Identification of neurovascular changes associated with cerebral amyloid angiopathy from subject-specific hemodynamic response functions

Abstract

Cerebral amyloid angiopathy (CAA) is a small-vessel disease preferentially affecting posterior brain regions. Recent evidence has demonstrated the efficacy of functional MRI in detecting CAA-related neurovascular injury, however, it is unknown whether such perturbations are associated with changes in the hemodynamic response function (HRF). Here we estimated HRFs from two different brain regions from block design activation data, in light of recent findings demonstrating how block designs can accurately reflect HRF parameter estimates while maximizing signal detection. Patients with a diagnosis of probable CAA and healthy controls performed motor and visual stimulation tasks. Time-to-peak (TTP), full-width at half-maximum (FWHM), and area under the curve (AUC) of the estimated HRFs were compared between groups and to MRI features associated with CAA including cerebral microbleed (CMB) count. Motor HRFs in CAA patients showed significantly wider FWHM (P = 0.006) and delayed TTP (P = 0.03) compared to controls. In the patient group, visual HRF FWHM was positively associated with CMB count (P = 0.03). These findings indicate that hemodynamic abnormalities in patients with CAA may be reflected in HRFs estimated from block designs across different brain regions. Moreover, visual FWHM may be linked to structural MR indications associated with CAA.

Introduction

Cerebral amyloid angiopathy (CAA) is a common age-related pathology, characterized by deposition of the β-amyloid peptide within the media and adventitia of the small cerebral blood vessels.¹ CAA is a leading cause of intracerebral haemorrhage (ICH), which is often detected at the later stages of disease progression.² In addition to being a leading cause of ICH, CAA is associated with Alzheimer's disease and cognitive impairment.³ Small infarcts may contribute to cognitive impairment observed in patients with CAA, although many of these infarcts are detectable only at autopsy.⁴ Currently, CAA can only be diagnosed in its later stages by demonstrating a characteristic pattern of ICH and microbleeds restricted to lobar locations of the brain.⁵ Additional, sensitive biomarkers are needed for diagnosis and prognosis in earlier stages of the disease.⁶

The distribution of CAA pathology is preferentially located in the posterior lobar regions,⁷ with the posterior cerebral artery perfusion territory showing decreased vascular reactivity to a visual stimulation task.⁸ Functional MRI (fMRI) based on blood oxygenation level-dependent (BOLD) contrast is sensitive to hemodynamic and metabolic changes associated with neural activity.^9,10 Reduced BOLD responses are evident in patients with CAA relative to healthy controls (HC),^11,12 thus indicating that the BOLD signal may be altered in the presence of CAA-related neurovascular injury. Characterizing these BOLD signal alterations and understanding their relation to CAA-related pathology is imperative for the development of fMRI biomarkers of CAA.

The hemodynamic response function (HRF) is the basis of BOLD fMRI, describing the temporal evolution of BOLD signal change to invoked neural activity.¹³ HRF characterization is important for understanding how dynamic changes in cerebral physiology result in an observable BOLD response. Estimating the HRF allows for the calculation of information pertaining to signal shape and timing features such as the time-to-peak (TTP), full-width at half-maximum (FWHM) and area under the curve (AUC), which provide information about the latency, duration and extent of the BOLD signal.¹⁴ Such features are not typically inferred from traditional analyses of the fMRI time course, where focus is placed on the magnitude of the response. In addition, HRF characterization is important for the prevention of statistical error. In the analysis of fMRI data, a pre-specified HRF convolved with the stimulus time course is often used to model the BOLD signal change.^13,15 Sensitivity to signal detection is dependent on an accurate HRF model. However, HRF alterations have been demonstrated in patients with cerebrovascular insult.^16,17 Currently, HRF estimation has not been performed in patients with CAA. Therefore, the unknown extent of HRF alteration in these patients may lead to decreased statistical power. Understanding how the HRF is altered in this patient cohort is essential for the continued use of fMRI in the investigation of CAA-related neurovascular changes.

Various methods for HRF estimation from fMRI data have been proposed,^18,19 most of which require an event-related paradigm, where briefly presented stimuli are interspersed with baseline periods of variable durations.^20–23 Although this design increases the accuracy with which the HRF can be estimated, it is at the cost of signal detection. Studies recruiting patient and aged cohorts may rely more on block designs, where stimuli are presented within an experimental framework optimized for signal detection rather than HRF estimation.²⁴ However, recent work by Shan and colleagues²⁵ demonstrated that, by using a curve fitting procedure, it is possible to model HRFs from block designs with good reproducibility. These findings are advantageous for clinical fMRI studies as block designs with optimal signal detection can continue to be implemented, while gaining novel information about the shape and timing features of estimated HRFs.

In the present study, we estimated HRFs from a block design in patients with a diagnosis of probable CAA and a healthy aged-matched population. HRFs were estimated from two different brain regions corresponding to the primary visual cortex (V1) and the primary motor cortex (M1). These regions were chosen based on the hypothesis that V1 would be more heavily affected by CAA pathology than M1. Moreover, both regions can be robustly activated by simple tasks. The first research aim was to establish whether HRFs estimated from this patient group are different to those obtained in a group of HC. Essentially, with this research aim we intended to determine whether there are specific features of the HRF that differ between CAA patients and HC. Our second research aim was to understand how HRF features are related to underlying structural pathology. To achieve this, we quantified number of cerebral microbleeds (CMBs) and white matter hyperintensity (WMH) volumes from structural MR images. These lesions, commonly implicated in CAA, have demonstrated to be associated with alterations in cerebral hemodynamics.^26,27 We characterized the relationship between these CAA-related structural indications and the calculated HRF features of TTP, FWHM and AUC in order to determine if any of these features are specifically altered in the presence of CAA pathology.

Methods

Participants

Participants were enrolled as part of a prospective longitudinal cohort study, the Functional Assessment of Vascular Reactivity (FAVR) Study. Participants were recruited and gave informed consent to participate in this study, which was approved by the University of Calgary Conjoint Health Research Ethics Board (CHREB). CHREB ethical guidelines were adhered to. For the present study, data were drawn from participants in a previously published study.¹² Data were reanalysed from 13 patients with a diagnosis of probable CAA using Boston criteria⁵ (mean age = 74.3 ± 9.2 years, 5 female) and 14 aged-matched HC participants (mean age = 68.4 ± 9.4 years, 6 female). We excluded five CAA patients and four controls from this re-analysis because noisy BOLD waveforms would have prevented optimal HRF estimation. This was required as the current HRF analyses were performed on an individual basis, rather than group-averaged data. A conservative activation threshold was implemented for all first-level analyses with participant inclusion limited to those with surviving voxels. Presenting symptoms of the CAA patients were ICH (n = 8), cognitive decline (n = 2) and transient focal neurological symptoms (n = 3). One patient was in remission for CAA-related inflammation. Non-CAA participants were recruited from the community and classified as HC based on a clinical history free from stroke or dementia, as verified in an interview by a neurologist. Cognitive testing and qualitative information relating to cerebrovascular health were acquired for all participants. These data included the Folstein Mini-Mental State exam²⁸ and reported history of coronary heart disease, atrial fibrillation, diabetes mellitus, hypertension and hypercholesterolemia. These measures were statistically compared between the two groups using the nonparametric Mann–Whitney U test and Fisher's exact test.

Data acquisition

All MRI data were acquired on a 3 T MR scanner (Signa VHi, General Electric Healthcare, Waukesha, WI) with a 12-channel head coil. For each participant, 180 sets of whole-brain echo-planar imaging data sensitized to BOLD contrast (TR = 2000 ms, TE = 30 ms, α = 70°, in-plane resolution = 3.75 × 3.75 mm, slice thickness = 4 mm) were acquired across two separate runs, with the motor and visual tasks performed in different runs. Both runs were completed within a single scan session. Structural imaging included a high-resolution 3D T₁-weighted structural image (TR = 6000 ms, TE = 2.5 ms, TI = 650 ms, α = 8°, 0.9 × 0.9 × 1 mm³), a 2D fluid-attenuated inversion recovery (FLAIR) sequence (TR = 9000 ms, TE = 140 ms, TI = 2250 ms, α= 90°, slice thickness = 3.5 mm), and a 2D T₂*-weighted gradient-recalled echo sequence (TR = 1200 ms, TE = 20 ms, α= 18°, slice thickness = 3.5 mm). Structural imaging protocols and analysis were consistent with standards for lesion detection in small vessel disease as recommended by an international working group.²⁹

For the visual task, participants were required to passively view a contrast-reversing black and white radial checkerboard (8 Hz). The checkerboard blocks were interspersed with a baseline condition consisting of a black screen. A central fixation cross was consistently present. A total of four visual blocks, each 40 s in duration and interspersed by 40 s rest blocks, were used in this block design. For the motor task, participants were visually cued to tap the index and middle fingers of their dominate hand at a frequency of 1.5 Hz. Similar to the visual task, there were four motor stimulation blocks each with a duration of 40 s and separated by a blank baseline condition of equal duration. Stimuli were presented using Presentation software (version 14.0, Neurobehavioral Systems, Albany, CA) and projected onto a screen located inside the bore of the scanner using a projector (Avotec, Inc., Stuart, FL, U.S.A.). Subjects viewed the screen using a mirror mounted on top of the head coil.

Data analysis

Individual activation maps. For the fMRI data, preprocessing and first-level statistical analyses were run in Statistical Parametric Mapping 8 (SPM8, Wellcome Trust Centre for Neuroimaging, London, UK) running on MATLAB (MATLAB Release R2013a, The MathWorks, Inc., Natick, MA). Pre-processing involved slice-timing correction, realignment and reslicing, and coregistration of the structural to the mean functional image for each participant. Visual inspection of coregistration was performed for each subject. In cases where poor coregistration was visually detected, an initial manual registration was performed before re-running the automatic algorithm in SPM8. Segmentation of the coregistered T₁-weighted structural image was then performed using the Unified Segmentation algorithm.³⁰ This algorithm simultaneously produces tissue segmentation and performs nonlinear transforms of the anatomical image into MNI space, yielding deformation fields. The inverse deformation fields were used to transform two atlas-based region-of-interests (ROIs) corresponding to the primary visual cortex V1 (Brodmann Area 17) and the primary motor cortex M1 (Brodmann Area 4) from MNI to subject space on an individual basis. The two anatomical ROIs were generated from a Brodmann area atlas using the Wake Forest University (WFU) Pickatlas running on MATLAB.³¹ Mask dilation was not performed, ensuring that the ROI remained within the cortical boundary of the atlas-defined Brodmann area. To ensure high spatial accuracy of activation, no spatial normalization or smoothing was performed on the functional images.

After transformation into native space, the anatomical ROIs were used to produce the functional ROIs for HRF estimation analyses. The functional ROIs consisted of all significantly activated voxels within the anatomical ROIs. To determine which voxels were significantly activated, first-level statistical modelling was performed on the motor and the visual data separately using the finite impulse response (FIR) basis functions. These basis functions allow for variability in the shape and timing parameters of the impulse response without imposing an a priori functional form. Using the FIR basis functions, a beta value was estimated at every TR within the 40-second stimulation periods. For every participant, contrast images were generated using an F-test to identify the effects of the stimulation conditions relative to baseline.

Two contrast images were produced for each participant; one inclusively masked with the V1 anatomical ROI and the other with the M1 anatomical ROI for the visual and motor task, respectively. All activation maps were thresholded at a false-discovery rate (FDR) corrected threshold of P < 0.05. The functional V1 and M1 ROIs, determined on a subject-specific basis, consisted of all suprathreshold voxels within the inclusively masked first-level activation maps. Because prior research has demonstrated that healthy subjects activate larger volumes of cortex than CAA patients for these sensory tasks,¹² the number of voxels in each functional ROI was recorded for every participant and considered as covariates in the statistical analyses described below.

For each subject, BOLD time-courses from all voxels within each functional ROI were averaged using the MarsBaR Region of Interest tool for SPM³² for V1 and M1 separately. This produced two averaged fMRI time-courses for every participant corresponding to V1 and M1, which were then used for HRF estimation.

HRF estimation. HRFs were estimated using routines running on MATLAB.²⁵ HRFs were estimated on a per-subject basis for both V1 and M1 separately using the averaged fMRI time-courses from the voxels in the final ROIs described above. Similar to prior work modelling the HRF,^14,33,34 the fMRI time course was modelled as the convolution of a HRF model and a stimulus function. The stimulus function here was a boxcar function corresponding to the implemented motor and visual tasks, the 40 s on/off design described above. The aim of the estimation analysis was to find the parameters of the HRF model that best fits each subject's fMRI time-course using a least-squares approach. The sum of two-gamma ( $\Gamma$ ) functions was implemented to model the hemodynamic response h(t) at time t, with six free parameters $A_i$ , ${\alpha }_i$ , ${\beta }_i$ :

$$h(t)=\sum _{i=1}^2(A_i\frac{t^{{\alpha }_i-1}{\beta }_i^{{\alpha }_i}{e}^{-{\beta }_{i}t}}{\Gamma ({\alpha }_i)})$$ (1)

These parameters ( $A_i$ , ${\alpha }_i$ , ${\beta }_i$ ) represent the variables whose values are free to vary in order to fit the function to the subject's fMRI data. These three parameters are estimated for two gamma functions in order to account for a positive BOLD response and undershoot. The parameter $A_i$ corresponds to the height of the response. The parameters ${\alpha }_i$ and ${\beta }_i$ model the onset latency and scale of the response, accounting for temporal delays in the onset time of the positive and negative components. Larger ${\alpha }_i$ values result in more delayed onset times and a smaller AUC. Previously, Shan and colleagues²⁵ demonstrated the robustness of this model for estimating the HRF from block designs using simulations with a ground-truth, and validated with fMRI data acquired across multiple time points for assessments of reproducibility. Comparisons of five different HRF models revealed that the sum of two-gamma functions model with six free parameters was optimal in terms of maximizing reproducibility and parameter accuracy. We therefore implemented this model in the present study to estimate subject-specific HRFs from our data. To initialize the fit, we used starting parameters as specified by Shan et al.²⁵ HRF parameter estimation was performed using a modified constrained Nelder–Mead Simplex algorithm on MATLAB. Upper and lower search bounds were set for each of the six parameters to allow for a broad range of physiologically relevant HRF shapes (see Shan et al.²⁵).

For every participant, the data were visually inspected for fitting failure. In cases where noisy data during stimulation epochs resulted in poor fitting, the noisy epochs were excluded from the analysis. To assess for group differences between CAA and HC participants in terms of goodness-of-fit, the residual sums-of-squares for each fit was entered into an analysis of covariance (ANCOVA) test. Motor and visual fits were compared in separate tests. As described above, the fitting procedure was performed for each participant on a single fMRI time-course obtained from the average of all voxel time-courses in the functional ROI. However, we anticipated a discrepancy between the two groups in terms of number of voxels in the functional ROIs. This could influence the noise present in the averaged fMRI time-courses, and therefore the goodness-of-fit. The numbers of voxels in each functional ROI were therefore entered as covariates in these ANCOVA tests.

The resultant subject-specific motor and visual curves from the best-fitting HRF parameters were created and the TTP, FWHM and AUC calculated from each. These quantities were then compared between brain regions (motor, visual) and group (CAA, HC) in a mixed ANCOVA test. Because of the observed number of participants reporting a positive history of hypertension, this was entered into the ANCOVA as one covariate. Age was entered as a second covariate, due to the noted difference in mean age between the CAA and HC groups (∼6-year difference). These two covariates were also entered into all post hoc ANCOVA tests. IBM SPSS Statistics was used for the statistical analyses (IBM Corp., Released 2011, IBM SPSS Statistics for Macintosh, Version 20.00, Armonk, NY).

Structural MRI. White matter hyperintensities and, for CAA patients only, CMBs were characterized and quantitatively assessed. The CMBs were identified from the ${\text{T}}_2^{*}$ -weighted gradient-recalled echo images by an individual rater. The total number of identified microbleeds across the whole brain was recorded for all CAA patients. WMH volumes were quantified from the FLAIR images for each participant using Quantamo (Cybertrials, Inc., Calgary, Canada). This custom software, incorporating components from the Insight Segmentation and Registration Toolkit (National Library of Medicine, Bethesda, MD), uses a threshold-based region growing approach to segment and calculate the volume of white matter lesions. All WMH volumes were normalized to the mean head size of the group on a per-subject basis using a method previously demonstrated to accurately estimate intracranial volume.³⁵ WMH volumes were compared between the two groups using an independent samples t-test. Bivariate correlation analyses were performed on the normalized WMH lesion volumes and number of CMBs and the HRF features of FWHM, TTP and AUC. These statistical analyses were performed used SPSS.

Results

Participants

A small group mean age difference was observed between the control and CAA patient groups (M = 68.4 and 74.3 years, respectively), however an independent samples t-test determined this difference to be non-significant, t(25) = 1.6, P = 0.11. Cognitive testing of the patients and healthy participants showed a difference in the distribution of Mini-Mental State Exam scores between the two groups (P = 0.01; M = 27.6 ± 2.8 for CAA patients, M = 29.5 ± 1.1 for HC). The number of participants reporting a positive history of coronary heart disease (N = 1 for CAA, N = 0 for HC), atrial fibrillation (N = 1 for CAA, N = 0 for HC) and diabetes mellitus (N = 1 for both groups) was low and not statistically different between groups. Histories of hypertension (N = 10 for CAA, N = 2 for HC) and hypercholesterolemia (N = 4 for CAA, N = 5 for HC) were more commonly reported. These distributions were therefore statistically compared between the two groups. Fisher's exact test was significant for hypertension, P = 0.002, however there was no difference between groups in terms of hypercholesterolemia, P = 0.55.

Individual activation maps

The first-level individual activation maps revealed activation within the anatomical ROIs for all participants in the CAA and HC groups. Variation was observed in the peak F-values across brain regions and between groups. However, contrasts directly comparing across groups were not performed, as the images were kept in native space and not spatially normalized to a common template image. The HC group consistently demonstrated a larger number of activated voxels than the CAA patients for both tasks. Figure 1(a) shows representative anatomical ROIs from the original atlas and after transformation into subject space. Figure 1(b) demonstrates representative activation maps inclusively masked with the anatomical ROIs in subject space. Because all images were kept in native space, voxel coordinates are not shown.

Figure 1.

For every participant, motor (a) and visual (b) ROIs were transformed from a Brodmann atlas in MNI space (upper row (a) and (b)) into subject space (lower row (a) and (b)). ROIs from the atlas are overlaid onto an SPM MNI single-subject template. The ROIs in subject space are overlaid onto the anatomical T₁-weighted image of a representative participant. Statistical parametric maps (SPMs) for contrasts demonstrating effects of the motor (c) and visual (d) task compared to baseline for a representative CAA patient and a healthy control. Sagittal and axial slices shown. Slices are at different levels due to differences in activation across participants. SPMs in (c) and (d) inclusively masked with anatomical ROIs in subject space, as demonstrated in (a) and (b), respectively. Statistical maps overlaid onto participant's own registered T₁ weighted anatomical image in native space. Colour bar indicates F-values. Maps thresholded at P < 0.05 FDR-corrected.

ROIs: regions-of-interest; CAA: cerebral amyloid angiopathy; FDR: false-discovery rate.

HRF estimation

Visual inspection of the fitted time courses revealed good fits for all participants and conditions, with one exception (CAA group, motor task). This poor fit appeared to be caused by noisy data and low BOLD signal change to stimulation for the later epochs (i.e. all except the first motor-movement block) in the run. These noisy fMRI data were therefore excluded from the fitting analysis, resulting in the removal of the last three motor stimulation epochs in the fitting analysis. Therefore, for this case, only the first motor activation epoch was modelled for HRF fitting. For the quantitative analyses comparing goodness-of-fit (controlling for number of voxels in the ROIs), the ANCOVA tests showed no group differences in residual sums-of-squares for both the motor task, F(1, 24) = 0.32, P = 0.60, and the visual task, F(1, 24) = 2.23, P = 0.15. This indicated that fit quality was equal across the patient and control groups.

For the estimated HRFs, there was variability in the six estimated parameters across participants for both motor and visual tasks, particularly for those controlling the height of the response ( ${\text{A}}_1$ and ${\text{A}}_2$ ). This variability was particularly evident for the visual HRFs estimated from the CAA group (see Table 1). The best-fitting parameters, averaged across both groups, are shown in Table 1. The visual and motor HRFs were created from the estimated parameters for each subject on an individual basis, allowing for the calculation of TTP, FWHM and AUC. These values, derived from the individual curves, were entered into the following statistical comparisons of HRF features. The individual curves were then averaged across the CAA and HC groups and shown in Figure 2. These averaged curves were used for Figure 2 display purposes only.

Table 1.

Hemodynamic response function parameter estimates.

	$A_1$	${\alpha }_1$	${\beta }_1$	$A_2$	${\alpha }_2$	${\beta }_2$
Motor
HC	0.4 ( ± 0.4)	5.7 ( ± 2.8)	1.2 ( ± 0.3)	0.3 ( ± 0.4)	12.5 ( ± 4.6)	1.0 ( ± 0.5)
CAA	0.5 ( ± 0.6)	7.3 ( ± 1.7)	1.3 ( ± 0.4)	0.4 ( ± 0.6)	13.0 ( ± 5.2)	1.0 ( ± 0.5)
Visual
HC	0.5 ( ± 0.9)	5.7 ( ± 1.2)	1.2 ( ± 0.3)	0.3 ( ± 0.9)	14.5 ( ± 5.6)	1.2 ( ± 0.3)
CAA	0.4 ( ± 0.6)	5.7 ( ± 2.1)	1.0 ( ± 0.4)	0.4 ( ± 0.8)	17.0 ( ± 5.6)	1.0 ( ± 0.4)

HC: healthy control; CAA: cerebral amyloid angiopathy.

The best-fitting parameters of the sum of two gamma functions, averaged across the two groups. These parameters are responsible for the height $(A_i)$ , shape $({\alpha }_i)$ and scale $({\beta }_i)$ of the response. Group means and standard deviations (in parentheses) shown.

Figure 2.

HRFs averaged across all participants in the CAA (blue lines) and HC (red lines) groups for the (a) motor (upper row) and (b) visual (lower row) regions. For the HC group, the visual HRFs showed a larger response than the motor HRFs (as demonstrated by the Y-axes). A difference in amplitude across the two tasks was less evident in CAA group. Error bars indicate standard error of the mean.

HRF: hemodynamic response function; CAA: cerebral amyloid angiopathy; HC: healthy control; SC: signal change.

For statistical analyses comparing HRF features of TTP, FWHM and AUC across the two groups, initial interrogation of the data revealed one CAA patient as an outlier in terms of visual TTP and FWHM. For these two HRF features, the values obtained for this patient was >3 standard deviations from both the mean TTP and FWHM for the CAA group, both in the positive direction. While no methodological issue such as movement or fitting error was identified as the underlying cause of these high values, these data were excluded from all statistical analyses to avoid violating the assumptions of the implemented statistical tests. This outlier was also excluded from the group-averaged data shown in Table I and Figure 2.

The mixed ANCOVA including repeated-measures factors of HRF feature (TTP, FWHM, AUC) and Task (motor, visual), independent-measures factor of either Group (CAA, HC), and covariates hypertension and age was tested to determine whether statistical assumptions had been violated. These assumptions include sphericity and homogeneity of variance and regression slopes. Sphericity is a concern affecting experimental designs with within-subjects factors. Because data points are collected from the same subjects, it is assumed that the differences between the within-subjects factors are equal in variance. For instance, when calculating the difference between TTP and FWHM for each subject, it is assumed that the variance of these differences is equal to the variance of the differences between TTP and AUC, and FWHM and AUC. The violation of this assumption results in an increase in the probability of a Type II error. Mauchly's Test of Sphericity was non-significant, indicating that this assumption had not been violated. Homogeneity of variance was assessed using Levene's Test for Equality of Error Variances. This test assessed error variance across groups. It was found to be non-significant at every level of the between-subjects factor, indicating that variance was equal across the two groups. Homogeneity of regression slopes is an assumption underpinning the ANCOVA test and assumes that the relationship between the covariate(s) and the dependent variable is equal across groups. This was assessed by individually testing the interaction terms between group (CAA, HC) and the covariates (age, hypertension). Both interactions were non-significant, indicating that this assumption had not been violated.³⁶

Overall, the mixed ANCOVA revealed two significant interactions, one for Task and Group, F(1, 22) = 5.1, P = 0.03 and one for HRF feature and Group, F(2, 44) = 3.5, P = 0.04. There were no significant effects of the covariates. Post hoc ANCOVAs were then performed to determine how the visual and motor HRFs differed between the two groups. These post hoc tests revealed that there were significant group differences in the HRFs estimated from the motor cortex in terms of TTP, F(1, 23) = 5.4, P = 0.03 and FWHM, F(1, 23) = 9.3, P = 0.006. For these motor HRFs, the TTP was longer and FWHM wider for the CAA group compared to the HC. The TTP of the motor HRFs for the HC group demonstrated a wider range of values than the CAA group, as indicated by the large error bars in the upper panel of Figure 2 and in panel B of Figure 3. For the visual HRFs, post hoc tests indicated a significant group difference for AUC, F(1, 23) = 11.9, P = 0.002. The HRFs estimated from the visual cortex of the CAA patients showed reduced AUC compared to the HC group, indicating attenuated HRFs in the visual cortex of the CAA group. The ANCOVAs comparing the two groups in terms of visual TTP and FWHM failed to reach significance. No post hoc test showed significant effects of the covariates, indicating that age and hypertension were unrelated to HRF features. Means and standard errors of the three HRF features for both CAA and HC groups are displayed in Figure 3.

Figure 3.

Bar graphs showing mean FWHM (a), TTP (b) and AUC (c) of visual and motor HRFs for both CAA and HC groups. P-values from post hoc one-way ANCOVAs. Error bars indicate standard error of the mean.

s: seconds, SC: BOLD signal change; FWHM: full-width at half-maximum; TTP: time-to-peak; AUC: area under the curve; HRF: hemodynamic response function; CAA: cerebral amyloid angiopathy; HC: healthy control; ANCOVA: analysis of covariance.

Structural MRI

The CMB count distribution and WMH volumes deviated from normality and were therefore log-transformed. These transformed values were entered into bivariate correlation analyses. The results from these analyses are shown in scatterplots in Figure 4. Overall, there was only one significant correlation found, between CMB count and FWHM for the visual HRF (r = .62, P = .03.). The corresponding correlation between CMBs and motor FWHM failed to reach significance (r = −.31, P = .33). No other significant correlations were found between CMBs and HRF features. The only correlation that appeared to be approaching significance was the AUC for visual HRF (r = .51, P = .09), but this was non-significant with the current sample size. The CAA patients (M = 1.1, SD = .38 mL) showed significantly greater WMH volumes than the HC (M = .49, SD = .33 mL), t(23) = 4.3, P < .0005. However, there were no significant correlations found between HRF feature and WMH.

Figure 4.

Scatter plots showing results of the bivariate correlation analyses between CMBs (log-transformed) and HRF features of FWHM, TTP and AUC for visual (a) and motor (b) HRFs.

s: seconds; SC; BOLD signal change; CMB: cerebral microbleed; HRF: hemodynamic response function; FWHM: full-width at half-maximum; TTP: time-to-peak; AUC: area under the curve; HRF: hemodynamic response function.

Discussion

In this study, we show that visual and motor HRFs estimated from block designs differ between patients with CAA and a healthy, aged-matched, control cohort. We found that the HRFs estimated from the visual cortex of the CAA group were reduced in AUC relative to the controls, and a wider FWHM was positively correlated with CMB count. It was hypothesized that patients with CAA would demonstrate HRF deviations from the control group for the visual, but not the motor HRFs. The motor HRFs estimated from the patient cohort, however, showed significantly delayed TTP and wider FWHM than the control group. This study is, to the best of our knowledge, the first to demonstrate that HRFs estimated from block designs are altered in a patient cohort and in particular, in patients with CAA.

Globally altered BOLD responses

We found here that the peak timing and width of the HRFs estimated from the motor cortex of patients with CAA significantly deviated from those estimated from the HC group. This outcome was unexpected as the motor cortex is removed from the posterior brain regions that are most heavily affected by CAA. One possible explanation is that the affected brain regions extend beyond the posterior circulation. CAA pathology most commonly affects the occipital lobes, although involvement of other cortical regions is also frequent. With no current imaging biomarker sensitive and specific to vascular amyloid load, determining the cerebral location most heavily affected by CAA is not possible in vivo. However, we can only speculate about a relationship between CAA pathology and altered motor HRFs, considering none of the HRF features estimated from this region correlated with CMB count or WMH volume. Another explanation is that neurovascular changes extend beyond the more heavily afflicted occipital region. Neurovascular-coupling changes in patients with localized cerebrovascular damage have found to extend beyond the affected brain region. These changes have been demonstrated in fMRI studies through BOLD signal magnitude and, in studies utilizing event-related designs, alterations in estimated HRFs. This has been demonstrated in patients with vascular obstruction,³⁷ stenosis,³⁸ aphasia due to stroke and neurodegeneration^17,39 and ischemia.¹⁶ Altered BOLD responses in these patients have implicated global changes in cerebral blood flow (CBF) and coupling between local hemodynamics and neural activity. In a study utilizing magnetoencephalography (MEG) and BOLD fMRI, neurophysiological and cerebrovascular responses to sensory stimulation were investigated in patients with a positive history of an ischaemic injury and HC.⁴⁰ Both neurophysiological and BOLD responses to median nerve electrical stimulation were intact for the control group. For the patient group, neurophysiological responses recorded with MEG were intact, however, suprathreshold BOLD responses to stimulation were undetected for many participants. Further investigation with transcranial Doppler during a hypercapnic challenge found a strong relationship between impaired vasomotor reactivity and BOLD activation in these patients. No relationship between BOLD activation and lesion site was found. Similarly, in an fMRI study of subcortical stroke patients, BOLD responses to a motor task were reduced in magnitude relative to a control group, despite a lack of motor impairment in the study participants.⁴¹ These studies, consistent with our motor HRFs reported here, demonstrate that BOLD responses removed from the preferentially affected brain region can be altered due to extended cerebral changes in the presence of vascular injury.

Locally altered BOLD responses

We found here that the AUC of the HRF to visual stimulation was significantly smaller in patients relative to controls. Prior studies have demonstrated decreased BOLD response magnitude in the visual cortex of CAA patients.^11,12 Here we did not report BOLD response magnitude as this had been already described by these prior studies, although it is important to note that AUC is related to both signal magnitude and FWHM. Without further interrogation, a reduced AUC could ambiguously be attributed to a small signal magnitude or narrow FWHM. Because we assessed FWHM and found no group differences between CAA and HC visual HRFs in terms of this, it can be assumed that the reduced visual AUC is due to a small BOLD response magnitude. However, we did not observe a relationship between visual AUC and CAA pathology as identified on structural MRI. The significant correlation between FWHM of the visual HRF and CMB count indicates that this HRF feature is related to CAA pathology, although it did not significantly differ between the two groups. These findings support an association between the HRF and CAA pathology, which demonstrates the utility of fMRI as a method for the visualization of CAA progression. These data may also benefit signal detection of future fMRI studies of CAA patients, where basis function/s flexible in terms of width may be necessary to accurately detect BOLD activation in patients with high CMB count.

While the BOLD signal magnitude is very commonly used as an indicator of neural activity, the present findings of reduced BOLD signal in the visual cortex of CAA patients could complicate this interpretation. Research conducted with transcranial Doppler ultrasound⁸ has suggested that a reduced BOLD signal to visual stimulation in CAA patients may be due to impaired vasoreactivity to increased metabolic demand, rather reduced neural activity alone. This is an important caveat that requires consideration for future work using fMRI as a technique to probe neural activity in patients with CAA and cerebrovascular disease in general. The use of quantitative fMRI approaches has been proposed to address this issue.⁴²

HRF characterization

In the present study, we show that HRFs estimated from block designs are altered in patients with CAA relative to controls, and that the FWHM of HRFs estimated from the visual cortex correlate with CAA pathology detected on structural MRI. Despite this, it remains to be determined if HRF alterations detected in patients with cerebrovascular injury affects signal detection in block designs. It is assumed that an accurate HRF model is essential for signal detection in event-related fMRI, and improved detection using optimized HRFs for patients with cerebrovascular injury has been demonstrated.⁴³ Alternately, recent evidence suggests that altered HRFs in the presence of vascular-related injury do not influence signal detection in block designs.⁴⁴ In this study, individualized HRFs were created from an event-related scan and subsequently used in a block design to model BOLD signal change within a general linear model (GLM). Compared to the canonical HRF, the use of the individualized HRFs did not significantly improve signal detection. These results pertain to the study group encompassing both cortical and subcortical stroke patients, however the authors noted that the largest differences were observed between individuals with cortical stroke lesions and HC. This observation is consistent with the results reported here, as vascular amyloid deposition in CAA largely affects the cortical, rather than subcortical brain regions.⁴⁵

Further work determining if block designs analysed with individualized HRFs improve signal detection in patients with CAA is warranted. When the estimation of individualized HRFs is not possible, the use of a cohort template such as the group-averaged CAA HRF presented here (Table 1) may improve signal sensitivity. The parameters in Table 1 can be inserted into equation (1) to create a cohort-specific template, which can be used in conjunction with fMRI processing software for the modelling of BOLD activation in CAA patients within a GLM. It is clear from the results of the present study that HRFs estimated from block designs are significantly different between HC and CAA patient groups. This provides support for prior work demonstrating that the BOLD signal is sensitive to CAA-related neurovascular changes^11,12 and the use of fMRI as a method for identifying CAA disease progression.

Limitations and future directions

In the present study, inter-subject variability was observed in HRF shape across all subjects and, for the CAA group, in CMB count and WMH. For the latter, Figure 4 demonstrates the inter-subject variability in CMB count. Variability in HRF shape was observed across both groups, as demonstrated by error bars in Figures 2 and 3. The variability observed across the patient group, particularly for the structural MR measures, may be associated with CAA severity. Severity ranges from localized amyloid deposition within a minor proportion of superficial cortical vessels in mild cases, to the majority of small vessels being affected in severe cases.⁴⁵ It may be advantageous for future work to divide CMB count and WMHs into lobar regions, or to stratify the CAA group by severity in order to reduce some of the inter-subject variability observed here, if a sufficient sample size permits. It may also be beneficial to estimate HRFs in a voxel-wise manner, rather than from the average of activated voxels within defined brain regions as performed here. HRFs estimated from individual voxels may benefit the patchy spatial distribution of CAA by providing more insight into how the HRF may vary within a discrete region of cortex in these patients.

Within-subject variance was particularly salient for the HC group in terms of AUC and TTP. Some of this intra-subject variability is not unexpected, in particular the AUC differences between the motor and visual tasks. Large BOLD responses arising from the visual cortex are common and may be explained by normal regional heterogeneity in CBF. Studies of healthy adults often report higher resting CBF in the visual cortex compared to whole brain values.^46,47 For much of the remaining variability observed here, nuisance effects such as subject differences in task compliance, image acquisition, and physiological noise are standard contributors to BOLD signal variance in fMRI and can be reduced with increasing sample size.

A limitation of our approach for HRF estimation is that we selected voxels for fitting using a GLM, modelling the BOLD signal change with FIR basis functions. Any method requiring the selection of a subset of voxels for further statistical analysis raises the issue of circularity through non-independent voxel selection.⁴⁸ We attempted to reduce this concern by limiting voxel selection to those falling within anatomical ROIs, and by implementing FIR basis functions, as these impose no assumptions regarding the shape of the response. Other methods based on deconvolution have been proposed, which are able to estimate the HRF at every voxel using a stimulus function alone.^18,49 It is currently unknown if these methods can be used in the context of block designs, however this is one avenue for further research.

Conclusion

In summary, we demonstrate here that HRF shape and timing features estimated from patients with a diagnosis of probable CAA differ to those obtained in a group of HC. In the motor cortex, we saw group differences in terms of time delay and width, while the visual cortex demonstrated smaller HRFs for the patient group. In the visual cortex only, we found that the FWHM of the HRF was positively correlated with CMB count. Estimating HRFs from block designs is a useful technique for studies investigating BOLD signal changes associated with cerebrovascular disease, however, further work is required to determine if individualized HRFs improve signal detection in patients with cerebrovascular injury.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The study was funded by operating grants from the Canadian Stroke Network, Heart and Stroke Foundation of Canada, and the Alzheimer Society of Canada. RJW received funding from the NSERC CREATE International and Industrial Imaging Training Program. GBP received funding from Campus Alberta Innovates and CIHR FDN-143290.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Authors' contributions

All authors made substantial contributions to editing the manuscript. Data were analysed and interpreted by RJW and GBP. RJW wrote the manuscript. SP and CRM collected the data and performed data analyses. CRM, BGG, RF and EES conceived and designed the study.