Association of midlife stroke risk with structural brain integrity and memory performance at older ages: a longitudinal cohort study

Abstract Cardiovascular health in midlife is an established risk factor for cognitive function later in life. Knowing mechanisms of this association may allow preventative steps to be taken to preserve brain health and cognitive performance in older age. In this study, we investigated the association of the Framingham stroke-risk score, a validated multifactorial predictor of 10-year risk of stroke, with brain measures and cognitive performance in stroke-free individuals. We used a large (N = 800) longitudinal cohort of community-dwelling adults of the Whitehall II imaging sub-study with no obvious structural brain abnormalities, who had Framingham stroke risk measured five times between 1991 and 2013 and MRI measures of structural integrity, and cognitive function performed between 2012 and 2016 [baseline mean age 47.9 (5.2) years, range 39.7–62.7 years; MRI mean age 69.81 (5.2) years, range 60.3–84.6 years; 80.6% men]. Unadjusted linear associations were assessed between the Framingham stroke-risk score in each wave and voxelwise grey matter density, fractional anisotropy and mean diffusivity at follow-up. These analyses were repeated including socio-demographic confounders as well as stroke risk in previous waves to examine the effect of residual risk acquired between waves. Finally, we used structural equation modelling to assess whether stroke risk negatively affects cognitive performance via specific brain measures. Higher unadjusted stroke risk measured at each of the five waves over 20 years prior to the MRI scan was associated with lower voxelwise grey and white matter measures. After adjusting for socio-demographic variables, higher stroke risk from 1991 to 2009 was associated with lower grey matter volume in the medial temporal lobe. Higher stroke risk from 1997 to 2013 was associated with lower fractional anisotropy along the corpus callosum. In addition, higher stroke risk from 2012 to 2013, sequentially adjusted for risk measured in 1991–94, 1997–98 and 2002–04 (i.e. ‘residual risks’ acquired from the time of these examinations onwards), was associated with widespread lower fractional anisotropy, and lower grey matter volume in sub-neocortical structures. Structural equation modelling suggested that such reductions in brain integrity were associated with cognitive impairment. These findings highlight the importance of considering cerebrovascular health in midlife as important for brain integrity and cognitive function later in life (ClinicalTrials.gov Identifier: NCT03335696).


Participant inclusion/ exclusion
VBM analysis was based on a final available sample of N = 566, TBSS on N = 548, and SEM on N = 775. Participants with no (N = 25) or inadequate quality T1-weighted scan (N = 1), missing (N = 12) or un-useable dMRI scan (N = 6), and missing Framingham stroke risk score at any of the five data waves (N = 196), were excluded from whole-brain analysis. Following careful inspection of all T1weighted, FLAIR and dMRI images by an old-age psychiatrist with experience in neuroimaging (KPE), scans were excluded from whole-brain analysis due to the following structural abnormalities: large cyst (N = 7), meningioma (N = 1), infarction or stroke (N = 4) obvious on a T1-weighted and FLAIR image. Incidental findings were further dealt with according to the FMRIB Centre internal protocol and processed by the associated neuroradiologists and neurologists without the involvement of the research team. At the time of scan, 18 participants reported a suspected stroke or TIA event. Out of these, 11 had a missing FSRS score (by definition the score can only be computed to predict a first stroke) and were removed from whole-brain analysis. Four participants had sub-optimal quality T1 images that could not be segmented by FAST (total N = 771), two by FreeSurfer (N = 773), and five by BIANCA (N = 770). One participant did not have a FLAIR scan therefore the WMH volume was imputed by regression (regressing WMH on age, sex, ethnicity, scanner type, years of education, socioeconomic status and FSRS between 2012-2013, to derive the regression coefficients, which can be used to calculate the WMH volume).

MRI acquisition and analysis
Cortical atrophy: Brain images were segmented and the total grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) volumes were extracted using FAST and FreeSurfer. FAST provides a partial volume image for each class, where each voxel contains a value in the range 0-1 that represents the proportion of that class's tissue present in that voxel (Zhang et al., 2001). Cortical atrophy was estimated by scaling the GM values for the total intracranial volume (GM + WM + CSF) resulting in percentage total GM volume. Total intracranial volume (TIV) was estimated with FreeSurfer version 5.3.

Hippocampal volume:
FreeSurfer is a set of tools for a fully automated structural imaging analysis and visualization of brain imaging data. The volume-based subcortical stream of version 5.3 was used to pre-process MRI volumes and label subcortical tissue classes using the DKT atlas (?h.aparc.DKTatlas40.annot) (Fischl et al., 2002). Hippocampal volume was used in the current analysis.
Grey matter density: Voxelwise analysis of GM was performed using FSL-VBM (Douaud et al., 2007) an optimised voxel-based morphometry (VBM) protocol (Good et al., 2001). Normalized biascorrected brain extracted images were grey matter segmented before being registered to the Montreal Neurological Institute (MNI) 152 standard space using FMRIB's non-linear registration tool (FNIRT (Andersson et al., 2007)). The images were averaged and flipped along the x-axis to create a left-right symmetric, study-specific GM template. All native GM images were non-linearly registered to this study-specific template and modulated to correct for local expansion (or contraction) due to the non-linear component of the spatial transformation. The modulated GM images were then smoothed with an isotropic Gaussian kernel with a sigma of 3 mm. Finally, a study-specific GM mask was used to enable accurate localization of results in an ageing sample.
White matter microstructure. Diffusion tensor imaging (DTI) quantifies the directionality and rate of diffusion of water molecules within different tissues and allows inferences about the structural integrity of WM tracts. When movement is anisotropic, such as in healthy myelinated fibres, diffusion is restricted perpendicular to the longitudinal axis of the fibre. DTI images were corrected for susceptibility-induced distortions using two b=0 scans (b-value 0 s/mm 2 ), acquired with opposing phase-encoding directions using the FSL 'topup' tool (Andersson et al., 2003). Eddy current-induced distortions and subject movement were corrected using the FSL 'eddy' tool . This employs a second order polynomial model and transforms each scan towards the Gaussian process predicted scan. It also identifies outlier slices (dropout) caused by movement during diffusion encoding . Slices were classified as outliers and replaced if the signal was found to be more than 3 SD from the Gaussian Process predicted slice. The volume was removed if over 10 slices were identified as outliers within a volume. The scan was excluded from analysis if more than five volumes were removed. Voxelwise statistical analysis of fractional anisotropy (FA) and mean diffusivity (MD) was carried out using tract-based spatial statistics ((TBSS) (Smith et al., 2006)). A tensor model was fitted to the pre-processed diffusion data using DTIFit part of FMRIB's Diffusion Toolbox (http://fsl.fmrib.oc.ac.uk/fsl/fdt) to create FA and MD maps. This fits a diffusion tensor model to the pre-processed diffusion data and then brain-extracts using BET (Smith, 2002). All participant's FA data were then aligned into a common space using FNIRT (Andersson et al., 2007), which uses a b-spline representation of the registration warp field (Rueckert et al., 1999). Next, the mean FA image was created and thinned to create a mean FA skeleton, which represents the centers of all tracts common to the group. Each participant's aligned FA data were then projected onto this skeleton. The same non-linear warps and skeleton projections calculated for FA were then applied to MD. The resulting skeletonized FA and MD data were fed into voxelwise cross-subject statistics.
White matter hyperintensity (WMH) volume. WMHs were automatically segmented on FLAIR images with FMRIB's Brain Intensity AbNormality Classification Algorithm (BIANCA; (Griffanti et al., 2016)). BIANCA is a fully automated, supervised method for WMH detection, based on the k-nearest neighbour algorithm. It classifies the image's voxels based on their intensity and spatial features, where the intensity features are extracted from FLAIR, T1 and fractional anisotropy (FA) images.
BIANCA offers options for weighting the spatial information, local intensity averaging and the choice of the number and location of training points. The following additional options were used: local average intensity within a 3D kernel of size = 3 voxels, MNI coordinates as spatial features with a weighting factor of 2, and 2000 lesion points and 10000 non-lesion points (avoiding the lesion border) for each of the 24 manually segmented images used as training dataset. The output image represents the probability per voxel of being WMH. The total WMH volume was calculated from voxels exceeding a probability of 0.9 of being WMH located within a white matter mask and then adjusted for the total intracranial volume to obtain a percentage WMH volume.

Framingham Stroke Risk score
Risk components were drawn from both questionnaires and clinical examinations at 5 waves between 1991 and 2013. Risk components were collected according to standard operating protocols (Kivimaki et al., 2012;Kaffashian et al., 2013) detailed below.
Biological measures and samples were collected according to standard operating protocols (Kivimaki et al., 2012;Kaffashian et al., 2013). Systolic blood pressure was measured twice in sitting position after five minutes rest with a Hawskley random-zero sphygmomanometer (phases 3 and 5; Lynjay Services Ltd, Worthing, UK) and OMROM HEM 907 (phase 7 onwards; Omron, Milton Keynes, UK).
The average of two readings was used in analysis. Atrial fibrillation (AF) and left ventricular hypertrophy (LVH) were identified on a standard 12-lead electrocardiogram analysis programme combined with manual review and the Minnesota code classification system for electrocardiographic findings (AF arrhythmias: 8-3-1; LVH high amplitude R-waves: 3-1) (Prineas et al., 1982). History of prior cardiovascular disease (CVD) was based on electrocardiogram and angiogram examinations at phases 1, 3 and 5 as well as from general practitioner or hospital records. Hypertensive medication use was self-reported. Venous blood was taken in the fasting state or at least five hours after a light, fat-free breakfast before undergoing a standard 2h oral glucose tolerance test. Glucose was measured in fluoride plasma by an electrochemical glucose oxidase method (Cooper, 1973) on YSI model 23A glucose analyser at Phase 3 (Alpert, 1976) and YSI model 2300 STAT PLUS analyser at Phase 5 onwards (Astles et al., 1996) (YSI Corporation, Yellow Springs, OH, USA). Diabetes mellitus was defined by a fasting glucose level of ≥ 7.0 mml/L, a 2hr post-load glucose level of ≥ 11.1 mml/L, self-reported diabetes diagnosed by a doctor or use of diabetes medication (Expert Committee on the and Classification of Diabetes, 2003). Participants were categorised as current cigarette smokers or past/non-smokers.

Assessment of cognition and premorbid functioning
The Hopkins Verbal Learning Test-Revised (HVLT-R) provides a measure of verbal learning and memory ability (1999; 2005), where the participant is required to learn a list of twelve words over the course of three trials, and recall or recognise them at increasing time intervals. The sum of items recalled after a 30-minute period (Total Delayed Recall; HVLT-DR (Wechsler, 2011)) was used for analysis.
The Test of Premorbid Functioning (TOPF) (Wechsler, 2011) consists of a list of seventy written words, which must be read aloud and is marked according to pronunciation. It is used to estimate an individual's level of intellectual functioning (full scale intelligence quotient) before the onset of injury or illness.