1.

Introduction

The interest in near-infrared spectroscopy (NIRS) for mapping regional brain activation associated with functional tasks is growing rapidly, as reflected by the increasing number of commercial systems.^1–3 Despite advantages in terms of safety, low cost, and portability, functional near-infrared spectroscopy (fNIRS) applications using continuous-wave (CW) technologies are limited in insolating signals from the brain. Substantial signal contamination can result from changes in hemodynamics in the extracerebral layer (ECL).^4,5 As a consequence, a number of techniques have been proposed to reduce the influence of ECL signal contributions. An increasingly popular approach is to collect light intensities at a short source–detector separation (i.e., $r_{\mathrm{SD}}\sim 1\mathrm{cm}$), which is predominately sensitive to scalp hemodynamics while simultaneously collecting data at a larger separation ($r_{\mathrm{SD}}\ge 3\mathrm{cm}$).⁶ Assuming the two signals contain the same ECL contributions, the short separation signal can be used as a regressor to filter the ECL interference from the main signal, improving the quality of the recovered brain signal.^6–8

An alternative strategy is to enhance the depth sensitivity of the measurements using time-resolved NIRS (trNIRS), which uses picosecond light pulses and fast detectors to record the distribution of times-of-flight (DTOF) of diffusely reflected photons.⁹ As DTOFs contain both time and intensity information, absorption changes at different depths can be resolved since photon arrival times are proportional to path lengths. The most popular depth-enhancing methods are based on calculating the statistical moments of a DTOF^10,11 or integrating the photon counts within time windows/gates.^12,13 In both cases, the goal is to focus on late-arriving photons since they have the greatest probability of interrogating the brain. Previous studies using layered tissue-mimicking phantoms, animal models, and human subjects have shown that trNIRS provides superior sensitivity to cerebral hemodynamics compared to conventional CW NIRS.^13–17 Moreover, studies conducted on healthy subjects and patients have shown that signals extracted using either approach provide superior correlation with hemodynamic changes in the brain.^9,18

Despite the enhanced depth sensitivity of trNIRS, signals weighted toward late-arriving photons can still be influenced by hemodynamic fluctuations in the ECL.¹⁹ This was demonstrated in a recent trNIRS study focused on measuring cerebrovascular reactivity in response to a global vascular stimulus (hypercapnia).²⁰ Contamination from the scalp was evident by distortions in the oxy- and deoxyhemoglobin signals recorded at $r_{\mathrm{SD}}=3$ and 4 cm that mirrored the recordings at $r_{\mathrm{SD}}=1\mathrm{cm}$. By applying higher moment analysis, these effects were substantially diminished but not completely removed.²⁰ Alternatively depth sensitivity can be further enhanced by subtracting higher moments of DTOFs recorded at two source–detector distances;²¹ however, the application of this approach is challenging due to the substantial reduction in contrast-to-noise ratio compared to analyzing individual DTOFs.²²

In this study, an alternative technique was investigated that combines trNIRS with a regression approach analogous to short source-detector measurements used in CW fNIRS studies.^6,7 Unlike CW NIRS that requires an additional channel to act as the regressors, regression can be applied to trNIRS data from a single detector by utilizing the depth information provided in the recorded DTOFs. In this application, a signal that is predominately weighted by the ECL can be extracted from a DTOF by focusing on early arriving photons and subsequently used as a regressor to remove ECL interference from signals with greater depth sensitivity, such as those obtained from higher moments or later gates. An additional advantage of trNIRS is that both the regressor and the dependent variable are extracted from data collected by the same detector. This avoids potential artifacts that can arise when the two signals are recorded by sensors located at different locations, considering the regression approach is based on the assumption that the regressor and dependent variable contain the same physiological nuisance signal. It has been demonstrated that the correlation between physiological signals recorded using separate sensors diminishes with an increasing distance between each sensor due to time delays and spatial variations in the hemodynamic properties of superficial tissue.^8,23

The ability to improve the isolation of the brain signal by adapting regression analysis to trNIRS was evaluated using data from two previous studies using trNIRS to measure oxygenation responses to hypercapnia²⁰ and functional activation.²⁴ The former was chosen because the rapid and relatively large increase in end-tidal carbon dioxide pressure ($\sim 15\mathrm{mmHg}$) elicited a substantial systemic effect, as indicated by a significant change in arterial blood pressure, in addition to the expected cerebral response. Differences in ECL and cerebral contributions were evident by comparing signals recorded at $r_{\mathrm{SD}}=1$ and 3 cm. Consequently, this data set was used to evaluate the ability of the regression approach to remove ECL contributions when applied to DTOFs recorded at $r_{\mathrm{SD}}=3\mathrm{cm}$. The functional study²⁴ involved a motor imagery task that invoked activation in motor planning regions and was included to demonstrate the feasibility of applying the regression approach to functional trNIRS data. Both applications involved two common approaches used to extract depth information from DTOFs, time gating and statistical moment analysis.^10,13 In addition to these applications, this study includes a sensitivity analysis conducted to select the appropriate trNIRS metrics for regression analysis.

2.

Methods

All experiments were approved by the Western University Health Sciences Research Ethics Board, which adheres to the guidelines of the Tri-Council Policy Statement for research involving humans. Written informed consent was obtained from each subject prior to the experiment. All subjects had no history of neurological or psychiatric disorders.

2.2.

Experimental Protocol: Hypercapnia

Data were obtained from a previous study that involved 11 healthy participants (three females, eight males, aged 25 to 36 $y$, $\text{mean}=28\pm 3$ $y$).²⁰ Subjects sat on a reclining chair while wearing a facemask connected to a computer-controlled gas delivery circuit (RespirAct™, Thornhill Research Inc, Toronto, Canada). A custom-designed probe holder was placed on the subject’s forehead and secured by a velcro headband. Two detection fiber bundles were placed at $r_{\mathrm{SD}}=1$ and 3 cm [Fig. 1(a)]. The experimental protocol consisted of three 2 min periods of hypercapnia in which end-tidal carbon dioxide pressure (${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$) was increased by 15 mmHg above each subject’s normocapnic ${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ value, as determined by the gas delivery circuit. The first period started two minutes after baseline recordings and each hypercapnia period was followed by 5 min of normocapnia.

Fig. 1

Schematic of the design and location of the probe holders used in (a) the hypercapnia protocol and (b) the functional activation protocol.

2.4.

Time-Resolved Depth Sensitivity—Theory

Two approaches for extracting depth information from DTOFs are by time gating and calculating the first three statistical moments: the zeroth moment (total number of photons, $N$), the first moment (mean time of flight, $⟨t⟩$), and the second central moment (variance, $V$). Higher moments are more sensitive to late-arriving photons due to the positive skewness of the DTOF.^10,11 With the former approach, each DTOF is divided into time gates, and the photon count is integrated within each gate.^12,29 Similar to higher moments, gates positioned on the tail of a DTOF provide the greatest depth sensitivity, whereas the gates located on the left side of the DTOF are more sensitive to more superficial layers. Based on these considerations, combinations of statistical moments and time gates can be used to extract signals with different depth sensitivities from a single DTOF.

Figure 2 shows an example of a theoretical DTOF and the corresponding sensitivity profiles of the signals extracted from various gates and moments. The DTOF was generated using the analytical solution to the diffusion approximation for a semi-infinite homogeneous medium for a source-detector separation of 3 cm, an absorption coefficient (${\mu }_{\mathbf{a}}$) of $0.1{\mathrm{cm}}^{-1}$, and a reduced scattering coefficient (${{\mu }_{\mathrm{s}}}^{\prime }$) of $10{\mathrm{cm}}^{-1}$.³⁰ The sensitivity profiles were also generated from the solution to the diffusion approximation for a semi-infinite homogeneous medium following the approach outlined by Kacprzak et al.³¹ Briefly, the medium was divided into a grid of $0.1\times 0.1\times 0.1{\mathrm{cm}}^3$ voxels and the signal response to a change in ${\mu }_{\mathrm{a}}$ of $0.01{\mathrm{cm}}^{-1}$ was recorded within each consecutive voxel. Next, the sensitivity factors for consecutive layers (each with a thickness of 1 mm) were generated by summing the sensitivity values for all voxels within a layer. The same optical properties used to generate the theoretical DTOF were used to define the background optical properties for calculating the sensitivity profiles.^15,21,32 Figure 2(a) shows the three statistical moments, as well as early and late gates ($\text{gate width}=250\mathrm{ps}$). The sensitivity profiles for these two gates and the three moments are shown in Fig. 2(b). Observable differences in these sensitivity profiles indicate that signals weighted differently by the ECL and brain can be extracted from data recorded using a single source–detector separation.

Fig. 2

(a) Theoretical DTOF generated for $r_{\mathrm{SD}}=3\mathrm{cm}$. Overlaid on the DTOF is the locations of early and late gates ($\text{width}=250\mathrm{ps}$) and the first three statistical moments. (b) Corresponding sensitivity profiles for the two gates and the three statistical moments generated using ${\mu }_{\mathrm{a}}=0.1{\mathrm{cm}}^{-1}$ and ${{\mu }_{\mathrm{s}}}^{\prime }=10{\mathrm{cm}}^{-1}$.

2.5.

Analysis of Recorded DTOFs

For each time series of DTOFs recorded at 760 and 830 nm, the first three statistical moments ($N$, $⟨t⟩$, and $V$) were calculated by setting the lower and upper integration limits based on arrival times corresponding to 1% of the peak of the DTOF. The change in each moment relative to its initial value was calculated to generate three time series (i.e., $\mathrm{\Delta }N$, $\mathrm{\Delta }⟨t⟩$, and $\mathrm{\Delta }V$) for the two wavelengths individually. Next, each DTOF was divided into 12 consecutive gates with a fixed width of 250 ps²⁹ and the time-varying change in attenuation ($\mathrm{\Delta }A$) was calculated for each gate (Fig. 3). Similar to moment analysis, the start of the first gate was positioned at the rising edge of the DTOF when the signal intensity reached 1% of the peak value.

Fig. 3

Position of 12 time gates ($\text{width}=250\mathrm{ps}$) used to extract attenuation signals from the DTOF. In this case, the DTOF was averaged over the 2 min baseline period for one subject from the hypercapnia study ($\lambda =760\mathrm{nm}$ and $r_{\mathrm{SD}}=3\mathrm{cm}$).

Time courses determined at 760 and 830 nm from the statistical moments and time gates were converted into the corresponding absorption changes $\mathrm{\Delta }{\mu }_{\mathrm{a}}(\lambda )$ using the sensitivity analysis described in Sec. 2.4. To improve the accuracy of the sensitivity profiles, an IRF (described in Sec. 2.1) was incorporated into the sensitivity profile calculations. The sensitivity factors were generated using average optical properties for the subjects in the hypercapnia study (760 nm: ${\mu }_{\mathrm{a}}=0.14\pm 0.02{\mathrm{cm}}^{-1}$, ${{\mu }_{\mathrm{s}}}^{\prime }=9.4\pm 1.2{\mathrm{cm}}^{-1}$; 830 nm: ${\mu }_{\mathrm{a}}=0.13\pm 0.02{\mathrm{cm}}^{-1}$, and ${{\mu }_{\mathrm{s}}}^{\prime }=8.1\pm 1.2{\mathrm{cm}}^{-1}$):²⁰

Eq. (1)

$$\mathrm{\Delta }{\mu }_a(\lambda )=\frac{\mathrm{\Delta }S}{S{F}_S},$$

where $\mathrm{\Delta }\mathrm{S}$ represents the signal change for each of the statistical moments ($\mathrm{\Delta }N$, $\mathrm{\Delta }⟨t⟩$, and $\mathrm{\Delta }V$) and time gates ($\mathrm{\Delta }A$), and ${\mathrm{SF}}_{\mathrm{S}}$ represents the corresponding sensitivity factor. The $\mathrm{\Delta }{\mu }_{\mathrm{a}}$ time courses were then converted to change in concentration of oxyhemoglobin ($\mathrm{\Delta }{C}_{\mathrm{HbO}}$) and deoxyhemoglobin ($\mathrm{\Delta }{C}_{\mathrm{Hb}}$):

Eq. (2)

$$\mathrm{\Delta }{\mu }_a(\lambda )=\ln (10)·({\epsilon }_{\mathrm{HbO}}(\lambda )·\mathrm{\Delta }{C}_{\mathrm{HbO}}+{\epsilon }_{\mathrm{Hb}}(\lambda )·\mathrm{\Delta }{C}_{\mathrm{Hb}}),$$

where ${\epsilon }_{\mathrm{HbO}}(\lambda )$ and ${\epsilon }_{\mathrm{Hb}}(\lambda )$ are the molar extinction coefficients for oxy- and deoxyhemoglobin, respectively.

2.6.

Regression Analysis

Regression analysis was based on the method proposed by Saager et al.,⁶ which was developed to isolate absorption trends in the lower layer of a two-layer turbid medium. The signal change in the lower layer (i.e., brain), $\mathrm{\Delta }{S}_{\mathrm{D}}$, was calculated according to

Eq. (3)

$$\mathrm{\Delta }{S}_D=\mathrm{\Delta }{S}_F-{\alpha }_{NF}·\mathrm{\Delta }{S}_N,$$

where $\mathrm{\Delta }{S}_{\mathrm{F}}$ is the signal change for a parameter with greater depth sensitivity, $\mathrm{\Delta }{S}_{\mathrm{N}}$ is the signal change for a parameter primarily sensitive to the ECL, and ${\alpha }_{\mathrm{NF}}$ is the scaling parameter obtained by fitting $\mathrm{\Delta }{S}_{\mathrm{N}}$ to $\mathrm{\Delta }{S}_{\mathrm{F}}$ using a least-squares criterion. For the hypercapnia data sets, regression was first applied using the $\mathrm{\Delta }N$ time series measured at $r_{\mathrm{SD}}=1\mathrm{cm}$ as the regressor ($\mathrm{\Delta }{N}_{1\mathrm{cm}}$), analogous to original approach proposed for CW NIRS data. Next, $\mathrm{\Delta }{S}_{\mathrm{N}}$ was extracted from DTOFs measured at $r_{\mathrm{SD}}=3\mathrm{cm}$ based on the criterion that the selected signal had $>97\%$ sensitivity to the ECL. This value was based on an average distance from scalp to brain of 14 mm as measured by magnetic resonance imaging.²⁰ Signals that satisfied this criterion were $\mathrm{\Delta }A$ for gates 1 to 3 [Fig. 4(a)]. The dependent signals ($\mathrm{\Delta }{S}_{\mathrm{F}}$) used in Eq. (3) were obtained for DTOFs measured at $r_{\mathrm{SD}}=3\mathrm{cm}$ and included the three statistical moments and the late gate [Fig. 4(b)]. The last gate was the twelfth for the hypercapnia data and the tenth for the activation data.

Fig. 4

(a) Sensitivity profiles of the signals used as the regressor (i.e., $N$ measured at $r_{\mathrm{SD}}=1\mathrm{cm}$ and the first three gates measured at $r_{\mathrm{SD}}=3\mathrm{cm}$). (b) Sensitivity profiles of the dependent signals measured at $r_{\mathrm{SD}}=3\mathrm{cm}$, which included the three statistical moments and the last gate (gate 12 for the hypercapnia data). The vertical gray line represents the average distance to the brain $d=14\mathrm{mm}$.

3.

Results

Figure 5(a) displays the average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ responses to the 2 min hypercapnia challenge. Data are presented for both source-detector separations and each of the three moments. These time courses illustrate how the dynamics of the hypercapnic response varied with $r_{\mathrm{SD}}$ and statistical moments. Included in each figure is the best fit of the hemodynamic model described in Sec. 2.7. Figure 5(b) shows the $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses obtained from the regression analysis using $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ as the regressor. These time courses show that the dynamics of the reconstructed hypercapnic response were improved for all three statistical moments, as demonstrated by the removal of the substantial residue signals that were observed after ${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ had returned to normocapnia. The first graph in Fig. 5(b) shows the average $\mathrm{\Delta }{\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ time course, which illustrates the highly reproducible hypercapnic response across subjects, and the corresponding $\mathrm{\Delta }S(\mathrm{t})$ generated from Eq. (4) using $\tau =27\mathrm{s}$, which was obtained from previous BOLD studies of dynamic CVR.^33,34 The theoretical bold signal is provided to indicate the signal expected from NIRS if there was no contamination from the ECL.

Fig. 5

(a) Average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ (red) and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ (blue) responses to a 2 min hypercapnic challenge (indicated by the gray shaded region). Time courses are presented for the signals measured at two source-detector separations: $r_{\mathrm{SD}}=1$ and 3 cm and were averaged across nine subjects. The best fit of the hemodynamic model to each averaged hemoglobin time course is illustrated by the black dashed line. (b) Corresponding hemoglobin signals obtained after regressing the $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ from the time courses measured at $r_{\mathrm{SD}}=3\mathrm{cm}$ for the three moments. Average recorded $\mathrm{\Delta }{\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ is presented in the left column, along with a theoretical $\mathrm{\Delta }S(\mathrm{t})$ generated for $\tau =27\mathrm{s}$.

Table 1 provides averaged best-fit values of $\tau$ and ssCVR, from the analysis of $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time series, derived for the individual moments measured at $r_{\mathrm{SD}}=3\mathrm{cm}$ [Fig. 5(a)] and following regression using $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ as the regressor [Fig. 5(b)]. The mean values of $\tau$ did not include two outliers that exceeded the upper boundary of the fitting range $\tau =250\mathrm{s}$ (same subjects for both $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$).

Table 1

Best fit estimates of the time constant (τ) and steady-state cerebrovascular reactivity (ssCVR) obtained from the analysis of the ΔCHbO, ΔCHb data derived from the three moments recorded at rSD=3  cm and the regression approach applied to each moment separately. Values are presented as average (standard deviation) across subjects.

Figure 6 presents average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ responses to hypercapnia derived from the first three gates for DTOFs recorded at $r_{\mathrm{SD}}=3\mathrm{cm}$ [Fig. 6(a)]. Similar to the times series derived from $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ (Fig. 5(a)), all time series exhibited noticeable residue signals after ${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ had returned to normocapnia, particularly for $\mathrm{\Delta }{C}_{\mathrm{HbO}}$. These residues were also evident in the $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ time courses derived from gate 12 at $r_{\mathrm{SD}}=3\mathrm{cm}$ [Fig. 6(b)] and, to a lesser extent, from $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ [Fig. 6(c)]. Similar to Fig. 5, $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses obtained after regression analysis using an early gate as the regressor exhibited improved dynamics of the reconstructed hypercapnic responses for both gate 12 and $\mathrm{\Delta }{V}_{3\mathrm{cm}}$. No significant difference was found in the goodness of fit using an early gate as the regressor compared to using $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ as the regressor.

Fig. 6

Average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ (red) and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ (blue) responses to a 2 min hypercapnic challenge (indicated by the shaded gray region) from data recorded at $r_{\mathrm{SD}}=3\mathrm{cm}$. (a) Responses from the first gates (gate # 1–3); (b) responses from the last gate (gate # 12) before and after applying regression using one of gates 1 to 3 as the regressor; and (c) responses derived from the variance signal before and after applying the same regression. All time courses were averaged across nine subjects, and shading surrounding each line represents the standard error of the mean. The best fits of the hemodynamic model to average hemoglobin signals are illustrated in each graph by the black dashed line.

Table 2 provides the best-fit estimates of $\tau$ and ssCVR from the analysis of $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses obtained with an early gate signal ($r_{\mathrm{SD}}=3\mathrm{cm}$) as the regressor and either gate 12 or $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ as the dependent variable. The ANOVA indicated that $\tau$ and ssCVR estimates for both dependent variables were independent of which early gate was chosen as the regressor. Table 2 includes the average $\tau$ and ssCVR values across the three early gates.

Table 2

Best fit estimates of the time constant (τ) and steady-state cerebrovascular reactivity (ssCVR) obtained from the analysis of the ΔCHbO, ΔCHb data derived from the regression approach based on combinations of early and late gates and by combining ΔV3  cm and an early gate. Values are presented as average (standard deviation) across subjects.

Figure 7 shows the average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses obtained from the functional data for early and late gates, as well as from the variance signal [Fig. 7(a)]. Similar to the hypercapnia data, differences in hemoglobin time courses can be observed for responses obtained for signals primarily sensitive to the ECL (i.e., early gate) and signals with greater depth sensitivity (i.e., late gate or $\mathrm{\Delta }{V}_{3\mathrm{cm}}$). In particular, the averaged hemoglobin time courses from the early gate exhibited a slow component that propagated throughout the 60 s period. The influence of this physiological signal was smaller but still observable in the $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time series derived from the late gate or $\mathrm{\Delta }{V}_{3\mathrm{cm}}$. Regression analysis using the early gate signal as the regressor further reduced the effect of the ECL component, resulting in $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses that were similar to time course predicted by the HRF. This was confirmed by the reduction of the values of the ${\chi }^2$ goodness of fit obtained for late gate or $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ after applying regression analysis; however, statistical significance was only achieved for $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ calculated using $\mathrm{\Delta }{V}_{3\mathrm{cm}}$.

Fig. 7

(a) Average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ (red) and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ (blue) response to functional activation (indicated by the shaded region) obtained for the early gate (mean of signal for gate # 2–3). (b) Regression analysis results obtained for the early gate (used as regressor) and the late gate (gate # 10). (c) Regression analysis results obtained for the early gate (used as regressor) and $\mathrm{\Delta }{V}_{3\mathrm{cm}}$. Shading surrounding each line represents the standard error of the mean. The best fit of the hemodynamic model to each average hemoglobin time course is illustrated by the black dashed line.

4.

Discussion

The main challenge with fNIRS studies involving adults is the substantial signal contamination from the ECL. By isolating late-arriving photons, trNIRS provides the ability to fundamentally improve depth sensitivity. This advantage is demonstrated by the temporal differences in the reconstructed hemoglobin time courses shown in Fig. 5(a). The magnitude of the residue signal following the hypercapnic challenge (i.e., after ${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ returned to normocapnia) diminished as the order of the statistical moment increased. This trend was more pronounced for $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ due to its greater sensitivity to scalp hemodynamics.²⁰ However, higher moments and late gates are still sensitive to the ECL, as indicated by the sensitivity factors shown in Figs. 2 and 3. Consequently, the time constant $\tau$, which characterizes the dynamics of cerebrovascular reactivity was larger than expected for $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ determined from $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ ($\tau =77\pm 45\mathrm{s}$) compared to functional magnetic resonance imaging studies (fMRI).^33,35 For healthy gray matter, $\tau$ should be in the range of 15 to 40 s, as illustrated by the theoretical response shown in Fig. 5(b).

This study focused on investigating if the depth sensitivity of trNIRS could be further improved by incorporating regression analysis. To emulate the method proposed by Saager and Berger⁶ for CW NIRS applications that include a short-separation channel, regression was first applied to trNIRS data recorded at $r_{\mathrm{SD}}=3\mathrm{cm}$ using the $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ as the regressor. The improvement to the hypercapnic response was demonstrated by the reduced residue signal observed in the recovery phase [Fig. 5(b)]. Again, the benefit was greater for $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ as indicated by the significant reduction in $\tau$ for the time series obtained from $\mathrm{\Delta }{⟨t⟩}_{3\mathrm{cm}}$ and $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ (Table 1). The corresponding $\tau$ values for $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ from all three moments were also lower after regression; however, only the results for $\mathrm{\Delta }{N}_{3\mathrm{cm}}$ reached statistical significance. In agreement with CW NIRS studies, these results confirm that applying regression to trNIRS data that includes a short-separation channel is beneficial, even for trNIRS metrics with enhanced depth sensitivity.

A further advantage of trNIRS is the possibility to apply a regression to data acquired at a single $r_{\mathrm{SD}}$ since early arriving photons should be predominately sensitive to the ECL. This is reflected in the sensitivity factors shown in Fig. 4(a) for the first three gates that span the initial rise of a typical DTOF. The contribution from the brain is expected to be $<3\%$ for all three gates assuming the distance to the brain is 14 mm.²⁰ To evaluate the feasibility of using an early gate as the regressor, this study focused on the two metrics that provided the greatest depth sensitivity (the variance and the 250-ps gate positioned at the end of the DTOF, both measured at $r_{\mathrm{SD}}=3\mathrm{cm}$) to act as the dependent variable in the regression analysis. In agreement with the results involving the short-separation channel, applying regression to signals obtained from the same DTOF substantially reduced the residue signal in the hypercapnia experiments [Figs. 6(b)–6(c)]. For both the late gate and $\mathrm{\Delta }{V}_{3\mathrm{cm}}$, $\tau$ for $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ was significantly lower compared to the values derived from each metric on its own, and the regression also reduced inter-subject variability (Table 2). Regression appeared to work better for $\mathrm{\Delta }{V}_{3\mathrm{cm}}$ compared to the late gate (i.e., $\tau =20\pm 18\mathrm{s}$ versus $37\pm 50\mathrm{s}$) likely due to the greater depth sensitivity of the former, but this difference did not reach statistical significance. Regression did not significantly improve the time constant for the corresponding $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses, which likely reflects the lower residue signal compared to $\mathrm{\Delta }{C}_{\mathrm{HbO}}$. In general, the average $\tau$ values following regression for both hemoglobin signals were within the expected range reported in fMRI studies.^33,35

No significant differences were found between fitting parameter estimates ($\tau$ and ssCVR) for regression involving the three early gates, suggesting that any of the gates, or some combination, would be adequate to act as the regressor. The lack of any difference indicates that selection of the time window for the early gate is not overly critical, provided it is predominately sensitive to the ECL. There may be incidences in which it would be prudent to select the earliest gate possible due to variations in ECL thickness between individuals and across the head.³⁶ In this study, the average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses from gate 1 appeared to exhibit greater variability compared to those generated from gates 2 and 3 [Fig. 6(a)]. This may have been caused by light contamination related to differences in probe-to-skin contact but it did not influence the overall fitting results. Interestingly, the average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses extracted from the early gates were not the same as those obtained from $\mathrm{\Delta }{N}_{1\mathrm{cm}}$. All three $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses from the early gates at $r_{\mathrm{SD}}=3\mathrm{cm}$ increased in response to hypercapnia, whereas the corresponding $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ decreased during the same period. One possible explanation is that $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ contained some brain signal considering the corresponding average $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time course exhibited a decrease near the onset of hypercapnia similar to that observed in the time courses with greater sensitivity to the brain. To investigate this possibility, time series were extracted for the first 250 ps of the $\mathrm{\Delta }{N}_{1\mathrm{cm}}$ data, and these resulted in similar $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses as shown in Fig. 5(a) (data not presented). A most likely explanation would be regional variations in scalp hemodynamics that led to different deoxyhemoglobin contributions measured by the two detectors.

The regression approach was also applied to previously acquired fNIRS data involving a motor imagery task. The average $\mathrm{\Delta }{C}_{\mathrm{HbO}}$ and $\mathrm{\Delta }{C}_{\mathrm{Hb}}$ time courses were improved by applying the regression approach as reflected by the removal of a slow frequency component observed in Figure 6 and the improved ${\chi }^2$ goodness of fit between the experimental data and the GLM, although this improvement did not reach statistical significance. One of the challenges with assessing the performance of the regression approach on activation data is that the magnitude of scalp signal changes likely varies considerably between subjects. This is in contrast to the hypercapnic challenge that was both global (i.e., can cause hemodynamic responses in multiple tissues including scalp) and fairly robust across participants due to the relatively large increase in ${\mathrm{P}}_{\mathrm{ET}}{\mathrm{CO}}_2$ ($14.2\pm 0.6\mathrm{mmHg}$).²⁰ It would be useful to apply the regression approach to larger functional trNIRS data sets to further confirm its value for improving the sensitivity to the brain. Lastly, this study implemented the linear least-squares method first proposed by Saager and Berger, and the technique would likely benefit from refinements such as incorporating adaptive filters.³

5.

Conclusions

In summary, this study demonstrated that applying regression analysis to trNIRS metrics with different depth sensitivities improved the characterization of dynamic cerebrovascular reactivity and the oxygenation responses during an activation task. The unique ability of trNIRS to extract the necessary information regarding the ECL from the same data set used to detect changes in cerebral oxygenation ensures the regressor truly reflects the ECL contribution to each probe. The challenge with trNIRS is the additional cost and complexity of the equipment—in this study, all measurements were acquired with a four-channel system. However, this bottleneck is likely to be overcome with the development of miniaturized laser sources and compact silicon photomultiplier devices,^16,37,38 which will enable multi-channel trNIRS systems to rival current CW NIRS devices.