The utility of lesion classification in predicting language and treatment outcomes in chronic stroke-induced aphasia

Abstract

Stroke recovery models can improve prognostication of therapy response in patients with chronic aphasia, yet quantifying the effect of lesion on recovery is challenging. This study aimed to evaluate the utility of lesion classification via gray matter (GM)-only versus combined GM plus white matter (WM) metrics and to determine structural measures associated with aphasia severity, naming skills, and treatment outcomes. Thirty-four patients with chronic aphasia due to left hemisphere infarct completed T1-weighted and DTI scans and language assessments prior to receiving a 12-week naming treatment. GM metrics included the amount of spared tissue within five cortical masks. WM integrity was indexed by spared tissue and fractional anisotropy (FA) from four homologous left and right association tracts. Clustering of GM-only and GM+WM metrics via k-medoids yielded four patient clusters that captured two lesion characteristics, size and location. Linear regression models revealed that both GM-only and GM+WM clustering predicted baseline aphasia severity and naming skills, but only GM+WM clustering predicted treatment outcomes. Spearman correlations revealed that without controlling for lesion volume, the majority of left hemisphere metrics were related to language measures. However, adjusting for lesion volume, no relationships with aphasia severity remained significant. FA from two ventral left WM tracts was related to naming and treatment success, independent of lesion size. In sum, lesion volume and GM metrics are sufficient predictors of overall aphasia severity in patients with chronic stroke, whereas diffusion metrics reflecting WM tract integrity may add predictive power to language recovery outcomes after rehabilitation.

Introduction

Aphasia is one of the most common and long-lasting sequelae of stroke, affecting approximately 30% of acute stroke survivors and persisting into the chronic post-stroke stage in 16-31% of these individuals (Engelter et al., 2006; Flowers et al., 2016). Aphasia typically results in a constellation of language deficits that greatly impair communication abilities and decrease functional independence (Boehme, Martin-Schild, Marshall, & Lazar, 2016; Tsouli, Kyritsis, Tsagalis, Virvidaki, & Vemmos, 2009). Despite the chronicity of aphasia, patients can continue to regain lost language skills in the years beyond stroke onset with targeted language treatment (Brady, Kelly, Godwin, Enderby, & Campbell, 2016). Nevertheless, variability in language recovery and response to treatment are hallmarks of the disorder (Code, Torney, Gildea-Howardine, & Willmes, 2010). As such, predictive models that provide prescriptive treatment recommendations for patients according to demographic, stroke and deficit profiles are critical for chronic aphasia care, yet such models do not yet exist.

A central barrier to accurate recovery prognostication is fully quantifying and incorporating lesion factors into predictive models (Price, Hope, & Seghier, 2017; Thye & Mirman, 2018). Historically, lesions in patients with aphasia have been described in terms of focal damage to anterior (e.g., Broca’s area) and/or posterior (e.g., Wernicke’s area) left hemisphere cortical gray matter (GM) regions. Such descriptions may be too simplistic given that infarct occurs along vascular territories—which involve many structures—rather than anatomical boundaries (Charidimou et al., 2014). For example, beyond cortical involvement, frank damage to or disconnection of white matter (WM) association pathways is common in stroke aphasia and has been associated with a variety of linguistic deficits in patients (e.g., Basilakos et al., 2014; Bonilha et al., 2017; Bonilha, Rorden, & Fridriksson, 2014; Breier, Hasan, Zhang, Men, & Papanicolaou, 2008; Del Gaizo et al., 2017; Geva, Correia, & Warburton, 2015; Gleichgerrcht et al., 2015; Griffis, Nenert, Allendorfer, & Szaflarski, 2017; Han et al., 2013, 2016; Harvey & Schnur, 2015; Harvey, Wei, Ellmore, Hamilton, & Schnur, 2013; Ivanova et al., 2016; Kummerer et al., 2013; Marchina et al., 2011; Marebwa et al., 2017; McKinnon et al., 2018; Rolheiser, Stamatakis, & Tyler, 2011; Rosso et al., 2015; Wilson et al., 2011). Given that language processing requires a network of interconnected brain regions, it has been suggested that WM integrity or combined GM and WM integrity may better predict aphasia recovery outcomes than GM metrics alone (Bonilha et al., 2014; Del Gaizo et al., 2017; Han et al., 2016; Marebwa et al., 2017; McKinnon et al., 2018; Mirman, Zhang, Wang, Coslett, & Schwartz, 2015; Naeser & Palumbo, 1994; Wilson et al., 2011; Xing et al., 2018). For example, Yourganov, Smith, Fridriksson, and Rorden (2015) found that a novel parcellation of the brain that considered spared tissue of both GM and WM outperformed GM-only parcellations in the prediction of aphasia syndromes in 98 patients. With regards to specific linguistic skills, Bonilha, Rorden, & Fridriksson (2014) found that even after controlling for cortical damage to left inferior frontal gyrus, pars triangularis (LIFGtri), structural disconnect of LIFGtri was independently associated with naming abilities in patients with chronic aphasia. Similarly, Xing et al. (2018) found that after accounting for demographic variables and overall left hemisphere lesion volume, naming skills in 45 patients with chronic aphasia were related to the integrity of 13 left intra-hemispheric white matter connections but were not associated with the extent of lesion in cortical nodes.

In addition to the aforementioned cross-sectional studies, modern computational models that utilize multimodal imaging datasets to link brain structure and language profiles show promise for improving prognostication of chronic aphasia recovery over time (Price et al., 2010). However, such models have not been utilized to predict response to evidence-based language treatment. Furthermore, the utility of GM-only versus combined GM plus WM (GM+WM) integrity metrics for predicting language abilities and—most crucially, treatment response—in chronic aphasia has not been fully explored. Additionally, most lesion mapping studies infer how damage results in impairment but the reverse inference—how the integrity of intact tissue relates to language function—can provide additional insights into the recovery capacity of patients in the chronic post-stroke stage (Charidimou et al., 2014; Price, Seghier, & Leff, 2010).

Therefore, the primary goals of the present investigation were to use multimodal structural metrics and clustering analyses to generate data-driven lesion classification schema and to subsequently compare the utility of GM-only versus combined GM+WM integrity metrics in predicting aphasia severity, naming abilities and response to naming treatment for word retrieval deficits (i.e., anomia) in chronic stroke patients. Anomia was selected as the deficit targeted in therapy because it is the most common impairment in chronic aphasia, is ubiquitous across aphasia subtypes and greatly affects the ability to participate in everyday conversation (Goodglass & Wingfield, 1997). We hypothesized that size and location would be the primary characteristics of patients’ lesions captured by k-medoids clustering, and that patients with similar lesions per these two dimensions would cluster together. We additionally predicted that combined GM+WM lesion classification would better predict language abilities than GM-only clusters. The second aim was to determine specific left and right hemisphere metrics that are associated with language outcomes. We predicted that the integrity of bilateral WM association tracts would relate to language outcomes independent of lesion volume whereas GM regions of interest (ROIs) would not.

Materials and Methods

Participants

Thirty-four individuals (24 males; mean age=61.91±10.84 years) with chronic stroke-induced aphasia (i.e., ≥ six months post-onset; mean time post-onset=61.18±86.49 months) due to left hemisphere ischemic infarct participated in the study. The primary eligibility requirement was presence of anomia. Exclusionary criteria included contraindications for MRI; active medical conditions that precluded study participation; history of neurological disease other than stroke; and history of multiple left hemisphere infarcts. All participants were premorbidly proficient in English and had normal or corrected-to-normal vision and hearing. Demographic and neurological case histories were obtained from medical records and study-specific questionnaires. Study protocols were executed in accordance with the institutional review boards of Boston University, Massachusetts General Hospital and Northwestern University. As needed, additional verbal and written explanations were utilized to ensure patient understanding of study protocols prior to obtaining their written informed consent.

Upon entering the study, all participants were administered the Western Aphasia Battery-Revised (WAB-R; Kertesz, 2006) and the Boston Naming Test (BNT; Kaplan, Goodglass, Weintraub, Segal, & van Loon-Vervoorn, 2001) to obtain measures of overall aphasia severity (per WAB-R Aphasia Quotient [AQ]) and naming impairment, respectively. A 180-item study-specific confrontation naming probe was utilized to supplement the BNT as a baseline measure of anomia, to determine treatment assignment and to measure therapy gains from the pre- to post-treatment time points.

Treatment Protocol

This study was conducted as part of a multi-site project (http://cnlr.northwestern.edu/) investigating the neurobiology of language recovery in patients with chronic aphasia. The study design was quasi-experimental as participants were pseudo-randomly assigned to enter either a treatment or a no-treatment/natural history branch of the study. Ultimately, 30 participants received modified semantic feature analysis-based treatment for anomia (Boyle, 2010; Kiran & Thompson, 2003) for four hours per week for up to 12 weeks or until criterion (i.e., ≥ 90% accuracy on trained items on two consecutive weekly probes) was met. Patients were trained on 36 items, split between two semantic categories (i.e., birds, vegetables, clothing or furniture). Category assignment was pseudo-counterbalanced across patients so that categories were assigned to patients who correctly named, on average, < 75% of within-category items across three pre-treatment naming probes. Sessions targeted training items via auditory and written feature judgment tasks, naming attempts before and after feature review, and generative naming tasks. Treatment success was determined by calculating the proportion of potential maximal gain (PMG) (Lambon Ralph, Snell, Fillingham, Conroy, & Sage, 2010) for each trained category using the following formula: $\frac{(AVGPostTxScore-AVGPreTxScore)}{(\mathrm{n}traineditems-AVGPreTxScore)}$, where AVG PreTx and AVG PostTx denote the averaged accuracy on trained items at pre-treatment and post-treatment, respectively. Averaged PMG across categories captured the degree to which patients’ naming of trained items improved secondary to therapy while accounting for their pre-treatment abilities. Alternative measures have been used in single-subject research designs to index treatment gains (e.g., Beeson & Robey, 2006; Howard, Best, & Nickels, 2015); we report individual scores for three such effect size measures in the Supplementary Materials. The full results of this treatment study are also reported in Gilmore, Meier, Johnson, and Kiran (2018).

MR Data Acquisition and Preprocessing

Patients underwent MR imaging on either a Siemens 3T Trio Tim at the Athinoula A. Martinos Center in Charlestown, MA or a 3T Siemens Prisma Fit at the Center for Translational Imaging in Chicago, IL. Participants completed T1-weighted sagittal imaging (TR/TE = 2300/2.91ms, T1 = 900ms, flip angle = 9°, FOV = 256×256mm, slice thickness = 1mm³, 176 sagittal slices) and a high resolution whole-brain DTI sequence (TR/TE = 900/92ms, T1 = 900ms, flip angle = 90°, FOV = 230×230mm, slice thickness = 1.98×1.98×2mm voxels, 70 interleaved slices with 60 gradient and 10 b0 volumes, b value = 1500 s/mm²).

Research assistants blinded to patients’ behavioral scores drew lesion maps slice-by-slice on T1 structural images using MRIcron (www.mccauslandcenter.sc.edu/crnl/mricron/). Frankly-damaged tissue was included while other structural abnormalities (e.g., enlarged ventricles due to ventricular dilation) were excluded from the manual lesion dissections. Spatial normalization of the T1-weighted images and lesion maps (in which lesioned voxels were preserved) to MNI space was performed in SPM12. Normalized lesion maps were filtered at 50%, and the volume of each map was calculated using in-house Matlab scripts.

Preprocessing of the DTI data was performed using a bespoke pipeline optimized for stroke data. First, diffusion images were denoised using principle components, and a b0 reference image was created from the mean of the non-diffusion weighted scans. The T1 structural image was skull-stripped and enantiomorphic warping of the intact right hemisphere into the damaged left hemisphere was performed. The modified T1 image was normalized to the mni_icbm152_T1_2009 template via nonlinear warping, and a pseudo-T2 image (created by inverting the T1 image contrast) was rigidly aligned to the b0 reference image. Next, the distortion field was calculated and eddy current distortion correction was performed. The affine matrix was applied to the b-vector file to yield rotated b-vectors. Eddy current corrected parameters were concatenated with the b0 distortion field and applied to the diffusion scans. Finally, the diffusion tensor was calculated using the nonlinear weighted positive definite tensor-fitting algorithm from Camino (http://camino.cs.ucl.ac.uk/) and then warped to MNI space to yield normalized scalar maps. For the current investigation, normalized fractional anisotropy (FA) maps were used to address study aims given that FA is the most commonly used scalar in both healthy and patient DTI studies to reflect integrity of white matter tract microstructure (Geva, Correia, & Warburton, 2011; O’Donnell & Westin, 2011).

Calculation of Spared GM ROIs and WM Tracts

Large GM ROI masks were generated by combining ROIs from the Automated Anatomical Labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002) that aligned with major anatomical boundaries of the lateral surface of the brain (see Table 1). As shown in Figure 1A, this procedure yielded six large ROIs: dorsolateral prefrontal cortex (DLPFC), inferior frontal (iFrontal), anterior and posterior temporal (aTemporal and pTemporal, respectively), parietal and occipital masks. The rationale for using large GM ROIs (as opposed to individual AAL ROIs) was three-fold: the ratio of participants to GM metrics was maximized in aim 1 analyses; the number of multiple comparison corrections in aim 2 analyses was minimized; and a comparable number of GM and WM structures were interrogated (see below).

Table 1.

Regions of interest (ROIs) included in large gray matter (GM) masks

Number in Figure 1	ROI mask	Automated Labeling Atlas (AAL) ROI
1	DLPFC	Superior frontal gyrus
2	DLPFC	Middle frontal gyrus
3	iFrontal	Inferior frontal gyrus (IFG), pars orbitalis
4	iFrontal	IFG, pars triangularis
5	iFrontal	IFG, pars opercularis
6	iFrontal	Insula
7	iFrontal	Rolandic operculum
8	iFrontal	Precentral gyrus
9	aTemporal	Superior temporal pole (TP)
10	aTemporal	Anterior superior temporal gyrus (STG)
11	aTemporal	Mid TP
12	aTemporal	Anterior middle temporal gyrus (MTG)
13	aTemporal	Anterior inferior temporal gyrus (ITG)
14	aTemporal	Anterior fusiform gyrus (FUSI)
15	pTemporal	Posterior STG
16	pTemporal	Posterior MTG
17	pTemporal	Posterior ITG
18	pTemporal	Posterior FUSI
19	Parietal	Postcentral gyrus
20	Parietal	Superior parietal lobule
21	Parietal	Inferior parietal lobule
22	Parietal	Supramarginal gyrus
23	Parietal	Angular gyrus
24	Occipital	Superior occipital cortex
25	Occipital	Mid occipital cortex
26	Occipital	Inferior occipital cortex
27	Occipital	Cuneal cortex
28	Occipital	Calcarine cortex
29	Occipital	Lingual gyrus

DLPFC = dorsolateral prefontal cortex; iFrontal = inferior frontal; aTemporal = anterior temporal; pTemporal = posterior temporal

Figure 1.

Anatomical masks. (A) Gray matter (GM) regions of interest (ROIs), including DLPFC (purple), iFrontal (green), aTemporal (red), pTemporal (navy blue), Parietal (cyan) and Occipital (yellow). Numbers denote ROIs in the AAL atlas, listed in Table 1. (B) Bilateral white matter (WM) tract masks, including—from left to right—the arcuate (in red), inferior fronto-occipital (in green), inferior longitudinal (in blue) and uncinate (in violet) fasciculi. (C) From left to right, the normalized T1 image, lesion map, and subject-specific GM and WM masks shown for a sample patient. Matrices below the masks represent 4×4mm swathes of tissue as a toy example of the procedure used to calculate spared tissue and fractional anisotropy (FA) per each subject-specific mask. In the far left and middle matrices, 1 = spared and 0 = lesioned voxels in the GM and WM masks, respectively; the font color corresponds to the mask from which spared voxels were extracted. The numbers in the third matrix (at right) reflect FA values per voxel in the same 4×4mm of WM tissue shown in the middle matrix.

To determine the integrity of canonical WM pathways implicated in lexical-semantics and naming, we extracted tract masks from the Johns Hopkins University White Matter Tractography atlas (Wakana, Jiang, Nagae-Poetscher, van Zijl, & Mori, 2004) that corresponded to the bilateral arcuate (AF), inferior fronto-occipital (IFOF), inferior longitudinal (ILF) and uncinate (UF) fasciculi (Figure 1B). Tract masks were thresholded with a probability value of 0.20, binarized and resampled to the resolution and dimensions of the FA maps. To ensure optimal image alignment, all lesion maps and WM tract masks were warped to the mni_icbm152_T1_2009 template via a 12-parameter affine transformation.

To determine the amount of spared GM and WM tissue, non-lesioned voxels were retained from the intersection of ROI and tract masks and each patient’s normalized lesion map. The number of spared voxels from each subject-specific ROI mask (excluding the occipital mask since it was mostly spared in the sample) and subject-specific tract mask were extracted and used in subsequent analyses (Figure 1C). Subject-specific tract masks were also multiplied by each patient’s FA map on a voxel by voxel basis. To decrease the potential for partial volume effects, only voxels with FA values ≥ 0.20 were averaged to yield mean FA (FA_mean) per tract mask.

Statistical Analysis

Statistical analyses were performed in R (R Core Team, 2017). For aim 1, we conducted two k-medoids cluster analyses using the ‘cluster’ package (Maechler, Rousseeuw, Struyf, Hubert, & Hornik, 2018) and Partitioning Around Medoids (PAM) algorithm to determine how patients clustered according to (1) GM-only metrics (i.e., number of spared voxels in the DLPFC, iFrontal, aTemporal, pTemporal and Parietal masks) versus (2) combined GM+WM metrics (i.e., number of spared voxels in GM masks plus spared voxels in and FA extracted from the left AF, IFOF, ILF and UF masks). Like k-means, k-medoids partitions a dataset into a set of k maximally-dissimilar groups (in this case patient subgroups) but is less sensitive than k-means to outliers and noise in the data (Kaufman & Rousseeuw, 1990). Cluster assignment was based on the Manhattan distance between the center of each cluster and other patient data points. Within the Manhattan distance formula, pairwise dissimilarities between data points are based on absolute distance (as opposed to the squared error distance); thus, cluster assignment is less influenced by outliers than when other distance formulas (e.g., Euclidean distances) are applied. Elbow plots were used to specify the number of clusters, and cluster assignments of individual participants from the two analyses (i.e., GM-only and GM+WM clustering) were obtained.

Given our hypothesis that overall lesion size would be reflected in the clustering results, we first performed linear regressions predicting total lesion volume from GM-only and GM+WM cluster membership. Next, to test whether clustering also captured lesion location, lesion subtraction plots (i.e., the pure subtraction of overlaid lesion maps of patients in one cluster from overlaid lesion maps of another cluster) were generated and visualized in MRIcron. Finally, linear regression models were constructed to predict aphasia severity (per WAB-R AQ), baseline naming skills (per BNT total correct [BNT_total]) and naming treatment outcomes (per PMG) from the categorical predictors of either GM-only or combined GM+WM clusters. These predictors were dummy-coded so that cluster 1 was set as the reference group in both types of analyses. The ‘gvlma’ package in R (Pena & Slate, 2014) was used to check model assumptions.

For aim 2, Spearman correlations between each behavioral metric (i.e., AQ, BNTtotal and PMG) and each left hemisphere structural metric (i.e., number of spared voxels in each left hemisphere mask and FA extracted from left WM tract masks) were conducted. The same analyses were then run as partial correlations controlling for lesion size. To assess the potential of right hemisphere structural compensation, correlations were also conducted between each language measure and FA extracted from the right WM tract masks, with and without controlling for left hemisphere lesion volume. These analyses were conducted to better understand the specific brain structures that may be most relevant for predicting different language outcomes. All correlations were corrected for multiple tests according to the false discovery rate (FDR; Benjamini & Hochberg, 1995).

To support our findings regarding pre-treatment structural correlates of naming therapy gains, we replicated all treatment-related statistical analyses for both aims by replacing PMG with effect sizes calculated according to recommendations from Beeson and Robey (2006). These results are reported in the Supplementary Materials. The dataset generated and analyzed during the current study is available in the OSF repository (DOI 10.17605/OSF.IO/3EG4X).

Results

As shown in Table 2, patients ranged in their overall aphasia severity with a mean ± SD WAB-R AQ of 62.25 ± 24.60 out of 100. Variability in baseline naming skills was observed, with a mean of 24.76 ± 20.19 out of 60 items named correctly on the BNT and an average of 32.72 ± 25.13% of correctly-named items on the study-specific probe. As stated previously, the full treatment results are available elsewhere (Gilmore et al., 2018). In brief, patients benefitted from therapy such that patients achieved, on average, approximately 43% of their potential naming gain following therapy. As illustrated by Figure 2 and shown in Table 2, total lesion volume also varied across the group, suggesting that a variety of lesion and language deficit profiles were represented in the sample.

Table 2.

Demographic, behavioral and lesion data for all patients

ID	Sex	Handedness	Age	MPO	Education	Lesion Volume (cc)	WAB-R AQ (/100)	BNT (/60)	Naming Probe (%)	PMG
P1	M	R	55	12	16	57.25	87.2	50	58.33	0.89
P2	F	L	50	29	16	249.93	25.2	1	0.99	0.00
P3	F	R	63	62	16	175.38	52.0	10	17.59	0.36
P4	M	R	79	13	16	84.78	74.1	52	67.96	1.00
P5	M	R	67	8	18	171.94	30.8	4	6.11	−0.07
P6	M	R	49	113	16	298.97	66.6	44	55.97	0.74
P7	M	R	55	137	16	181.97	48.0	6	14.07	0.27
P8	F	R	71	37	16	11.66	95.2	45	59.07	0.83
P9	F	R	53	12	16	76.55	80.4	37	64.81	0.83
P10	M	R	78	22	18	32.11	92.1	41	33.70	0.39
P11	M	R	68	104	12	186.85	40.0	1	2.78	0.03
P12	M	L	42	18	13.5	12.13	92.7	43	56.94	0.98
P13	F	R	64	24	13	96.93	64.4	41	40.56	0.77
P14	F	R	71	74	12	189.31	87.2	43	56.48	0.28
P15	M	R	61	152	16	163.49	74.3	54	52.22	0.98
P16	F	R	70	152	16	69.64	78.0	24	48.33	0.76
P17	M	R	80	22	18	89.03	28.9	1	7.78	0.22
P18	F	R	48	14	16	164.33	13.0	0	0.00	0.42
P19	M	R	65	16	18	247.59	11.7	0	0.37	0.06
P20	M	R	62	12	16	100.02	65.4	1	7.22	0.11
P21	M	R	60	24	16	172.81	45.2	6	5.19	0.04
P22	M	R	69	170	16	183.45	40.4	3	6.85	0.14
P23	F	R	76	33	18	184.39	37.5	2	2.22	0.06
P24	F	R	64	115	12	127.70	58.0	15	20.56	0.20
P25	M	R	62	15	12	76.65	56.0	21	35.74	0.42
P26	M	R	49	49	12	87.59	85.5	53	68.61	1.00
P27	M	R	81	11	12	51.70	73.8	24	40.56	n/a
P28	M	R	49	67	12	317.07	32.3	3	5.00	0.00
P29	M	R	39	18	16	26.22	71.3	36	47.22	0.23
P30	M	L	64	13	12	34.15	79.6	41	45.93	0.63
P31	M	L	62	21	16	1.57	91.5	42	74.26	n/a
P32	M	R	68	21	13.5	80.28	82.5	33	31.48	n/a
P33	M	R	58	23	14	186.52	61.8	10	11.94	0.37
P34	M	R	53	467	17	120.82	94.0	55	65.74	n/a
AVG			62.18	61.18	15.09	126.79	62.25	24.76	32.72	0.43
SD			10.89	86.49	2.11	82.13	24.60	20.19	25.13	0.36

M=male; F=female; L=left-handed; R=right-handed; MPO=months post-onset; WAB-R AQ=Western Aphasia Battery-Revised Aphasia Quotient; BNT=Boston Naming Test; PMG=Proportion of potential maximal gain; AVG=average; SD=standard deviation; cc = cubic centimeters

Figure 2.

Patient lesion data. Lesion overlay and lesion volume mean (denoted by $\stackrel{‒}{\mathrm{x}}$) in cubic centimeters (cc) and distribution for n = 34 patients.

Table 3 summarizes descriptive statistics regarding the amount of spared tissue in left hemisphere masks and FA in bilateral tract masks. On average, the most damaged GM and WM structures were the iFrontal ROI and left AF and UF whereas the most spared regions and tracts were DLPFC and the left IFOF and ILF. As expected, FA was significantly lower in the left compared to the right hemisphere association tracts (F(4,60) = 54.210, Pillai’s trace = 0.828, p < 0.001) (see Table 3).

Table 3.

Summary of ROI and tract mask results

		FA (mean±SD)
Mask	LH %spared (mean±SD)	LH	RH	F(1,63)	p-value
DLPFC	93.02±11.20	.	.	.	.
iFrontal	67.77±27.69	.	.	.	.
aTemporal	75.60±19.19	.	.	.	.
pTemporal	78.61±17.76	.	.	.	.
Parietal	77.92±24.25	.	.	.	.
AF	57.11±28.45	0.353±0.052	0.496±0.047	132.86	<0.001
IFOF	74.01±16.35	0.375±0.043	0.409±0.028	14.46	<0.001
ILF	77.70±17.47	0.365±0.048	0.394±0.028	7.709	0.0072
UF	57.84±31.31	0.322±0.037	0.443±0.026	250.02	<0.001

LH=left hemisphere; RH=right hemisphere; SD=standard deviation; DLPFC = dorsolateral prefontal cortex; iFrontal = inferior frontal; aTemporal = anterior temporal; pTemporal = posterior temporal; AF = arcuate fasciculus; IFOF = inferior fronto-occipital fasciculus; ILF = inferior longitudinal fasciculus; UF = uncinate fasciculus

Aim 1: GM-Only versus GM+WM Clustering

Elbow plots for the GM-only k-medoids analysis indicated that four patient clusters were most representative of the data. As shown in Figure 3A, the first and second clustering dimensions captured, respectively, 54.7% and 22.9% of the variance in GM integrity. The regression predicting lesion volume from cluster membership trended towards significance (F(1, 32) = 4.116, p = 0.051, R² = 0.086). As shown in Figure 3B, though, a clear difference in lesion volume was observed between patients within clusters 1 and 2. Indeed, cluster membership predicted lesion volume in a regression including only these patients (F(1, 15) = 54.210, p < 0.001, R² = 0.769), such that patients in cluster 1 had significantly smaller lesions than individuals in cluster 2. Given that patients within clusters 1 and 2 maximally differed along the x-axis, we concluded that the first dimension primarily captured lesion size. Therefore, we hypothesized that the second dimension reflected lesion location and would differentiate patients in clusters 3 and 4. As shown in Figure 3C, lesion subtraction plots indicated that patients in cluster 3 had primarily ventral lesions whereas patients in cluster 4 had primarily dorsal lesions.

Figure 3.

GM-only clustering. (A) Results of the k-medoids analysis using GM-only ROI metrics. P1-34 correspond to single patients, Dim = Dimension. (B) Lesion volume per cluster in cubic centimeters (cc). (C) Lesion subtraction plots of patients in cluster 3 (in orange) versus cluster 4 (in green). Brighter colors reflect greater overlap of patients’ lesions.

Four patient clusters were also retained from the combined GM+WM data. As shown in Figure 4A, the two primary dimensions captured 46.7% and 18.2% of the variance in combined GM and WM integrity. Considering the GM-only clustering results, we hypothesized that the first and second GM+WM dimensions captured lesion size and location, respectively. Indeed, as shown in Figure 4B, GM+WM cluster membership predicted total lesion volume (F(1, 31) = 8.145, p = 0.008, R² = 0.183), where patients in clusters 1 and 3 had smaller lesions than did patients in clusters 2 and 4. Again, lesion map subtraction plots of patients in cluster 1 versus cluster 3 (Figure 4C) and patients in cluster 2 versus 4 (Figure 4D) illustrate that patients in clusters 1 and 4 had primarily ventral lesions whereas patients in clusters 2 and 3 had primarily dorsal lesions. Thus, we established that the two primary dimensions captured in both the GM-only and GM+WM clusters were lesion size and location.

Figure 4.

GM+WM clustering. (A) Results of the k-medoids analysis using combined GM and WM metrics, Dim = Dimension. (B) Lesion volume per cluster in cubic centimeters (cc). (C) Lesion subtraction plots of patients in cluster 1 (in yellow) versus cluster 3 (in vermillion). (D) Lesion subtraction plots of patients in cluster 2 (in blue) versus cluster 4 (in pink). Brighter colors reflect greater overlap of patients’ lesions.

The predictive utility of the GM-only versus GM+WM clusters differed by language measure. Specifically, both clustering types significantly predicted overall aphasia severity per WAB-R AQ, yet the predictive power of the GM-only lesion classification (F(3, 29) = 12.260, p < 0.001, R² = 0.559) was higher than the combined GM+WM clustering (F(3, 29) = 9.562, p < 0.001, R² = 0.497). As shown in Figure 5A, within the GM-only predictive model, patients within cluster 2 (β = −48.432, t = −5.363, p < 0.001), cluster 3 (β = −29.075, t = −3.428, p = 0.002) and cluster 4 (β = −31.932, t = −4.396, p < 0.001) had significantly lower WAB-R AQ scores than patients in cluster 1. Within the GM+WM model, patients in cluster 2 (β = −36.610, t = −4.720, p < 0.001) and cluster 4 (β = −36.940, t = −3.831, p < 0.001)—but not cluster 3 (β = −11.100, t = −1.225, p = 0.230)—had significantly more severe aphasia than patients in cluster 1.

Figure 5.

Language measures, including A) aphasia severity per WAB-R AQ, B) naming skills per BNT total correct and C) treatment gains per (PMG) plotted by GM-only and GM+WM clusters. *** p < 0.001, ** p < 0.01, * p < 0.01, n.s. = not significant

In contrast to the AQ models, GM+WM clustering provided greater explanatory power of pre-treatment naming skills per BNT_total (F(3, 29) = 4.653, p = 0.009, R² = 0.325) than did the GM-only clusters (F(3, 29) = 3.396, p = 0.031, R² = 0.260). As shown in Figure 5B, compared to patients in GM-only cluster 1, patients in cluster 2 (β = −26.533, t = −2.790, p = 0.009) and cluster 4 (β = −17.533, t = −2.292, p = 0.029) had significantly lower BNTtotal scores, while patients in cluster 3 trended towards lower BNT_total scores than patients in cluster 1 (β = −17.333, t = −1.940, p = 0.062). Within the combined GM+WM model, patients in cluster 2 (β = −24.933, t = −3.412, p = 0.002) and cluster 4 (β = −24.133, t = −2.657, p = 0.013) had significantly lower BNT_total scores than patients in cluster 1, whereas naming skills between patients in clusters 1 and 3 did not significantly differ (β = −12.500, t = −1.465, p = 0.154).

Critically, GM+WM clustering significantly predicted treatment outcomes as measured by PMG (F(3, 26) = 3.036, p = 0.047, R² = 0.260) but the GM-only clusters did not (F(3, 26) = 1.773, p = 0.177, R² = 0.170) (see Figure 5C). Within the GM+WM model, patients in cluster 2 (β = −0.376, t = −2.571, p = 0.016) and cluster 4 (β = −0.445, t = −2.483, p = 0.020) had significantly lower treatment outcomes than patients in cluster 1; patients in clusters 1 and 3 similarly benefitted from naming therapy (β = −0.210, t = −1.174, p = 0.251). While the predictive utility of these models cannot be directly, statistically compared (as would be the case in nested models), the amount of explained variance indicates that the GM-only model best predicts overall aphasia severity, while baseline naming skills and naming treatment outcomes are best predicted by combined GM+WM metrics.

Aim 2: Relationships between Structural Metrics and Language Outcomes

As shown in the second column of Table 4, without controlling for lesion volume, significant positive correlations were found between WAB-R AQ and all left hemisphere GM and WM metrics (range: r = 0.393 – 0.721, p = 0.024 – < 0.001 after FDR correction). Similarly, without the lesion volume covariate, significant relationships were found between the majority of left hemisphere GM and WM metrics and BNT_total (range: r = 0.39 – 0.53, p = 0.035 – 0.002) (see Table 4, third column) and PMG (range: r = 0.34 – 0.64, p = 0.035 – < 0.001) (Table 4, last column). As the only exceptions, spared voxels in the aTemporal and pTemporal masks were not associated with either BNT or PMG, and FA_mean in left UF was not related to PMG. No significant relationships between language measures and right hemisphere tract FA were found (see the bottom four rows of Table 4).

Table 4.

Correlations between structural metrics and language variables, with and without controlling for total lesion volume

	Aphasia Severity (WAB-R AQ)		Naming Skills (BNT Total)		Treatment Gains (PMG)
Measure	Corr.	Partial corr.	Corr.	Partial corr.	Corr.	Partial corr.
DLPFC (%sp)	0.512**	−0.088(n.s.)	0.416*	0.014(n.s.)	0.557**	0.111(n.s.)
iFrontal (%sp)	0.657***	0.229(n.s.)	0.435*	0.041(n.s.)	0.432*	−0.117(n.s.)
aTemporal (%sp)	0.600***	0.256(n.s.)	0.270(n.s.)	−0.113(n.s.)	0.289(n.s.)	−0.169(n.s.)
pTemporal (%sp)	0.494**	0.238(n.s.)	0.231(n.s.)	−0.036(n.s.)	0.196(n.s.)	−0.143(n.s.)
Parietal (%sp)	0.629***	0.138(n.s.)	0.435*	0.022(n.s.)	0.406*	−0.161(n.s.)
Left AF (%sp)	0.721***	0.371(n.s.)	0.492**	0.144(n.s.)	0.642***	0.333(n.s.)
Left IFOF (%sp)	0.641***	0.252(n.s.)	0.507**	0.211(n.s.)	0.638***	0.370(n.s.)
Left ILF (%sp)	0.567**	0.243(n.s.)	0.423*	0.155(n.s.)	0.506**	0.211(n.s.)
Left UF (%sp)	0.641***	0.261(n.s.)	0.392*	0.018(n.s.)	0.454*	0.040(n.s.)
Left AF (FA)	0.518**	0.174(n.s.)	0.384*	0.108(n.s.)	0.502**	0.259(n.s.)
Left IFOF (FA)	0.393*	0.427(n.s.)	0.441*	0.441*	0.513**	0.560**
Left ILF (FA)	0.465**	0.468(n.s.)	0.527**	0.507*	0.624***	0.658***
Left UF (FA)	0.454**	0.410(n.s.)	0.348*	0.267(n.s.)	0.212(n.s.)	0.117(n.s.)
Right AF (FA)	−0.129(n.s.)	−0.169(n.s.)	−0.123(n.s.)	−0.137(n.s.)	−0.161(n.s.)	−0.234(n.s.)
Right IFOF (FA)	0.186(n.s.)	0.323(n.s.)	0.250(n.s.)	0.332(n.s.)	0.184(n.s.)	0.281(n.s.)
Right ILF (FA)	0.128(n.s.)	0.425(n.s.)	0.166(n.s.)	0.349(n.s.)	0.016(n.s.)	0.202(n.s.)
Right UF (FA)	0.250(n.s.)	0.107(n.s.)	0.301(n.s.)	0.207(n.s.)	0.197(n.s.)	0.005(n.s.)

WAB-R AQ=Western Aphasia Battery-Revised Aphasia Quotient; BNT=Boston Naming Test; PMG=Proportion of potential maximal gain; Corr. = Correlation; %sp = percentage of spared mask tissue; FA = fractional anisotropy; DLPFC = dorsolateral prefontal cortex; iFrontal = inferior frontal; aTemporal = anterior temporal; pTemporal = posterior temporal; AF = arcuate fasciculus; IFOF = inferior fronto-occipital fasciculus; ILF = inferior longitudinal fasciculus; UF = uncinate fasciculus. An FDR correction via the Benjamini-Hochberg (BH) method was performed to correct for multiple comparisons; significance levels reflect adjust p-values.

^***p < 0.001

^**p < 0.01

^*p < 0.05, n.s. = not significant

By contrast, partial correlations controlling for lesion volume revealed no significant relationships remained between AQ and any integrity metrics from either hemisphere (Table 4, second column). When controlling for lesion volume, relationships between BNT_total and FA in the left IFOF (partial r = 0.441, p = 0.046) and FA in the left ILF (partial r = 0.507, p = 0.025) remained significant (see Figure 6A and Table 4, third column). Similarly, partial correlations with lesion volume revealed remaining significant associations between PMG and FA in the left IFOF (partial r = 0.560, p = 0.006) and FA in the left ILF (partial r = 0.658, p < 0.001) (see Figure 6B and Table 4, last column). As shown in the bottom four rows of Table 4, no significant partial correlations between either BNT or PMG and FA in the right WM tracts were found.

Figure 6.

Associations between language measures and structural metrics, controlling for lesion volume, including (A) BNT total correct and fractional anisotropy (FA) in the left inferior fronto-occipital fasciculus (IFOF) and inferior longitudinal fasciculus (ILF) and (B) proportion of potential maximal gain (PMG) and FA in the left IFOF and ILF.

Discussion

In the present study, we investigated the predictive power of different types of gray and white matter integrity metrics for understanding language skills and naming treatment outcomes in patients with chronic stroke-induced aphasia. One novelty of this study was the use of multimodal imaging in classifying stroke patients according to structural indices of the lesioned hemisphere. Via the cluster analyses, we found that 13 separate GM and WM metrics reflecting the integrity of left hemisphere structures involved in language processing could be boiled down into simple, clinically-accessible classification schemas. Specifically, according to the GM-only metrics, patients were classified as having either small, large, or medium-sized lesions, the latter which further split into ventral versus dorsal involvement. Clustering of combined GM+WM measures resulted in four lesion classifications that each carried information about lesion size and reflected the dorsal/ventral bifurcation of the middle cerebral artery (i.e., small ventral, large ventral, small dorsal, and large dorsal lesions). These results reflect a marriage of methods researchers use to index brain damage after stroke (e.g., percentage of spared/lesioned tissue, DTI measures) and ways in which clinicians typically qualify patient lesions (i.e., general estimate of size and gross site of damage).

Furthermore, as referenced above, the cluster results also indicated such structural metrics carry non-dissociable lesion size and location information. Recently, Thye and Mirman (2018) explicitly set out to determine the relative contribution of lesion size versus location in predicting overall aphasia severity and various linguistic deficits. In this paper, the authors found that overall aphasia severity and naming abilities were mainly predicted by lesion size whereas speech recognition and speech production skills were best predicted by lesion location. Similarly, in the present study, none of the partial correlations between WAB-R AQ and integrity metrics were significant; thus, it can be inferred that total lesion volume also adequately captured overall aphasia severity in our patient sample. This finding aligns with other studies (see Watila & Balarabe, 2015 for review) and is not surprising given that receptive and expressive language tasks comprise WAB-R AQ, and lesions in the aphasic population span frontal, temporal and parietal regions implicated in these abilities. Analogous to Thye and Mirman's (2018) lesion location finding, we found that naming abilities—and naming treatment outcomes—were associated with overall lesion volume, but we also found that the integrity of specific left hemisphere structures (i.e., ILF and IFOF) was a critical, independent determinant of these language measures. Subtle differences between the present findings and Thye and Mirman (2018) may lie in differences in the neuroimaging methods employed in each study (e.g., a whole-brain versus ROI approach). Future work should aim to further disentangle the relative impact lesion location and size (as well as other structural factors) have on stroke recovery in general and response to specific therapies in particular.

Consistent with our hypotheses, we also found that GM+WM cluster membership was a stronger predictor of pre-treatment naming skills and better indicator of the potential to improve word retrieval abilities after treatment than GM-only clustering. The fact that only diffusion measures explained additional variance in pre-treatment naming and treatment outcomes indicates that DTI scalars capture different, critical information about structural integrity in chronic stroke. In particular, the integrity of the left ILF and IFOF—as reflected by FA—were related to both pre-treatment naming and treatment outcomes. The importance of structural connectivity of left temporal regions in the preservation of lexical-semantic and naming skills in patients with chronic aphasia has been highlighted in recent cross-sectional studies (Gleichgerrcht et al., 2015; Han et al., 2013; Harvey & Schnur, 2015; Ivanova et al., 2016; Yourganov, Fridriksson, Rorden, Gleichgerrcht, & Bonilha, 2016). Controlling for demographic variables and total lesion volume, McKinnon et al. (2018) found that the integrity of a posterior segment of the left ILF (per axonal water fraction, a diffusion metric of axonal density) was associated with semantic paraphasias, independent of overall left hemisphere damage and the degree of cortical necrosis in regions inferior to the perisylvian fissure. Our findings cohere with this result, given that spared tissue in the aTemporal and pTemporal ROIs was not related to baseline naming or treatment outcomes and therefore would not alter the relationship between these language metrics and left ILF and left IFOF FA. With regards to treatment outcomes, Bonilha, Gleichgerrcht, Nesland, Rorden, and Fridriksson (2016) used structural connectome methods and found that temporal lobe connections were particularly critical for anomia therapy success. In a recent study, McKinnon et al. (2017) used diffusion keurtosis imaging and found neuroplastic changes in the left ILF of patients who underwent semantic feature-based anomia treatment.

One surprising finding of the current study is that no measure of the left AF remained significant for any language measure after controlling for overall lesion volume. Damage to the left AF has often been implicated in naming impairments in patients with aphasia (e.g., Geva et al., 2015; Han et al., 2016; Ivanova et al., 2016; Marchina et al., 2011; van Hees et al., 2014; Wang, Marchina, Norton, Wan, & Schlaug, 2013). However, the lack of left AF findings may be related to the nature of the anomia therapy, which directly targeted semantic feature re-learning and therefore relied heavily on ventral stream processes. It therefore stands to reason that baseline integrity of the left IFOF and left ILF—more than the AF—would predict a patient’s ability to regain semantic knowledge that results in improved naming skills.

Similarly, none of the right hemisphere tract metrics predicted language skills independent of lesion volume. Researchers have postulated that one mechanism of stroke recovery is structural and functional compensation by the non-damaged hemisphere. Indeed, Forkel et al. (2014) found that greater volume of the right AF derived from acute diffusion images predicted greater degree of recovery from aphasia six months post-stroke onset. Yet, similar to the present study, they found no significant relationships between FA of right hemisphere tracts and aphasia recovery (Forkel et al., 2014). It may be that the integrity of right hemisphere structures is most crucial for recovery of language skills in the acute and subacute phases of left hemisphere stroke but not as critical by the chronic stage. Our findings align with this interpretation, given that no significant positive or negative associations between language measures and right hemisphere metrics were found.

While the results of the current study are compelling, some limitations must be acknowledged. First, the variance in naming and treatment outcomes explained by the lesion classification systems was low, and some p-values approached the 0.05 mark. Although the present sample represents one of the largest samples of patients in the aphasia treatment neuroimaging literature, it is possible the cluster analyses suffered from lack of power. Thus, future studies should endeavor to replicate these findings with a larger sample. Moreover, given that proximal anatomical regions are often damaged together following stroke (Charidimou et al., 2014), a primary rationale for creating large GM ROIs was to reduce noise in the spared tissue calculations, thereby increasing the robustness of subsequent statistical analyses. Nevertheless, it is possible that some granularity in quantifying cortical GM damage was lost and that more finegrained cortical metrics would predict language outcomes independent of overall lesion volume. Along the same vein, the manner in which gray matter integrity was measured in the present study did not allow for inclusion of right hemisphere gray matter metrics. As such, future investigations that include bilateral measures of GM integrity (e.g., cortical thickness) may improve the predictive power of brain structure indices.

Conclusions

To our knowledge, this is the first investigation to directly compare the prognostic utility of specific GM and WM measures for predicting naming treatment success in patients with chronic aphasia. We found that GM-only and GM+WM clustering adequately predicted overall language impairment (i.e., aphasia and anomia severity) but only the combined GM+WM classification predicted the potential of benefitting from anomia treatment. Furthermore, while it may be that overall lesion volume can serve as a proxy for aphasia severity in chronic patients, diffusion metrics reflecting the integrity of left ventral WM association pathways are critical predictors of naming abilities and treatment success. These results, therefore, suggest that the inclusion of diffusion measures of WM pathways into aphasia recovery models would improve prognostication of treatment response in patients with chronic aphasia. As neuroimaging techniques advance and large databases of patients increase in size, future research must endeavor to determine how best to quantify and incorporate structural integrity metrics into computational models so that the likelihood of recovery is maximized for patients with aphasia.