Study population

This study uses previously collected data from a cross-sectional imaging study on cSVD [14]. We included patients with clinically manifest cSVD, consisting of patients with lacunar stroke or mild vascular cognitive impairment (mVCI) [14]. Lacunar stroke was defined as an acute lacunar stroke syndrome with a compatible recent small subcortical infarct on clinical brain MRI. If no such lesion was visible on MRI, or if no acute clinical brain MRI scan was performed, established clinical criteria for lacunar stroke syndrome were used [15, 16]. Stroke patients were included at least 3 months post-stroke to avoid acute stroke phase changes. Inclusion criteria for mVCI consisted of subjective cognitive complaints, objective cognitive impairment in at least one cognitive domain determined by neuropsychological assessment, and vascular lesions on brain MRI that suggested a link between the cognitive deficit and cSVD: moderate to severe WMHs (Fazekas score deep > 1 and/or periventricular > 2) or mild WMH (Fazekas score deep = 1 and/or periventricular = 2) combined with lacune(s) or microbleeds [17, 18]. Exclusion criteria included a suspicion of a neurodegenerative disease (e.g., Alzheimer’s disease), severe cognitive impairment (defined as a clinical dementia rating score > 1 or a Mini-Mental State Examination (MMSE) score of < 20), and other central nervous system diseases or contraindications for MRI. Lacunar stroke patients were also excluded in case of a symptomatic carotid stenosis of ≥ 50% or a possible cardioembolic source (e.g., atrial fibrillation). Participants were recruited from the Maastricht University Medical Centre and Zuyderland Medical Centre, the Netherlands. The study has been approved by the Medical Ethics Committee of the Maastricht University Medical Centre (Dutch Trial Register; NTR3786). All participants gave written informed consent.

General and health characteristics

The presence of cardiovascular risk factors (i.e., hypertension, diabetes mellitus, hypercholesterolemia, smoking) was obtained from self-reported medical history and medication use. Body mass index (BMI) was calculated as weight in kilograms divided by height in meter squared. Educational level was registered and categorized based on the Dutch classification system of Verhage [19].

Brain MRI acquisition

All patients underwent brain imaging on a 3-T MRI scanner (Achieva TX, Philips Healthcare, Best, The Netherlands) using a 32-element head coil suitable for parallel imaging. For anatomical segmentation a T1-weighted sequence (TR/TI/TE = 8.3/800/3.8 ms; field of view (FOV) 256 × 256 × 160 mm3; 1.0-mm cubic voxel size) and a T2-weighted fluid-attenuated inversion recovery (FLAIR) sequence (TR/TI/TE = 4800/1650/299 ms; FOV 256 × 256 × 180 mm3; 1.0-mm cubic voxel size) were performed.

DCE-MRI, providing measures of BBB integrity, was performed as described before [20]. Briefly, we used a dual-time resolution DCE-MRI to capture the fast signal changes during bolus arrival (fast sequence) and accurately measure the whole-brain tissue signal changes (slow sequence). Pre-contrast scans of both the fast and slow sequences were acquired, and a quantitative pre-contrast T1 relaxation time map was calculated to convert the contrast-enhanced signal intensities to concentrations in tissue. Subsequently, contrast agent (gadobutrol; dose 0.1 mmol/kg body weight) was injected.

IVIM imaging, providing measures of microvascular perfusion and the perivascular clearance system, was performed as described before [14]. In brief, a Stejskal-Tanner diffusion weighted (DW) spin echo single shot echo planar imaging pulse sequence (diffusion sensitization in anterior–posterior direction, b values 0, 5, 7, 10, 15, 20, 30, 40, 50, 60, 100, 200, 400, 700, and 1000 s/mm2) was used, which included an inversion recovery pulse (TI = 2230 ms) to suppress the cerebrospinal fluid.

MRI analysisBrain segmentation

Grey and white matter were automatically segmented on T1-weighted images (Freesurfer software) [21]. WMHs were automatically segmented on FLAIR images [22] and manually corrected by a trained investigator, and infarcts were excluded. For the current study, we restricted our main analysis to physiological MRI measures in the normal appearing white matter (NAWM), as the NAWM is regarded as ‘tissue at risk’, and multiple previous studies have shown early pathophysiological alterations in the NAWM in patients with cSVD [23,24,25]. As a secondary analysis, we also investigated physiological MRI measures in WMH.

DCE-MRI analysis

Analysis of the DCE-MRI data consisted of pharmacokinetic modelling and subsequent histogram analysis, as described previously [26]. The contrast agent concentration in tissue was calculated by using the relative signal enhancement and the quantitative T1 maps, and the vascular input function was derived from the superior sagittal sinus [27]. The graphical Patlak method was applied to the tissue concentration curves over time in a voxel-wise manner to derive the BBB leakage rate Ki (min−1) maps. The mean Ki of the NAWM and WMH was calculated. Additionally, the leakage volume (VL) was calculated using the histogram method as described before [23], representing the fractional volume of BBB leaking tissue that can be detected (i.e., the spatial extent of leakage). As Ki and VL provide complementary information on BBB permeability, both variables were included in the CCA model as (inverse) measures of the NVU function ‘regulation of BBB integrity’.

IVIM MRI analysis

Pre-processing of DW images consisted of corrections for head displacements and spatial distortion (echo planar imaging and eddy current distortions) (FSL) [14]. A three-component IVIM (3C-IVIM) model was assumed, which describes the effects of the microcirculation, interstitial and perivascular fluid, and microstructural integrity on the DW signal. This 3C-IVIM model was fitted to the DW signal decay curves in a voxel-wise manner using the state-of-the-art physics-informed neural network fitting approach as described recently [28]. Although altered microstructural integrity may be a consequence of diminished NVU function, it does not provide a direct measure of NVU function and is therefore further disregarded. The extracted NVU function measures of interest included the microvascular diffusivity (D*), along with its corresponding microvascular perfusion volume fraction (f) and the intermediate volume fraction (fint). These 3C-IVIM measures were averaged within the NAWM and WMH. The D* reflects the fast directional changes in the microcirculatory bloodstream and depends on the microvascular blood velocity and the architecture of the microvascular bed. The f represents the volume of blood flowing through the capillaries. The microcirculation-related parameters D* and f were included in the CCA model as measures of the NVU function ‘regulation of CBF’. The fint represents the interstitial fluid either between the parenchymal cells or within perivascular spaces, and is proposed to be indicative of the perivascular clearance system [29]. fint was therefore included in the CCA analysis as an (inverse) measure of the NVU function ‘regulation of perivascular clearance pathways’.

Neuropsychological assessment

We performed an extensive neuropsychological assessment covering three main cognitive domains: memory, executive function, and information processing speed. To examine memory, we performed the Rey Auditory Verbal Learning Test (RAVLT) (immediate recall, delayed recall, and delayed recognition) and the Digit Span Forward (subtest of Wechsler Adult Intelligence Scale (WAIS)-III) [30, 31]. To assess executive function, we used the Stroop Color-Word Test (SCWT) interference score (time of part 3 minus mean time of parts 1 and 2), Trail Making Test (TMT) interference score (time of part B minus time of part A), Category (animals and professions) and Letter Fluency, Letter-Number Sequencing (subtest of WAIS-III), and Digit Span Backward (subtest of WAIS-III) [31,32,33,34,35]. For examination of information processing speed, we performed the Symbol Substitution-Coding (subtest of WAIS-III), TMT part A, and SCWT parts 1 and 2 [31,32,33]. The scores of tests with higher scores representing worse performance (i.e., SCWT and TMT) were inverted, to simplify interpretation. All neuropsychological test scores were transformed into sample-based z scores (by dividing the difference between the individual raw test score and the overall group sample mean by the overall group sample standard deviation). For additional analysis, cognitive domain compound scores were determined by averaging the z scores of all tests within one domain. When one test score was missing, compound scores were calculated from the scores of the remaining tasks. In two patients, more than one test score in the executive domain was missing; for these patients, no reliable domain score could be calculated.

Canonical correlation analysisGeneral concepts of CCA

CCA identifies sources of common variation and aims to correlate two multivariable datasets. CCA finds linear combinations of the variables of each set, such that the correlation between the linear combination of the one set and the linear combination of the other set is maximal. This results in multiple, uncorrelated, canonical modes. The number of canonical modes that is extracted is equal to the minimal number of variables in either of the variable sets [12].

The linear combination of the variables within one set, or the (latent) canonical variate, is computed from the weighted sum of the original variables as indicated by the canonical weight. Canonical loadings represent the correlation between a variable and its corresponding canonical variate. Canonical loadings are used for the interpretation of the nature of the relationship and explaining underlying constructs [36, 37]. Moreover, signs of the canonical loadings yield directional information about the variables’ contributions to the canonical relationship.

The canonical correlation is the Pearson correlation between the canonical variates of each set. Canonical correlations were squared to compute the proportion of variance shared by the linear composites of the two variable sets (the latent canonical variates). Using the canonical loadings, we computed the proportion of variance in the variable set which is extracted by its canonical variate. Additionally, a measure of redundancy was computed, which reflects how redundant one set of variables is, given the other set of variables. It provides a measure of the ability of a set of variables to explain variation in the other set of variables [37].

CCA for NVU function and cognitive performance

In this study, we investigated the correlation between NVU function in the NAWM and cognitive performance (both latent variates, composed of a set of measurable variables). NVU function variables (Ki, VL, f, D*, and fint) were fed to the CCA model along with the 13 cognitive test scores. All variables were standardized by z score transformation (z = (subject value – population mean) / population SD). Five canonical modes were extracted (equal to the minimal number of variables within either of the variable sets; in this case NVU variables), and for each estimated mode, p values were calculated using random permutation testing (Wilk’s lambda, 1000 permutations). For each statistically significant mode, we obtained the Pearson correlation, the proportion of shared variance between the two latent canonical variates, and the variables’ canonical loadings. The proportion of the variables’ explained variance by its canonical variate and the redundancy were also obtained.

Additional analyses

Additional analyses were performed to assess result robustness. Firstly, CCA was repeated with all variables individually adjusted for age, sex, and educational level, and additionally for relative brain volume and relative WMH volume, using linear regression, to determine whether this had any impact on the results. Next, CCA was repeated with 3 cognitive domain z scores instead of 13 individual test scores to evaluate whether results remained similar.

As a secondary analysis, we investigated the correlation between NVU function in WMH and cognitive performance. Analyses were performed with SPSS software (v28.0; IBM, Chicago) and R (v4.3.3; R Foundation for Statistical Computing, Vienna). A p value lower than 0.05 was considered statistically significant.