2.1.

Subjects

Seventeen healthy adults (6 females, 11 males, $\mathrm{age}=40.1\pm 11.1$ y.o.; $\text{mean}\pm \mathrm{SD}$, $\text{range}=24$ to 57 y.o.) participated in this study. Their handedness was assessed with the Edinburgh Handedness Inventory:⁵⁰ 16 subjects were right-handed. Data were obtained according to the standards of the internal review board of the Research and Development group, Hitachi, Ltd. All subjects received a detailed explanation of the measurements to be performed and provided the informed consent before participating in this experiment. Unexpected technical problems occurred, and the affected data (one and two samples for spatial WM and motor control tasks, respectively) were excluded in data analysis.

2.2.

Task Paradigms

Each subject was measured in two sessions as shown in Fig. 1. Each session lasted for 15 min. All tasks were designed using the Platform of Stimuli and Tasks software (Hitachi Ltd., Central Research Lab.) and presented on a monitor put in front of subjects. The first session involved 15 trials of spatial WM task and 10 trials of motor control task [Figs. 1(a) and 1(b)]. As in previous spatial WM studies,⁵¹ subjects were asked to encode the position of four red squares among displayed eight squares within 1.5 s (target stimulus) and to maintain the spatial information for 7 s. During the memory maintenance period, subjects kept their sight on a fixation cross appearing on the black background screen. The monitor displayed a red square among eight squares (the probe stimulus) after the maintenance period. In addition to evoking WM, subjects were required to switch into decision-making mode to retrieve information and respond.^52,53 Subjects needed to indicate whether the position of the red square in the probe stimulus was identical with the target stimulus [Fig. 1(a)] within 2 s. Since subjects gave responses by pressing buttons, we were aware of motor-evoked activation in the motor cortex.^2,54 In order to estimate the impact of button pressing (motor) alone, a motor control task (with no information encoding) was given following the spatial WM task (still in session 1). The motor control was designed to always show all the red squares in the target stimulus and a fixation cross during the maintenance period. In the probe stimulus, we showed white squares on the left and red on the right (or vice versa) and asked the subjects to press button according to the location of the white squares [e.g., press left button if the white squares were on the left, see Fig. 1(b)]. The intertrial interval was randomized from 16 to 24 s. In the second session, the resting state without any target and probe stimulus was used as another control task [Figs. 1(c) and 1(d)]. While resting, subjects fixed their sight on a fixation cross for the first 8 to 9 min [similar to the period for 15 trials; Fig. 1(c)] and closed their eyes for the remaining time [Fig. 1(d)]. The end of the second session was marked by soft beeping.

Fig. 1

Two measurement sessions (spatial WM, motor control; eyes-open, eyes-closed RS). Each session consisted of 25 trials; 15 trials for spatial WM/eyes-open RS and 10 trials for motor control/eyes-closed RS. (a) In the spatial WM task, subjects responded the probe stimulus by matching the position of red square in the target stimulus. (b) Meanwhile, the target stimulus was not required to be encoded and retrieved because subjects were merely required to answer which side majorly showed white squares during motor control. In the RS, subjects rested while (c) kept their eyes opened and (d) then closed their eyes.

2.3.

Measurements

Prefrontal measurements were acquired using an ETG-4000, a dual wavelength (695 and 830 nm) fNIRS system (Hitachi Medical Corporation, Tokyo, Japan), with a sampling rate of 10 Hz. One plane probe [Fig. 2(a), $3\times 11$ S-D arrangement; 17 emitters and 16 detectors] with a 52-channel system was put on the subject’s frontal lobe. The coordinates of the measured regions (i.e., channels) were estimated in the middle of each S-D pair and illustrated by the template of spatial registration [Fig. 2(b)].^55,56 Channels were categorized into four major regions based on the macroanatomy classification: bilateral postcentral, bilateral premotor, bilateral temporal, and frontopolar. The peripheral measurement was simultaneously conducted using multichannel PPG device (McLean Hospital, Massachusetts). This device was equipped with Nellcor type D-YS (ear-clip sensor) and DS-100A pulse oximeter (finger-clip sensor) probes with dual wavelengths (660 and 920 nm) and sampling rate 31.25 Hz.⁵⁷ Ear-clip probes were attached to both ears and a finger-clip sensor was put on the left index finger [Fig. 2(c)].

Fig. 2

(a) A plane of fNIRS probe holder with $3\times 11$ configuration of source-detector (b) measured 52 channels that were spatially registered in frontal cortex. The frontal cortex was classified into four major channel regions: bilateral postcentral (i.e., white-colored channels), bilateral premotor (black-colored channels), bilateral temporal (blue-colored channels), and frontopolar (red-colored channels). The fNIRS probe was put on the subject’s forehead whereas the PPG with controller box was connected to ear- and finger-clip probes fixed in subject’s ears and left index finger.

2.4.

Data Preprocessing and Analysis

Analyses were computed using MATLAB (The MathWorks, Inc.) and plug-in-based analysis platform, Platform for Optical Topography Analysis Tools (POTATo, developed by Hitachi Central Research Laboratory).⁵⁸ Measured data from prefrontal and peripheral sites were both initially converted to the product of hemoglobin concentration changes and optical path length ($\mathrm{\Delta }C·L$) for three signal types ${\mathrm{O}}_2\mathrm{Hb}$, HHb, and Hb-total following the modified Beer–Lambert equation. ^2,59 Figure 3(a) shows the flowchart of the analysis summary in which three analysis steps were performed.

Fig. 3

(a) Overall summary of data analysis and (b) detail flow-chart for GLM-based systemic denoising process.

2.4.1.

Processing without systemic denoising

The continuous and channelwise fNIRS signal was bandpass filtered (0.01 to 0.15 Hz; zero-phase Butterworth).⁵¹ The activation was estimated on the basis of the hemodynamic response function (HRF) in GLM analysis

Eq. (1)

$$y=x\beta +ϵ$$

where $y\in R^{L\times ch}$ ($L$ = length of time points, ch = channel numbers) is fNIRS signals. $x\in R^{L\times N}$ is the regressor matrix ($N$ = number of regressor) designed by a constant, convolution between the boxcar function and HRF functions (i.e., canonical two-gamma, temporal and dispersion derivatives). The boxcar function corresponds to the time-piece of events (i.e., memory encoding and button pressing) as shown in Fig. 4(a). $\beta \in R^{N\times ch}$ is the estimated linear regressor of $x$ by the least square error, normally distributed $ϵ\in R^{L\times ch}$ with 0 mean and ${\sigma }^2$ variance:⁶⁰

Eq. (2)

$$h(t)=\frac{t^{{\tau }_p}{e}^{-t}}{({\tau }_p)!}-\frac{t^{({\tau }_p+{\tau }_d)}{e}^{-t}}{A({\tau }_p+{\tau }_d)!},$$

where $h$ is the time ($t$) function of canonical two-gamma HRF,⁶¹ ${\tau }_p$ is the parameter of first peak delay, ${\tau }_d$ is the parameter of second undershoot peak delay, and $A$ is the amplitude ratio between the first and second peaks. We used the typical parameters of 6, 10, and 6 for ${\tau }_p$, ${\tau }_d$, and $A$, respectively. Figure 4(b) shows the example of a regressor matrix. The model significance was statistically evaluated as follows:

Eq. (3)

$$T_{\mathrm{stat}}=\frac{c^{\prime }\beta }{\sqrt{{\hat{\sigma }}^2{c}^{\prime }{(x^{\prime }x)}^{-1}c}},$$

Eq. (4)

$${\hat{\sigma }}^2=\frac{{(y-x\beta )}^{\prime }(y-x\beta )}{L-p},$$

where $T_{\text{stat}}$ is the $t$-statistic value, $c$ is the contrast vector to infer the effect of canonical two-gamma HRF (i.e., $c$ = [0 1 0 0 0 0 0]), ${\hat{\sigma }}^2$ is the sum of squared error values corrected by $L-p$ degree of freedom ($p$ = rank $x$).^62–64 The channel significance was statistically evaluated in random effect group analyses using a one-sample $t$-test of estimated $\beta$-values for the HRF regressor against zero. The strength of the phenomenon was assessed using the parameter of Cohen’s $d$ effect size. The subject-level analysis evaluated the HRF $T_{\text{stat}}$ with $L-p$ degree of freedom.

Fig. 4

(a) Construction of neuronal activation regressors based on HRF (red plot for canonical HRF) and its derivatives (temporal and dispersion derivatives; blue and black plots, respectively) convolved with two event-designs of memory encoding and button pressing. (b) An example of regressor matrix design ($x$) with orders of constant, encoding canonical HRF, encoding temporal, encoding dispersion, button-pressing canonical, button-pressing temporal, and button-pressing dispersion derivatives.

3.1.

Brain Activation in the Processing without Systemic Denoising

We found the significantly increased ${\mathrm{O}}_2\mathrm{Hb}$ activations ($p<0.05$; Holm–Bonferroni corrected) only in the spatial WM task but not in any control tasks [Figs. 5(a)–5(d)]. For the spatial WM task, the increase of ${\mathrm{O}}_2\mathrm{Hb}$ activation was significantly seen in both middle frontal (Broadmann area/BA 10, 45, 46, and 11) and bilateral temporal supramarginal areas, with the strongest activation observed in the right ventrolateral prefrontal cortex (VLPFC, channel 24, right BA 45, $T_{\text{stat}}=9.37$). HHb activation was also investigated in group analysis; however, there was no significant decrease observed in any tasks. Oppositely, we found minor significant HHb increases in the lower frontopolar only in the spatial WM task. We confirmed the significance of individual data. All subjects (i.e., 16) showed the significantly increased ${\mathrm{O}}_2\mathrm{Hb}$ activation ($p<0.05$) in right BA 45 (i.e., channels 24). Seven out of 16 subjects showed the significantly decreased HHb activation ($p<0.05$) in right BA 45 as performing the spatial WM task. From these results, we classified three response types toward the spatial WM task in BA 45: (1) seven subjects having both significant ${\mathrm{O}}_2\mathrm{Hb}$ increase and HHb decrease, (2) eight subjects having significant ${\mathrm{O}}_2\mathrm{Hb}$ and HHb increases, and (3) a subject having significant ${\mathrm{O}}_2\mathrm{Hb}$ increase and null HHb decrease. This suggested that the ${\mathrm{O}}_2\mathrm{Hb}$ response was more uniform compared to the HHb response during the spatial WM task. Despite the various hemodynamic responses in the spatial WM task, neither significant ${\mathrm{O}}_2\mathrm{Hb}$ increase nor HHb decrease was found during control tasks. The finding that the spatial WM task alone showing significant activation suggested that (1) the presence of spontaneous LF systemic noise (and any other noises) does not evoke false-positive results in the group-analysis of control tasks and (2) fNIRS signal during the spatial WM task does reflect encoding-evoked response, neither the sham encoding (i.e., motor control task) nor the motor-related activation (i.e., button-pressing).

Fig. 5

$T_{\text{stat}}$ maps of ${\mathrm{O}}_2\mathrm{Hb}$ (A1–D1) and HHb (A2–D2) activations for spatial WM (A1–2), motor control (B1–2), eyes-open (C1–2), and eyes-closed RS (D1–2) before denoising. Black encircled channels indicate the channels with significant activation ($p<0.05$; Holm–Bonferroni correction) in one-sample $t$-test.

3.2.

Temporal Correlation between Prefrontal and Peripheral Signals

Figure 6 shows the ${\mathrm{O}}_2\mathrm{Hb}$ and HHb correlations between the fNIRS and shifted peripheral ear signals from a representative subject. The fNIRS signal came from channel 24 located in the right BA 45 in approximate. Because $r_{\text{max}}$ for the ${\mathrm{O}}_2\mathrm{Hb}$ signal [i.e., 0.57; Fig. 6(a)] is greater than the threshold, ${\mathrm{O}}_2\mathrm{Hb}$ denoising was done in that channel. $r_{\text{max}}$ for the HHb signal is low [i.e., 0.03; Fig. 6(b)]; thus, the HHb signal of channel 24 was left untouched (red-dotted line). Figure 7 presents the subject-average correlation (inverse $z$-transform) between channelwise fNIRS and peripheral PPG signals (${\mathrm{O}}_2\mathrm{Hb}$ and HHb) for all tasks. There was no significant task effect on both (${\mathrm{O}}_2\mathrm{Hb}$ and HHb) prefrontal-periphery correlation and optimum delay (${\tau }_{\mathrm{opt}}$) using any peripheral sources. The ${\mathrm{O}}_2\mathrm{Hb}$ ${\tau }_{\mathrm{opt}}$ was more positive (i.e., peripheral signal was ahead of the prefrontal signal, $+0.8\pm 1.5\mathrm{s}$; $\text{mean}\pm \mathrm{S}.\mathrm{D}.$) than HHb ${\tau }_{\mathrm{opt}}$ ($-0.6\pm 2.2\mathrm{s}$; $\text{mean}\pm \mathrm{S}.\mathrm{D}.$). ${\mathrm{O}}_2\mathrm{Hb}$ signals measured in premotor, temporal, and frontopolar regions had the significantly higher correlation with peripheral signals measured in ears than in fingers [Fig. 7(e)]. Stronger correlation in the frontopolar was significantly observed compared to premotor (post-hoc Tukey–Kramer test; $p<0.05$) for ${\mathrm{O}}_2\mathrm{Hb}$ prefrontal and peripheral ear signals. Meanwhile, peripheral finger signals showed strong relationship with ${\mathrm{O}}_2\mathrm{Hb}$ signals of postcentral and frontopolar regions compared to premotor and temporal regions (post-hoc Tukey–Kramer test; $p<0.05$). The selection of peripheral source unlikely influenced HHb correlations [Figs. 7(A3)–7(D4)] and there was no observed spatial effect (interregions ANOVA, $p>0.05$). The lower correlation of HHb prefrontal and peripheral ear signals was significantly observed ($\mathrm{CI}>0.999$) in all regions compared to the ${\mathrm{O}}_2\mathrm{Hb}$ correlations [i.e., red-filled and blue-filled boxplots in Fig. 7(e)]. ${\mathrm{O}}_2\mathrm{Hb}$ and HHb correlations between premotor-temporal and peripheral finger signals insignificantly differed. However, postcentral and frontopolar regions showed higher ${\mathrm{O}}_2\mathrm{Hb}$ correlation compared to HHb correlation toward the peripheral finger signal [i.e., red-void and blue-void boxplots in Fig. 7(e)]. These suggested that (1) HHb signal might be less affected by peripheral shared systemic noise, (2) peripheral ear signals shared more similar components with prefrontal signals than peripheral finger signals, and (3) the effect of systemic noise on ${\mathrm{O}}_2\mathrm{Hb}$ signal was inhomogeneous over the PFC.

Fig. 6

Representative temporal traces of (a) ${\mathrm{O}}_2\mathrm{Hb}$ and (b) HHb prefrontal (solid black line, left $y$-axis scale) and peripheral ear (blue dashed-line, right $y$-axis scale) signals illustrating high ${\mathrm{O}}_2\mathrm{Hb}$ ($r=0.57$) and low HHb ($r=0.03$) temporal correlations in the middle frontal cortex of channel 24 during the spatial WM task. Shown peripheral ear signal had already been shifted ($+0.7\mathrm{s}$ and $-4.3\mathrm{s}$ for ${\mathrm{O}}_2\mathrm{Hb}$ and HHb signals, respectively) according to the corresponding maximum correlation ($r_{\text{max}}$). Denoising was performed resulting in the low correlation ($r=-0.13$) between ${\mathrm{O}}_2\mathrm{Hb}$ denoised (red dotted-line, left $y$-axis scale) and peripheral ear signals. HHb denoising was not performed due to lower correlation than 0.3 of threshold. Shaded-gray area indicated the task period of spatial WM (target stimulus and delay intervals, 8.5 s in total).

Fig. 7

Subject-average ${\mathrm{O}}_2\mathrm{Hb}$ (A1–D2) and HHb (A3–D4) correlation maps ($r_{\text{max}}$; inverse $z$-transform) of prefrontal and peripheral signals measured from ears (A1–D1, A3–D3) and finger (A2–D2, A4–D4) during spatial WM (A1–4), motor control (B1–4), eyes-open (C1–4), and eyes-closed RS (D1–4). ${\mathrm{O}}_2\mathrm{Hb}$ correlation variances across subjects (E) showing significant differences of peripheral sources (red-filled and red-void boxplots for ear and finger signals, respectively) in premotor, temporal, and prefrontal cortices – $p<0.001$^(***), $p<0.01$^(**) for paired-sample $t$-test. There was no observed effect of peripheral sources in HHb correlations of all regions (blue-filled and blue-void boxplots for ear and finger signals, respectively).

3.3.

Efficiency of Denoising

Denoising performance was compared in all signals (i.e., ${\mathrm{O}}_2\mathrm{Hb}$ and HHb) and all tasks (i.e., spatial WM and other control tasks) using the current peripheral-GLM method (i.e., ear, finger, and both signals) and PCA 95. For signal reduction parameter, the task effect was initially evaluated. The signal reduction was comparable across tasks in all signal types and methods. Figure 8 shows subject variances in ${\mathrm{O}}_2\mathrm{Hb}$ and HHb reductions using several denoising methods corresponding to four major regions. There was no spatial effect (ANOVA, $p>0.05$) on ${\mathrm{O}}_2\mathrm{Hb}$ and HHb reductions using PCA 95. Performance of ${\mathrm{O}}_2\mathrm{Hb}$ reduction using the ear-GLM denoising was spatially related where the highest reduction happened in the frontopolar ($34\pm 16\%$, task-average max. in left BA 10/11 by $45\pm 23\%$; $\text{mean}\pm \mathrm{S}.\mathrm{D}.$). Peripheral-GLM denoising using ear signals had comparable ${\mathrm{O}}_2\mathrm{Hb}$ noise reduction to PCA 95 in all regions except premotor (post-hoc Tukey–Kramer test). By using both ear and finger signals, the denoising performance insignificantly differed compared to PCA 95. The performance of peripheral finger regressor was worse compared to the dual ear-finger regressors in all regions. HHb prefrontal and any peripheral signals presented low correlation in all regions (i.e., no spatial effect); thus, HHb signals were likely left unprocessed. Therefore, denoising performance of HHb peripheral-GLM method was significantly lower than PCA 95.

Fig. 8

Subject variances of systemic noise reduction in (a) ${\mathrm{O}}_2\mathrm{Hb}$ and (b) HHb signals neglecting the task effect in four channel regions: postcentral, premotor, temporal, and prefrontal cortices performed by ear-GLM denoising (black-filled boxplot), finger-GLM denoising (black-void boxplot), dual peripheral-GLM denoising (red-filled boxplot), and PCA 95 (red-void boxplot) – $p<0.001$^(***), $p<0.01$^(**), $p<0.05$^(*) for ANOVA and post-hoc Tukey–Kramer test. Because of low correlations between prefrontal and peripheral signals in HHb parameter, HHb signals were mostly not denoised. Therefore, the performance of peripheral-GLM denoising in HHb signal was significantly lower than PCA 95.

We consistently found no activation and no substantial effect size change of ${\mathrm{O}}_2\mathrm{Hb}$ and HHb activations in control tasks after denoising using any methods. The denoising effect on activation changes was only observed in the spatial WM task. Figure 9 shows Cohen’s $d$ effect size maps of ${\mathrm{O}}_2\mathrm{Hb}$ and HHb activations during the spatial WM task after denoising. Denoising decreased the significances of ${\mathrm{O}}_2\mathrm{Hb}$ increase ($p<0.05$; Holm–Bonferroni correction; black-circled channels) in group-analysis. The most decrease was performed by the dual-peripheral-GLM denoising [Fig. 9(D1)] showing more right lateralization (i.e., BA 45; channel 24). By using only ear signals as noise regressors, we still observed ${\mathrm{O}}_2\mathrm{Hb}$ significance in the surrounding BA 45 compared to the finger-GLM and dual peripheral-GLM denoising. Regarding the low reduction of ${\mathrm{O}}_2\mathrm{Hb}$ signal by finger-GLM denoising [Fig. 8(a)], the decreased significance of ${\mathrm{O}}_2\mathrm{Hb}$ activation was unexpectedly observed. Furthermore, the change of HHb activation was not observed after denoising using any methods [Figs. 9(A2)–9(E2)]. These results were expected in the peripheral-GLM denoising due to low signal reduction; yet, PCA 95 presenting comparable signal reduction in both ${\mathrm{O}}_2\mathrm{Hb}$ and HHb signals also showed the insignificant change of HHb activation. There was no significant difference of channelwise activation (i.e., ${\mathrm{O}}_2\mathrm{Hb}$ and HHb) across methods.

Fig. 9

Cohen’s $d$ effect size maps of ${\mathrm{O}}_2\mathrm{Hb}$ (A1–E1) and HHb (A2–E2) activations for the spatial WM task before denoising (A1-2), after ear-GLM denoising (B1–2), finger-GLM denoising (C1–2), dual peripheral-GLM denoising (D1–2), and PCA 95 (E1–2). Black encircled channels indicate the significant activation channels ($p<0.05$; Holm–Bonferroni correction).

4.1.

Direct Versus Multimodality Interpretations

Some physiological measurements, such as respiration rate by a pneumatic belt or arterial blood pressure by cuff,^45,65 require additional interpretations to translate them into blood-related LF signals, before used as noise regressors in fNIRS studies. In our study, we directly measured the blood-related LF signals in peripheral sites using PPG, concurrently with the prefrontal fNIRS measurement. The peripheral LF signal is so-called “catch-all” signal that represents the combination of LF systemic physiological noises from various origins. It may include the fluctuations in heart rate (HR), arterial blood pressure (ABP), respiration (or downsampled version of them). Under the assumption that these signals affect both brain and peripherals based on vascular relations, the blood-related physiological signals can be used directly to denoise functional signals (i.e., fNIRS or fMRI) without knowing exactly the origins of these low-frequency oscillations (LFOs). The detailed benefits of such direct measurement can be found in the previous work of Tong et al. through simultaneous fMRI-NIRS measurement.^49,66

4.2.

Regressor Suitability

Cooper et al.⁶⁷ previously used frontal and temporal lobes fNIRS signals as physiological regressors for fMRI data. However, an fNIRS signal with 3 cm of S-D distance is a mixture of both neuronal and nonneuronal signals from cortical and extracerebral layers, respectively. Directly using those as models could regress out the real neuronal signal. Concurrent finger measurements have been attempted to obtain the neuronal-free (i.e., systemic) signal as a noise regressor.^49,66 However, noise regressor obtained from the ear has the following advantages: (1) the blood in ear is supplied by external carotid artery, which is close to the internal carotid artery (i.e., blood supply to brain); (2) the ear is easy to measure by our clip-equipped PPG probe with superior signal-to-noise ratio. As shown in Fig. 7, peripheral ear signals had the significantly higher correlation with prefrontal signals than those from fingers resulting in greater noise reduction (Fig. 8). This confirms our hypothesis that closer vascular distance may lead to high similarities between the signals. Despite the less signal reduction, the significance of ${\mathrm{O}}_2\mathrm{Hb}$ activation in group analysis decreased after finger-GLM denoising. This suggested that denoising using less correlated signals could possibly regress out some signal portions. It was difficult to evaluate whether over-processing (i.e., spurious improved activation localization) actually happened or not. Therefore, the highly correlated noise model (e.g., peripheral ear signals) is more preferable for denoising.

In addition to sources of systemic noise regressor, we also investigated the regressor suitability between two signal types (${\mathrm{O}}_2\mathrm{Hb}$ versus HHb). The correlation between HHb prefrontal and peripheral (either ear or finger) signals is significantly lower than that of ${\mathrm{O}}_2\mathrm{Hb}$ signals (Fig. 7). There are two arguments that may explain this HHb characteristic. First, HHb is less affected by the systemic noise. For example, according to Franceschini et al.,¹⁸ heart rate change significantly modulates the arterial compartment and is highly reflected in ${\mathrm{O}}_2\mathrm{Hb}$; this is not the case for HHb in venous vasculature. Therefore, the HHb signal may be less prone to those systemic physiological fluctuations. This hypothesis is also consistent with the results reported by Sato et al.⁶⁸ with the finding of lower correlation between HHb signal at brain and systemic skin blood flow (SBF) measured using laser Doppler flowmeter. Second, neither peripheral nor systemic SBF recordings are sufficient to model systemic noise in HHb signal. Even HHb extracerebral was found having low coherence with HHb brain.⁶⁹ Previous studies reported that HHb brain is less sensitive and inconsistently responds toward cerebral blood flow change.^70,71 In the current results, we found nonuniform HHb responses during the encoding interval across subjects resulting in the insignificance of HHb activation. The individual variances may implicate this issue. According to the above arguments, a global systemic variation in ${\mathrm{O}}_2\mathrm{Hb}$ signal is a better noise regressor.

4.3.

Systemic Denoising and Its Comparison to Other Denoising Methods

Katura et al.⁷² demonstrated the causal relationship between fNIRS signals and cardiovascular systemic noises modeled using HR and mean ABP during resting state. The study focused on a single channel signal around the posterior-superior temporal lobe and reported that the contribution of cardiovascular systemic noise up to 35% in ${\mathrm{O}}_2\mathrm{Hb}$ and Hb-total signals. The remaining 65% portion was still unidentified noise components. In this study, we sought to isolate and remove systemic noise identified from peripheral sites. The reduction caused by systemic denoising (i.e., peripheral ear in ${\mathrm{O}}_2\mathrm{Hb}$ signal), or also called as systemic noise contribution, was found to be around 25% to 34% in average and 39% to 58% in maximum (frontopolar channel 36; BA 10/11) across tasks.

From our results, we found that the systemic noise contribution was equal for any tasks. The spatial variability of systemic noise contribution was also observed over PFC as previously reported.⁷³ Zhang et al.⁷⁴ showed that LFO highly accumulated in ${\mathrm{O}}_2\mathrm{Hb}$ signal over the frontopolar regions. This phenomenon may be explained by the fact that low signal sensitivity in the regions with big vessels (causing high absorption and low photon escape^75,76). While the frontopolar is covered by the relatively small vessels (0.9 to 1.1 mm in diameter),⁷⁷ the vessels in temporal regions are slightly greater than 2 mm.⁷⁸ Apart from the above arguments, the task-evoked systemic noise might be colocalized in the draining scalp veins.^46,49

In order to quantify the comparison between current and previous methods, we performed PCA 95 in the same datasets. PCA 95 is a well-known method to eliminate motion artifact with the assumption of less contribution of noise toward the signal. The arbitrary threshold of contribution cut-off (e.g., 95% and 80%) always become a limitation of PCA. Performance of PCA in signal reduction was uniform over PFC while the performance of peripheral-GLM denoising was spatially influenced. The ${\mathrm{O}}_2\mathrm{Hb}$ signal reduction was comparable between (dual) peripheral-GLM denoising and PCA. After denoising, the significance of ${\mathrm{O}}_2\mathrm{Hb}$ activation in group analysis was decreased by all the methods. There was no channelwise difference across methods in both ${\mathrm{O}}_2\mathrm{Hb}$ and HHb activations despite the high reduction of HHb signal by PCA 95. This could suggest that (1) the current peripheral-GLM denoising performed comparably to PCA 95 in ${\mathrm{O}}_2\mathrm{Hb}$ signals and (2) the HRF model for HHb signals might not be optimum. Uga et al.⁷⁹ reported that the optimization of ${\tau }_p$ parameter could improve the observations of ${\mathrm{O}}_2\mathrm{Hb}$ increase and HHb decrease. Even though the performance of peripheral-GLM denoising was equal to PCA, our current method could specifically remove systemic (i.e., nonneuronal) components rather than determining the assumption of signal contribution as PCA. Furthermore, PCA performance is influenced by the availability of channel number. Peripheral GLM denoising is potentially applicable for the required minimum measurement system.

Despite the potential results, we were still aware of the extracerebral (e.g., scalp) signal in the denoised signals. The scalp effect has been reported to be significant in PFC.⁸⁰ One of the well-known methods to eliminate extracerebral signal is MD S-D separation.³⁶ The characteristics of systemic denoising and MD S-D separation can be different from each other. External carotid artery supplying blood to ear shares the same upstream vasculature (i.e., common carotid artery) with internal carotid artery for brain supply. Any fluctuations in common carotid artery could affect downstream vessels in both brain and ears. Thus, the ear measurement is a reasonable interpretation of global systemic noise. In opposite, the MD S-D separation directly measures the regional changes in the superficial layer corresponding to nearby fNIRS channels. We believe that global systemic noise is also confined in the regional superficial signal due to similar blood vasculature as mentioned above. The impact of global systemic noise on regional superficial signal is still unknown. Furthermore, both systemic and extracerebral denoising have not yet been done together in fNIRS data analysis. We hope to address this issue in the future studies.

4.4.

Task-Related Effect on Peripheral Ear Signals

There are two issues concerned with this subdiscussion. First, we have observed the high correlation between the ${\mathrm{O}}_2\mathrm{Hb}$ frontopolar and peripheral ear signals for all tasks. The correlation distributions were also similar regardless of tasks (Fig. 7). Tachtsidis et al. and Sato et al.^68,81 evidenced the high correlation of frontal lobe ${\mathrm{O}}_2\mathrm{Hb}$ fNIRS with global changes of mean blood pressure and SBF, respectively. Kohno et al.⁴⁰ also reported the effect of SBF contamination on the wide area of the forehead. This may implicate the high correlation between ${\mathrm{O}}_2\mathrm{Hb}$ peripheral ear and SBF. This relationship is likely associated with the same vascular source of the common carotid artery.

Second, we explored the task-related signals from peripheral measurements as shown in Figs. 10(A1)–10(D2). The averaged waveforms were different for all tasks. As performing the same task, ear and finger signals presented difference waveforms. Li et al.⁸² demonstrated the sensitivity difference on systemic manipulation responses (e.g., passive leg raising, paced breathing) in the ear, finger, and toe. We confirmed that there was no significant difference in ${\mathrm{O}}_2\mathrm{Hb}$ and HHb activations between eyes-open and eyes-closed RS controls in both peripheral ear and finger signals. The significant task-related effect was observed in both prefrontal and peripheral (i.e., ear and finger) signals during the spatial WM task [Figs. 10(A1)–10(A3)]. Such effect was also reported in other tasks: verbal WM, finger tapping, visual checker-board, and anagram tasks.^36,68,81 This finding is also consistent with the previous work of Kirilina et al.,⁴⁶ where they found task-related activations in the superficial layer and concluded that “physiological origin of the systemic artifact is a task-evoked sympathetic arterial vasoconstriction followed by a decrease in venous volume.” These results comprehensively suggested that the tasks also simultaneously triggered the global systemic noises, which is consistent with the neurovascular coupling theorem arguing the local systemic regulation.^4,83,84

Fig. 10

Subject-average ear (A1-D1), finger (A2–D2), and right BA 45 (channel 24, A3–D3) signals (red and blue lines for ${\mathrm{O}}_2\mathrm{Hb}$ and HHb signals, respectively, with their standard errors) during (a) spatial WM, (b) motor control, (c) eyes-open, and (d) eyes-closed RS. The gray-shaded interval indicates the task period (target stimulus and delay intervals, 8.5 s in total). $P<0.01$^(***), $p<0.05$^(**), and $p<0.1$^(*) for one-sample $t$-test of ${\mathrm{O}}_2\mathrm{Hb}$ increase and HHb decrease activations ($\beta$ of encoding canonical HRF).

Since the peripheral signals have task-evoked effects, there is a concern about the denoising method using the peripheral signals. We argue that the signal from the ears is a pure systemic physiological artifact. It does not contain any neuronal “contamination,” even it shows task-evoked changes. These changes are likely induced by global systemic changes as a result of performing tasks. Therefore, these peripheral signals are “noises” and have to be removed, regardless of their waveforms. Moreover, their waveforms (i.e., global systemic changes) should be different from the real brain activations, which is neuronal and regional [Figs. 10(A3)–10(D3)]. These global systemic changes are likely coupled and propagated into the blood-related change (i.e., flow, volume, and oxygenation) that affects the different brain and facial regions at the different time. Thus, in this study, in addition to accurately record the non-neuronal noise from ears, we adaptively remove it by temporally matching with each fNIRS channel before subtraction. As a result, we found the significant ${\mathrm{O}}_2\mathrm{Hb}$ increase in DLPFC/VLPFC during the spatial WM task (Fig. 9) as previous studies^85–88 even after denoising. This demonstrates that removing physiological noise with task-related waveform, we did not remove the task activation.