domain or frequency domain. However, the recorded reference signal tends to be contaminated as well; regressing it out would thus attenuate the fNIRS value, which is undesirable. Moreover, reference signals are not available for other interferences such as the low-frequency oscillation. Another approach for interference elimination is to use digital filters; however, the frequency spectrum of most interferences overlaps with that of the brain activity signal; hence, the interferences can only be partially removed by this method.
More recently, principal component analysis (PCA) [
We have used a Monte Carlo method of a five-layered model of the human adult head to simulate the brain activity experiments. Monte Carlo modeling has been extensively utilized in the biomedical fields. Different approaches have been used to develop the Monte Carlo methods such as white Monte Carlo [
The present paper introduces a new approach to physiological interference cancellation from fNIRS recordings. The inclusion of the reference channel has succeeded in ensuring that one component is primarily determined by physiological interference. We have used Monte Carlo simulations to assess the use of independent component analysis in global interference removal from fNIRS brain activity data. Least squares algorithm was further employed to determine the amplitude of the independent components. Our results have shown that the combination of independent component analysis and least squares can reduce global interference within the measurements of
Studies of independent component analysis of fNIRS data with a virtual channel have demonstrated that the physiological interference in the fNIRS signals can be substantially suppressed. These results were evaluated with multidistance fNIRS measurement using Monte Carlo simulation. Least squares has been applied to deal with the separating components calculated with ICA and the results show that brain activity response can be separated in fNIRS signals since the method is effective to remove global interference induced not only by heartbeat and respiration but also by low-frequency oscillation and very-low-frequency oscillation, and other correlated interferences between superficial and deep layer. The advantage of ICA-LS in multidistance fNIRS measurement compared with other possible methods also arises from its convenient implementation; it neither requires an auxiliary measurement instrument nor the dependence on a priori knowledge of the global interference frequency. The experiment also indicated that, by introducing the virtual noise channel properly, the brain activity response can be extracted even in the strong noise background. Thus, the methodology has clear potential for use in fNIRS brain-computer interface measurement.
The authors are grateful for the support from the National Science Foundation of China (Grant no. 61201017), the Natural Science Foundation of Heilongjiang Province (Grant no. QC2011C097), China Postdoctoral Science Foundation (Grant no. 2013M531027), Heilongjiang Postdoctoral Fund (Grant no. LBH-Z12093), the Fundamental Research Funds for the Central Universities (Grant no. HIT.NSRIF.2013010), and OBHL, UK. They also thank colleagues in China, Italy, UK, and Japan for their helpful comments on this work.