White Matter Hyperintensities Contribute to Language Deficits in Primary Progressive Aphasia

Abstract

Objective:

To determine the contribution of white matter hyperintensities (WMH) to language deficits while accounting for cortical atrophy in individuals with primary progressive aphasia (PPA).

Method:

Forty-three individuals with PPA completed neuropsychological assessments of nonverbal semantics, naming, and sentence repetition plus T2-weighted and fluid-attenuated inversion recovery scans. Using three visual scales, we rated WMH and cerebral ventricle size for both scan types. We used Spearman correlations to evaluate associations between the scales and scans. To test whether visual ratings—particularly of WMH—are associated with language, we compared a base model (including gray matter component scores obtained via principal component analysis, age, and days between assessment and MRI as independent variables) with full models (ie, the base model plus visual ratings) for each language variable.

Results:

Visual ratings were significantly associated within and between scans and were significantly correlated with age but not with other vascular risk factors. Only the T2 scan ratings were associated with language abilities. Specifically, controlling for other variables, poorer naming was significantly related to larger ventricles (P = 0.033) and greater global (P = 0.033) and periventricular (P = 0.049) WMH. High global WMH (P = 0.034) were also correlated with worse sentence repetition skills.

Conclusion:

Visual ratings of global brain health were associated with language deficits in PPA independent of cortical atrophy and age. While WMH are not unique to PPA, measuring WMH in conjunction with cortical atrophy may elucidate more accurate brain structure–behavior relationships in PPA than cortical atrophy measures alone.

White matter hyperintensities (WMH) are signal abnormalities that appear brighter than the surrounding tissue on T2-weighted MRIs. The pathogenesis of WMH is postulated to be vascular in nature, with WMH often cited as one marker of cerebral small vessel disease (SVD) burden (Paradise et al, 2018; Rensma et al, 2018; Wardlaw et al, 2013). Scientists have theorized that endothelial dysfunction, which is caused by factors such as hypertension, causes a breakdown of the blood–brain barrier, which in turn causes damage to neurons and glial cells that manifest as WMH (Grueter and Schulz, 2012; Hase et al, 2018; Joutel and Chabriat, 2017; Lin et al, 2017; Schmidt et al, 2011; Wardlaw et al, 2013). The most commonly cited causes of WMH include chronic hypoperfusion, impaired cerebrovascular reactivity, and breakdown of the blood–brain barrier. However, some researchers (eg, Joutel and Chabriat, 2017; Lin et al, 2017; Weller et al, 2015) have proposed alternative pathophysiological mechanisms of WMH, such as dysfunction of ogliodendrocytes and the glymphatic system.

While WMH are not specific to any one clinical population, individuals with dementia are known to exhibit vascular pathology, occurring in 61% of cases of frontotemporal dementia and 82% of cases of Alzheimer disease (Toledo et al, 2013). In recent years, increased attention has focused on the interplay between primary pathologies in dementia and markers of SVD burden as measured on neuroimaging, such as WMH (Hase et al, 2018; Koncz and Sachdev, 2018; Liu et al, 2018; Mortamais et al, 2013; Santos et al, 2017). In particular, WMH have been shown to predict an increased risk of both vascular dementia and Alzheimer disease (Debette et al, 2007; Liu et al, 2018; Rensma et al, 2018). Although findings from specific studies have been mixed, meta-analytic findings have indicated that WMH are related to the severity of cognitive deficits, and predict future cognitive decline, in individuals with dementia, independent of other cerebrovascular disease risk factors such as hypertension and age (Debette and Markus, 2010).

Despite emerging associations between WMH and deficits in certain dementias (eg, vascular dementia, Alzheimer disease), relatively little is known about the unique impact that WMH have on individuals with primary progressive aphasia (PPA). PPA is a neurodegenerative disorder with diverse underlying pathologies. It is clinically characterized by a gradual decline in language skills (with relative initial sparing of other cognitive skills) due to progressive cell loss primarily in the left temporal, parietal, and/or frontal cortices (Diehl-Schmid et al, 2014; Gorno-Tempini et al, 2011; Mesulam, 1982; Vandenberghe, 2016).

There are three recognized variants of PPA: semantic variant PPA (svPPA), characterized by gradual deterioration of semantic representations manifesting as deficits in single-word comprehension and expression; logopenic variant PPA (lvPAA), characterized by deficits in phonological short-term memory resulting in difficulty with naming and repetition, especially for multisyllabic words and sentences; and nonfluent agrammatic variant PPA (nfaPPA), characterized by effortful and poorly articulated speech output with impaired syntactic production and comprehension (Gorno-Tempini et al, 2011).

Although the effect of WMH on language deficits in individuals with PPA remains relatively unexplored, findings of a recent study by our group (Odolil et al, 2019) indicated that the severity of WMH predicts the rate of decline in naming abilities of individuals with PPA. However, Odolil et al (2019) did not interrogate the relative contributions of cortical atrophy and WMH to naming or other language deficits in individuals with PPA. Moreover, it is unclear whether measuring WMH in individuals with PPA provides information that is unique from that provided by other measures of overall brain health (eg, cerebral ventricle size).

Thus, we had two main aims in the current study. The first aim was to examine how visual ratings performed on T2-weighted and fluid-attenuated inversion recovery (FLAIR) scans of WMH and overall brain health (captured by cerebral ventricle size) are related in individuals with PPA. We hypothesized that visual ratings performed on both scan types and on different scales would be strongly, positively correlated, but that the strongest correlations would be between ratings from different WMH scales performed on the same scan type. The second aim was to determine whether visual ratings are related to individuals’ language abilities, accounting for cortical atrophy and nuisance variables. We hypothesized that ventricle size ratings would add minimal predictive power to cortical atrophy models, whereas WMH ratings would be strong predictors of individuals’ language deficits.

METHOD

Participants

We retrospectively analyzed prospectively collected data from 43 individuals (27 women; 16 men; M age = 68.71 ± 8.78 years) who had been seen at Johns Hopkins Hospital between November 2007 and March 2017. All 43 individuals had been diagnosed with PPA based on language and cognitive testing, medical history, and imaging consensus criteria (Gorno-Tempini et al, 2011). The individuals had no history of other neurologic disorders or contraindications for MRI. Neuropsychological assessments of nonverbal semantics, spoken and written verb and noun naming, sentence repetition and reading skills, and spelling abilities were used to characterize the individuals’ deficits and to classify the particular variant of PPA with which they had been diagnosed (ie, 11 svPPA, 13 lvPPA, 15 nfaPPA, 4 unclassifiable). Information regarding demographics and vascular risk factors for the individuals with PPA is provided in Table 1.

TABLE 1.

Summary of Demographics and Vascular Risk Factors

	n	M (SD)	Range
Age (in years)	43	68.71 (8.78)	50–85
Handedness (right:left)	38:5	—	—
Education (in years)	43	15.83 (2.57)	12–20
Hypertension status (yes:no)	42	23:19	—
Diabetes mellitus status (yes:no)	42	5:37	—
Hypercholesterolemia status† (yes:no)	42	27:15	—

Data reported as M ± SD unless otherwise indicated.

^†Hypercholesterolemia data were not available for one individual.

The study protocol was approved by the institutional review board of Johns Hopkins University and was performed according to the ethical guidelines of the Declaration of Helsinki and its later amendments. All participants provided informed written consent before enrolling in the study.

Language Assessment

Language abilities examined in this study included nonverbal semantic association, oral noun naming, and sentence repetition. Because the individuals in our study had participated in different previous studies, some of them had been administered different assessments. For most of the participants, nonverbal semantic association skills had been assessed with the 14-item Pyramids and Palm Trees Test (PPTT; Breining et al, 2015), spoken noun naming had been assessed with the 30-item Boston Naming Test (BNT; Saxton et al, 2000), and sentence repetition skills had been assessed with the 5-item Sentence Repetition Test (Hillis, 2011). Scores obtained from the language tests are provided in Table 2.

TABLE 2.

Language Assessment Data

Test	n	M (SD)	Range
Pyramids and Palm Trees Test (%)	26	86.54 (22.34)	7.14–100.00
svPPA	6	65.48 (31.11)	7.14–92.86
lvPPA	7	84.69 (23.12)	35.71–100.00
nfaPPA	12	97.02 (4.78)	85.71–100.00
unclassifiable	1	—	100.00
Boston Naming Test (%)	25	50.80 (34.12)	3.33–100.00
svPPA	4	35.83 (25.44)	20.00–73.33
lvPPA	9	30.74 (26.66)	3.33–80.00
nfaPPA	9	81.85 (17.25)	53.33–100.00
unclassifiable	3	37.78 (45.99)	3.33–90.00
Sentence Repetition Test (%)	22	47.27 (37.31)	0.00–100.00
svPPA	5	40 (31.62)	0.00–80.00
lvPPA	8	30 (35.46)	0.00–80.00
nfaPPA	7	71.43 (27.95)	20.00–100.00
unclassifiable	2	50 (70.71)	0.00–100.00

The number of individuals who were administered specific assessments varies between tests because these individuals participated in different previous studies.

lvPPA = logopenic variant of primary progressive aphasia. nfaPPA = nonfluent agrammatic variant of primary progressive aphasia. svPPA = semantic variant of primary progressive aphasia.

Neuroimaging Data

All of the individuals had undergone clinical MRIs that included T2-weighted and FLAIR sequences, acquired on either a 1T (n = 1), 1.5T (n = 18), or 3T (n = 24) Siemens magnet. The average time interval between the language assessment and imaging was 47.33 ± 64.70 days. A multi-atlas FLAIR library was used to segment the individuals’ FLAIR images into 143 regions of interest (ROIs) for different tissue types (ie, white matter [WM], gray matter [GM], cerebrospinal fluid, and non-brain tissue) (Wu et al, 2019). Automated segmentations for each individual were visually assessed using ROIEditor software (www.mristudio.org/wiki/user_manual/roieditor). Data from three individuals were excluded from further analysis due to low resolution of their FLAIR images and poor segmentation.

For the remaining 40 individuals, GM volumes in the 12 left-hemisphere ROIs most often atrophied in individuals with PPA (ie, superior, middle, and inferior frontal and temporal; and orbital, precentral, postcentral, supramarginal, angular, and fusiform gyri) were extracted in each individual’s native space. GM volumes for each individual were standardized by dividing each regional volume by the individual’s cerebral volume (ie, sum of telencephalon, diencephalon, mesencephalon, and metencephalon volumes).

We used the ventricular and WM disease scales from the Cardiovascular Health Study (CHS; Longstreth et al, 1996; Manolio et al, 1994) to assess the size of the cerebral ventricles and the severity of the WMH, respectively. Scores on the ventricular size scale (hereafter referred to as CHS Ventricles) range from 0 (ie, normal size) to 9 (ie, severely enlarged ventricles). Scores on the CHS WM scale (hereafter referred to as CHS WM) also range from 0 (ie, no WMH) to 9 (ie, large, confluent WMH). We also assessed the WMH using a scale devised by Fazekas et al (1987). This scale contains two subscales: one to assess periventricular hyperintensities (PVH; hereafter referred to as Fazekas PVH), with scores ranging from 0 (no hyperintensities) to 3 (irregular PVH extending into the deep WM), and one to assess deep white matter hyperintensities (DWMH; hereafter referred to as Fazekas DWMH), with scores ranging from 0 (no hyperintensities) to 3 (large, confluent areas). The WMH were identified according to consensus standards from Wardlaw et al (2013); specifically, signal abnormalities in the WM (but not the subcortical GM or brainstem) that presented as hyperintense areas on T2-weighted and FLAIR scans without cavitation were rated.

Four of the authors (E.L.M., B.L.B., S.M.S., and E.B.G.) independently completed visual ratings on each scale for all individuals in our study. Prior to completing ratings, the authors underwent training on using the visual scales by reading the source material, rating two independent sets of sample images, and conferring with experts in cerebrovascular disease (A.V.F. and A.E.H.). For consistency with the previous literature—including the original studies detailing the scales, consensus standards detailed in Wardlaw and colleagues (2013), and previous work by our group (Odolil et al, 2019)— we performed ratings on both T2 and FLAIR images for each individual. The rating assignments were counterbalanced so that each individual was rated by two raters using each scale for both scan types (ie, T2 vs FLAIR). Interrater reliability was determined using Cohen weighted kappa coefficients for the ratings that we performed on each scan type (ie, two ratings per scale, per individual). Consensus ratings were made when ratings performed on the same scan type differed by >1 point between the two raters. Ratings that differed between the two raters by 1 point were averaged.

Statistical Analysis

All of the statistical analyses were performed using R software (www.R-project.org). Given the variability in scan acquisition parameters, we conducted two sets of analyses to ensure that the visual ratings were not unduly influenced by the different scanners that had been used. Specifically, we conducted Wilcoxon rank sum tests to determine if the visual ratings and GM volumes differed between the scans collected on the 1.5T versus 3T magnets. We then performed Spearman correlation analyses between these two variables (ie, visual ratings and GM volumes) and scan resolution (per voxel size in mm³). Because the presumed origin of WMH is vascular in nature, we additionally tested for any relationships between the visual ratings and vascular risk factors, including sex, hypertension status, diabetes status, hypercholesterolemia status (via Wilcoxon rank sum tests), and age (via Spearman correlations).

To examine whether the visual ratings of T2 and FLAIR scans of cerebral ventricle size and WMH are related in individuals with PPA, we conducted Spearman correlations between ratings using the same scan type (ie, T2 or FLAIR) and between ratings on the same scale/subscale (ie, CHS Ventricles, CHS WM, Fazekas PVH, and Fazekas DWMH), performed on different scans. We corrected for multiple tests by adjusting alpha according to the Benjamini and Hochberg (1995) false discovery rate (FDR) at P < 0.05 for four correlations of ratings from the same scale, performed on different scans, and for two sets of six correlations of ratings from the same scan, performed using different scales.

To examine whether visual ratings of WMH and ventricle size are related to the language abilities of individuals with PPA when controlling for cortical atrophy and other variables, we conducted a series of analyses. First, we entered the GM volumes from the 12 ROIs into a principal component analysis. We used a principal component analysis to reduce the large number of factors into a manageable set of variables while preserving unique information about gyral atrophy. Specifically, a principal component analysis produces weighted scores reflecting regions of the brain that covary in terms of atrophy. Unlike volumes that are extracted from a coarser brain parcellation (eg, split by lobe), this procedure captures the extent to which language network atrophy varies by region, with some regions within a given lobe being more atrophied (eg, left inferior frontal gyrus) in individuals with PPA and others relatively spared (eg, left superior frontal gyrus). Components with eigenvalues >1.0 were retained, and single-subject factor loadings per component were extracted. The component scores were then entered into backward stepwise regression models predicting each language variable.

Next, we used hierarchical regression to compare reduced (ie, base) to full models for each language variable (ie, nonverbal semantics, spoken noun naming, and sentence repetition). Specifically, for each language variable, the base model included the independent variables of age, days between assessment and MRI, and GM components retained from the backward stepwise regression. To assess whether WMH and ventricular metrics added predictive power to the base model, we added each visual rating separately to the base model in a series of eight linear regressions (ie, one per visual rating). An FDR correction was applied to the multivariable model P values across all linear regressions for each language variable. Finally, we compared the full factor models to the corresponding base model using an ANOVA. An FDR correction was also applied to the P values across each set of ANOVAs. At each stage of the analysis, checks on regression model assumptions were performed using the car (Fox and Weisburg, 2011), gvlma (Peña and Slate, 2006), and MASS (Venables and Ripley, 2002) packages in R.

RESULTS

Scans and visual ratings for two sample participants are shown in Figure 1. The weighted kappa coefficients denoting interrater reliability across all within-scan ratings are reported in Table 3. All participants were rated as having some degree of ventricular dilation (ie, CHS Ventricles scores >0) in both the T2-weighted (median: 5 out of 9) and the FLAIR (median: 4 out of 9) images. Nearly all of the individuals (ie, 41/43 or 95.35%) were rated as exhibiting some degree of WMH (ie, scores on the WMH scales >0) on the CHS WM (T2 median: 2.5, FLAIR median: 2.5 out of 9), Fazekas PVH (T2 median: 1.5, FLAIR median: 1 out of 3), and Fazekas DWMH (T2 median: 1, FLAIR median: 1 out of 3) scales.

FIGURE 1.

Example scans and visual ratings for two individuals with primary progressive aphasia. A. Participant had mild WMH and relatively small ventricles. B. Participant had severe WMH and greatly enlarged ventricles. We used the CHS Ventricles scale to estimate ventricle size (ie, “Ventricles” column) and the CHS WM scale to estimate WMH (ie, “WM” column). We used the Fazekas subscales to estimate PVH WMH (ie, “PVH” column) and DWMH WMH (ie, “DWMH” column). CHS = Cardiovascular Health Study. DWMH = deep white matter hyperintensities. FLAIR = fluid-attenuated inversion recovery. PVH = periventricular hyperintensities. WM = white matter. WMH = white matter hyperintensities.

TABLE 3.

Interrater Reliability of the Visual Ratings Performed on Both Scan Types According to Cohen Weighted Kappa (κ_w) Coefficients

	T2		FLAIR
Rating Scale	κw	P	κw	P
CHS Ventricles	0.806	< 0.001	0.827	< 0.001
CHS WM	0.717	< 0.001	0.670	< 0.001
Fazekas PVH	0.479	0.002	0.295	0.045
Fazekas DWMH	0.269	0.044	0.607	< 0.001

CHS = Cardiovascular Health Study. DWMH = deep white matter hyperintensities. FLAIR = fluid-attenuated inversion recovery. PVH = periventricular hyperintensities. WM = white matter.

We found no significant relationships between visual ratings and magnet strength (FDR-corrected P > 0.439) or scan resolution (FDR-corrected P > 0.813) (see Supplemental Table 1 in Supplemental Digital Content). We also found no significant relationships between cortical ROI volumes and magnet strength (FDR-corrected P > 0.083) or scan resolution (FDR-corrected P > 0.170) (see Supplemental Table 2 in Supplemental Digital Content). There also were no significant differences in visual ratings between women and men (FDR-corrected P = 0.883 across tests) or between individuals with versus without hypertension (FDR-corrected P > 0.837), diabetes (FDR-corrected P > 0.237), or hypercholesterolemia (FDR-corrected P = 0.951 across tests) (see Supplemental Table 3 in Supplemental Digital Content). However, increased age was significantly related to higher ratings on all of the scales (range: r = 0.441–0.598, FDR-corrected P < 0.007) except for the FLAIR CHS WM (r = 0.250, FDR-corrected P = 0.125) and the FLAIR Fazekas PVH (r = 0.300, FDR-corrected P = 0.072) ratings. The latter finding confirms the necessity of controlling for age in the following regression models.

Visual ratings performed separately on the T2 and FLAIR scans were positively correlated for the CHS Ventricles (r = 0.893, P < 0.001), CHS WM (r = 0.586, P < 0.001), Fazekas PVH (r = 0.504, P < 0.001), and Fazekas DWMH (r = 0.657, P < 0.001) scales. As expected, there were strong, positive associations between WM ratings performed on the same scan type and weaker associations between the other scales, particularly between WMH and ventricle size ratings performed on FLAIR scans (Table 4).

TABLE 4.

Spearman Correlations Between Ratings From Different Scales on the Same Scan Type

Rating Scale	CHS Ventricles	CHS WM	Fazekas PVH	Fazekas DWMH
CHS Ventricles	—	FLAIR: 0.286	FLAIR: 0.332*	FLAIR: 0.468**
CHS WM	T2: 0.747***	—	FLAIR: 0.836***	FLAIR: 0.787***
Fazekas PVH	T2: 0.627***	T2: 0.846***	—	FLAIR: 0.743***
Fazekas DWMH	T2: 0.505**	T2: 0.768***	T2: 0.767***	—

^*Significant at P < 0.05.

^**Significant at P < 0.01.

^***Significant at P < 0.001.

^{^}0.08 < > 0.05, corresponding to an FDR correction (P < 0.05) performed across the correlations conducted for each scan type.

CHS = Cardiovascular Health Study. DWMH = deep white matter hyperintensities. FLAIR = fluid-attenuated inversion recovery. PVH = periventricular hyperintensities. WM = white matter.

The principal component analysis of the left-hemisphere GM ROIs resulted in four principal components (PCs) that explained 73.2% of the variance in the data. As shown in Figure 2, ROIs that highly loaded (at ≤–0.50 or ≥0.50) together onto components were generally in close spatial proximity. Ventral temporal regions positively loaded onto PC1, and superior and middle frontal gyri negatively loaded onto PC1, reflecting a pattern within the sample of higher inferior temporal and lower frontal GM volumes. Parietal regions positively loaded onto PC2. Regions that positively loaded onto PC3 included the inferior frontal gyrus and the primary motor and sensory cortices. The orbital gyrus as well as the middle and superior temporal gyri loaded onto PC4.

FIGURE 2.

Loadings of gray matter regions of interest volumes onto principal components. Red indicates positive loading; blue indicates negative loading. AG = angular gyrus. FuG = fusiform gyrus. GM = gray matter. IFG = inferior frontal gyrus. ITG = inferior temporal gyrus. L = left. MFG = middle frontal gyrus. MTG = middle temporal gyrus. OG = orbital gyrus. PC = principal component. PoCG = postcentral gyrus. PrCG = precentral gyrus. ROI = region of interest. SFG = superior frontal gyrus. SMG = supramarginal gyrus. STG = superior temporal gyrus.

Nonverbal Semantics Regression

The backward stepwise regression predicting the PPTT scores from the four GM PCs was not significant (F_1,22 = 2.257, P = 0.147, adjusted R² = 0.052), and only PC2 (β = 0.636, SE = 0.424, t = 1.502, P = 0.147) was retained. The final base model predicting PPTT from PC2, number of days between assessment and MRI, and age also was not significant (F_3,20 = 0.117, P = 0.468, adjusted R² = −0.016). Individually adding each of the visual ratings to the base model did not significantly improve the model’s predictive power (F_4,19 < 1.897, FDR-corrected P = 0.617 across models, adjusted R² range = −0.060–0.135). Supplemental Tables 4 and 5 (in Supplemental Digital Content) provide complete results from the series of full models.

Naming Regression

The backward stepwise regression predicting the BNT scores from the four GM PCs resulted in a significant model (F_4,18 = 8.710, P < 0.001, adjusted R² = 0.584) in which PC2 (β = 3.706, SE = 1.253, t = 2.958, P = 0.008), PC3 (β = −6.414, SE = 1.645, t = −3.900, P = 0.001), and PC4 (β = 7.649, SE = 1.697, t = 4.508, P < 0.001) were significant predictors of naming, but PC1 (β = −2.308, SE = 1.400, t = −1.648, P = 0.117) was not. The base model predicting naming was also significant (F_6,16 = 7.743, P < 0.001, adjusted R² = 0.648). Controlling for other variables within the model, PC2 (β = 3.957, SE = 1.159, t = 3.414, P = 0.004), PC3 (β = −4.561, SE = 1.715, t = −2.661, P = 0.017), PC4 (β = 8.438, SE = 1.648, t = 5.119, P < 0.001), and days between assessment and scan dates (β = −0.055, SE = 0.024, t = −2.273, P = 0.037) were significantly associated with naming, but PC1 (β = −1.855, SE = 1.398, t = −1.326, P = 0.203) and age (β = 0.081, SE = 0.157, t = 0.512, P = 0.616) were not. Unexpectedly, within this model, better naming was associated with lower volumes of regions that highly loaded onto PC3 (ie, left inferior frontal, precentral, and postcentral gyri).

To determine if this finding reflected a previously documented finding that individuals with nfaPPA have more intact naming skills but greater frontal atrophy than individuals with other variants of PPA (Gorno-Tempini et al, 2004; Migliaccio et al, 2016; Race et al, 2013), we ran a follow-up regression model predicting BNT scores by PPA variant. This model was significant (F_3,19 = 5.753, P = 0.006, adjusted R² = 0.393): Individuals with nfaPPA demonstrated better naming than individuals with svPPA (β = −14.250, SE = 4.977, t = −2.863, P = 0.010), lvPPA (β = −15.375, SE = 4.064, t = −3.783, P = 0.001), and unclassifiable PPA (β = −13.667, SE = 5.503, t = −2.484, P = 0.023).

After correcting for multiple tests, the series of regression analyses predicting naming from the four GM PCs, days between assessment and scan dates, age, and visual ratings were significant for all models including T2 (Table 5) and FLAIR (Table 6) ratings. After correcting for multiple comparisons, the addition of T2 CHS Ventricles (FDR-corrected P = 0.033), T2 CHS WM (FDR-corrected P = 0.033), and T2 Fazekas PVH (FDR-corrected P = 0.049) ratings improved the predictive power of the base model, whereas the addition of FLAIR CHS Ventricles ratings (FDR-corrected P = 0.057) approached significance in improving the model’s predictive utility. Compared to the base model in naming abilities, the greatest increase in explained variance resulted from the addition of the T2 CHS WM ratings (from 64.78% to 88.49% of the adjusted explained variance).

TABLE 5.

Naming Regression Models With T2 Visual Ratings as Predictor Variables

	Multivariate Model
Model	(df) F	FDR-corr. P	Adj. R2	Univariate Predictors	P
#1	(7,15) 11.360	<0.001***	0.767	PC1: β = −2.032, SE = 1.138, t = −1.785	0.095
				PC2: β = 3.696, SE = 0.946, t = 3.906	0.001**
				PC3: β = −4.623, SE = 1.394, t = −3.316	0.005**
				PC4: β = 7.415, SE = 1.382, t = 5.366	<0.001***
				Age: β = 0.254, SE = 0.140, t = 1.809	0.090
				Days Assess–MRI: β = −0.048, SE = 0.020, t = −2.434	0.028*
				T2 CHS Ventricles: β = −2.814, SE = 0.927, t = −3.034	0.008**
#2	(7,13)† 22.970	<0.001***	0.885	PC1: β = −1.195, SE = 0.843, t = −1.417	0.180
				PC2: β = 4.145, SE = 0.773, t = 5.360	<0.001***
				PC3: β = −2.850, SE = 1.033, t = −2.758	0.016*
				PC4: β = 6.923, SE = 1.097, t = 6.310	<0.001***
				Age: β = 0.245, SE = 0.109, t = 2.250	0.042*
				Days Assess–MRI: β = −0.073, SE = 0.014, t = −5.262	<0.001***
				T2 CHS WM: β = −2.183, SE = 0.662, t = −3.295	0.006**
#3	(7,15) 10.120	<0.001***	0.744	PC1: β = −0.831, SE = 1.254, t = −0.663	0.517
				PC2: β = 4.344, SE = 0.999, t = 4.347	<0.001***
				PC3: β = −4.309, SE = 1.465, t = −2.941	0.010*
				PC4: β = 7.146, SE = 1.489, t = 4.801	<0.001***
				Age: β = 0.127, SE = 0.135, t = 0.938	0.363
				Days Assess–MRI: β = −0.051, SE = 0.021, t = −2.483	0.025*
				T2 Fazekas PVH: β = −5.095, SE = 1.927, t = −2.644	0.018*
#4	(7,15) 6.873	0.001**	0.651	PC1: β = −1.566, SE = 1.417, t = −1.106	0.286
				PC2: β = 4.086, SE = 1.159, t = 3.525	0.003**
				PC3: β = −4.344, SE = 1.717, t = −2.530	0.023*
				PC4: β = 6.968, SE = 2.131, t = 3.270	0.005**
				Age: β = 0.152, SE = 0.170, t = 0.894	0.385
				Days Assess–MRI: β = −0.052, SE = 0.024, t = −2.178	0.046*
				T2 Fazekas DWMH: β = −3.053, SE = 2.826, t = −1.080	0.297

Univariate predictors in each numbered model include a visual rating metric plus predictors from the base model.

^†The degrees of freedom for this model differ from those of the other models because two participants who were extreme outliers (ie, P33 and P39 with, respectively, 8.81 and 4.53 times the group’s mean Cook’s distance) were removed from the analysis.

^*Significant at P < 0.05.

^**Significant at P < 0.01.

^***Significant at P < 0.001.

CHS = Cardiovascular Health Study. Days Assess-MRI = number of days between assessment and MRI dates. df = degree of freedom. DWMH = deep white matter hyperintensities. FDR = false discovery rate. PC = principal component. PVH = periventricular hyperintensities. WM = white matter.

TABLE 6.

Naming Regression Models With FLAIR Visual Ratings as Predictor Variables

	Multivariate Model
Model	(df) F	FDR–corr. P	Adj. R2	Univariate Predictors	P
#1	(7,15) 9.495	<0.001***	0.730	PC1: β = −1.909, SE = 1.225, t = −1.559	0.140
				PC2: β = 3.497, SE = 1.032, t = 3.388	0.004**
				PC3: β = −4.190, SE = 1.509, t = −2.776	0.014*
				PC4: β = 7.503, SE = 1.494, t = 5.022	<0.001***
				Age: β = 0.207, SE = 0.147, t = 1.402	0.181
				Days Assess–MRI: β = −0.055, SE = 0.021, t = −2.602	0.020*
				FLAIR CHS Ventricles: β = −1.952, SE = 0.806, t = −2.422	0.029*
#2	(7,15) 6.618	0.001**	0.641	PC1: β = −2.058, SE = 1.432, t = −1.437	0.171
				PC2: β = 4.243, SE = 1.218, t = 3.484	0.003**
				PC3: β = −4.271, SE = 1.764, t = −2.420	0.029*
				PC4: β = 7.902, SE = 1.781, t = 4.436	<0.001***
				Age: β = 0.119, SE = 0.165, t = 0.719	0.483
				Days Assess–MRI: β = −0.051, SE = 0.025, t = −2.084	0.055^
				FLAIR CHS WM: β = −0.748, SE = 0.888, t = −0.842	0.413
#3	(7,15) 6.257	0.001**	0.626	PC1: β = −1.929, SE = 1.472, t = −1.311	0.210
				PC2: β = 4.013, SE = 1.215, t = 3.302	0.005**
				PC3: β = −4.335, SE = 1.988, t = −2.180	0.046*
				PC4: β = 8.192, SE = 1.967, t = 4.164	<0.001***
				Age: β = 0.099, SE = 0.179, t = 0.555	0.587
				Days Assess–MRI: β = −0.055, SE = 0.025, t = −2.205	0.043*
				FLAIR Fazekas PVH: β = −0.649, SE = 2.611, t = −0.249	0.807
#4	(7,15) 6.282	0.001**	0.627	PC1: β = −1.852, SE = 1.439, t = −1.287	0.218
				PC2: β = 3.986, SE = 1.196, t = 3.333	0.005**
				PC3: β = −4.505, SE = 1.773, t = −2.541	0.023*
				PC4: β = 8.226, SE = 1.818, t = 4.525	<0.001***
				Age: β = 0.112, SE = 0.188, t = 0.594	0.561
				Days Assess–MRI: β = −0.053, SE = 0.025, t = −2.143	0.049*
				FLAIR Fazekas DWMH: β = −0.863, SE = 2.646, t = −0.326	0.749

Univariate predictors in each numbered model include a visual rating metric plus predictors from the base model.

^*Significant at P < 0.05.

^**Significant at P < 0.01.

^***Significant at P < 0.001.

^{^}0.08 < > 0.05.

CHS = Cardiovascular Health Study. Days Assess-MRI = number of days between assessment and MRI dates. df = degree of freedom. DWMH = deep white matter hyperintensities. FDR = false discovery rate. PC = principal component. PVH = periventricular hyperintensities. WM = white matter.

Sentence Repetition Regression

The backward stepwise regression predicting the Sentence Repetition Test scores from the four GM PCs resulted in a significant model (F_1,18 = 5.675, P = 0.028, adjusted R² = 0.198); the only GM predictor retained in the model was PC2 (β = 0.907, SE = 0.381, t = 2.382, P = 0.028). The base model predicting sentence repetition from PC2 (β = 1.019, SE = 0.384, t = 2.653, P = 0.017), age (β = 0.009, SE = 0.040, t = 0.218, P = 0.830), and days between assessment and scan dates (β = −0.008, SE = 0.005, t = −1.568, P = 0.136) approached significance (F_3,16 = 2.779, P = 0.075, adjusted R² = 0.219). After correcting for multiple tests, the series of full regression models predicting sentence repetition from PC2, age, days between assessment and scan dates, and visual ratings were significant for the models that included T2 CHS WM, T2 Fazekas PVH, and T2 Fazekas DWMH ratings (Tables 7 and 8). While the addition of T2 CHS WM ratings (FDR-corrected P = 0.034) significantly improved the predictive power of the base model, the addition of T2 Fazekas PVH (FDR-corrected P = 0.071) and T2 Fazekas DWMH (FDR-corrected P = 0.071) ratings only trended toward improving the base model. The greatest increase in explained variance in sentence repetition abilities compared to the base model resulted from the addition of the T2 CHS WM ratings (from 21.9% to 52.5% of the adjusted explained variance).

TABLE 7.

Sentence Repetition Regression Models With T2 Visual Ratings as Predictor Variables

	Multivariate Model
Model	(df) F	FDR–corr. P	Adj. R2	Univariate Predictors	P
#1	(4,15) 3.279	0.065^	0.324	PC2: β = 0.796, SE = 0.377, t = 2.113	0.052^
				Age: β = 0.052, SE = 0.043, t = 1.194	0.251
				Days Assess–MRI: β = −0.005, SE = 0.005, t = −0.998	0.334
				T2 CHS Ventricles: β = −0.529, SE = 0.283, t = −1.867	0.082
#2	(4,15) 6.258	0.029*	0.525	PC2: β = 0.791, SE = 0.307, t = 2.576	0.021*
				Age: β = 0.086, SE = 0.038, t = 2.235	0.041*
				Days Assess–MRI: β = −0.006, SE = 0.004, t = −1.425	0.175
				T2 CHS WM: β = −0.728, SE = 0.216, t = −3.364	0.004**
#3	(4,15) 4.344	0.045*	0.413	PC2: β = 0.841, SE = 0.340, t = 2.472	0.026*
				Age: β = 0.037, SE = 0.036, t = 1.015	0.326
				Days Assess–MRI: β = −0.005, SE = 0.005, t = −1.000	0.333
				T2 Fazekas PVH: β = −1.342, SE = 0.535, t = −2.507	0.024*
#4	(4,15) 4.246	0.045*	0.406	PC2: β = 1.053, SE = 0.335, t = 3.139	0.007**
				Age: β = 0.054, SE = 0.039, t = 1.381	0.188
				Days Assess–MRI: β = −0.006, SE = 0.004, t = −1.440	0.170
				T2 Fazekas DWMH: β = −1.225, SE = 0.499, t = −2.455	0.027*

Univariate predictors in each numbered model include a visual rating metric plus predictors from the base model.

^*Significant at P < 0.05.

^**Significant at P < 0.01.

^***Significant at P < 0.001.

^{^}0.08 < > 0.05.

CHS = Cardiovascular Health Study. Days Assess–MRI = number of days between assessment and MRI dates. df = degree of freedom. DWMH = deep white matter hyperintensities. FDR = false discovery rate. PC = principal component. PVH = periventricular hyperintensities. WM = white matter.

TABLE 8.

Sentence Repetition Regression Models With FLAIR Visual Ratings as Predictor Variables

	Multivariate Model
Model	(df) F	FDR–corr. P	Adj. R2	Univariate Predictors	P
#1	(4,15) 2.569	0.081	0.248	PC2: β = 0.722, SE = 0.443, t = 1.629	0.124
				Age: β = 0.038, SE = 0.045, t = 0.837	0.415
				Days Assess–MRI: β = −0.006, SE = 0.005, t = −1.072	0.301
				FLAIR CHS Ventricles: β = −0.344, SE = 0.271, t = −1.272	0.223
#2	(4,15) 3.291	0.065^	0.325	PC2: β = 1.101, SE = 0.360, t = 3.060	0.008**
				Age: β = 0.026, SE = 0.038, t = 0.677	0.508
				Days Assess–MRI: β = −0.005, SE = 0.005, t = −1.030	0.319
				FLAIR CHS WM: β = −0.372, SE = 0.198, t = −1.875	0.080
#3	(4,15) 2.889	0.073^	0.285	PC2: β = 1.048, SE = 0.368, t = 2.848	0.012*
				Age: β = 0.030, SE = 0.040, t = 0.756	0.462
				Days Assess–MRI: β = −0.006, SE = 0.005, t = −1.169	0.261
				FLAIR Fazekas PVH: β = −0.760, SE = 0.485, t = −1.568	0.138
#4	(4,15) 2.805	0.073^	0.275	PC2: β = 1.009, SE = 0.370, t = 2.727	0.016*
				Age: β = 0.039, SE = 0.043, t = 0.911	0.377
				Days Assess–MRI: β = −0.005, SE = 0.005, t = −1.021	0.324
				FLAIR Fazekas DWMH: β = −0.861, SE = 0.576, t = −1.496	0.155

Univariate predictors in each numbered model include a visual rating metric plus predictors from the base model.

^*Significant at P < 0.05.

^**Significant at P < 0.01.

^***Significant at P < 0.001.

^{^}0.08 < > 0.05.

CHS = Cardiovascular Health Study. Days Assess-MRI = number of days between assessment and MRI dates. df = degree of freedom. DWMH = deep white matter hyperintensities. FDR = false discovery rate. PC = principal component. PVH = periventricular hyperintensities. WM = white matter.

DISCUSSION

We investigated the impact of WMH and ventricular dilation on disorder symptomology in individuals with PPA. We found significant associations between qualitative, visual ratings of these brain phenomena, conforming to the notion that WM changes and ventricle enlargement co-occur in neurodegenerative disease. Visual ratings—particularly WMH ratings—were associated with language deficits in individuals with PPA, accounting for age, days between assessment and scan dates, and cortical atrophy. To our knowledge, the present study represents the first attempt to examine the contributions of established brain pathology (ie, cortical atrophy) and structural metrics of global brain health on language deficits in individuals with PPA.

Before turning to the study’s main results, we briefly address the GM findings. Consistent with the PPA literature (Botha et al, 2015; Gorno-Tempini et al, 2004, 2011; Mesulam et al, 2009; Migliaccio et al, 2016; Race et al, 2013; Vandenberghe, 2016; Win et al, 2017), naming abilities in individuals with PPA were related to the integrity of widespread left perisylvian cortex, including frontal, inferior parietal, and superior and middle temporal gyri. Unexpectedly, better naming was associated with lower volumes of frontal regions. This finding was explained by better naming in individuals with nfaPPA than in individuals classified with other PPA variants. Sentence repetition abilities, on the other hand, were solely related to the integrity of the left inferior parietal volumes. This finding is consistent with the literature linking phrase- and sentence-level repetition deficits to atrophy in the left temporoparietal cortex, which is a core clinical feature of lvPPA (Gorno-Tempini et al, 2008, 2011; Leyton et al, 2015; Rohrer et al, 2013).

The main findings of our study confirmed that visual ratings of WMH and ventricular dilation are related to language deficits in individuals with PPA, even when controlling for other factors (eg, age and cortical atrophy) that are known to influence language abilities in this population. These results are consistent with those of previous studies (eg, Holz et al, 2017; Seo et al, 2012; van den Berg et al, 2018; van der Vlies et al, 2013; Verdelho et al, 2010) indicating that WMH in particular predict cognitive deficits and decline in dementia. Our study also demonstrated that greater WMH—and, to a lesser extent, ventricular enlargement—negatively impact language skills, associations that are explored less often than the impact of WMH on nonlinguistic cognitive deficits but are reported in poststroke samples (Basilakos et al, 2019; Wright et al, 2018; Yatawara et al, 2018). Diffusion-weighted imaging and tractography studies (Agosta et al, 2013; Catani et al, 2013; D’Anna et al, 2016; Elahi et al, 2017; Galantucci et al, 2011; Marcotte et al, 2017; Powers et al, 2013; Schwindt et al, 2013) have demonstrated that WM degeneration is present in individuals with PPA and likely contributes to language deficits across the PPA variants. Odolil and colleagues (2019) also found that greater WMH predict a faster decline over time in naming abilities in individuals with PPA. Future work aimed at disentangling relationships between language loss and progressive change in cortical atrophy and WM degeneration is warranted.

Our study also addressed issues regarding measurement of structural abnormalities that are visible on neuroimaging. In general, we found that visual ratings of WMH were better independent predictors of language deficits than visual ratings of ventricle size. The pathogenesis of ventricular enlargement is likely nonvascular in origin and more tightly linked to neuronal volume loss (Schmidt et al, 2011); thus, enlarged ventricles may reflect pathological mechanisms that are less distinct from cortical atrophy than WMH.

Across models, we also found that visual ratings that were performed on T2 scans better predicted language than ratings that were performed on FLAIR scans. FLAIR scans are particularly susceptible to enhanced PVH due to artifacts related to increased ventricle size and not WM pathology per se (Todd et al, 2018). As such, T2 ratings may have reflected actual WM disease more accurately—which was overall more predictive of language skills—than FLAIR ratings.

Furthermore, although global WMH (as reflected by T2 CHS WM ratings) were associated with both naming and sentence repetition abilities, PVH were additionally associated with naming deficits and DWMH were not. This finding is consistent with the assertion that PVH and DWMH may impact cognition differently and may reflect different disease mechanisms (Griffanti et al, 2018; Schmidt et al, 2011; Seo et al, 2012). For example, Schmidt et al (2011) hypothesized that smooth PVH (eg, caps surrounding the horns of the lateral ventricles, smooth halos surrounding the ventricles) have a nonvascular origin and are in fact caused by a disruption of the ependymal lining, unlike DWMH, which are vascular in origin. Findings from other studies have indicated that PVH have a negative impact on executive functioning in particular, whereas similar relationships between cognition and DWMH have not been found (Bombois et al, 2007; de Groot et al, 2000, 2002; Debette et al, 2007; Seo et al, 2012; van den Heuvel et al, 2006). While the importance of PVH versus DWMH cannot be definitively deduced from such studies, measuring PVH and DWMH separately is justified in individuals with PPA as in other populations.

Currently, the clinical utility of SVD metrics—and WMH in particular—in the diagnosis and treatment of individuals with PPA is unknown. As highlighted previously, WMH are common and are present not only in clinical populations, but also in ostensibly healthy, older individuals (de Leeuw et al, 2001; Longstreth et al, 1996). Therefore, WMH cannot be used for the differential diagnosis of PPA. However, measuring WMH in conjunction with cortical atrophy may be critical in explaining deficit patterns and variations in the rate of decline in individuals with PPA.

Vascular pathology exists in many individuals with dementia, and WMH were present in >95% of the current sample of individuals with PPA. Consequently, it stands to reason that the presence of WMH would be a frequent observation in the general PPA population. A related, and highly relevant, question pertains to the mechanisms that underlie the presence of WMH and relationships between WMH and language deficits in individuals with PPA. As posited by other researchers (Debette and Markus, 2010; Weller et al, 2015), it could be that WM tracts deteriorate independently of other pathology in dementia; thus, WMH represent an additional, separate factor contributing to the clinical expression of PPA symptoms. Alternatively, pathological changes in PPA could directly interact with WMH—for example, Wallerian degeneration due to cortical atrophy could result in greater WM changes—thus resulting in a worsening of clinical symptoms. Disentangling these mechanisms is outside the scope of the present study but represents an important avenue for future work aimed at characterizing the full impact that WMH have on disease progression in PPA.

Regarding treatment of PPA, no gold standard therapy currently exists to prevent expression or slow progression of the disorder. However, vascular risk factors (eg, hypertension, diabetes) are treatable and modifiable (Paradise et al, 2018), suggesting that early detection and management of such risk factors may be important for prophylactic care in individuals with PPA. To counter that point, however, in the present study, we found no significant relationships between the severity of WMH and hypertension, diabetes, or hypercholesterolemia status. It may be that these gross, dichotomous measures were unable to capture the mediating influence that such factors may have on the relationship between WMH and language. Prospective studies should include more precise measures of vascular risk factors in order to further explore their potential relationships with WMH in individuals with PPA.

Study Limitations and Future Directions

While our investigation elucidated important findings regarding the impact of WMH and ventricular dilation on language in individuals with PPA, our approach had some limitations. For one, our study was a retrospective analysis of data that had been collected for different studies; consequently, only a subset of the full sample was administered the same language measures. Furthermore, due to the necessity of assessing a wide range of cognitive–linguistic skills in these individuals, the measures used to assess nonverbal semantics and sentence repetition in our sample were brief and likely did not fully capture these skills. Nonetheless, the sentence repetition test that we chose is part of a standardized cognitive battery that is widely used in the dementia literature. In the present sample, variance in performance on this measure was also noted despite the limited number of assessment items. On the other hand, performance on the PPTT was highly right skewed; >75% of our sample performed within normal limits (ie, scores of 12–14). The lack of variance in PPTT scores likely contributed greatly to the dearth of significant results pertaining to predictions of nonverbal semantics.

The goal of our study was to assess the effects of WMH on language deficits in individuals with PPA. However, if WMH in the present sample were truly symptomatic of SVD (which is outside the scope of the present paper and cannot be confirmed), then relationships between WMH severity and vascular risk factors would be expected (although see earlier for an alternative explanation regarding the dearth of such associations). Other markers of SVD burden, such as dilated perivascular spaces, lacunes, and cerebral microbleeds, however, have also been linked to cognitive impairment and decline and dementia (Hase et al, 2018) but were not assessed in the present study. It has been suggested that combined cerebrovascular disease risk factors lead to stronger predictors of outcomes in individuals with dementia than single predictors alone (Paradise et al, 2018; Rensma et al, 2018). As such, the creation of a weighted score including multiple SVD markers may further increase our understanding of how these metrics relate to language impairment and decline in individuals with PPA.

Another study limitation pertains to the rating scales themselves. Visual, qualitative scales are subjective, gross measures of brain health (Caligiuri et al, 2015). In the present study, interrater reliability was shown to be strong for the ventricular ratings and moderate for the CHS WM ratings but minimal to moderate for the Fazekas ratings across scales and scan types. However, we attempted to mitigate the effects of interrater variability on our data by using consensus ratings by multiple authors. The influence of lower interrater reliability on the results, though, cannot be entirely discounted in our study, especially for the T2 Fazekas DWMH and FLAIR Fazekas PVH ratings. Furthermore, by design, the scales used in our study lack specificity regarding the location of structural abnormalities. Thus, a comparison of semi-automated quantitative measures of WMH and ventricular dilation to the present visual, qualitative ratings is warranted.

Most crucially, the sample size in our study was relatively small (especially compared to other studies of healthy individuals and individuals with other forms of dementia) and may not reflect the PPA population; thus, validation of the current findings with a larger sample is necessary. Future investigation of how WMH impact brain function (eg, resting state connectivity) could also result in a more complete picture of how such metrics contribute to language deficits and decline in individuals with PPA.

CONCLUSION

In this study, we found that the severity of WMH is associated with naming and sentence repetition deficits in individuals with PPA. Consequently, identifying the extent of WMH in individuals with PPA may assist clinicians and clinical researchers in delineating unique deficit profiles in specific individuals. Furthermore, given that no treatment exists for mitigating neuronal loss and subsequent language deficits in individuals with PPA, early identification of vascular and other risk factors associated with WMH could be critical in order to stem the compounding effects of WMH on other disease processes.

Glossary

CHS

Cardiovascular Health Study

DWMH

deep white matter hyperintensities

FDR

false discovery rate

FLAIR

fluid-attenuated inversion recovery

GM

gray matter

lv

logopenic variant

nfa

nonfluent agrammatic variant

PC

principal component

PPA

primary progressive aphasia

PPTT

Pyramids and Palm Trees Test

PVH

periventricular hyperintensities

ROI

region of interest

sv

semantic variant

SVD

small vessel disease

WM

white matter

WMH

white matter hyperintensities