Participants between the ages of 18 and 65 were recruited for this study. Participants were required to be free of MR-contraindicators, concurrent substance abuse, have normal or corrected-to-normal vision, and speak fluent English. Because study procedures included a gas-inhalation challenge (see Section 2.4), participant selection was limited to non-smokers. Participants did not have histories of respiratory or pulmonary problems, cerebral vascular issues, or cardiac problems. Participants were required to have a score greater than 21 on the telephone interview for cognitive status [45]. Thirty-one participants in total met the inclusion criteria.

Twelve MS patients meeting the above criteria were recruited from the Clinical Center for Multiple Sclerosis at the University of Texas Southwestern Medical Center. Eleven patients had a diagnosis of relapsing-remitting MS and one patient had a diagnosis of secondary-progressive MS. Patients were required to be at least 1 month past their most recent exacerbation and their last corticosteroid treatment. Patients were recruited who did not report a history of optic neuritis. Patients without a history of optic neuritis were specifically selected so as to limit additional variability from attributed to severe, anterior visual pathway damage/dysfunction (e.g., such as that resulting from conduction block) and potential visual impairment. All MS patients’ vision was normal or corrected-to-normal. Two patients withdrew or declined to undergo the gas challenge (total n = 10).

Nineteen HC participants were recruited from the Dallas-Fort Worth Metroplex via email, posted flyers, and word-of-mouth. These participants were evaluated for the general inclusion/exclusion criteria described above. Three HCs did not undergo the scanning protocol because of exclusions discovered after study enrollment (e.g., concussion history revealed after pre-screening, incidental MR finding). Two HCs withdrew or declined to undergo the gas challenge. During imaging processing (see Section 2.5), one HC’s functional images failed to appropriately register to their anatomical image after multiple attempts, so this person was excluded. Thirteen HCs (n = 13) remained for subsequent analyses. These participants were closely age- and sex-matched to the MS patients (see Table 1).

Study procedures were approved by the University of Texas Southwestern Medical Center Institutional Review Board. Recruitment numbers were approximated based upon previous research showing sufficient power to demonstrate group changes in calibrated fMRI (cfMRI) contrasts with similar sample sizes [22,23]. Participants meeting inclusion criteria were asked to refrain from caffeine use at least two hours before their scheduled appointment time, e.g., [47]. They were also asked not to consume alcohol on the same calendar day before their scheduled appointment. Participants gave written informed consent before undergoing procedures and were compensated for their time. Participants underwent functional and structural neuroimaging on a Philips 3-Tesla magnet (Philips Medical Systems, Best, The Netherlands) with an 8-channel SENSE radiofrequency head coil. Foam padding was placed around the head to minimize motion during MRI scan acquisition. Participants completed standard neuropsychological tests (e.g., Brief Repeatable Battery of Neuropsychological tests [44]) and self-report measures regarding their general health and symptomology (i.e., SF-36 [48], Modified Fatigue Impact Scale (MFIS, [49]); see Table 2 for a complete list of model variables).

Dual-echo pseudocontinuous arterial spin labeling (pCASL) and BOLD images (together referred to as dual-echo images) were acquired using an interleaved echo scanning protocol see [7,52]. Together, the perfusion (Echo 1) and BOLD-weighted (Echo 2) images along with biophysical modeling procedures allowed for estimation of CMRO₂ and a neural-vascular coupling coefficient (n, see [8]) associated with steady-state, neural stimulation [5,7]. One task run of dual-echo imaging data and one gas-challenge run of dual-echo imaging data were collected using the following parameters: Echo 1: labeling duration 1650 ms, labeling flip angle 18°, labeling gap = 63.5 mm, 3.44 × 3.44 × 5 mm voxel, repetition time (TR) = 4000 ms, echo time (TE) = 14 ms, 1525 ms post-label delay, 0 mm slice gap. Echo 2: 90° flip angle, 3.44 × 3.44 × 5 mm voxel, TR = 4000 ms, TE = 40 ms, 0 mm slice gap. Total scan time for the visual stimulation task = 600 s (72 dual-echo dynamics). Total scan time for the gas challenge = 624 s (75 dual-echo dynamics).

Estimations of CMRO2 and n were based upon the Davis model of BOLD signal change [6,7]: (1) Δ S S 0 = M ( ( 1 − Δ CBF CBF 0 ) ∝ − β ( Δ CMRO 2 CMRO 2 | 0 ) β ) where ∆x/x₀ denotes a change from baseline, α is an empirically derived constant linking cerebral blood flow and cerebral blood volume, and β is an empirically derived constant related to vascular exchange and susceptibility of deoxyhemoglobin at specific field strengths (e.g., [53,54,55]). We assumed α = 0.38 [56] and β = 1.3 [52]; these values were chosen because they have been shown to be sensitive to group differences in neurophysiology [22,23]. Also, these values have previously demonstrated group-equivalence in the estimation of M, e.g., [22,23]. M is a subject-specific scaling factor dependent upon the washout resting deoxyhemoglobin see [8]. M was estimated in each participant, using the gas challenge detailed below.

The measurement of BOLD, CBF, and M allows for the estimation of CMRO₂. Here, ∆CMRO₂ reflects the visual task-related change in neurometabolism of oxygen from resting baseline: (2) Δ CMRO 2   CMRO 2 | 0 =   ( 1 − Δ BOLD BOLD 0 M ) 1 / β ( Δ CBF CBF 0 ) 1 −   α / β where ∆x/x₀ reflects percent change of signal during task compared to resting baseline. With the estimation of ∆CMRO₂, n, may also be estimated: (3) n =   Δ CBF CBF 0 Δ CMRO 2 CMRO 2 | 0 thus, n reflects per unit output of ∆CBF per unit input of ∆CMRO₂ see [8].

Participants completed a visual stimulation task during dual-echo task imaging. This task was chosen for two reasons. First, differences in the functional response to visual stimulation have been observed in MS visual cortex see [42,57]. Second, because this task required minimal effort, group differences in performance were not expected to be a factor.

Participants were trained on the task before entering the MR environment. During the task, participants focused on a fixation cross at the center of their visual field. Participants were required to respond via bilateral, thumb-button press when a change in the luminance of the fixation cross occurred. This task was used in order to control the center of the participants’ visual field [22,23,58]. Change in luminance was jittered and occurred every 2, 3, 4, or 6 s. Visual stimulation occurred in a block format. There were 6 visual stimulation task blocks consisting of 60 s of continual annulus flickering in the participants’ near-foveal visual field. Annuli alternated at orthogonal orientations (0 to 90°) to avoid neural adaptation [58]. Alterations occurred at a constant frequency of 8 Hz because both electrochemical neural activity and BOLD signal have been shown to peak at this frequency, potentially yielding the greatest signal-to-noise estimates, e.g., [59,60]. Rest blocks were jittered at 32, 34, 36, 38, and 40 s intervals (see Figure 1).

Participants also completed a gas-challenge in order to estimate M. Participants breathed 4 min of room air (~0.03% CO₂: 21% O₂: 78% N₂) and 6 min of an iso-oxic, CO₂ solution (5% CO₂: 21% O₂: 74% N₂) during dual-echo imaging. Each participant was fitted with a two-way, non-rebreathing valve/mouthpiece and a nose clip. Baseline end-tidal CO₂ (EtCO₂), O₂ saturation, breath rate, and heart rate measures were collected. After the 4 min of room air breathing, a valve was opened to release the CO₂ solution from a Douglas airbag which then flowed into the participants’ breathing apparatus [22,23]. The CO₂ inhalation lasted 6 min.

Hypercapnic challenge, via the inhaled 5% CO₂ solution, increases global CBF, but probably has no or a minimal depressant effect on oxygen metabolism, e.g., [61,62,63]. Hypercapnia acts to wash out local baseline concentrations of deoxyhemoglobin, yielding a local maximum estimate of resting BOLD signal. Potential changes to oxygen metabolism due hypercapnic challenge have not been shown to appreciably alter the estimation of M as relationships between hypercapnia-derived M and M derived from non-hypercapnic techniques show high correspondence [64].

Task and gas-challenge Echo 1 and Echo 2 data were processed in analysis of functional neuroimages (AFNI [65]) and the Functional MRI of the Brain Software Library (FSL [66]). Data were transformed into cardinal planes. Anomalous data points in each voxel time series were then attenuated using an interpolation method based upon the average signal. Data were volume registered to correct for motion to the fourth functional volume of each dataset’s (task or gas challenge) Echo 2 sequence using a heptic polynomial interpolation method. CBF was estimated from Echo 1 images using the surround subtraction method [67]. Dual-echo BOLD data were also interpolated by pairwise averaging of temporally adjacent images.

For the visual stimulation task, Echo 2 data were linearly registered (12 degrees-of-freedom) to each participant’s anatomical data using AFNI’s align_epi_anay.py program. The transformation matrix from this registration was then applied to Echo 1 data, placing these two datasets in the same space. For gas-challenge data, a binary mask was created for functional voxels in Echo 2 to aid in co-registration. This mask was then registered to the respective participant’s anatomical space using the align_epi_anay.py program. Gas-challenge Echo 2 and Echo 1 data were also aligned to the mask which was registered in native anatomical space. After alignment, Echoes 1 and 2 data from both the visual task and gas challenge were visually inspected for registration errors. One HC participant failed to register correctly after multiple attempts and was discarded from further analyses. Echoes 1 and 2 data from the visual task and gas challenge were then spatially smoothed using a Gaussian kernel (FWHM = 8 mm) and high-pass filtered (0.0039 Hz).

Preprocessed data from Echoes 1 and 2 in the visual stimulation task were analyzed via generalized linear modeling of task versus rest periods using a boxcar reference function. This modeling quantified task-related CBF and BOLD changes from baseline. BOLD and CBF beta-values were scaled to each voxel’s resting baseline signal and were multiplied by 100, yielding percent signal change estimates from baseline (∆BOLD and ∆CBF). Data were averaged from a visual (functional) region of interest (ROI) comprised of overlapping ∆BOLD and ∆CBF suprathreshold signals within occipital lobe (see Structural and Functional ROI; [22,23]). ∆BOLD, ∆CBF, ∆CMRO₂, and n results extracted from the functional region of interest were taken as the visual-evoked signals (i.e., veBOLD, veCBF, veCMRO₂, and ven).

For the gas challenge, resting baseline BOLD and CBF signals during room air breathing were averaged for each voxel time-series (BOLD₀ and CBF₀). The first two minutes of hypercapnia BOLD and CBF time-series were discarded to allow participants’ blood flow to stabilize on the CO₂ solution, e.g., [22,23]. The last four minutes of hypercapnia BOLD and CBF time-series were averaged to yield BOLD_hc and CBF_hc respectively. Average values were extracted from a functional region of interest (see Structural and Functional ROI) using overlapping BOLD_hc and CBF_hc suprathreshold signals within occipital lobe, and were used to calculate M, using the following equation: (4) M = BOLD hc −   BOLD 0 BOLD 0 ( 1 − ( 1 + CBF hc −   CBF 0 CBF 0 ) α − β ) where (x_hc−x₀)/x₀ reflects percent change in signal from normocapnic to hypercapnic states, normalized by the signals during normocapnia and multiplied by 100. Once M was estimated, ∆CMRO₂ and n were also estimated (see Equations (2) and (3); see Figure 2) within a functional region of interest (see Structural and Functional ROI).

First, the magnetization-prepared rapid acquisition gradient-echo (MPRAGE) data were processed to create a native-space, occipital ROI. The skull was removed using an automated command, separating parenchyma and cerebral spinal fluid from the skull. An intensity based automated segmentation algorithm was used to delineate primarily white matter, grey matter, and cerebral spinal fluid voxels yielding a partial volume estimate of each tissue type, for each voxel. A grey matter mask was then created, retaining voxels with only a greater than or equal to grey matter partial volume estimate of 80%. A structural ROI of occipital lobe was manually delineated on each participant’s MPRAGE image. These were drawn in native space because native space analyses tend to allow for more sensitive patient-control contrasts [68]. The structural ROI was drawn using gyral and sulcal landmarks and encompassed most of occipital cortex including calcarine sulcus, cuneus, and occipital portions of lingual gyrus. Several anatomical landmarks were used in the demarcation of this ROI (parieto-occipital sulcus, occipital pole, pre-occipital notch). Within the anatomically defined occipital lobe, only voxels with partial volume estimates of grey matter (≥80%) were retained. These final masks were down-sampled to the functional voxel size.

A visual task functional ROI was created within the structural ROI described above to estimate veBOLD, veCBF, veCMRO₂, and ven (see Figure 3). This procedure eschewed noise from inactive voxels, e.g., [68]. Voxels comprising each participant’s functional ROI were the overlapping top 5% of BOLD and top 5% of CBF t-values obtained from the generalized model, within the structural ROI. This ensured that average veBOLD and veCBF estimates were being derived from the same, task-responsive voxels and that veCMRO₂ and ven were derived in voxels with both CBF and BOLD task-related increases (see Figure 3).

veCMRO₂ was calculated voxel-wise within the functional ROI using ∆BOLD, ∆CBF, M (which was extracted from functional ROI described below). ven was then calculated similarly. The final product of these analyses was average positive veBOLD, veCBF, and veCMRO₂, and ven extracted from the functional ROI (see Figure 3).

Because the gas challenge data differed in occipital coverage compared to the visual task data, M was estimated ex situ. To create a functional ROI for the gas challenge, ∆BOLD_hc/BOLD₀ and ∆CBF_hc/CBF₀ maps were thresholded and extracted from the structural ROI detailed above. The criteria for retention of a voxel within these maps required that the voxel was within the top 15% (top 20% for one participant) of ∆BOLD_hc/BOLD₀ and ∆CBF_hc/CBF₀ voxels in the structural ROI, and that these ∆BOLD_hc/BOLD₀ and ∆CBF_hc/CBF₀ voxels overlapped. This procedure ensured complementary maximum ∆BOLD_hc/BOLD₀ and ∆CBF_hc/CBF₀ signals in the retained voxels. Average ∆BOLD_hc/BOLD₀ and ∆CBF_hc/CBF₀ signals were extracted from this ROI and M was calculated (see Equation (4)).

One T₁-weighted MPRAGE image was acquired for each participant: 160 slices, TE = 3.7 ms, repetition time TR = 8.1 ms, sagittal slice orientation, 1 × 1 × 1 mm³ voxel, 12° flip angle. SIENAX [15,69] was used to obtain measures of grey matter, white matter, and total brain volume normalized by participant’s head size. This technique uses partial volume estimation to calculate volume of differing tissue types (see Figure 4B,C). Further, this technique takes into account lesioned tissue, as demarcated by lesion masks (see below), in order to avoid misclassification of this tissue. The final products of these analyses were scaled estimates of each participant’s grey matter, white matter, and total brain volume (mm³).

A T₂ fluid attenuated inversion recovery (FLAIR) scan was also acquired for each participant: 33 slices, TE = 125 ms, TR = 11,000 ms, no slice gap, transverse slice orientation, 0.45 × 0.45 × 5.00 mm³ voxel, 120° refocusing angle. FLAIR images were used to estimate the extent of gross lesion burden for each participant. Hyperintense voxels were demarcated using in-house MATLAB code based upon slice-wise, signal intensity (i.e., voxels that were ≥1.25 SD over the slice mean intensity). Next, lesions were manually delineated from the hyperintense tissue by two trained researchers (L.H., S.F.). Manual delineation ruled out false positives in lesion classification due to fat signals, motion, ventricular edge effects, skull, or signal inhomogeneites [70]. Lesion burden was estimated by extracting the number of voxels that were demarcated by the automated and manual procedures. Inter-rater agreement of lesion burden was calculated using a Dice ratio (κ) of the lesion burden estimates made by the two researchers on a sample of several subjects [71]. After the researchers were trained on lesion classification, inter-rater agreement was found to be high, κ = 0.89; where κ > 0.70 is generally thought to reflect excellent inter-rater agreement [72]. Lesion burden was quantified using absolute (total mm³ of lesioned tissue; see Figure 4E) and relative scales (percent of total mm³ of lesioned tissue scaled by uncorrected white matter volume in mm³). Spatially distinct lesion count was also obtained by counting the number of non-touching lesions for each subject (see Figure 4F), e.g., [73]. A lesion was required to have at least 3 mm³ volume in order to be added to the total lesion count. Thus, the final products of these analyses were absolute lesion volume, relative lesion volume, and spatially distinct lesion count.

DTI images were acquired using a single-shot, echo-planar imaging sequence with a Sensitivity Encoding parallel imaging scheme (reduction factor = 2.3), 112 × 112 matrix, field of view = 224 × 224 mm² (nominal resolution of 2 mm), 65 slices (0 mm gap), slice thickness = 2 mm, TR = 7.78 s, TE = 97 ms. The diffusion weighting was encoded along 30 independent orientations [74] and the b value was 1000 s/mm². Imaging time was 5 min and 15 s. Two HCs did not undergo DTI (nHC = 11).

Automatic Image Registration [75] was performed on raw diffusion-weighted images to correct distortion caused by eddy currents. Six elements of the 3 × 3 diffusion tensor were determined by multivariate least-squares fitting. The tensor was diagonalized to obtain three eigenvalues (λ_1–3) and eigenvectors (v_1–3). Standard tensor fitting was conducted with DTIStudio [76] to generate the most common DTI-derived diffusion characteristics, fractional anisotropy (FA), axial diffusivity (AD), mean diffusivity (MD), and radial diffusivity (RD).

DTI measurements were obtained at the skeletons of the white matter using FSL [77] to alleviate partial volume effects with tract-based spatial statistics (see Figure 4F–H) [77]. Participant FA maps were registered nonlinearly to the EVE single-subject FA template [78,79,80] for better alignment with a digital white matter atlas (JHU ICBM-DTI-81) [81]. Registered FA maps of all subjects were averaged to generate a mean FA map, from which an FA skeleton mask was created. Skeletonized FA images of all subjects were obtained by projecting the registered FA images onto the mean FA skeleton mask. Skeletonized AD, MD, and RD metrics were obtained by applying the same registration, projection, and skeletonization procedures. We extracted skeleton-wide averages of each DTI metric (i.e., AD, FA, MD, RD), wherein an average of each metric is calculated across all voxels within the white matter skeleton (see Figure 4A).

All analyses were performed on distributions free of outliers (≥±2 SD from group mean for simple group comparisons, ≥±3 MAD from group median for classification modeling see [82]). Binary logistic regression was used for classifying MS status. A description of model variables can be found in Table 2. The accuracies of these models were computed as the proportion of correct classification outcomes over all outcomes. Accuracy was chosen as the metric of interest because it combines sensitivity and specificity in binary classification analysis by taking into account both true positives and true negatives relative to all outcomes. We used resampling-based hypothesis testing to examine both within-sample and out-of-sample classification of patient status see [83]. Because we used relatively conservative analytic techniques, inherently reducing the likelihood of Type I error and increasing the generalizability of our results, the criterion for a rejection of the null hypothesis was not corrected for multiple comparisons and all models were evaluated at the field-standard α = 0.05. We also denote which hypothesis tests survived Benjamini-Hotchberg correction (Table 4; Figure 7).

Within-sample classification analyses obtained bias-corrected and accelerated (BCa) bootstrapped-resampled (B = 10,000) 95% confidence intervals of the accuracy of binary logistic regression models. The BCa procedure was used because it is robust to both skewness and sampling bias in the bootstrap distribution [84]. To avoid unstable classification, we stratified all resamples to match the original sample’s constitution of patients and controls, 56.5% and 43.5%, respectively. If the BCa-derived 95% confidence interval did not contain a value at or below 0.50 (binary chance), this would demonstrate the measure’s accuracy was significantly greater than chance to classify MS patients and HCs.

Out-of-sample classification analyses used a leave-one-out cross-validation approach [85]. This technique used training and sample iterations to test the ability of the model derived from the training set to predict an observation in the test (out-of-sample) set, thus, circumventing problems of sample bias, model over fitting, and lending a true predictive element to these analyses. Briefly, the leave-one-out cross validation (LOOCV) approach fitted N models, where N was proportional to our sample size. Each model was trained on N-1 samples and then the accuracy of the training model was assessed on the left-out sample. The N accuracies were then averaged to attain a representative and generalizable measure of the average out-of-sample classification accuracy. Permutation based p-values (5000 permutations) were computed to assess the significance of the LOOCV-derived accuracy statistics. The test permuted patient status labels and recomputed the accuracy of the model at each iteration, thus building the null distribution. The p-values were calculated from the percentage of the accuracy estimates of the permuted samples that were better than actual LOOCV-derived accuracy statistic of each model. This procedure was slightly modified according to Ojala and Garriga [86].

MS patients (92.75 ± 1.11%) did not significantly differ from HCs (94.86 ± 0.44%) on accuracy on the visual stimulation secondary task, t(10.54) = −1.76, p = 0.108. Patients (492.06 ms ± 31.15) also did not significantly differ from HCs (487.19 ms ± 24.10) on their average correct response time to press the button on the secondary task, t(16.22) = 0.12, p = 0.903.

MS and HCs did not significantly differ in breath rate, end-tidal CO₂, heart rate, or O₂ saturation at baseline or during CO₂ solution breathing (all ps > 0.05; see Table 3). We tested whether MS patients differed in their CBF response to the CO₂ solution ((CBF_hc−CBF₀)/CBF₀) and M in their respective gas challenge ROIs within occipital lobe see [87]. MS patients did not significantly differ in CBF response to the CO₂ solution (167.48 ± 19.8%) compared to HCs (146.90 ± 14.64%), t(15.70) = 0.83, p = 0.417. MS patients (3.88 ± 0.48%) did not significantly differ in M compared to HCs (5.11 ± 0.39%), t(18.90) = −1.98, p = 0.062.

MS patients (1.12 ± 0.77%) did not significantly differ from HCs (1.18 ± 0.66%) on veBOLD response to visual stimulation, t(19.18) = −0.60, p = 0.555. MS patients (4.08 ± 0.35) did not show significant changes in ven compared to HCs (4.23 ± 0.23), t(16.16) = −0.35, p = 0.731. MS patients (48.06 ± 12.58%) had significant decreases in veCBF compared to HCs (92.68 ± 17.29%), t(19.76) = −2.09, p = 0.050. MS patients (9.59 ± 0.90%) also showed significant decreases in veCMRO₂ compared to HCs (17.85 ± 1.97%), t(16.45) = −3.81, p = 0.002 (see Figure 5).

Measures are ranked on original accuracy and presented in Table 4. Accuracy and smoothed density distributions for the significant and bottom 5 measures can be found in Figure 6.

Predictors presented in Figure 7 are ranked on LOOCV-derived accuracy.