Effects of Processing Methods on fNIRS Signals Assessed During Active Walking Tasks in Older Adults

Abstract

Functional near infrared spectroscopy (fNIRS) is a noninvasive optics-based neuroimaging modality successfully applied to real-life settings. The technology uses light in the near infrared range (650–950nm) to track changes in oxygenated (HbO2) and deoxygenated hemoglobin (Hb) obtained from measured light intensity using light-tissue interaction principles. fNIRS data processing involves artifact removal and hemodynamic signal conversion using modified Beer-Lambert law (MBLL) to obtain Hb and HbO2, reliably. fNIRS signals can get contaminated by various noise sources of physiological and non-physiological origins. Various algorithms have been proposed for the elimination of artifacts from frequency selective filters to blind source separation methods. Hemodynamic signal extraction using raw intensity measurements at multiple wavelengths based on MBLL usually involves apriori knowledge of certain conversion parameters such as molar extinction coefficients (ε) and differential path length factor (DPF). Different sets of conversion parameters dependent upon wavelength, chromophores, and age have been reported. Variation in processing algorithms and parameters can cause differences in Hb and HbO2 extraction which can in turn change study outcomes. Using fNIRS, we have previously shown significant increases in oxygenation in the prefrontal cortex from Single-Task-Walking (STW) to Dual-task-Walking (DTW) conditions in older adults due to greater cognitive demands inherent in the latter condition. In the current study, we re-analyzed our data and determined that although using different conversion parameters i.e. ε and age dependent DPF and filter cut-off frequencies at 0.14 and 0.08Hz including those designed to remove confounding effects of Mayer waves had caused some linear increases or decreases on the extracted Hb and HbO2 signals, those effects were minimal in task related comparisons and hence, the overall study outcomes.

I. Introduction

Functional near infrared spectroscopy (fNIRS) is a noninvasive brain imaging method that is utilized in various research and clinical applications providing information on cognitive functioning in attention, memory, language, motor, sensorimotor and executive function domains under more realistic, everyday settings [1–14]. This is partially due to the fact that the technology can lend itself to safe, portable, comfortable, affordable, easy to engineer and apply designs allowing brain monitoring in various conditions, including Single-Task-Walking (STW) and Dual-Task-Walking; the latter task imposes greater demands on the attention system and requires increased involvement of the prefrontal cortex [12–21]. fNIRS is an optics-based method where tissue is irradiated by constant amplitude light in the near infrared (NIR) range (650–950nm) at two or more wavelengths through the skin [7,22]. The back scattered light from the tissue is captured by light detectors where the change in its attenuation carries information on the concentration of the absorbers within the tissue during the monitoring time. Since main absorbers in the tissue within the NIR range are oxygenated (HbO2) and deoxygenated hemoglobin (Hb), fNIRS can provide information on hemodynamic changes related to cortical activity based on neurovascular coupling.

In order to obtain hemodynamic signals from the raw intensity measurements efficiently and reliably, several processing stages have to be performed based on physics principles of light-tissue interaction, signal and noise characteristics and apriori knowledge or assumption on various conversion parameters. Most fNIRS system applications use modified Beer-Lambert law (MBLL) to convert the intensity measurements to hemodynamic changes in Hb and HbO2 which requires apriori knowledge of various parameters such as molar extinction coefficient, ε, and path length of light, L, within the tissue [11,22–24]. Different groups of researchers have previously studied and published values of ε for different tissue chromophores, such as Hb, HbO2 and water at different wavelengths using experimental methods [24–29]. These existing, published ε values differ from each other slightly which can result in different values of Hb and HbO2 separation [30–32]. Such differences can affect repeatability and reproducibility of the fNIRS studies if performed using different ε values which can in turn affect overall study outcomes depending on the task and fNIRS features implemented. However, even though other parameters required in hemodynamic conversion such as optical path length related ones are recently being mentioned in fNIRS publications, most of the time ε values used are overlooked [2,9].

Similarly, optical path length, L, is required to convert intensity measurements to Hb and HbO2 which is approximated as having a linear relationship with light source and detector distance, d by a coefficient called differential pathlength factor (DPF). As previously studied by many researchers [33–39], DPF has been shown to be dependent upon wavelength and the thickness and optical properties of the head layers, therefore can change with age and head location from frontal to temporal, etc. Many researchers including us have used a constant DPF value for all wavelengths and ages before, which was approximated for adult human head as DPF=6 [2,9]. Others, have used wavelength dependent DPF values but taken as the same for all adult ages even though it can differ within the studied population [40–42]. Recently, researchers have started using wavelength and age dependent DPF values using the formula provided in [2,9,43,44] obtained by fitting the previously published DPF values to a nonlinear equation. Though, the formula is suggested for the ages up to 70 and for frontal head regions and may still not fully reflect individual optical and geometric differences of the head.

A very important aspect in fNIRS data processing requires a reliable separation of task-related cortical hemodynamic signals from possible systemic activity confounders that may exist in Hb and HbO2 measurements. Since the technology measures changes in hemodynamics, extracted Hb and HbO2 carries information not only due to cortical activity in relation to neurovascular coupling but also originating from systemic signals such as heart pulsation, respiration and blood pressure. The frequency content of these systemic signals lies within ~1Hz for heart pulsation, ~0.3Hz for respiration signal and ~0.1Hz for blood pressure related oscillations called Mayer waves [45–52]. Mayer waves are related to oscillations in the sympathetic vasomotor tone of arterial blood vessels and hence sometimes referred to as vasomotion [53]. Since cognitive activity related hemodynamic signals take ~10seconds to evolve [11,47], elimination of Mayer waves poses a challenge due to frequency overlapping whereas heart pulsation and respiration can easily be eliminated using frequency selective filters. Furthermore, since Mayer waves are associated with blood pressure, postural changes such as standing or walking under single or dual-task conditions can increase their spectral power [49,54]. Mayer waves appear both in raw intensity measurements and also in the converted hemodynamic signals around similar frequency ranges. However, since intrinsically the origin of such oscillations is related to hemodynamic variables, it was suggested that the algorithms to eliminate such interferences is best performed on hemodynamic signals [48].

Even though frequency content of Mayer waves can overlap with cognitive activity related hemodynamics, low-pass filters have been widely used in fNIRS applications for their removal in many studies either applied on the raw intensity measurements or after Hb and HbO2 conversion [9–12,42,55]. Several other advanced techniques have also been proposed such as average waveform subtraction, principal or independent component analysis, wavelet filtering and filters using short source-detector separation channels as regressors and independent measures of the systemic fluctuations [2,9,.45–48,56–64]. Specifically, if the study design is an event-related one and stimulus related evoked hemodynamic response function (HRF) is of interest, it can easily get buried under Mayer waves generating false negatives or false positives [60]. In such cases usually other processing methods based on use of multi distance fNIRS or other separate sensor recordings with blind source separation or adaptive filtering techniques were suggested [58–64]. In addition, if the task is a block design of enough time length where usually mean value of the hemodynamic signals within each task block is used as a feature to compare task conditions, such operations can further help in the reduction of Mayer wave oscillations within the recordings.

The current study was specifically designed to investigate the effects of different fNIRS conversion parameters and processing methods on Hb and HbO2 signals associated with cognitive activity assessed during STW and DTW conditions in older adults. In this study, we investigated the effects of these aforementioned conversion parameters and Mayer wave oscillations on a previously collected, analyzed and published fNIRS data set in terms of changes in Hb and HbO2 signal values and overall study’s statistical outcomes [12]. fNIRS data was collected from the prefrontal cortex (PFC) of an elderly population while they were performing a single (STW) and dual task walking (DTW) paradigm. The conversion parameters for ε were adopted from Prahl [24] and the DPF value was selected as being equal to 6 for all wavelengths and ages. We have used a low-pass filter with cut-off frequency of 0.14Hz to eliminate higher frequency oscillations such as respiration on the raw intensity measurements from the cognitive activity related hemodynamic signals. Furthermore, since it was a block design and the feature used in statistical comparisons of STW and DTW (walking while performing a cognitive task) was the mean value of the HbO2 and Hb blocks, separately, we assumed that the effect of Mayer waves would be reduced. Here, we investigated the effects of using i) other published values of ε from Cope [25] and Ziljstra et al. [26], ii) age and wavelength dependent DPF [44] and iii) low-pass filters of lower cut-off frequency to eliminate Mayer waves before and after hemodynamic conversion on the values of Hb and HbO2 [48] and overall statistical comparisons based on linear mixed effects model (LMEM) between task conditions. We compared these three effects separately and altogether with our earlier findings. Our results suggest that for our selected task, subject population and used features, none of these tested values or algorithms applied separately or all together have changed the overall statistical outcomes: DTW was significantly higher in HbO2 as compared to STW. Though there were very minor changes in the extracted Hb and HbO2 values in some cases as compared to earlier values, the differences were along the same direction. resulting in very high significant correlation with earlier values.

The paper is organized as follows, in the next section we will provide background on fNIRS processing such as MBLL conversion and prior study specifics. In section III, we will explain the separate and combined tests we have used to evaluate the effects of ε, DPF and Mayer waves on hemodynamics together with the limitations of the study. Then in section IV, results of these separate and combined tests will be presented together with discussions and the limitations of the study. Finally, we will conclude the study in section V.

II. Background

A. fNIRS Measurements and Hemodynamic Conversion

Typically, fNIRS systems contain a wearable sensor piece to be placed on the skin surface holding the light sources with selected center wavelengths within the near infrared range (650–950nm) to shine light at constant amplitude to the tissue medium to be monitored and light sources at a certain distance, d, to the light sources to detect light after it has interacted with the tissue. Variation in the detected light intensity carries information on the changes in the tissue chromophore concentrations where the main absorbers within the NIR wavelength range are oxygenated (HbO2) and deoxygenated hemoglobin (Hb). This relationship is expressed using modified Beer-Lambert law (MBLL) [11,25] as follows:

$$I^{\lambda }=I_o^{\lambda }{10}^{-\left({\epsilon }_{HbO2}^{\lambda }{C}_{HbO2}+{\epsilon }_{Hb}^{\lambda }{C}_{Hb}\right)L^{\lambda }+G^{\lambda }}$$ (1)

where I^λ is the measured and $I_o^{\lambda }$ is the incident light intensity at wavelength λ, C_HbO2 and C_Hb are the concentrations of the main chromophores within the near infrared range, HbO2 and Hb, respectively. The parameters ${\epsilon }_{HbO2}^{\lambda }$ and ${\epsilon }_{Hb}^{\lambda }$ are the wavelength and chromophore dependent molar extinction coefficients for HbO2 and Hb as given in Figure 1 using values provided in [24] and L^λ is the wavelength dependent path length of the light within the tissue which is related to the source-detector separation, d by the wavelength and age dependent DPF, as L^λ = d DPF^λ. The wavelength dependent factor G^λ is related to scattering and sensor and tissue geometry, assumed to be constant throughout the recording session.

Fig. 1.

Molar extinction coefficients in the NIR range

From the intensity measurements, change in wavelength dependent optical density (ΔOD^λ) is obtained by dividing the intensity measurement, I^λ to its value, $I_b^{\lambda }$, at a certain time point, selected usually during a baseline condition e.g. resting or a relaxation period. Use of ΔOD^λ reduces the need for the unknowns of incident light intensity, $I_o^{\lambda }$ and scattering parameter G^λ but provides a measure related to changes in hemoglobin concentrations (ΔC_HbO2 and ΔC_Hb) throughout the recording session relative to the selected baseline region [25] as follows:

$$\Delta \mathit{\text{O}}{D}^{\lambda }=-lo{g}_{10}\left(\frac{I^{\lambda }}{I_b^{\lambda }}\right)=\left({\epsilon }_{HbO2}^{\lambda }\Delta C_{HbO2}+{\epsilon }_{Hb}^{\lambda }\Delta C_{Hb}\right)L^{\lambda }$$ (2)

In fNIRS applications usually, measurements at two wavelengths (λ1 and λ2) are selected as one below and one above the isosbestic point (~800nm shown with a circle in Figure 1) to focus light absorption more to Hb or HbO2, respectively for their spectroscopic separation. Then, by using the ΔOD^λ values at those two wavelengths as given in equation (2), changes in hemoglobin concentrations (ΔC_HbO2 and ΔC_Hb) can be found by simple matrix inversion as follows:

$$\left[\begin{array}{c}\Delta \mathit{\text{O}}{D}^{\lambda 1}\\ \Delta \mathit{\text{O}}{D}^{\lambda 2}\end{array}\right]=\left[\begin{array}{cc}{\epsilon }_{HbO2}^{\lambda 1}dDP{F}^{\lambda 1}& {\epsilon }_{Hb}^{\lambda 1}dDP{F}^{\lambda 1}\\ {\epsilon }_{HbO2}^{\lambda 2}dDP{F}^{\lambda 2}& {\epsilon }_{Hb}^{\lambda 2}dDP{F}^{\lambda 2}\end{array}\right]\left[\begin{array}{c}\Delta C_{HbO2}\\ \Delta C_{Hb}\end{array}\right]$$ (3)

Notice that, matrix inversion and hence extraction of changes in hemoglobin concentrations is possible only with apriori knowledge of conversion parameters, the wavelength and chromophore dependent molar extinction coefficients (${\epsilon }_{HbO2}^{\lambda 1}$, ${\epsilon }_{Hb}^{\lambda 1}$, ${\epsilon }_{HbO2}^{\lambda 2}$, and ${\epsilon }_{Hb}^{\lambda 2}$) and the wavelength and age dependent DPF (DPF^λ1 and DPF^λ1).

B. Prior Study

1). Participants

Participants (n=83) were community-residing older adults (age= 78.05±6.37), enrolled in “Central Control of Mobility in Aging” (CCMA), who underwent the combined dual-task walking fNIRS protocol [12]. Exclusion criteria were: dementia, current or history of severe neurological or psychiatric disorders, inability to ambulate independently, significant loss of vision and/or hearing, and recent or anticipated medical procedures that may affect ambulation. Demographics on participants such as age, gender and general health was also collected as provided in [12]. Written informed consents were obtained in-person and approved by the IRB.

2). Dual-Task Walking Protocol

In the original protocol [12], there were two single task conditions: 1) STW and 2) Cognitive (Alpha). In STW participants were asked to walk around the electronic walkway at their “normal pace” for three consecutive loops. In Alpha, participants were required to stand still while reciting alternate letters of the alphabet (e.g., A C E…) for 30-sec out loud. In DTW participants were instructed to walk around the walkway for three consecutive loops at their normal pace while reciting alternate letters of the alphabet. Participants were instructed to pay equal attention to both tasks to minimize task prioritization effects. The results confirmed the increased involvement of the prefrontal cortex in DTW compared to STW and also demonstrated that prefrontal cortex activation efficiency could be improved after practice [12]. To reduce complexity and the number of comparisons in the current study we used the baseline Hb and HBO2 in STW and DTW conditions as outcomes for comparison between conversion parameters and filter cut-off frequencies.

3). fNIRS System

We have used fNIRS Imager 1100 (fNIRS Devices, LLC, Potomac, MD) which collects data from the PFC at a sampling rate of 2Hz and equipped with standard analog, anti-aliasing filters [65]. The fNIRS sensor consists of 4 LED light sources and 10 photodetectors, which cover the forehead using 16 voxels, with a source-detector separation of 2.5 cm. The light sources on the sensor (Epitex Inc. type L4X730/4X805/4X850–40Q96-I) contain three built-in LEDs having peak wavelengths at 730, 805, and 850 nm. The photodetectors (Bur Brown, type OPT101) are monolithic photodiodes with a single supply transimpedance amplifier. We implemented a standard sensor placement procedure as previously described [2,15].

4). Processing and Parameters Used

Raw data at 730 and 850nm wavelengths were inspected for excessive noise, saturation or dark current conditions. To eliminate possible respiration, heart rate signals and unwanted high frequency noise raw intensity measurements at 730 and 850nm were low-pass filtered with a finite impulse response filter of cut-off frequency at 0.14Hz [12]. Saturation or dark current conditions were excluded. Changes in HbO2 and Hb concentrations relative to separate proximal 10-second baselines were calculated from the artifact-removed raw intensity measurements at λ1=730 and λ2=850 nm using MBLL with DPF=6 for all wavelengths and ages with molar extinction coefficients adapted from [24] as:

$$\begin{gathered} {\epsilon }_{HbO2}^{730}=0.390,{\epsilon }_{Hb}^{730}=1.1022 \\ {\epsilon }_{HbO2}^{850}=1.058,{\epsilon }_{Hb}^{850}=0.69132 \end{gathered}$$ (4)

Individual mean HbO2 and Hb data were extracted separately for each experimental condition (STW and DTW) using all available data from the 16 channels of the first trial.

III. Tests and Parameters

Here we will explain, the tests we have carried out to evaluate the effects of conversion parameters and use of Mayer wave elimination techniques on HbO2 and Hb conversion, separately and in combination. All the tests were carried out using custom codes built in Matlab 2017a (Mathworks, Inc. Natick, MA).

A. Test 1: Molar Extinction Coefficient, ε Value Comparison

There are various studies in the literature carried out for the extraction of molar extinction coefficients at various wavelengths and for different tissue chromophores using experimental methods [24–29]. Some of those are presented in tabular form and the others as plots. There are very few studies that compared the effects of these ε values on the Hb and HbO2 separation [30–32] where it has been mentioned that even very small discrepancies in ε values from different sources can result in 5–25% relative errors in quantification of Hb and HbO2 [31].

Here we tested the effects of using different ε values from 1a) Prahl (prior study values [24]), 1b) Cope [25] and 1c) Ziljstra et al. [26] on hemodynamic conversion while keeping DPF=6 for all wavelengths and ages on the filtered raw intensity measurements with a low-pass filter of cut-off frequency fc=0.14Hz as used in the prior study. The parameters and algorithms used are summarized in Table I where test 1a corresponds exactly to the prior study values.

TABLE I:

Test 1 Values

Test	ε values	DPF	Filter
1a	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.14Hz before
1b	${\epsilon }_{HbO2}^{730}=0.4383, {\epsilon }_{Hb}^{730}=1.3029$ ${\epsilon }_{HbO2}^{850}=1.1596, {\epsilon }_{Hb}^{850}=0.7861$	6 for all ages and λ	Low-pass filter fc=0.14Hz before
1c	${\epsilon }_{HbO2}^{730}=0.508, {\epsilon }_{Hb}^{730}=1.308$ ${\epsilon }_{HbO2}^{850}=1.012, {\epsilon }_{Hb}^{850}=0.768$	6 for all ages and λ	Low-pass filter fc=0.14Hz before

B. Test 2: Age and Wavelength Dependent DPF Evaluation

It is necessary to know the DPF in order to convert raw intensity measurements to hemodynamic changes where use of wrong values can cause cross-talk between HbO2 and Hb [22,27,34,43,44]. It has been shown before that DPF depends on the source-detector separation, d, and optical properties of the medium which makes it intrinsically dependent on wavelength also. Yet, many researchers, like us have used a constant value of DPF=6 for all wavelengths before [9]. Duncan et al [43] showed that DPF also depends on age where this dependence was formulated as a power law function with parameters given for four selected wavelengths. Following this work and using previously published values, Scholkmann et al. [44] derived a nonlinear formula for age and wavelength dependent DPF where age dependence was taken as a power law function and wavelength dependence was modeled as a cubic equation as follows:

$$DPF\left(\lambda ,A\right)=\alpha +\beta A^{\gamma }+\delta {\lambda }^3+\eta {\lambda }^2+\zeta \lambda$$ (5)

where λ is the wavelength and A is the age. After using the previously published data sets and robust nonlinear least squares fitting with the least absolute residuals method, the parameters in the formula were found as: α = 223.3,

$$\beta =0.05624,\gamma =0.8493,\delta =-5.723\times 10-7,\eta =0.001245,$$

and ζ = −0.9025. It was mentioned in [44] that since these values were extracted using previously published values obtained for populations of age between 0 and 70 and collected from frontal and frontotemporal head regions, they may be appropriate to use only for those ages and measurement locations.

Using the formula in equation (5), we have tested the effect of wavelength and age dependent DPF on our measurements. However, even though our measurements were collected from the PFC and hence matching the head location suggestions for the use of formula in [44], age-wise, our study population were out of the suggested ranges (all of the participants had ages >70). Therefore, we have tested the age and wavelength dependent DPF on hemodynamic conversion in two ways, once by using the formula in (5) for age at the highest suggested limit A=70 and next using the actual ages of the participants in equation (5) even though they are outside the suggested limits. In all tests, we have obtained DPF for wavelengths at λ1=730 and λ2=850 nm as it was implemented in the fNIRS system used. The tests and values are summarized in Table II. Note that test 2a corresponds to our prior study values. Again, we kept the other parameters like ε values and filters for Mayer wave elimination in the new tests the same as in the original study before.

TABLE II:

Test 2 Values

Test	ε values	DPF	Filter
2a	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.14Hz before
2b	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	DPF(λ, A) for A=70	Low-pass filter fc=0.14Hz before
2c	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	DPF(λ, A) A=actual age	Low-pass filter fc=0.14Hz before

C. Test 3: Mayer Wave Elimination

Mayer waves can exist in fNIRS recordings as spontaneous oscillations of ~0.1Hz which can overlap with the frequency content of cognitive activity related hemodynamic changes causing misinterpretation in the outcomes. Nevertheless, in many fNIRS applications, Mayer waves are suppressed using frequency selective filters with cut-off frequency ~0.08Hz to take into account filter imperfections and possible variations in Mayer wave characteristics, even though it may suppress some of the signal content too [9–12,42,55], In those prior studies, filters were applied before or after the hemodynamic conversion, respectively since Mayer waves can appear around the same frequency ranges both on raw intensity measurements or hemodynamic signals. It is also suggested in [48] that such systemic artifacts e.g. Mayer waves are best eliminated after hemodynamic conversion since physiologically they are additive to Hb and HbO2 signals.

Previously, the low-pass filter we have used in our analysis was mainly to suppress possible respiration, heart pulsation and high frequency noise with a cut-off frequency, fc=0.14Hz and it was applied on raw intensity measurements. Since we used a block design as our protocol (continuous walking in STW and DTW) and mean value of hemodynamics in each block was selected as the feature in statistical comparisons, we have assumed that Mayer waves would also be suppressed during the time averaging process. Here, we applied a low-pass filter with a cut-off frequency fc=0.08Hz as in other fNIRS studies on the raw intensity measurements and on the hemodynamic signals in separate tests. We implemented a finite impulse response (FIR) low-pass filter of 100 tap-length using the windowing method with Hamming window corresponding to the command ‘fir1’ in Matlab. It provided −6dB attenuation (half the passband gain) at fc=0.08Hz and −22dB attenuation at f=0.1Hz. The parameters used in these tests are as summarized in Table III. We have compared these filtering outcomes applied before (Test 3b) and after (Test 3c) hemodynamic conversion with each other and with our original findings (Test 3a).

TABLE III:

Test 3 Values

Test	ε values	DPF	Mayer wave
3a	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.14Hz before
3b	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.08Hz before
3c	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.08Hz after

D. Test 4: Combined Effects Comparison

Prior tests are designed to evaluate the effects of conversion parameters and Mayer wave filters on hemodynamics and overall statistics in comparison to our earlier findings, separately. Next, we evaluated the combined effects of select measures such as the ones generating the most difference compared to prior outcomes as summarized in Table IV. Hence, we used once the earlier parameters (Test 4a), then the same molar extinction coefficients from Prahl with wavelength and age (using the actual age of the participants) dependent DPF and Mayer wave filter of fc=0.08Hz applied after conversion (Test 4b) and finally, the molar extinction coefficients from Cope [25], with age and wavelength dependent DPF with the same Mayer filter as in the second case (Test 4c). Note that, in this test, the combined effects of the processing methods were evaluated not only on the mean feature but also on an additional feature, the slope of the data in STW and DTW blocks.

TABLE IV:

Test 4 Values

Test	ε values	DPF	Filter
4a	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	6 for all ages and λ	Low-pass filter fc=0.14Hz before
4b	${\epsilon }_{HbO2}^{730}=0.390, {\epsilon }_{Hb}^{730}=1.1022$ ${\epsilon }_{HbO2}^{850}=1.058, {\epsilon }_{Hb}^{850}=0.69132$	DPF(λ, A) A=actual age	Low-pass filter fc=0.08Hz after
4c	${\epsilon }_{HbO2}^{730}=0.4383, {\epsilon }_{Hb}^{730}=1.3029$ ${\epsilon }_{HbO2}^{850}=1.1596, {\epsilon }_{Hb}^{850}=0.7861$	DPF(λ, A) A=actual age	Low-pass filter fc=0.08Hz after

E. Hemodynamic and Statistical Comparisons

In all the tests from 1 through 4, using the given parameters and methods explained in Tables I through IV, we obtained the HbO2 and Hb signals for all available conditions (STW and DTW), channels and subjects. Then we extracted the feature, mean value for HbO2 and Hb blocks and additionally the slope in Test 4, separately for all conditions, channels and subjects as we have used before in our prior analysis [12]. Using these values, we compared the prior findings with the evaluated test results in terms of the descriptive values (mean±standard deviation) of HbO2 and Hb, correlation with prior values and overall statistical comparison of STW and DTW conditions.

Using all participant, channel and condition values Pearson correlation coefficient R with its p value between prior study and the individual test carried out for HbO2 and Hb values was found. As for the statistical comparison between STW and DTW conditions we have used before Linear Mixed Effects Model (LMEM) to compare task conditions (STW and DTW). Here, we used the same statistical test and compared the outcomes with the prior study results and the newly applied test findings.

IV. Results and Discussion

A. Test 1: Molar Extinction Coefficient Value Comparison

The results of Test 1a, 1b and 1c to investigate the effects of molar extinction coefficients, ε, on HbO2 and Hb outcomes in different tasks are summarized in Table V. In terms of descriptive values, Test 1b which is the results obtained by using ε values as provided by Cope [25] generated lower HbO2 and higher Hb as compared to the original results obtained using Prahl’s values [24] in both STW and DTW tasks on the average. Notably, the trend in descriptive values for the task conditions such as the increase in HbO2 and the decrease in Hb from STW to DTW was still intact even though Cope’s values had reduced the reverse relationship between HbO2 and Hb. Whereas Test 1c values obtained by using ε values of Ziljstra et al. [26] resulted in increase in HbO2 and further decrease in Hb in both tasks where the reverse relationship between HbO2 and Hb was magnified. Again, the trend between the task conditions was clearly present even improved specifically in Hb. Nevertheless, the correlations between Test 1b and Test 1a (the original) and Test 1c and Test 1a was highly significant with Pearson correlation coefficient being R≈0.99 in all tasks suggesting that the difference in HbO2 and Hb values due to the use of different molar extinction coefficient sets in hemodynamic conversion was along the same directions.

TABLE V:

Test 1- Descriptive Values and Correlation Results

		Test 1a	Test 1b	Test 1c
Descriptive Values	HbO2-STW	0.24±1.05	0.21±1.07	0.26±1.24
	HbO2-DTW	1.11±1.61	1.01±1.49	1.26±1.89
	Hb-STW	−0.21±1.09	−0.17±0.91	−0.21±0.89
	Hb-DTW	−0.43±1.31	−0.37±1.09	−0.52±1.04
Correlation with Test1a	HbO2-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	HbO2-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-STW		R=0.99,p<0.0001	R=0.98,p<0.0001
	Hb-DTW		R=0.99,p<0.0001	R=0.97,p<0.0001
Test 1 - LMEM Task Effect Results (STW vs DTW)
STW vs DTW	Estimate	SE	95%CI	P
Test 1a-HbO2	−0.8691	0.0480	−0.9634 to −0.7749	<0.0001
Test 1b-HbO2	−0.7921	0.0444	−0.8793 to −0.7048	<0.0001
Test 1c-HbO2	−0.9954	0.0565	−1.1063 to −0.8846	<0.0001
Test 1a-Hb	0.2280	0.0469	0.1360 to 0.3201	<0.0001
Test 1b-Hb	0.1992	0.0391	0.1225 to 0.2759	<0.0001
Test 1c-Hb	0.3196	0.0375	0.2460 to 0.3932	<0.0001

The LMEM results on the task effect for HbO2 indicated that DTW was significantly higher than STW and for Hb, it resulted in significantly lower values in all Tests. It may be due to the fact that, the cognitive effort required to perform STW and DTW had a large enough contrast in hemodynamic responses on the PFC that the overall difference between tasks was still highly significantly different in all Tests even though there were subtle differences in the descriptive values between Tests. In [31], the percent error between HbO2 and Hb values obtained by using ε values by Prahl, Cope and Zijlstra et al., [24–26] in comparison was reported to be between 5–25% where similar differences can be obtained using our descriptive values. Note that, such differences in Hb and HbO2 conversion with the use of different ε value sets can be more pronounced when event related task designs with single trial analysis outcomes (e.g. in brain computer interface studies [3,41]) are of interest or when the cognitive contrast between task conditions or subject groups are rather small. However, in our block design study with the use of mean value as a feature and a task generating large enough cognitive contrast, our overall results in the statistical analysis are not affected in the worst case (Test 1b) and even improved in the other (Test1c). Similarly, if the reverse relationship between HbO2 and Hb is of interest such as in some artifact removal methods [66], then use of different ε value sets becomes more important since they can improve (Test 1c using [26]) or reduce (Test 1b using [25]) this effect.

B. Test 2: Age and Wavelength Dependent DPF Evaluation

The results of Test 2a, 2b and 2c for the use of wavelength and age dependent DPF in comparison to prior study outcomes are summarized in Table VI. In these tests for comparison purposes ε values are kept the same as in earlier study adopted from Prahl (Test 2a). Use of actual age (Test 1c) and Age=70 (Test 1b) did not result in any difference in terms of any statistical comparisons. In fact, the difference in HbO2 and Hb values by the use of wavelength and age dependent DPF as compared to using DPF=6 for both wavelengths as in our prior study was along the same lines with very highly significant correlation coefficient, R=0.99 in all. More importantly, the trend in HbO2 and Hb for STW and DTW task conditions was kept as the LMEM statistical results indicated that DTW resulted in significantly higher HbO2 and significantly lower Hb as compared to STW in all DPF tests. Only, the reverse relationship between HbO2 and Hb is reduced in Test 1b and 1c as compared to earlier results. Note that, our study was designed and applied on subject population composed of older adults, age ≥70 and main hypothesis was to investigate task effects. Hence, the age distribution was relatively small and more homogenous. If comparisons are to be made between different age groups involving children, adolescents, adults and elderly, then caution should be taken and appropriate DPF values should be used.

TABLE VI:

Test 2- Descriptive Values and Correlation Results

		Test 2a	Test 2b	Test 2c
Descriptive Values	HbO2-STW	0.24±1.05	0.21±1.07	0.21±1.07
	HbO2-DTW	1.11±1.61	1.05±1.64	1.05±1.64
	Hb-STW	−0.21±1.09	−0.17±0.84	−0.17±0.84
	Hb-DTW	−0.43±1.31	−0.41±0.99	−0.41±0.99
Correlation with Test 2a	HbO2-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	HbO2-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-DTW		R=0.99,p<0.0001	R=0.99, <0.0001
Test 2 - LMEM Task Effect Results (STW vs DTW)
STW vs DTW	Estimate	SE	95%CI	P
Test 2a-HbO2	−0.8691	0.0480	−0.9634 to −0.7749	<0.0001
Test 2b-HbO2	−0.8372	0.0492	−0.9339 to −0.7406	<0.0001
Test 2c-HbO2	−0.8372	0.0492	−0.9339 to −0.7406	<0.0001
Test 2a-Hb	0.2280	0.0469	0.1360 to 0.3201	<0.0001
Test 2b-Hb	0.2316	0.0356	0.1616 to 0.3016	<0.0001
Test 2c-Hb	0.2316	0.0356	0.1616 to 0.3016	<0.0001

C. Test 3: Mayer Wave Elimination

The results of Test 3a, 3b and 3c where low-pass filters with different cut-off frequencies applied before and after hemodynamic conversion are summarized in Table VII below. Here, in these tests we kept ε and DPF values the same as in the prior study (Test 3a). All of the results in terms of descriptive values, correlations and LMEM results were very similar in all tests, hemodynamic parameters and task conditions showing almost no effect. This may be due to the use of a block design in our study with longer task epochs (~ minutes) than event related designs with shorter task lengths (~seconds) and because of the use of mean value as the feature in comparisons.

TABLE VII:

Test 3- Descriptive Values and Correlation Results

		Test 3a	Test 3b	Test 3c
Descriptive Values	HbO2-STW	0.24±1.05	0.25±1.06	0.24±1.05
	HbO2-DTW	1.11±1.61	1.12±1.61	1.11±1.61
	Hb-STW	−0.21±1.09	−0.21±1.10	−0.20±1.10
	Hb-DTW	−0.43±1.31	−0.44±1.31	−0.43±1.31
Correlation with Test 3a	HbO2-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	HbO2-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
Test 3 - LMEM Task Effect Results (STW vs DTW)
STW vs DTW	Estimate	SE	95%CI	P
Test 3a-HbO2	−0.8691	0.0480	−0.9634 to −0.7749	<0.0001
Test 3b-HbO2	−0.8681	0.0482	−0.9626 to −0.7735	<0.0001
Test 3c-HbO2	−0.8697	0.0481	−0.9641 to −0.7753	<0.0001
Test 3a-Hb	0.2280	0.0469	0.1360 to 0.3201	<0.0001
Test 3b-Hb	0.2250	0.0471	0.1325 to 0.3176	<0.0001
Test 3c-Hb	0.2281	0.0470	0.1358 to 0.3205	<0.0001

We note that filtering may have different effects on individual time course data points as compared to the mean value used. To illustrate this point, we present an example of single fNIRS channel recording from a participant in Figure 2 during the time course of STW (top plot) and DTW (bottom plot). In all the plots, solid line is the unfiltered (NF) original data, dashed line is Test3a result, dotted line is Test 3b output and dashed-dot line is Test 3c outcome. In the legends of the plots the mean value (m) of the overall time course data is also presented. Even though, we present here an example of a rather high noise case from our data set, all of the algorithms smoothed the data satisfactorily. In addition, although there may be some cyclic signal ~0.1Hz range still appearing in the filtered signal specifically for Test 3a output as compared to much more smoothed Test 3b and 3c results due to lower cut-off frequency, the difference in the selected feature (the mean value of the time course), between all the filters used is minimal and around two orders of magnitude smaller than the actual mean value of the signal in both STW and DTW cases. It can also be seen that on the mean value DTW generates more brain activity (increase in HbO2) in the PFC as compared to STW, as it was found highly significant in all filtering test outcomes (Table VII).

Fig.2:

Example subject and channel HbO2 data before and after various filters for STW and DTW conditions (m=mean value of HbO2)

Prior studies in the literature such as [45] and [46] have discussed significant degradation in fNIRS outcomes due to the confounding effects of Mayer waves whereas our results based on the mean value of HbO2 and Hb parameters during STW and DTW tasks did not show any difference with the application of different noise removal methods for Mayer wave elimination. In [45], the data was generated partly as a simulation where synthetic evoked hemodynamic response functions (HRF) were added to a real resting state fNIRS data containing Mayer waves at different signal to noise ratios (SNR) in order to identify the contribution of Mayer waves to evoked HRF extraction. As expected low SNR (high power Mayer waves) had a strong effect on the degradation of single trial, evoked HRFs. The effect of Mayer waves was not as pronounced when SNR was high (low power Mayer waves). Our study was based on a block design where continuous performance of the task could have produced more pronounced and hence higher levels of activations as compared to evoked hemodynamic responses (HRFs) as in [45] which could have improved the SNR in our study (causing low Mayer waves power as compared to cortical activity signals). In [46], the task was a block design where fNIRS data was collected simultaneously with other sensor measurements providing heart rate, blood pressure, scalp blood flow, and respiration to be used in a wavelet coherence based denoising algorithm to eliminate physiological noise including Mayer waves from fNIRS recordings. Their results indicated that sensitivity of fNIRS signal to task related cortical activities were increased and became statistically significant after physiological noise removal which was not the case before noise elimination. Here, our results using the previously implemented filters (Test 3a) revealed significant differences in HbO2 levels in STW compared to DTW. Furthermore, these task-related changes in the fNIRS-derived hemodynamic response were not affected by the use of different filter cut-offs applied before or after hemodynamic conversion as in Test 3b and 3c suggesting that the significance found in Test 3a was not diminished by Mayer waves effects for the selected task and population. Again, it could be due to the fact that STW and DTW tasks generated enough contrast resulting in robust SNR where the power of Mayer waves was low as compared to cortical signal levels. Nevertheless, since we did not collect other separate measurements as in [46], it was not possible for us to directly compare the two studies.

It was also noted in the literature that posture can cause changes in blood pressure and hence the amount of Mayer waves in fNIRS recordings [54,55] which can mask the cognitive activity related changes. Our study was conducted in upright position throughout (standing baselines and walking tasks), hence the presence and amount of Mayer waves is not expected to change from baseline to task conditions which may be one reason that Mayer wave elimination using different methods did not create any difference in the overall result. Hence, it is important to select appropriate baselines for different task conditions to eliminate the possibility that Mayer wave effects, induced by postural changes, will not mask the effects of cognitive activity on HbO2 and Hb signals. Another possible explanation of these test results can be that the averaging operation to extract the HbO2 and Hb features for each subject, task and channel in our prior block design study (Test 3a) may have contributed to the reduction of cyclic Mayer waves that may have been present in the measurements, in turn generating similar results to other methods (Test 3b and c). The Mayer waves effect can be more pronounced in event related designs with single trial applications where features such as maximum value is used. Using a lower cut-off frequency (0.08Hz) and applying the filter before or after hemodynamic conversion for the elimination of Mayer waves should still be tested in future studies with different task protocols and conditions to fully investigate its effects on the outcomes.

D. Test 4: Combined Effects Comparison

In prior tests we have evaluated the effects of different published ε value sets, wavelength and age dependent DPF and various low-pass filters with different cut-off frequencies applied before and after conversion, separately. Even though these factors may not result in significant differences in the overall outcomes when evaluated separately, in combination their effects may reach significance levels. In order to evaluate their combined effects on the study outcomes we tested different combinations in Tests 4a, 4b and 4c where the results are summarized in Table VIII. Use of Cope’s ε values [25] (Test 4c) or the original Prahl’s values [24] (Test4b) together with age and wavelength dependent DPF [44] and applying a filter with cut-off frequency of 0.08Hz after conversion for Mayer wave elimination did not change the overall LMEM statistical comparison between STW and DTW task conditions as compared to the original, prior study results (Test 4a). As can be seen in Table VIII, the change in HbO2 and Hb values in different combined effect tests were in the same direction with very high statistically significant correlations. In general, use of age and wavelength corrected DPF and lower cut-off frequency filter (Test 4b) reduced HbO2 values and increased Hb values making the inverse relationship between them smaller as compared to the original results (Test 4a) which is more pronounced with the use of Cope’s ε values (Test 4c). However, in all cases the contrast between STW and DTW conditions were intact and statistically significant. To evaluate consistency among different test outcomes, intraclass correlations estimates [67] and their 95% confidence intervals with their significance value were calculated as shown in Table VIII. All results showed excellent reliability.

TABLE VIII:

Test 4- Descriptive Values and Correlation Results

		Test 4a	Test 4b	Test 4c
Descriptive Values	HbO2-STW	0.24±1.05	0.21±1.04	0.19±0.96
	HbO2-DTW	1.11±1.61	1.03±1.6	0.93±1.48
	Hb-STW	−0.21±1.09	−0.16±0.82	−0.14±0.68
	Hb-DTW	−0.43±1.31	−0.39±0.96	−0.33±0.80
Correlation with Test 4a	HbO2-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	HbO2-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-STW		R=0.99,p<0.0001	R=0.99,p<0.0001
	Hb-DTW		R=0.99,p<0.0001	R=0.99,p<0.0001
Test 4 - LMEM Task Effect Results (STW vs DTW)
STW vs DTW	Estimate	SE	95%CI	P
Test 4a-HbO2	−0.8691	0.0480	−0.9634 to −0.7749	<0.0001
Test 4b-HbO2	−0.8172	0.0480	−0.9114 to −0.7229	<0.0001
Test 4c-HbO2	−0.7445	0.0444	−0.8316 to −0.6573	<0.0001
Test 4a-Hb	0.2280	0.0469	0.1360 to 0.3201	<0.0001
Test 4b-Hb	0.2230	0.0349	0.1546 to 0.2915	<0.0001
Test 4c-Hb	0.1945	0.0290	0.1375 to 0.2515	<0.0001
Test 4 – Intraclass Correlations and Reliability analysis
Comparison	Variable-Task	Correlation	95%CI	P
Test 4a-4b	HbO2-STW	0.992	0.988 to 0.995	<0.0001
Test 4a-4b	HbO2-DTW	0.996	0.994 to 0.998	<0.0001
Test 4a-4b	Hb-STW	0.959	0.936 to 0.974	<0.0001
Test 4a-4b	Hb-DTW	0.947	0.918 to 0.966	<0.0001
Test 4a-4c	HbO2-STW	0.982	0.972 to 0.989	<0.0001
Test 4a-4c	HbO2-DTW	0.990	0.985 to 0.994	<0.0001
Test 4a-4c	Hb-STW	0.900	0.846 to 0.936	<0.0001
Test 4a-4c	Hb-DTW	0.883	0.821 to 0.925	<0.0001
Test 4b-4c	HbO2-STW	0.996	0.994 to 0.997	<0.0001
Test 4b-4c	HbO2-DTW	0.996	0.994 to 0.998	<0.0001
Test 4b-4c	Hb-STW	0.984	0.975 to 0.990	<0.0001
Test 4b-4c	Hb-DTW	0.984	0.974 to 0.990	<0.0001

In the prior section, specific trends in the time course of STW and DTW conditions (Figure 2) can be observed in their slope, as reported in detail in our prior work [12]. Findings based on Test 4a parameters revealed that the HbO2 slope within DTW trials was positive suggesting increased activity and hence elevated effort during the course of the task. In contrast, the slope of HbO2 during STW trials was negative indicating that the task became easier and required less brain resources due to increased automaticity. Similar findings on the increase in HbO2 data during DTW condition as compared to STW were also found and reported previously in our other data sets [15]. In the context of this work, we analyzed the effects of different processing parameters and algorithms on the slope of STW and DTW trial blocks of the same data set used in this study.

Similar to our prior work [12], using the timing of the walk features obtained from the electronic walkway, we extracted the 6 consecutive straight walk data epochs and used their mean value to obtain the slope of the data within each trial data block of STW and DTW conditions, separately. LMEM results on the slope feature in terms of within task blocks for HbO2 data only (to be concise) extracted for Tests 4a (original findings in [12]), 4b and 4c are summarized in Table IX. These results suggested that using different processing parameters for extinction coefficients, DPF values and filter types for Mayer wave elimination as selected in this study did not change overall statistical outcomes in STW and DTW conditions on HbO2 data using the slope as the outcome feature. This finding is consistent with the prior findings reported herein using the mean values for STW and DTW as the outcome feature.

TABLE IX:

LMEM Results For Within Task Blocks For STW And DTW Conditions Using HbO2 Slope Feature

STW	Estimate	SE	95%CI	P
Test 4a	−0.0578	0.0054	−0.0685 to −0.0471	<0.0001
Test 4b	−0.0508	0.0051	−0.0608 to −0.0408	<0.0001
Test 4c	−0.0451	0.0047	−0.0544 to −0.0359	<0.0001
DTW	Estimate	SE	95%CI	P
Test 4a	0.0341	0.0057	0.0228 to 0.04536	<0.0001
Test 4b	0.0380	0.0055	0.0272 to 0.04886	<0.0001
Test 4c	0.0353	0.0051	0.0253 to 0.0453	<0.0001

E. Limitations of the Study

There are various aspects to consider in fNIRS processing, from model parameters to suppression of different types of artifacts (e.g., motion artifacts, skin blood flow, systemic signals). While different algorithms have been developed to suppress the aforementioned artifacts [45,46,68–70], the current study examined only the effects of systemic artifact removal algorithms on Hb and HbO2 signals assessed during active walking. Future studies should examine scalp blood flow, respiration and head or jaw movement related artifacts that may possibly be generated by talking and/or walking tasks and the utility of different algorithms proposed to remove such artifacts from the extracted cortical hemodynamic signals.

V. Conclusion

In this study, we have evaluated the effects of different conversion parameters such as molar extinction coefficients, ε, and differential path length factor, DPF, and different filter applications for Mayer wave suppression on hemodynamic signal values and statistical outcomes. We have shown that for selected block design type protocol (continuous application of stimulus) and features extracted from the hemodynamic variables such as the mean value and slope, the overall statistical comparisons between conditions were not materially different. Task-related effects on the values and change in HbO2 and Hb among the different conversion parameters and filters were in the same direction and highly consistent. This finding maybe attributed, at least in part, to the following: 1) the task conditions (STW vs. DTW) have excellent reliability and validity as shown in numerous studies [12–19, 71,72]; 2) the task effect is very robust due to much greater cognitive demands that are inherent in DTW compared to STW resulting in higher oxygenation levels in the former compared to the latter task condition; 3) as noted earlier, the block design allows for averaging over many data points per task given the temporal resolution of fNIRS (.5sec). Use of age and wavelength dependent DPF values did not change the overall statistical outcomes in our data set, but it should be noted that our study population was composed of elderly of age >70years and original hypotheses were not aimed to test cognitive activity differences in different age populations. If task effect is to be compared across subject groups of different ages, use of age and wavelength corrected DPF can be more appropriate. Use of different filters before or after hemodynamic conversion with cut-off frequency adjusted to suppress Mayer waves also yielded very minor differences in hemodynamic signals and statistical outcomes as compared to prior findings. It may be due to the fact that this was a block design study and the time averaging within blocks was successfully facilitating the reduction of possible Mayer waves as it was assumed before. The effects of conversion parameters and Mayer wave elimination algorithms can be tested in the future in event related designs where single trial outcomes are of importance, for select hemodynamic features such as maximum value and in data sets where different age groups existed. Nevertheless, it is important in fNIRS studies to report on all the parameters used in hemodynamic conversion including ε and DPF values and other algorithms used in data processing in order to guarantee repeatability and reproducibility of the results.