Heterogeneity of cerebral white matter lesions and clinical correlates in older adults

Abstract

Background and Purpose:

Cerebral white matter signal abnormalities (WMSA) are a significant radiological marker associated with brain and vascular aging. However, understanding their clinical impact is limited due to their pathobiological heterogeneity. We determined whether use of robust reliable automated procedures can distinguish WMSA classes with different clinical consequences.

Methods:

Data from generally healthy participants aged > 50 years with moderate or greater WMSA were selected from the Human Connectome Project-Aging (n = 130). WMSA were segmented on T1 imaging. Features extracted from WMSA included total and regional volume, number of discontinuous clusters, size of non-contiguous lesion, contrast of lesion intensity relative to surrounding normal appearing tissue using a fully automated procedure. Hierarchical clustering was used to classify individuals into distinct ‘classes’ of WMSA. Radiological and clinical variability was evaluated across the individual WMSA classes.

Results:

Class I was characterized by multiple, small, lower-contrast lesions predominantly in the deep WM; class II by large, confluent lesions in the periventricular WM; and class III by higher-contrast lesions restricted to the juxtaventricular WM. Class II was associated with lower myelin content than the other two classes. Class II was more prevalent in older subjects and was associated with a higher prevalence of hypertension and lower physical activity levels. Poor sleep quality was associated with a greater risk of class I.

Conclusions:

We classified heterogeneous subsets of cerebral white matter lesions into distinct classes that have different clinical risk factors. This new method for identifying classes of WMSA will be important in understanding the underlying pathophysiology and in determining the impact on clinical outcomes.

INTRODUCTION

Cerebral white matter signal abnormalities (WMSA) on magnetic resonance imaging (MRI) are a prevalent consequence of brain and vascular aging,¹ and are often associated with neurological complications.² WMSA represent complicated mechanisms including neurodegeneration, small vessel disease, low flow associated with large artery disease, embolism, and impaired waste clearance via the perivascular glymphatic transport.^2-5 While some WMSA progress rapidly and mediate cognitive decline, other manifestations seem to be relatively benign and silent. It is supposed that the impact of WMSA on brain function depends on their pathophysiological mechanism, but these white matter lesions are typically examined at a gross level. For example, prior studies have differentiated between ‘punctate’ and ‘confluent’ lesion patterns, spatial location (periventricular vs., deep or lobar) and tissue damage (with diffusion tensor imaging or magnetization transfer imaging).^6-10 Prior work by Maillard and colleagues described a ‘penumbra’ of white matter lesions that quantifies the health of the tissue surrounding the WMSA and may be used for advanced characterization of lesion.¹¹ However, minimal prior work has integrated a range of robust quantitative WMSA features to determine a specific WMSA ‘class’.

It is likely that mechanically, WMSA can be understood based on the anatomy and physiology of small vessels. The medullary arteries arising from the pial artery penetrate the white matter in a straight course and run toward the margins of the lateral ventricle, giving off a few long side branches, and terminating in an arborizing pattern.^{12, 13} The basal arteries arising from the middle cerebral artery perforate the base of cerebrum directly towards periventricular white matter.^{12, 13} Given that a border zone is formed between the medullary and basal arterioles that concentrate in the periventricular white matter, hypoperfusion in small vessels can cause ischemic tissue damage primarily in the periventricular white matter.¹⁴ On the other hand, the proximal portion of the medullary arteries has relatively few side branches. They are highly compliant and possess affordable perivascular spaces in their stream, which are advantageous for the perivascular transport of cerebrospinal fluid (CSF).^13-15 Thus, WMSA patterns that are confined to the deep white matter may result from deterioration in functional properties of vessels.

A range of clinical factors can impose different classes of burdens on different sections of white matter arteries.^{10, 16} Hence, individuals may show differing characteristics of WMSA according to their own clinical factors, and subsequently have distinct outcomes. We describe here a novel automated procedure for quantification of more comprehensive WMSA features and use these parameters to classify individuals into distinct WMSA classes to determine whether such composite classifications could be differentially linked to clinical features.

MATERIALS AND METHODS

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request. Data used in this work were acquired as part of the Human Connectome Project Lifespan Study and are available for download online.

Study population

Data for this study were obtained from the local participants enrolled into the Massachusetts General Hospital site of the Human Connectome Project Aging (HCP-A). Following the original effort in the young adults,¹⁷ this ongoing study is recruiting ‘typically aging’ adults (>35 years) without significant clinical diagnosis or brain diseases, but allowing participants with prevalent health conditions such as hypertension, diabetes, dyslipidemia, and obesity. The participants are all physically and cognitively healthy. As with detailed HCP-A protocols published recently,¹⁸ all potential participants are screened for cognitive abilities, and for other exclusionary health conditions, including major psychiatric disorders, stroke, brain tumors, and Parkinson’s disease. At the time of this study, the site enrolled 233 individuals who met eligibility criteria, completed full sessions of cross-sectional study stream and had successful image preprocessing. Study procedures were approved by the Washington University institutional review board and conducted according to the principles of the Declaration of Helsinki. Informed consent was acquired from all participants after assessing their capacity to consent.

Clinical profiles

At the baseline visit, participants were interviewed with various classes of questionnaires about demographic information, basic health status, medications currently being taken, trauma, sleep quality and physical activity. Blood laboratory data were obtained after overnight fasting status and included HbA1c, high sensitivity C-reactive protein (hsCRP), glucose, blood urea nitrogen (BUN), creatinine, and lipid profiles. Vascular health/burden factors included hypertension, diabetes, dyslipidemia, smoking, body mass index, blood pressure and pulse pressure. Framingham risk score (FRS), which stratifies a 10-year cardiovascular disease risk, was also measured with a combination of some of these risk factors and laboratory data as described.¹⁹ Sleep patterns in the last month were evaluated with the Pittsburgh Sleep Quality Index (PSQI) and poor sleeper was defined as an individual with a PSQI score ≥ 5.²⁰ Recent physical activity levels were informed with the International Physical Activity Questionnaire (short version of IPAQ). IPAQ provided the time spent in sitting, specific activity, and total activity by combination with a weighing factor for walking (3.3), and moderate (4.0) and vigorous (8.0) activities.²¹ Montreal Cognitive Assessment (MoCA) was done to assess the cognitive function.

Brain MR Imaging acquisition and preprocessing

The detailed MRI scan protocols were reported previously.²⁰ In brief, all participants were scanned on a 3-Tesla scanner (Siemens Prisma, Erlangen, Germany) using the Siemens 32-channel Prisma head coil. A multi-echo MPRAGE pulse sequence for T1-weighted scan and a variable-flip-angle turbo-spin-echo sequence (SPACE) for T2-weighted scan were utilized to achieve a higher signal to noise ratio (SNR) of the individual. A higher bandwidth (744 Hz/Px) and volumetric navigators were also applied for attenuating susceptibility-induced distortions and motion correction, respectively: 0.8 mm isotropic voxels; sagittal field of view = 256 x 240 x 166 mm; matrix size = 320 x 300 x 208 slices; for the T1, TR/TI=2500/1000 ms, TE=1.8/3.6/5.4/7.2 ms, flip angle = 8 deg, water excitation employed for fat suppression, up to 30 TRs allowed for motion-induced reacquisition; for the T2, TR/TE=3200/564 ms, turbo factor = 314, up to 25 TRs allowed for motion-induced reacquisition.²² Raw structural image data were preprocessed for reconstruction and segmentation using Freesurfer image analysis suite (surfer.nmr.mgh.harvard.edu/, version 6.0.0). We followed the technical details for structural image preprocessing as described in prior publications.^23-25 WMSA were segmented on T1 images with a fully-automated method based on their intensity and spatial information.^{24, 26} This approach may systematically underestimate the lesion volume by selecting only the most hypointense component of the lesions visible on T2/FLAIR,²⁷ but these measures are highly correlated with volumes from traditional FLAIR based imaging hyperintensities.²⁶ The T1- and T2-weight scans were also prepared for co-registration using FMRIB’s Linear Registration Tool and voxel-based calculation. Demyelination has been noted as a histological correlate of white matter lesions.²⁸ We generated T1/T2 ratio map, which can provide a quantitative metric related to tissue damage and has been linked to myelin content in previous work.^{29, 30}

Quantification of WMSA features

We assessed WMSA features which exhibit variation across individuals and could therefore provide information about differential causes and/or consequences. These variables included total and regional volume, number of non-contiguous lesion, size of non-contiguous lesion, contrast of lesion intensity relative to surrounding normal appearing tissue, and total burden of WMSA relative to the total white matter volume. Given the different clinical features and pathophysiological bases of WMSA depending on the relative location from the lateral ventricle,^{7, 8} WMSA were regionally classified into the juxtaventricular (within 3 mm from the ventricular surface), periventricular (between 3 and 13 mm) and deep (13 mm or further) white matter lesions.⁷ Each mask for regional localization was created using Freesurfer procedures of lateral ventricle segmentation, dilatation of 3 mm and 13 mm and subtracted and combined with resulting volumes (Figure 1A). The results yielded volumes with respect to number of voxels, where one voxel corresponded to 1 mm³ of brain tissue. The regional preference was also measured with periventricular to deep WMSA volume ratio (pdr). Since the multiplicity and volume of each lesion might be distinct across classes, we sub-segmented WMSA into cluster with contiguous voxels (size threshold = 1 mm), which share a common row, column, or slice (Figure 1B). Freesurfer function (mri_volcluster) automatically provided the cluster segmentation process and calculated the number of WMSA clusters (le.no). The volume per WMSA (le.vol) was calculated as a total WMSA volume divided by cluster number, which would provide information about the contiguous versus punctate dominance for each individual. The intensity data of WMSA is another essential measure to characterize the nature of WMSA. Since the intensity data from closely neighboring voxels (within 2 mm) undergo a highly localized normalization of values,³¹ a ratio of bordering WMSA values may provide a metric of the degree of localized damage. Hence, we further segmented WMSA and surrounding WM into WMSA core, inner layer (1 mm within the WMSA border), outer layer (1 mm away from the border), and normal appearing white matter (NAWM). This was performed by dilating or eroding the WMSA region by one voxel. Lesion contrast (le.ct) was calculated with an intensity ratio of the outer and inner layers of regionally segmented WMSA (Figure 1C). Also, given the decrease in total WM volume with brain aging, WMSA burden in each individual was obtained with a formula of WMSA volume * le.ct * 100 / total WM volume. The values of T1/T2 ratio were extracted from the WM regions labeled as WMSA in the T1/T2 ratio map.

Figure 1. WMSA variable quantification procedure.

(A) Segmentation and regional localization procedures of WMSA in whole brain slices. Each mask for regional localization was created using Freesurfer procedures of lateral ventricle segmentation, dilatation of 3 mm and 13 mm and subtracted and combined with resulting volumes. (B) Clustering procedure for WMSA number with non-contiguous voxels and each lesion volume measurement. Discrete lesions are labeled as different colors. (C) Segmentation of WMSA, border (inner and outer layers of WMSA) and NAWM. Lesion contrast is calculated with the average of intensity ratios of outer and inner layers of regionally segmented WMSA. T.vol: total WMSA volume, JV.vol: juxtaventricular WMSA volume, PV.vol: periventricular WMSA volume, D.vol: deep WMSA volume, le.no: number of non-contiguous lesion, le.vol: volume of each non-contiguous lesion, pdr: periventricular to deep WMSA volume ratio, le.ct: WMSA lesion contrast

WMSA classification and statistical analysis

In order to focus on the individual classification for adults with WMSA and to minimize the bias towards total volume as opposed to the feature classification, we selected older adults aged > 50 years with moderate or greater WMSA. Cluster analysis was performed with various WMSA variables to determine the dominant individual class. As the collinearity between variables can cause problems in clustering algorithms by giving an irrelevant or redundant weight, it is generally recommended to put an independent variable or variable index formed by combining variables into a cluster model. Of all acquired WMSA variables, we eliminated ones that showed high correlations between each other using Pearson correlation analysis (r > 0.5). Then, a cluster dendrogram was constructed through hierarchical clustering analysis with Euclidean distance and ward linkage method. While raw WMSA features were utilized as main variables for clustering model for clinical usefulness and convenience, key components derived from principal component analysis were also used for model validation. The accuracy of principal component analysis for predicting clustering analysis-based individual classes was evaluated using a multinomial logistic regression analysis. The total and regional WMSA volumes were log-transformed due to their skewed distribution. Categorical variables were analyzed by Pearson Chi-Square test. For continuous variables, either ANOVA and the Tukey test or the Kruskal-Wallis rank sum test and the Dunn test were performed after the Shapiro-Wilk test. In order to identify independent clinical factors associated with individual class, multinomial logistic regression analysis was done with variables that yielded significant differences in univariate analyses and showed the lowest Akaike information criterion (AIC) in the model. For every analysis, P values <0.05 were determined as statistically significant. R software (version 3.5.3, R Foundation) was used for statistical analyses.

RESULTS

Study population and individual classification for WMSA

Initially, 233 subjects were enrolled. A scatter plot was drawn with age and total WMSA volume to select older adult subjects with moderate or greater WMSA (x = age, y = total WMSA volume). A total of 130 study subjects (age: 71 [61-81] years, men: 47.7%) were selected with the criteria of age > 50 years and total WMSA volume > y value corresponding to x = 50 (Figure I in the online-only Data Supplement). Taking into account differing degrees of correlations between variables (Figure II in the online-only Data Supplement), we utilized four variables (le.no, le.vol, pdr, and le.ct) for further analysis, which were considered to reflect key features of WMSA and showed a weak collinearity (r < 0.5). Hierarchical cluster analysis identified three clusters with a height of 35 on the dendrogram (Figure 2A). Although there are general patterns for the clusters being related to total volume, there is a great deal of overlap across classes suggesting that the classification is not a simple proxy for total volume. (Figure III in the online-only Data Supplement). Meanwhile, three classes were differentiated from one or another on the 3D scatter plot with le.no, le.vol, pdr, and le.ct (Figure 2B). Principal component analysis reduced the variables for classification to two components, PC1 and PC2, which were a function of the different contribution of WMSA features (Figure 2C). The model came up with the high accuracy of 92.3% for predicting individual classes, and the 2D plot of PC1 and PC2 provided a clear line between three WMSA classes (Figure 2D).

Figure 2. WMSA classification procedure.

(A) Cluster dendrogram by hierarchical clustering analysis. Dendrogram shows three clusters with four factors including le.no, le.vol, pdr and le.ct with Euclidean distance and Ward linkage method. (B) Formation of three classes in 3D scatter plots based on hierarchical clustering analysis. Each class is specifically located in domains with le.vol, le.no, pdr, and le.ct. (C) Multinomial logistic regression modeling based on principal component analysis. (D) The 2D plot with PC1 and PC2 clearly draw the line between the classes. le.no: number of non-contiguous lesion, le.vol: volume of each non-contiguous lesion, pdr: periventricular to deep WMSA volume ratio, le.ct: WMSA lesion contrast

Distinct WMSA characteristics between three individual classes

The class probability was measured using multinomial logistic regression models with key WMSA variables. When we plotted the predictions for individual classes from the logit model using le.no, le.vol, pdr and le.ct, the individuals with higher le.no and lower le.ct were more likely to have class I over any of the other two, and those with higher le.vol and pdr were more likely to have class II, while for those with lower le.vol but higher le.ct, class III were more likely to occur (Figure 3). To further validate distinct characteristics of each individual class, we examined statistical differences in total and regional WMSA volume, lesion number, each lesion volume, contrast, and WMSA burden between three classes (Table 1). When comparing the volume metrics, class I showed a relatively higher frequency of WMSA in deep WM, whereas class II showed higher frequencies of juxtaventricular and periventricular WMSA. Class III exhibited lower WMSA volumes limited to juxtaventricular WM. When lesion number and each lesion volume were analyzed, many small scattered lesions were noted in class I, whereas multiple larger lesions were found in class II. Class III showed only a few restricted lesions. Lesion contrast analysis demonstrated the lowest contrast in class I. Meanwhile, class II showed relatively higher contrast and class III represented highest contrast. Finally, class II showed highest WMSA burden, whereas classes I and III showed lower burden due to low WMSA volume or low lesion contrast. Examination of the lesion prevalence maps demonstrated that individual classes could be classified as small, punctate, scattered, and deep WM lesion class (class I); large, patch, irregular and periventricular WM lesion class (class II); and mild, restricted, and juxtaventricular WM lesion class (class III) (Figure 4). Class II showed lower T1/T2 ratio values in WMSA of all WM areas, but especially in the periventricular area, whereas class I had relatively higher values and class III had intermediate values (Table 1; Figure IV in the online-only Data Supplement).

Figure 3. Association between key WMSA variables and individual class risk.

WMSA class was predicted from a multinomial model with le.no + le.vol + pdr + le.ct for a hypothetical subject with WMSA. The stacked-area effect plots display relationships between le.no (A), le.vol (B), pdr (C), le.ct (D), and the predicted probabilities of each class. le.no: number of non-contiguous WMSA lesions, le.vol: volume for each noncontiguous lesion, pdr: periventricular to deep WMSA volume ratio, le.ct: WMSA lesion contrast

Table 1.

Clinical and radiological characteristics of WMSA classes

	Class I	Class II	Class III	P value	Difference
Frequency, (%)	55 (42.3)	32 (24.6)	43 (33.1)
Age, years	65.0 [59.0;77.0]	82.5 [76.0;86.0]	70.0 [58.5;76.5]	<0.001	II>I,III
Men, (%)	31 (56.4)	14 (43.8)	17 (39.5)	0.223
Race, (%)				0.143
African	4 (7.3)	5 (15.6)	9 (20.9)
Asian	4 (7.3)	1 (3.1)	3 (7.0)
White	47 (85.5)	26 (81.3)	31 (72.1)
WMSA number	294.0 [245.0;363.5]	264.0 [207.0;313.0]	184.0 [150.5;209.5]	<0.001	I>II>III
WMSA size, mm3	12.36 [10.68;16.51]	29.21 [22.45;43.63]	14.94 [12.20;17.70]	<0.001	II>I,III
WMSA volume, log10
Total	3.55 [3.45;3.72]	3.88 [3.68;4.07]	3.40 [3.33;3.53]	<0.001	II>I>III
Juxtaventricular	3.22 [3.12;3.33]	3.55 [3.43;3.69]	3.23 [3.11;3.31]	<0.001	II>I,III
Periventricular	3.03 [2.86;3.20]	3.44 [3.33;3.77]	2.70 [2.54;2.85]	<0.001	II>I>III
Deep	2.90 [2.74;3.10]	2.83 [2.46;3.11]	2.58 [2.42;2.75]	<0.001	I,II>III
P-D volume ratio	1.33 [0.96;1.91]	5.06 [3.63;7.13]	1.45 [0.80;2.33]	<0.001	II>I,III
WMSA contrast	1.11 [1.10;1.13]	1.15 [1.13;1.18]	1.18 [1.16;1.20]	<0.001	III,II>I
WMSA burden, (%)	0.30 [0.24;0.40]	0.74 [0.50;1.09]	0.25 [0.21;0.31]	<0.001	II>I,III
T1/T2 Ratio in WMSA	2.05 [1.88;2.318]	1.55 [1.37;1.65]	1.69 [1.45;1.92]	<0.001	II<III<I
Hypertension, (%)	20 (36.4)	20 (62.5)	17 (39.5)	0.047	II>I,III
Diabetes, (%)	7 (12.7)	3 (9.4)	4 (9.3)	0.827
Dyslipidemia, (%)	19 (34.5)	14 (43.8)	19 (44.2)	0.553
Current smoker, (%)	4 (7.2)	1 (3.1)	2 (4.7)	0.687
BMI, kg/m2	24.8 [22.7;29.4]	26.3 [24.3;29.4]	25.6 [22.9;29.6]	0.465
SBP, mmHg	129.1 ± 16.5	140.1 ± 17.7	129.4 ± 17.7	0.018	II>I,III
DBP, mmHg	78.6 ± 12.0	78.8 ± 11.1	77.6 ± 14.1	0.692
Pulse pressure, mmHg	47.0 [41.0;59.0]	57.0 [50.5;74.5]	48.0 [42.5;63.0]	0.005	II>I,III
Framingham risk	7.0 [2.0;12.5]	10.0 [6.0;16.0]	6.0 [2.0;11.0]	0.044	II>I,III
HbA1c	5.4 [5.2;5.7]	5.5 [5.2;5.6]	5.3 [5.2;5.7]	0.881
FBS, mg/dL	102.0 [96.0;109.0]	100.0 [93.0;107.0]	102.0 [95.0;106.0]	0.549
hsCRP, mg/L	1.5 [0.7;3.2]	1.9 [0.8;3.1]	1.0 [0.5;2.7]	0.450
BUN, mg/dL	15.0 [13.0;19.0]	16.0 [13.0;20.0]	15.0 [13.0;18.0]	0.643
Creatinine, mg/dL	0.9 [0.8;1.0]	0.9 [0.8;1.1]	0.9 [0.8;1.0]	0.649
Triglyceride, mg/dL	94.0 [68.0;128.0]	89.0 [60.0;124.0]	104.0 [62.0;134.0]	0.825
Total Cholesterol, mg/dL	191.5 ± 41.5	196.2 ± 23.4	200.7 ± 46.6	0.270
LDL-C, mg/dL	109.5 ± 33.6	111.2 ± 22.4	120.9 ± 39.3	0.112
HDL-C, mg/dL	62.0 [50.0;67.0]	58.0 [51.0;82.0]	55.0 [50.0;72.0]	0.702
PSQI score	4.0 [3.0;6.0]	3.0 [2.0;4.5]	5.0 [3.0;6.0]	0.060
Poor sleeper, (%)	24 (43.6%)	8 (25.0%)	24 (55.8%)	0.028	I,III>II
Physical activity
Sitting, min/d	35.0 [21.0;49.0]	35.0 [24.5;49.0]	35.0 [21.0;56.0]	0.660
Walking, hr/wk	4.7 [2.5;7.0]	2.0 [0.7;6.4]	3.5 [1.8;7.0]	0.112
Moderate, hr/wk	2.0 [0.0;4.0]	0.7 [0.0;2.5]	1.0 [0.0;3.0]	0.133
Vigorous, hr/wk	1.0 [0.0;3.8]	0.0 [0.0;0.9]	0.0 [0.0;1.5]	0.001	I>II,III
Total, MET-min/wk	2179.5 [1426.5;3813.0]	1078.5 [452.3;2003.7]	1386.0 [881.3;214.0]	0.003	I>III>II
MoCA	27.0 [25.0;28.0]	26.0 [24.0;26.5]	27.0 [25.0;28.0]	0.057

Data are represented as numbers and frequencies for categorical variables and mean ± standard deviation or median with interquartile range for continuous variables. WMSA: white matter signal abnormalities, BMI: body mass index, SBP: systolic blood pressure, DBP: diastolic blood pressure, FBS: fasting blood glucose, hsCRP: high sensitivity C-reactive protein, BUN: blood urea nitrogen, LDL-C: low density lipoprotein-cholesterol, HDL-C: high density lipoprotein-cholesterol, PSQI: Pittsburg sleep quality index, LOC: loss of consciousness, MET: metabolic equivalent, MoCA: Montreal Cognitive Assessment, P value by Pearson Chi-Square test, ANOVA followed by Tukey test, or Kruskal-Wallis rank sum test followed by the Dunn test

Figure 4. WMSA distribution patterns in individual WMSA classes.

The map was generated by summed volumes from individuals of each class. Colors in each slice show lesion prevalence within the group (0-50%).

Clinical factors associated with individual classes

We examined clinical factors to determine whether these measures were differentially associated with the WMSA classes (Table 1). Class II could be clearly distinguished from the other two classes due to older age, higher frequency of hypertension, higher SBP, higher pulse pressure, higher score of FRS and lower level of total physical activity. In contrast, poor sleep quality was more frequently associated with class I and class III than class II. Class II performed marginally worse on the MoCA, but there was no significant difference across classes after controlling for age and sex. In a multinomial logistic regression analysis, age, hypertension, sleep quality, and total physical activity were identified as independent factors associated with the risk of individual class (Table 2). Older age, higher rate of hypertension and lower physical activity were independently associated with class II risk, and poorer sleep quality was associated with higher risk of class I and III. After controlling for age, systolic blood pressure and pulse pressure was positively associated with total WMSA volume in class II. Poor sleep quality score was positively associated with total WMSA volume in class I. Total physical activity was negatively associated with total WMSA volume in the relatively younger age of class II (Figure V in the online-only Data Supplement).

Table 2.

Odds for predicting individual WMSA classes in a multinomial logit model

	Class I vs. Class III		Class II vs. Class III		Class II vs. Class I
	OR [95% CI]	P value	OR [95% CI]	P value	OR [95% CI]	P value
Age, yr	0.997 [0.977-1.017]	0.759	1.133 [1.104-1.164]	<0.001	1.137 [1.108-1.167]	<0.001
Hypertension	1.109 [0.503-1.445]	0.798	1.856 [1.254-2.750]	0.002	1.674 [1.129-2.483]	0.010
FRS	1.039 [0.955-1.131]	0.375	1.015 [0.919-1.119]	0.774	0.976 [0.891-1.070]	0.608
PSQI score	0.942 [0.782-1.135]	0.532	0.691 [0.520-0.918]	0.011	0.733 [0.552-0.973]	0.032
Total PA, MET·min/week	1.000 [0.999-1.000]	0.105	0.999 [0.999-1.000]	0.093	0.999 [0.998-0.999]	0.016

FRS: Framingham risk score, PSQI: Pittsburg sleep quality index, PA: physical activity, MET: metabolic equivalent

Discussion

This study aimed to implement an automated procedure to quantitatively define heterogeneous ‘classes’ of WMSA and how those classes relate to clinical features in typically aging middle aged and older adults. Each individual class displayed distinct characteristics of WMSA with respect to their volume, regional preference, multiplicity, contrast, burden, and myelin content. Within-lesion heterogeneity could be quantified to differentiate clinical causes of white matter damage. When the total WMSA volume alone was considered for classification (mild-intermediate-severe), the severe group shared some risk factors and phenotypes of class II, but other groups failed to display distinct WMSA features, myelin loss and clinical risk factors (Table I in the online-only Data Supplement). Thus, the novel classification system provides a classification system that is more clinically sensitive than simple staging of white matter disease, likely due to capture of distinct classes of etiologically relevant features.

The core idea behind the regional localization of WMSA is that the WMSA development would be influenced by different features of vascular anatomy and function in the WM. The basal perforators and medullary arteries terminate with no functional anastomoses with each other in the periventricular area.^{12, 13} Since lipohyalinosis or fibrinoid necrosis generally occurs in the terminal part of the arterioles, the periventricular area is vulnerable to small vessel disease and resulting hypoperfusion.^{32, 33} In contrast, deep WM is resistant to hypoperfusion.¹⁴ While medullary arteries give off fewer side branches, they are highly compliant and have sufficient adventitial spaces in their proximal segment, suggesting that the proximal and side zones of long medullary arteries might be irrigated by perivascular transport of CSF.^{12, 13} The juxtaventricular WMSA are characterized by the continuity to ventricular surface and unique shapes such as a cap, rim and halo. The histopathological findings suggest that the CSF spillover into tissue via disrupted ependymal lining or the impaired drainage of interstitial fluid via dysfunctional arterioles works on the pathogenesis of juxtaventricular lesion.^{34, 35} The involvement of different pathobiological bases may lead to differences in radiological phenotype of WMSA. Our classification process engaged automatic and comprehensive methods to characterize the heterogeneous subsets of WMSA.

Since individuals succumb to a certain phenoclass of WMSA according to their own clinical factors, individual WMSA class could be discriminated with one or another with respect to various radiological features. While class II generally showed highest WMSA burden, class I, class II, and class III represented peripheral preference, central preference, and ependymal preference in the lesion location, respectively. Lesions relevant to impaired perivascular transport are more likely to be linearly distributed along the penetrating arteries and those in the axial cut are shown as small, punctate and multiple in the deep WM. Lesions relevant to hypoperfusion in small arteries are distributed around the arborized terminal branch and those in the axial cut are shown initially in a fuzzy, irregular pattern, and extend into patchy and confluent forms.^{12, 13} It turned out that class I was characterized by small scattered lesions in the deep WM, and class II by large irregular lesions in the periventricular WM. Furthermore, pathophysiological differences might explain the reason for differing contrasts of WMSA. Class I showed lower contrast than class II and III. The contrast difference across classes is likely due to subtle WM damage in class I, ischemic necrosis in class II, and interstitial fluid increase in class III.⁵ T1/T2 ratio map provides a quantitative estimate of tissue damage and this measure has been previously linked to myelin content in the WM.^{29, 30} The decrease in myelin content was most prominent in class II, moderate in class III, and very mild in class I. Different levels of myelin content in each class might lead to different clinical outcomes.

Consideration of clinical factors helps to predict the individual risk of WMSA class and to understand the pathophysiology. Each class was associated with different patterns of demographic and risk factors. It is known that age effects are distinct across regional locations of WMSA, and the periventricular location of WMSA is more strongly associated with age and clinical factors related to aging than the deep area.³⁶ In the present study, older age was significantly associated with class II with a regional preference to the periventricular area. Hypertension is more relevant to small vessel pathology than other vascular risk factors.³⁷ While lipohyalinosis is a diffuse pathology in the small arterioles of the white matter, its impact would be greatest in the periventricular areas that are highly vulnerable to hypoperfusion.^{14, 38} As expected, class II was associated with a higher rate of hypertension than other atherosclerotic risk factors. In contrast, deep and juxtaventricular white matter areas are more likely sensitive to CSF flow dynamics and functional derangement of cerebral arteries. Brain waste products and interstitial fluid are drained by the glymphatic pathway that is driven by pulsation of compliant arterioles.^{15, 39} Oscillatory activity during sleep is associated with oscillations in blood and interstitial fluid volumes,⁴⁰ and sleep disturbances such as insomnia and obstructive sleep apnea have been associated with arteriolar dysfunction and subsequent development of white matter lesions.⁴¹ Our observation that poor sleep quality was closely related with class I and class III suggests that the glymphatic pathway might partly mediate the risk of these classes. In addition, lower levels of physical activity tended to have a higher risk of class II. Physical activity would confer a favorable effect on vascular health in many different ways, but its optimal intensity and class for age group remain to be determined. Taking into account different radiological features and underlying anatomical properties, classes I and III are anticipated to have different risk factors. However, we failed to find out significant clinical factors from the current dataset. Given the borderline associations between clinical/epidemiological factors and neuroimaging measures, we believe that additional investigation in a larger sample may be necessary to uncover additional defining features. Additionally, future investigation using inflammatory, hormonal, metabolic, and genetic markers may unveil more direct associations with the lesion classifications described here.

There are some limitations in our study. The current study examined ‘typically’ aging adults. Whether this criterion would still be effective in other specific populations with higher propensity to vascular disease and dementia warrants further research. Second, because of the limitation of a cross-sectional study, our study could not determine the outcome of WMSA and its impact on cognitive outcome. Future longitudinal studies may demonstrate distinct outcomes of each WMSA class. Additionally, we do believe that the classes are likely related to other markers of small vessel disease. However, we did not acquire protocols optimal for quantification of other markers such as microbleeds, perivascular space, and lacunes, and to our knowledge, software is limited for robust segmentation and quantification of these features. Nonetheless, our study is of great strength based on the automatic quantification for lesion heterogeneity, the blend of a wide range of clinical and radiological factors, and consideration of biological background.

Conclusions

Our quantification procedure for within-lesion heterogeneity identified three distinct classes that could provide more specific links to clinical variability. This new method for identifying classes of white matter lesions will be important in understanding the underlying pathophysiology and in determining the impact on brain function and clinical outcomes.

Abbreviations

HCP-A

Human Connectome Project Aging

IPAQ

International Physical Activity Questionnaire

MCA

middle cerebral artery

MoCA

Montreal Cognitive Assessment

NAWM

Normal appearing white matter

PCA

principal component analysis

PSQI

Pittsburgh Sleep Quality Index

WMSA

white matter signal abnormalities