2.1.

Participant Demographics

This study was approved by the Conjoint Health Research Ethics Board of the University of Calgary. Two groups of participants were recruited for this study. The control group included 30 participants (15M, 15F, age: $35.77\pm 14.77$ years) who had not experienced a concussion in at least the past year. The concussion group included 15 participants (4M, 11F, age: $42.93\pm 15.04$ years, time since injury: $2.6\pm 0.91$ months). Concussion participants were recruited from an outpatient brain injury clinic at an academic hospital seen from June 2020 to March 2022. They were referred to this brain injury clinic by practitioners in the community for treatment of persistent post-concussive symptoms (PPCS); therefore, this concussion sample is highly selective. Concussion was diagnosed by a medical practitioner based on the definition from the American Congress of Rehabilitation Medicine (ACRM).⁷ PPCS was diagnosed by a specialist using the ICD-10 criteria.²³ All concussion participants included in this study were confirmed to have experienced a concussion within the last 6 months (Table 1). Participants were included in the study if they were within 6 months of their concussion diagnosis, were currently experiencing PPCS after their injury, were between the ages of 18 and 65 years, had no pre-existing neurological disorders, and currently were not using psychoactive drugs or medication.

Table 1

Demographics of PPCS participants included in the study.

2.2.

Data Acquisition

We acquired fNIRS deoxygenated (Hb) and oxygenated (HbO) hemoglobin data using the NIRx NIRScoutX (Berlin, Germany) at a sampling rate of 3.90625 Hz. The fNIRS optodes (sources and detectors) were placed $\sim 30$ to 40 mm apart to maximize the signal obtained from the brain.²⁴ The optodes were placed on the scalp according to the EEG 10-20/10-5 system.^25,26 This allowed for accurate placement of the fNIRS channels on the scalp to cover the DLPFC for functional brain mapping.²⁶ Another set of detectors were placed $\sim 8\mathrm{mm}$ from the source optodes to form the short distance channels. The optodes were arranged as shown in Fig. 2. Fiber positions on the DLPFC were estimated using the fNIRS optodes’ location decider v2.2 with the specificity threshold set at “30%.”²⁷ fNIRS optodes were calibrated to ensure a good signal quality using the calibration feature on the NIRx NIRStar software.

2.3.

Task

A resting state task and a paced visual serial addition task²⁸ (PVSAT) were completed for this study. Participants were trained on the task before data collection began. Participants were seated ∼75 cm from a screen that projected images to identify the task to be completed. First, they completed an 8 min rest period with the participant being asked to focus on a cross in the middle of the screen. Next, they completed the PVSAT (Fig. 1). The PVSAT was presented in a block design (12 blocks). Participants were given a target number (9, 10, or 11) before each task block. For each block, they were presented with a single digit number on the screen that was replaced at an inter-stimulus interval of 1 s. Participants were asked to add the number currently presented on the screen to the number previously presented. If the two numbers add to the target number, they were asked to press the “right” arrow key on the keyboard. If they do not add to the target number, they were asked to press the “left” arrow key.

Fig. 1

Example of the PVSAT. The participant was given a target number and was asked to add numbers and respond by tapping an arrow key if the numbers added to the target or not.

2.4.

Data Analysis

Intensity data were processed using the Homer2 software package²⁹ in MATLAB (The Mathworks, Natick, Massachusetts) and following the steps. The intensity data recorded via the NIRStar software were converted into the Homer2 “.nirs” format to be further processed. The intensity data extracted from the “.nirs” file were converted to delta optical density. Movement related artifacts were removed from the data using the Homer2 spline motion correction algorithm²⁹ with a Savitzky–Golay smoothing filter. Optical density data were then converted to hemoglobin concentration using the modified Beer–Lambert law^29,30 and the age-dependent differential pathlength factor,³¹ and bandpass filtering (0.001 to 1.9 Hz) was performed. Next, we performed the physiological regression of the short channels.^32,33 This was done using the equation posited by Saager and Berger with the replacement of the alpha value with the value of the correlation between the long and short channel time series. Finally, we calculated interhemispheric coherence on the time series output of the short channel regression using the MATLAB wavelet coherence function. This function uses a Morlet wavelet as its basis for the coherence calculation with a moving average filter to smooth across time and frequency. Although the coherence calculation was done on the whole frequency band (0.001 to 1.9 Hz), we extracted the coherence values between 0.01 and 0.06 Hz for further analysis.

To study the impact of averaging the interhemispheric coherence data from different channel pairs (colored numbers in Fig. 2), we obtained data from channel pairs in the DLPFC, with eight source–detector pairs in each of the left and right DLPFC. We undertook coherence analysis in which no channel pairs were averaged, resulting in eight coherence values from eight source–detector pairs. We applied the same analysis to the data averaged from two nearest pairs, resulting in four coherence values per person (colored lines in Fig. 2); four nearest pairs, resulting in two coherence values per person (colored ovals in Fig. 2); and all eight pairs, resulting in one coherence value per person.

Fig. 2

fNIRS channel locations in the DLPFC. Sources are shown in red and detectors in blue with the lines between them representing the channels. Short channels are indicated with the light blue circle surrounding the source optode. The channels are numbered and color-coded to represent the channels on either hemisphere selected for the coherence calculation when no averaging and when two channels were averaged together. The color-coded ovals represent the four channels averaged together on each hemisphere.

2.5.

Statistical Analysis

A linear mixed model was used to assess group differences (control versus concussion) based on the coherence value with participant “id” listed as a random factor.³³ Model assumptions were verified using the “performance” package³⁴ in RStudio (version 4.2.0),³⁵ which reports the model fit, collinearity, homogeneity of variance, and normality of residual and random effects. To alleviate bias due to an unbalanced dataset, we performed a permutation test ($n=1000$) on all variations of signal combinations and compared the model statistics.

The most common metric to interpret statistical models is the $p$-value,³⁶ which provides insight into whether an effect is present or, alternatively stated, whether it is “statistically significant.”³⁷ It provides insight into whether there is a difference between two groups, in our case controls and concussion. The effect size is suggested to be accompanied alongside the $p$-values to allow for interpretation of the strength between variables, which provides insight into the magnitude of the effect.^37–39 The magnitude of effect sizes is labeled, either small (${\eta }^2=0.0099$), medium (${\eta }^2=0.0588$), or large (${\eta }^2=0.1379$).^39,40 Effect sizes that are considered “small” may still be important as variances from unmeasured variables may have decreased what might have been a medium or large effect.⁴⁰ It is expected that a $p$-value less than the threshold for statistical significance ($p<0.05$) coupled with a medium effect size would give confidence that the concussion group could be differentiated from the controls.

To determine the impact of averaging, we used both the $p$-value and effect size as criteria when comparing group (control versus concussion) differences. The significance of the model effects was evaluated with the Satterthwaite approximation for degrees of freedom, an alpha level of 0.05 was used for all statistical tests. Descriptive statistics data were calculated to show the mean, standard deviation (SD), coefficient of variation (CV), and maximum coherence value for each channel pair and group. All statistical analyses were performed in RStudio (version 4.2.0).³⁵

3.1.

Channel Pair Differences

Preliminary analysis revealed that there were no groups by channel pair interactions either during PVSAT or the resting state. Statistical differences were noted between individual channel pairs with a medium effect size (${\eta }^2>=0.058$)^39,40 during PVSAT [$F(\mathrm{7,280})=4.63$, $p<0.001$, ${\eta }^2=0.1$] when no averaging was performed (Table 4) and [$F(\mathrm{3,120})=3.33$, $p<0.05$, ${\eta }^2=0.06$] when only two channel pairs were averaged. There was no significant difference between the channel pairs during PVSAT when observing the four channel averaged data [$F(\mathrm{1,39})=0.25$, $p=0.62$, ${\eta }^2<0.01$]. Differences during the resting state showed similar responses with a medium effect size [$F(\mathrm{7,280})=5.54$, $p<0.001$, ${\eta }^2=0.12$] when no averaging was performed and a large effect [$F(\mathrm{3,120})=7.86$, $p<0.001$, ${\eta }^2=0.14$] when only two channel pairs were averaged, but no effect when four channel pairs were averaged [$F(\mathrm{1,40})=0.11$, $p=0.74$, ${\eta }^2<0.01$].

Table 4

Difference between channel pairs during the resting state and PVSAT sorted by the effect size.

3.2.

Channel Pair Averaging

We found statistical differences between groups when observing the different averaging strategies (Table 5). It was determined that averaging all eight fibers into one amplitude-time series before undertaking coherence analysis gives no significant difference between groups during PVSAT [$F(\mathrm{1,43})=3.932$, $p=0.056$, ${\eta }^2=0.084$] and resting state [$F(\mathrm{1,43})=0.371$, $p=0.58$, ${\eta }^2=0.009$], whereas averaging less channel pairs tends to improve the discrimination. During PVSAT, the group difference was significant with no averaging [$F(\mathrm{1,43})=8.419$, $p=0.010$, ${\eta }^2=0.164$] and when averaging two channel pairs [$F(\mathrm{1,43})=6.435$, $p=0.023$, ${\eta }^2=0.13$]. Averaging four channel pairs did not show a group difference [$F(\mathrm{1,43})=3.202$, $p=0.079$, ${\eta }^2=0.069$]. Resting state analysis did not show any group differences for no averaging [$F(\mathrm{1,43})=3.743$, $p=0.062$, ${\eta }^2=0.08$], when averaging two channel pairs [$F(\mathrm{1,43})=0.64$, $p=0.438$, ${\eta }^2=0.015$], and averaging four channel pairs [$F(\mathrm{1,43})=0.067$, $p=0.802$, ${\eta }^2=0.002$]. The effect size between averaging strategies during the resting state was small, ranging from (0.009 to 0.08), and medium to large during PVSAT (0.084 to 0.16). With resting state and PVSAT data, the effect size was largest with no averaging.

Table 5

Difference between groups (concussed and controls) during the resting state and PVSAT based on the permutation test (n=1000).

4.1.

Group Differences Between Concussion and Controls