Strategic infarct location for post-stroke cognitive impairment: A multivariate lesion-symptom mapping study

Abstract

Lesion location is an important determinant for post-stroke cognitive impairment. Although several ‘strategic’ brain regions have previously been identified, a comprehensive map of strategic brain regions for post-stroke cognitive impairment is lacking due to limitations in sample size and methodology. We aimed to determine strategic brain regions for post-stroke cognitive impairment by applying multivariate lesion-symptom mapping in a large cohort of 410 acute ischemic stroke patients. Montreal Cognitive Assessment at three to six months after stroke was used to assess global cognitive functioning and cognitive domains (memory, language, attention, executive and visuospatial function). The relation between infarct location and cognition was assessed in multivariate analyses at the voxel-level and the level of regions of interest using support vector regression. These two assumption-free analyses consistently identified the left angular gyrus, left basal ganglia structures and the white matter around the left basal ganglia as strategic structures for global cognitive impairment after stroke. A strategic network involving several overlapping and domain-specific cortical and subcortical structures was identified for each of the cognitive domains. Future studies should aim to develop even more comprehensive infarct location-based models for post-stroke cognitive impairment through multicenter studies including thousands of patients.

Keywords

Cognitive impairment infarct location ischemic stroke multivariate lesion-symptom mapping support vector regression

Introduction

Stroke is an important cause of cognitive impairment and dementia.¹ The impact of stroke on cognition depends on location.² Based on observations from small case series, it is traditionally taught that infarcts in, e.g., the internal capsule, caudate nucleus and angular gyrus can cause severe cognitive impairment and dementia.¹ Recent studies using voxel-based lesion-symptom mapping (VLSM)³ have provided further evidence for the role of strategic infarcts in post-stroke cognitive impairment and identified strategic regions for global cognition⁴ and cognitive domains including memory,^5,6 language,^7–10 visuospatial^11,12 and executive functions.^13,14 However, a comprehensive map of strategic brain regions for post-stroke cognitive impairment is still lacking due to limitations in sample size and methodology.

Recently, new lesion-symptom mapping methods^15–17 have emerged to overcome the main limitation of VLSM, namely its mass-univariate approach that does not take into account intervoxel correlations (which are typically present because lesion distribution follows the vascular tree), which can result in significant displacement of the strategic lesion locus. These new lesion-symptom mapping (LSM) methods can determine independent contributions of strategic voxels/regions in multivariate models, resulting in higher sensitivity and better accuracy.¹⁸

In the current study, we aim to provide a more complete picture of strategic brain regions for post-stroke impairment, focusing on global cognitive functioning, and memory, language, attention, executive and visuospatial function. For this purpose, we perform the largest LSM study to date, including over 400 stroke patients, and apply the recently developed multivariate SVR-based voxelwise¹⁷ and ROI-based¹⁹ methods to relate infarct location to cognitive functioning.

Methods

Subjects

Participants were patients of the ongoing Chinese University – Stroke Registry Investigating Cognitive Decline (CU-STRIDE) study.²⁰ The CU-STRIDE study recruited 1013 consecutive acute stroke/TIA patients, who were admitted to the Prince of Wales Hospital in Hong Kong between 2009 and 2010, aiming to investigate mechanisms of cognitive decline over five years. This study was approved by the Joint Chinese University of Hong Kong – New Territories East Cluster Clinical Research Ethics committee. Study procedures were performed in accordance to the Declaration of Helsinki of 1975 and its later revisions. Written informed consent was obtained from all participants who were deemed to be capable of giving informed consent, and written informed consent from proxy was obtained for participants diagnosed with dementia. The inclusion and exclusion criterion of the CU-STRIDE trial were described previously.²⁰ To associate acute ischemic lesions (AILs) and post-stroke cognition in this study, we further excluded the patients without visible AILs on magnetic resonance imaging (MRI) or computed tomography (CT) at baseline, and those who were not available for cognitive assessment of MoCA at three to six months after stroke. Patients with prior ischemic stroke or TIA were also included if they met the aforementioned criteria. Finally, 410 patients were included in this study (See Supplementary Figure S1, the flow-chart of the patient recruitment).

Neuropsychological assessment

The cognitive functions were assessed using the Hong Kong version of Montreal Cognitive Assessment (MoCA).²¹ Five cognitive domain scores were calculated using a method published previously,²² including memory (delayed recall, orientation, digit span forward), language (animal picture naming, sentence repetition), attention (serial 7s, digit vigilance), executive function (digit span backward, trail-making test, word similarities, category fluency) and visuospatial function (cube draw, clock draw).

Imaging

Non-contrast brain CT was performed for all patients on arrival to the accident and emergency departments.²⁰ Brain MRI was performed within one week of hospital admission on patients for whom we were not able to classify the stroke subtypes based on CT and other clinical parameters.²⁰ All MRI examinations were performed on a 1.5 T scanner (Sonata; Siemens Medical, Erlangen, Germany) or a 3.0 T scanner (Achieva 3.0T TX Series; Philips Medical System, Best, the Netherlands) using standard protocols. The applied MRI sequences in the proposed study included diffusion-weighted imaging (DWI), and axial spin echo T1-weighted fast field echo, and their imaging parameters were previously described.²⁰

Generation of lesion maps

AILs were manually delineated on either MRI (n = 307) or CT (n = 103). An AIL was defined by the presence of a hyperintense MRI DWI lesion with corresponding hypointensity in the apparent diffusion coefficient map, or hypodense lesions on CT that were relevant to the acute neurological signs and symptoms.²⁰ The DWI sequences were first linearly registered to the T1 sequences of the patients (if available) and were further registered to the 1-mm T1 MNI-152 (Montreal Neurological Institute) template.²³ For the patients without T1 sequences, the DWI image was registered directly to the MNI-152 template with optimized parameters. The registration procedure was performed by elastix, with a linear registration followed by a non-linear registration.²⁴ Considering that most of the included patients were elderly subjects, an age-specific MRI template²⁵ was used as an intermediate before the final registration to MNI-152 space. The resulting transformations were combined into a single transformation that was subsequently used to transform the corresponding lesion maps to the MNI-152 template. The same registration method was applied to the CT scans using a registration algorithm that was designed and validated for this purpose.²⁶ The image registration in this study was performed using tailor-made software called RegLSM, which was jointly developed by HJK and LZ; RegLSM will be made public at http://lsm.isi.uu.nl/. Rigorous quality checks of the registration results were performed by comparing the location of the lesion maps on native scan to that on MNI-152 template. Manual correction of the mapped lesions was performed when necessary.

Statistical analysis

Support vector regression-based lesion symptom mapping (SVR-LSM)¹⁷ was performed to determine the association between AIL location and post-stroke cognition measured by MoCA. This recently developed multivariate lesion-symptom mapping method takes into account the intervoxel correlations and has previously been shown to provide higher spatial accuracy than the commonly used mass-univariate LSM methods.¹⁷ In addition, it allows continuous measures of behavior as the dependent variables, which has broader applications than other multivariate methods where the neuropsychological measurements have to be dichotomized into binary variables. Total MoCA and subscores for cognitive domains were norm-corrected for age, gender, education year and the presence of prior TIA or ischemic stroke. A correction for total lesion volume was performed by weighting the lesioned voxels in inverse proportion to the square root of the lesion size, which captures the intuitive notion that the damage in a particular voxel is more informative for smaller lesions than for larger lesions. This method directly controls relationships between total lesion size, stroke distribution and behavior scores, and provides better control than the regression of behavioral scores on lesion volume.¹⁷ To achieve an SVR-LSM model with reliable performance for our lesion-behavior data, we adjusted the original nonlinear SVR-LSM approach to a linear model and performed feature selection to exclude noise voxels prior to model training. These adjustments significantly improved the prediction accuracy of the SVR-LSM model, and made it more suitable for our study where small lacunar infarcts were included and the lesion distribution was more dispersed. (See details in Supplementary Method, Supplementary Table S1 and Supplementary Figure S3). The parameter training was realized by weighing the prediction performance of the norm-corrected MoCA scores and the reproducibility of the voxel-wise weight coefficients in the linear SVR-LSM model. Statistical inference was performed by shuffling the observations of norm-corrected MoCA scores to create pseudo weight coefficients, and the significance level of each voxel was calculated by counting the number of pseudo weights larger than the real weight in 5000 permutations. The voxels with permutation-based p < 0.01 were treated as significant voxels in the multivariate analysis. For further details of the SVR-LSM methodology please refer to Supplementary Method.

Subsequently, we performed an assumption-free region of interest (ROI)-based analysis also using support vector regression.¹⁹ The ROIs were defined by the AAL atlas and ICBM-DTI-81 white matter tract atlas in MNI-152 space. In line with the SVR-LSM analysis, we also applied feature selection on these ROIs before entering their regional lesion volumes as independent variables in the linear SVR model. The candidate ROIs were damaged in at least four patients and associated with the norm-corrected MoCA scores in univariate models, where the association was measured by Pearson correlation (p < 0.05). Correction for total lesion volume was performed by including total lesion volume as an independent variable (together with the regional volumes of the candidate ROIs from feature selection). The parameter training and statistical inference of SVR-ROI analysis were similar to that of the SVR-LSM analysis (see Supplementary Method and Supplementary Table S2). The ROIs with permutation-based p < 0.01 were treated as significant in the multivariate analysis.

Results

Clinical characteristics of the patients in this study are provided in Table 1. The median AIL volume was 2.32 ml, indicating that the majority of patients had relatively small acute lesions rather than large infarcts. Table 1 shows the patients' performance on the global and domain measures of the MoCA. The distribution of acute ischemic lesions in the study cohort is illustrated by the lesion prevalence map in Figure 1(a). The voxels displayed were damaged in at least four patients. Lesion prevalence was higher in the right hemisphere than the left hemisphere, and higher in subcortical regions than in cortical regions. In addition, Figure 1(b) illustrated the lesion size topologies²⁷ of the study cohort, where cortical lesions were generally larger than the subcortical lesions.

Table 1.

Characteristics of the study cohort.

Characteristics	Study cohort (n = 410)
Demographic characteristics
Age, mean ± SD (years)	68.6 ± 10.4
Education, mean ± SD (years)	5.9 ± 4.7
Female, n (%)	163 (39.8)
Handedness
Right, n (%)	396 (96.6)
Left, n (%)	7 (1.7)
Ambidextrous, n (%)	7 (1.7)
Stroke subtype
Large-artery atherosclerosis, n (%)	170 (41.4)
Small-artery occlusion, n (%)	113 (27.6)
Cardioembolism, n (%)	79 (19.3)
Others, n (%)	48 (11.7)
Vascular risk factors
Smoking, n (%)	62 (15.1)
Hypertension, n (%)	285 (69.5)
Diabetes mellitus, n (%)	168 (41.0)
Prior TIA, n (%)	6 (1.5)
Prior ischemic stroke, n (%)	46 (11.2)
Lesion measures
Median acute infarct volume, ml (range)	2.32 (0.06–440.38)
Cognitive measures (max.): MoCA score
Global (30), mean ± SD	20.5 ± 5.7
Memory (12), mean ± SD	7.9 ± 2.5
Language (5), mean ± SD	4.3 ± 0.9
Attention (4), mean ± SD	3.0 ± 1.2
Executive (5), mean ± SD	2.4 ± 1.5
Visuospatial (4), mean ± SD	2.4 ± 1.3

MoCA: Montreal cognitive assessment.

Figure 1.

Lesion prevalence map (a) and lesion size topographies (b). (a) Voxels that are damaged in at least four patients are projected on the 1 mm MNI-152 template (Z coordinates: −33, −11, 0, 9, 17, 28, 40). Bar indicates the number of patients with a lesion for each voxel. (b) Lesion size topographies in ml for each voxel lesioned in at least four patients. Bar indicates the median lesion volume a patient would have, given that the specific voxel is lesioned. The lesion maps are shown in neurological convention (left is on the left).

Multivariate voxel-based lesion-symptom mapping

SVR-LSM identified significant associations between total MoCA and several clusters of voxels mainly located in the left basal ganglia, left and right frontal, left parietal and left occipital cortex and white matter (WM) (Figure 2). The location of tested and significant voxels is further specified in Supplementary Table S3. The number of patients that the significant associations were based on was also provided in Supplementary Figure S5. The regions where more than 10% of the tested voxels remained significant for global cognition in SVR-LSM were the left angular gyrus, the left caudate, left pallidum, the left anterior corona radiata (ACR), left anterior limb of internal capsule (AIC), left external capsule (EC), left posterior thalamic radiation (PTR), left superior fronto-occipital fasciculus (SFOF) and left tapetum (part of corpus callosum) (Supplementary Table S3).

Figure 2.

Results of multivariate lesion-symptom mapping. Voxel-wise associations between the presence of a lesion and cognitive functioning were determined using support vector regression (SVR-LSM). This multivariate approach assesses the inter-voxel correlations and identifies areas (in this case voxels) which have an independent contribution to the outcome measure. These associations are corrected for age, gender, education year and prior TIA or ischemic stroke. Significance of the clusters is shown in color ranging from yellow (−log10p=2, p = 0.01) to red (−log10p>3, p < 0.001). In order to visualize the voxels that were included in each step of the analyses, voxels that were associated with cognition in the univariate analyses, but not in the multivariate analyses, are shown in light blue (corresponding to the step referred to as ‘feature selection’ in the methods) and the remaining voxels are shown in dark blue. Uncolored voxels were not included in any step of the analyses because such lesions were found in <4 individuals. The slice showing the infratentorial regions varies across different domain to best visualize the significant clusters. The figures are shown in neurological convention (left is on the left).

Regarding the five cognitive domains, significant clusters generally covered the left basal ganglia (especially the left caudate), the left AIC, left frontal cortex and left frontal WM (Figure 2, Supplementary Table S3). There are also some discrepant regions for different domains. More left dorsal cortical regions were involved in language than the other domains. Moreover, more right temporal and left occipital cortical regions and less basal ganglia structures were involved in language. In addition, exclusively left hemispheric lesions were associated with memory and executive function, and right parietal regions were distinctively involved in attention and visuospatial function. Finally, several clusters of lesions in the cerebellum were exclusively associated with attention function.

Multivariate region-of-interest-based analyses using support vector regression

Next, we analyzed the correlations between the regional infarct volumes and MoCA using the multivariate region of interest-based analyses (SVR-ROI). The regions of interest with a significant association between lesion volume and MoCA are shown in Tables 2 and 3. For ease of interpretation, these significant ROIs are also shown in Figure 3, labeled from yellow (p < 0.01) to red (p < 0.001). The number of patients (Table 4 and Supplementary Figure S6) and the lesion size (Table 4) on which these significant associations were based on were also provided.

Table 2.

Results of region of interest-based SVR models with regional AIL volumes (mL) as independent variables and global cognition as outcome.

Region	MoCA (Global)
Region	beta	p
Anterior limb of internal capsule L	0.0014	<0.0002
Body of corpus callosum	0.0012	<0.0002
Posterior corona radiata L	0.0016	0.0006
Inferior fronto-occipital fasciculus L	0.0009	0.0008
Genu of corpus callosum	0.0011	0.0012
External capsule L	0.0011	0.0012
Superior fronto-occipital fasciculus L	0.0015	0.0014
Inferior frontal gyrus (orbital) L	0.0005	0.0016
Cingulum (white matter) L	0.0007	0.0016
Tapetum L	0.0009	0.0020
Angular L	0.0008	0.0022
Inferior frontal gyrus (triangular) L	0.0008	0.0036
Posterior thalamic radiation L	0.0009	0.0036
Insula L	0.0010	0.0042
Pallidum L	0.0014	0.0068
Posterior orbitofrontal cortex L	0.0004	0.0070
Caudate L	0.0012	0.0072
Middle temporal gyrus L	0.0005	0.0082
Anterior corona radiata L	0.0010	0.0088

Note: The brain regions are defined by the AAL atlas and ICBM DTI-81 white matter tract atlas. Only the significant regions with p < 0.01 from SVR-ROI analysis are listed (in ascending order from the lowest p to the highest p). For the candidate ROIs that survived feature selection but not the SVR-ROI analyses, their weight coefficients and p-values of the SVR models are shown in Supplementary Table S4. The weight coefficient (beta) indicates the relative contribution of the regional AIL volume of an ROI to a specific norm-corrected MoCA score. L left.

MoCA: Montreal cognitive assessment; AIL: acute ischemic lesions.

Table 3.

Results of region of interest-based SVR models with regional AIL volumes (mL) as independent variables and norm-corrected MoCA domains as outcomes.

Region	Memory		Language		Attention		Executive		Visuospatial
Region	Beta	p	Beta	p	Beta	p	Beta	p	Beta	p
IFGop L			0.0335	0.0080			0.0002	0.0030
IFGtri L	0.0016	0.0016
OFCpost L	0.0008	0.0064
Insula L					0.0070	0.0056	0.0002	0.0014
MOL L	0.0013	0.0078	0.0386	0.0012
SPG L	0.0017	0.0062
Angular L	0.0015	0.0012	0.0457	0.0002*
STG L	0.0011	0.0080
MTG L	0.0010	0.0048	0.0263	0.0030
MTP R			0.0326	0.0042
Caudate L	0.0027	0.0010					0.0003	0.0026	0.0044	0.0080
Putamen L	0.0022	0.0096					0.0002	0.0070
Pallidum L	0.0028	0.0034					0.0003	0.0066
GCC					0.0080	0.0026
BCC			0.0419	0.0098	0.0082	0.0040			0.0044	0.0016
AIC L	0.0030	<0.0002*			0.0100	0.0006*	0.0003	<0.0002*	0.0053	0.0010
PIC L	0.0029	0.0050
RIC L							0.0002	0.0078
ACR L	0.0021	0.0028			0.0093	0.0006*
SCR L	0.0028	0.0060
SCR R			0.0469	0.0072
PCR L	0.0026	0.0024
PTR L	0.0016	0.0068	0.0420	0.0052			0.0002	0.0002*
EC L	0.0019	0.0026			0.0082	0.0004*	0.0003	<0.0002*	0.0037	0.0028
Cingulum^a L							0.0001	0.0042
SLF L					0.0088	0.0066	0.0003	0.0058
SFOF L	0.0033	0.0006*			0.0122	0.0002*	0.0003	<0.0002*	0.0051	0.0042
IFOF L	0.0018	0.0006*			0.0073	0.0004*	0.0002	0.0018	0.0034	0.0034
Tapetum L					0.0073	0.0004*
Total IV									0.0027	0.0060

Note: The brain regions are defined by the AAL atlas and ICBM DTI-81 white matter tract atlas. Only the significant regions with p < 0.01 from SVR-ROI analyses are listed, and the blank cells indicate non-significant results in the feature selection with univariate analyses or the SVR-ROI analyses. For the candidate ROIs that survived feature selection but not the SVR-ROI analyses, their weight coefficients and p-values of the SVR models are shown in Supplementary Table S4. The weight coefficient (beta) indicates the relative contribution of the regional AIL volume of an ROI to a specific norm-corrected MoCA score. It should be noted that the betas are only comparable between the ROIs of a domain-specific SVR model, and the varied magnitude of the betas across different domains results from the different lesion-symptom associations in their corresponding SVR models.

The most significant ROIs with p < 0.001.

White matter of cingulum defined by ICBM-DTI-81 atlas.

IV: infarct volume; L left; R: right; ACR: anterior corona radiata; AIC: anterior limb of internal capsule; BCC: body of corpus callosum; EC: external capsule; GCC: genu of corpus callosum; IFGop: inferior frontal gyrus (opercular); IFGtri: inferior frontal gyrus (triangular); IFOF: inferior fronto-occipital fasciculus; MOL: middle occipital lobe; MTG: middle temporal gyrus; MTP: middle temporal pole; OFCpost: posterior orbitofrontal cortex; PCR: posterior corona radiata; PIC: posterior limb of internal capsule; PTR: posterior thalamic radiation; RIC: retrolenticular part of internal capsule; SCR: superior corona radiata; SFOF: superior fronto-occipital fasciculus; SLF: superior longitudinal fasciculus; SPG: superior parietal gyrus; STG: superior temporal gyrus.

Figure 3.

Results of multivariate region of interest-based analyses using support vector regression (SVR). The ROIs where the regional infarct volume was statistically associated with the cognitive functions are colored from yellow (−log10p = 2, p = 0.01) to red (−log10p > 3, p < 0.001). ROIs that were associated with cognition in the univariate analyses but not in the multivariate analyses are shown in light blue (corresponding to the step referred to as ‘feature selection’ in the methods). The names of the significant ROIs were attached above and the corresponding p-value and weight coefficient (beta) are shown in Table 2 for global cognition and Table 3 for cognitive domains. The figures are shown in neurological convention (left is on the left).

Table 4.

Number of patients with a median lesion volume for each of the significant ROIs from the SVR-ROI analyses.

ROI	MoCA scores the ROI was associated with (p < 0.01)	Patients with lesion (n)^a	Median regional lesion volume (ml [% of ROI])^b	Median total lesion volume (ml)^c
SCR L	Memory	99	0.19 (2.16)	3.62
SCR R	Language	82	0.38 (4.30)	8.59
Caudate L	Global, memory, executive, visuospatial	80	0.13 (1.84)	3.32
PIC L	Memory	78	0.29 (10.42)	2.49
PCR L	Global, memory	77	0.12 (2.33)	4.22
Putamen L	Memory, executive	76	0.15 (1.92)	3.31
BCC	Global, language, attention, visuospatial	69	0.09 (0.50)	14.76
AIC L	Global, memory, attention, executive, visuospatial	56	0.14 (4.72)	4.76
Insula L	Global, attention, executive	54	0.11 (0.77)	7.13
EC L	Global, memory, attention, executive, visuospatial	53	0.14 (3.50)	4.83
Pallidum L	Global, memory, executive	51	0.08 (3.28)	3.33
RIC L	Executive	51	0.12 (4.44)	4.05
SLF L	Attention, executive	50	0.18 (1.92)	14.90
GCC	Global, attention	45	0.11 (1.02)	15.76
SFOF L	Global, memory, attention, executive, visuospatial	40	0.08 (15.43)	5.77
MOL L	Memory, language	31	0.33 (1.23)	15.88
ACR L	Global, memory, attention	29	0.28 (3.69)	15.88
IFOF L	Global, memory, attention, executive, visuospatial	29	0.05 (2.52)	14.76
PTR L	Global, memory, language, executive	27	0.29 (4.49)	12.05
STG L	Memory	26	0.26 (1.22)	19.30
MTG L	Global, memory, language	24	0.60 (1.52)	13.78
SPG L	Memory	23	0.19 (1.24)	14.47
IFGtri L	Global, memory	22	0.39 (1.96)	15.09
IFGop L	Language, executive	21	0.40 (5.26)	18.15
Angular L	Global, memory, language	21	0.13 (1.23)	15.88
Tapetum L	Global, attention	17	0.08 (11.71)	15.42
IFGorb L	Global	11	0.07 (1.13)	19.91
Cingulum L	Global, executive	10	0.30 (8.16)	15.07
OFCpost L	Global, memory	6	0.41 (9.30)	42.88
MTP R	Language	6	1.90 (20.65)	26.05

Note: The brain regions are defined by the AAL atlas and ICBM DTI-81 white matter tract atlas. Only the significant regions with p < 0.01 from SVR-ROI analysis are listed (in descending order from the ROI that most frequently overlapped with lesions to the ROI that least frequently overlapped with lesions).

Number among 410 included patients that had a lesion that overlapped with the specified ROI in AAL atlas and ICBM DTI-81 white matter tract atlas.

The median regional lesion volume within the specific ROI for the patients who had a lesion that overlapped with this ROI. The regional lesion volumes (in ml) and their ratios (in percent) of the corresponding ROI size were both provided.

The median total lesion volume (in ml) a patient would have, given that the patient had a lesion that overlapped with this ROI.

L: left; R: right; ACR: anterior corona radiata; AIC: anterior limb of internal capsule; BCC: body of corpus callosum; EC: external capsule; GCC: genu of corpus callosum; IFGop: inferior frontal gyrus (opercular); IFGorb: inferior frontal gyrus (orbital); IFGtri: inferior frontal gyrus (triangular); IFOF: inferior fronto-occipital fasciculus; MOL: middle occipital lobe; MTG: middle temporal gyrus; MTP: middle temporal pole; OFCpost: Posterior orbitofrontal cortex; PCR: posterior corona radiata; PIC: posterior limb of internal capsule; PTR: posterior thalamic radiation; RIC: retrolenticular part of internal capsule; SCR: superior corona radiata; SFOF: superior fronto-occipital fasciculus; SLF: superior longitudinal fasciculus; SPG: superior parietal gyrus; STG: superior temporal gyrus.

Total MoCA was associated with infarcts located in the left basal ganglia and left frontal, temporal, parietal and occipital cortical and white matter structures. Infarct volume in the left AIC, body of corpus callosum (BCC), left posterior corona radiata (PCR) and left inferior fronto-occipital fasciculus (IFOF) were the most significantly associated (p < 0.001) with total MoCA (Table 2). Although total lesion volume was significantly associated with total MoCA in the univariate analysis, it did not remain significant in the multivariate SVR-ROI analyses as shown in Supplementary Table S4.

Regarding the significant regions for the cognitive domains (Table 3 and Figure 3), lesions in the left AIC, which were most significant for global cognition, were also significantly associated with memory (p < 0.001), attention (p < 0.001), executive function (p < 0.001) and visuospatial function (p = 0.001). Total lesion volume was only associated with visuospatial function in the multivariate SVR-ROI analyses (Table 3). In addition, most of the discrepant regions of cognitive domains found in SVR-LSM were reproduced in the SVR-ROI based analyses (Figure 3). Firstly, more left dorsal cortical regions (temporal, parietal and occipital) were involved in memory and language. Secondly, language function involved less basal ganglia structures than the other domains. Finally, exclusive left hemispheric regions were associated with memory and executive function.

Discussion

In this study, we applied multivariate voxelwise lesion-symptom mapping (SVR-LSM) and multivariate region of interest-based analyses (SVR-ROI) to investigate the spatial relationship between infarct location and post-stroke cognitive functioning measured after three to six months. The findings of this study confirm the relevance of infarct location, corroborate several previously identified strategic brain regions (e.g. the anterior limb of the left internal capsule and the basal ganglia) and provide a more complete map of strategic brain regions for post-stroke impairment in global cognitive functioning, memory, language, attention, executive functioning, and visuospatial functioning.

The current study has several strengths. A main strength is the large sample size as it is the largest study to date to investigate the relation between infarct location and post-stroke cognitive impairment. A second main strength is the use of the state of the art VLSM methods to identify strategic brain regions. We applied a recently developed multivariate lesion-symptom mapping approach (SVR-LSM), which has better spatial accuracy than univariate VLSM.¹⁷ Importantly, this method takes into account the intervoxel correlations regarding lesion status, where univariate VLSM does not. Brain lesions due to ischemic stroke are not randomly distributed but these lesions follow the vascular tree (i.e. voxels within the vascular territory of a specific artery are strongly intercorrelated regarding lesion status). Consequently, univariate lesion-symptom mapping methods are prone to some degree of displacement regarding the identified strategic brain region.¹⁶ In a direct comparison of univariate VLSM and SVR-LSM on hypothetical and real lesion data, SVR-LSM had higher sensitivity for detecting lesion-deficit relations.¹⁷ Third, we used both voxelwise analyses, which provide very high spatial resolution, and ROI-based analyses, which take into account cumulative lesion burden in a specific structure. An additional strength is that both the SVR-LSM and SVR-ROI analyses were hypothesis-free and lesions anywhere in the brain were considered, including infratentorial lesions. Fourth, well-defined and validated MoCA subscores were used to assess not only global cognitive functioning, but memory, language, attention, executive functioning, and visuospatial functioning, and determine their anatomical correlates as well.²²

Previous studies on stroke cohort have demonstrated an association between infarct location and global cognition.^{1,2,4,28–30} These studies identified the basal ganglia,³⁰ internal capsule,² thalamus,² corpus callosum,² angular gyrus,³¹ cingulate cortex,⁴ and frontal subcortical regions²⁸ as strategic regions for global post-stroke cognitive impairment. Our findings corroborate most of these strategic regions and identify several additional strategic regions for global cognitive impairment. More specifically, our results corroborate the left angular gyrus,¹ the left caudate and pallidum,⁴ and several left hemispheric white matter tracts (anterior internal capsule,^1,2 anterior corona radiata,⁴ external capsule,⁴ and posterior thalamic radiation³²) and the corpus callosum as strategic regions. Additionally, the left superior fronto-occipital fasciculus and tapetum (part of corpus callosum) were identified as strategic regions for global post-stroke cognitive impairment, along with several other regions which are summarized in Figure 3 and Table 2. The association between infarct volume and total MoCA was strongest for the anterior limb of the internal capsule.

This study additionally provides a more complete map of strategic regions for post-stroke impairment in five specific cognitive domains (Figure 3 and Table 3). Cognitive domain-specific strategic regions overlapped in some regions as for example the anterior limb of the left anterior capsule and the left fronto-occipital fasciculi were associated with all cognitive domains except language. On the other hand, there were also some clear distinctions between cognitive domain-specific strategic regions. For example, (1) language was most strongly (but not exclusively) associated with left hemispheric cortical regions (including the left inferior frontal gyrus, left middle temporal gyrus, left middle occipital lobe, left angular gyrus), and not with the basal ganglia; (2) memory and executive functioning were associated with a widely distributed network of almost exclusively left hemispheric cortical and subcortical regions; (3) attention was exclusively associated with a left and right hemispheric white matter tract network and the left insula; (4) visuospatial functioning involved a distributed left and right hemispheric subcortical network and right temporo-parietal cortical regions (the latter were only observed in the voxelwise analyses, see Supplementary Table S3) and was the only domain which was associated with total infarct volume in the multivariate models.

Of note, although lesion coverage and average lesion size was highest for the right cerebral hemisphere (Figure 1), the significant anatomical substrates of global and domain-specific cognitive functioning were mostly located in the left cerebral hemisphere in both the multivariate voxel-based (Figure 2) and ROI-based (Figure 3) analyses. The predominant associations between MoCA global scores and left hemispheric lesions can be explained by the fact that the MoCA is made up of predominately verbal items, and therefore its performance draws heavily on the language-dominant left hemisphere. However, at the level of specific cognitive domains, the associations are consistent with the expected brain–behavioral relationships with language loading predominantly on the left hemisphere and visuospatial functions loading more strongly on the right hemisphere. In addition, the anatomical substrates for memory and attention were also mostly located in the left hemisphere because they were assessed verbally (e.g. word list, verbally presented numerical strings). The number of patients and/or the lesion size on which these associations were based on were provided in Supplementary Figure S5 for the SVR-LSM results and in Table 4 and Supplementary Figure S6 for the SVR-ROI results. These data, in addition to the significance level from the analyses, could be used as a supplementary reference to assess the statistical power of the identified strategic regions for post-stroke cognitive functioning. As the lesion distribution was very dispersed in this study (especially in the left hemisphere), the significant voxels from the SVR-LSM analyses were generally based on less than 10 patients (Figure S5). Nevertheless, we have performed sufficient optimizations for the SVR-LSM models to achieve the best statistical power of the analyses based on our lesion-behavior data. In addition, the strategic regions identified by the complementary SVR-ROI analyses were based on more patients (Table 4 and Figure S6) as the lesions were evaluated as cumulative lesion burden in a specific region instead of voxel-wise basis. Importantly, the results of the SVR-LSM and the SVR-ROI are clearly converging. The statistical power in terms of the patients on which these associations were based on can only be further improved by including more patients in the study with sufficient lesion coverage throughout different regions in the brain.

It should also be noted that brain regions that are crucially involved in cognitive functioning but are rarely affected by infarcts would not be identified as strategic regions for post-stroke cognitive impairment. Therefore, our findings should not be interpreted as an exhaustive map of anatomical correlates of their corresponding cognitive domain, but rather as a map of strategic regions in which infarcts are likely to result in post-stroke cognitive impairment. Such information is of clinical value, as this helps clinicians to understand why a patient with a strategic lesion suffers from post-stroke cognitive impairment. The next step towards translating these findings to clinical practice is to develop prediction models for post-stroke cognitive impairment based on stroke location. Such a prediction model could be used to identify patients at risk of developing cognitive impairment at an early stage in order to initiate adequate rehabilitation strategies at the earliest possible stage. The development of such a prediction model would require the inclusion of thousands of stroke patients and can probably only be achieved through multicenter collaborations.

Regarding the methods, it should be noted that in the five previous SVR-LSM studies,^33–37 a nonlinear model was used without feature selection, whereas we used a linear model with feature selection. These previous studies focused on patients with isolated left hemispheric stroke and studied cognitive processes involved in language. Because the lesion distribution in the current study was completely different (i.e. patients with lesions anywhere in the brain were included, see Figure 1 and Supplementary Figure S2) and we included many more patients with small lacunar infarcts than the SVR-LSM methodology paper,¹⁷ the original nonlinear SVR-LSM approach performed very poorly in predicting MoCA scores (Supplementary Table S1). Then we selected four candidate SVR-LSM models (nonlinear and linear SVR-LSM with and without feature selection) and compared their prediction performance and reproducibility on the same training and testing data. As can be appreciated from the results and information in the online supplementary materials, the linear model with feature selection outperformed the other models, which is why this model was selected for the current study in statistical inference (See Supplementary Method for more details). Of note, this study is the first to directly compare the linear and non-linear models with and without feature selection in SVR-LSM. The adjustments to the SVR-LSM method that are described in this study extend the use of SVR-LSM from patients with relatively large, unilateral cerebral hemispheric lesions, to map lesion-behavior associations in lesion data that is dispersed throughout the entire brain.

There are several limitations to this study that should be taken into account. Despite this being the largest lesion-symptom mapping study in post-stroke cognitive impairment to date, there was still insufficient lesion coverage in some brain regions to be included in the analyses. As for the before mentioned development of prediction models, overcoming this limitation can only be achieved by further increasing sample size, which will probably require (international) multicenter collaborations with thousands of patients. It should be noted that we included many patients with relatively small infarcts (median volume 2.3 ml), which may in part be due to some of the exclusion criteria of the CU-STRIDE cohort (e.g. patients who died before assessment of who were incommunicable were not included). Furthermore, patients with small infarct may have been more likely to undergo MRI or follow-up CT imaging to confirm the diagnose, which was an additional inclusion criterium for the current study. Consequently, some of our findings may not be generalizable to patients with large cerebral infarcts. More specifically, it is possible that in patients with large infarcts, the size of the lesion becomes more relevant in predicting post-stroke cognitive impairment, whereas in our cohort, lesion size was not independently associated with total MoCA, nor with four out of five cognitive domains. Second, we used the MoCA instead of detailed neuropsychological testing to assess global cognitive functioning and cognitive domains performance. Further studies with detailed neuropsychological testing should be conducted to substantiate the findings in this study.

In conclusion, this study provides a more complete picture of strategic brain regions for post-stroke cognitive impairment by using recently improved lesion-symptom mapping methods in a large cohort of stroke patients. The maps on strategic brain regions for global and domain-specific cognitive impairment may help clinicians to understand and predict the cognitive impact of ischemic strokes. Future studies with even larger sample size are needed to develop individualized infarct location-based prediction models for post-stroke cognitive impairment.

Footnotes

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The work described in this paper was partially supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No.: CUHK 14113214). HJK was financially supported by the project Brainbox (Quantitative analysis of MR brain images for cerebrovascular disease management), funded by the Netherlands Organisation for Health Research and Development (ZonMw) in the framework of the research programme IMDI (Innovative Medical Devices Initiative); project 104002002.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Authors' contributions

LZ: data collection, image registration, statistical analysis, data interpretation, manuscript preparation. JMB: statistical analysis, data interpretation, manuscript preparation. LS: study design. WL: data collection, lesion delineation, data interpretation. HJK: image registration. WWCC: data collection. JMA: data collection. RKLL: data collection. TWHL: data collection. AYLL: data collection. GJB: study design, commenting on drafts. VM: study design. AW: study design, data interpretation, manuscript preparation.

Supplementary material

Supplementary material for this paper can be found at the journal website:

References

Gorelick

Scuteri

Black

et al.

Vascular contributions to cognitive impairment and dementia: a statement for healthcare professionals from the American Heart Association/American Stroke Association. Stroke 2011; 42: 2672–2713.

Biesbroek

Weaver

Biessels

. Lesion location and cognitive impact of cerebral small vessel disease. Clin Sci 2017; 131: 715–728.

Bates

Wilson

Saygin

et al.

Voxel-based lesion-symptom mapping. Nat Neurosci 2003; 6: 448–450.

Munsch

Sagnier

Asselineau

et al.

Stroke location is an independent predictor of cognitive outcome. Stroke 2016; 47: 66–73.

Glascher

Tranel

Paul

et al.

Lesion mapping of cognitive abilities linked to intelligence. Neuron 2009; 61: 681–691.

Biesbroek

van Zandvoort

Kappelle

et al.

Distinct anatomical correlates of discriminability and criterion setting in verbal recognition memory revealed by lesion-symptom mapping. Hum Brain Mapp 2015; 36: 1292–1303.

Walker

Schwartz

Kimberg

et al.

Support for anterior temporal involvement in semantic error production in aphasia: new evidence from VLSM. Brain Lang 2011; 117: 110–122.

Dronkers

Wilkins

Van Valin

Jr. et al.

Lesion analysis of the brain areas involved in language comprehension. Cognition 2004; 92: 145–177.

Butler

Lambon Ralph

Woollams

. Capturing multidimensionality in stroke aphasia: mapping principal behavioural components to neural structures. Brain 2014; 137: 3248–3266.

10.

Biesbroek

van Zandvoort

Kappelle

et al.

Shared and distinct anatomical correlates of semantic and phonemic fluency revealed by lesion-symptom mapping in patients with ischemic stroke. Brain Struct Funct 2016; 221: 2123–2134.

11.

Ten Brink

Matthijs Biesbroek

Kuijf

et al.

The right hemisphere is dominant in organization of visual search-A study in stroke patients. Behav Brain Res 2016; 304: 71–79.

12.

Biesbroek

van Zandvoort

Kuijf

et al.

The anatomy of visuospatial construction revealed by lesion-symptom mapping. Neuropsychologia 2014; 62: 68–76.

13.

Turken

Whitfield-Gabrieli

Bammer

et al.

Cognitive processing speed and the structure of white matter pathways: convergent evidence from normal variation and lesion studies. Neuroimage 2008; 42: 1032–1044.

14.

Kalenine

Shapiro

Buxbaum

. Dissociations of action means and outcome processing in left-hemisphere stroke. Neuropsychologia 2013; 51: 1224–1233.

15.

Smith

Clithero

Rorden

et al.

Decoding the anatomical network of spatial attention. Proc Natl Acad Sci U S A 2013; 110: 1518–1523.

16.

Mah

Husain

Rees

et al.

Human brain lesion-deficit inference remapped. Brain 2014; 137: 2522–2531.

17.

Zhang

Kimberg

Coslett

et al.

Multivariate lesion-symptom mapping using support vector regression. Hum Brain Mapp 2014; 35: 5861–5876.

18.

Karnath

Smith

. The next step in modern brain lesion analysis: multivariate pattern analysis. Brain 2014; 137: 2405–2407.

19.

Yourganov

Fridriksson

Rorden

et al.

Multivariate connectome-based symptom mapping in post-stroke patients: networks supporting language and speech. J Neurosci 2016; 36: 6668–6679.

20.

Yang

Wong

Wang

et al.

Risk factors for incident dementia after stroke and transient ischemic attack. Alzheimers Dement 2015; 11: 16–23.

21.

Wong

Xiong

Kwan

et al.

The validity, reliability and clinical utility of the Hong Kong Montreal Cognitive Assessment (HK-MoCA) in patients with cerebral small vessel disease. Dement Geriatr Cogn Disord 2009; 28: 81–87.

22.

Lam

Middleton

Masellis

et al.

Criterion and convergent validity of the Montreal cognitive assessment with screening and standardized neuropsychological testing. J Am Geriatr Soc 2013; 61: 2181–2185.

23.

Fonov

Evans

McKinstry

et al.

Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. Neuroimage 2009; 47: S102.

24.

Klein

Staring

Murphy

et al.

elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging 2010; 29: 196–205.

25.

Rorden

Bonilha

Fridriksson

et al.

Age-specific CT and MRI templates for spatial normalization. Neuroimage 2012; 61: 957–965.

26.

Kuijf

Biesbroek

Viergever

et al.

Registration of brain CT images to an MRI template for the purpose of lesion-symptom mapping. Multimodal Brain Image Anal 2013; 8159: 119–128. .

27.

Sperber

Karnath

. Topography of acute stroke in a sample of 439 right brain damaged patients. Neuroimage Clin 2016; 10: 124–128.

28.

Kandiah

Wiryasaputra

Narasimhalu

et al.

Frontal subcortical ischemia is crucial for post stroke cognitive impairment. J Neurol Sci 2011; 309: 92–95.

29.

Zhang

Yang

et al.

Incidence and risk factors of cognitive impairment 3 months after first-ever stroke: a cross-sectional study of 5 geographic areas of China. J Huazhong Univ Sci Technolog Med Sci 2012; 32: 906–911.

30.

Narasimhalu

Wiryasaputra

Sitoh

et al.

Post-stroke subjective cognitive impairment is associated with acute lacunar infarcts in the basal ganglia. Eur J Neurol 2013; 20: 547–551.

31.

Benson

Cummings

. Angular gyrus syndrome simulating Alzheimer's disease. Arch Neurol 1982; 39: 616–620.

32.

Zhuang

Wen

Zhu

et al.

White matter integrity in mild cognitive impairment: a tract-based spatial statistics study. Neuroimage 2010; 53: 16–25.

33.

Mirman

Zhang

Wang

et al.

The ins and outs of meaning: Behavioral and neuroanatomical dissociation of semantically-driven word retrieval and multimodal semantic recognition in aphasia. Neuropsychologia 2015; 76: 208–219.

34.

Xing

Lacey

Skipper-Kallal

et al.

Right hemisphere grey matter structure and language outcomes in chronic left hemisphere stroke. Brain 2016; 139: 227–241.

35.

Lacey

Skipper-Kallal

Xing

et al.

Mapping common aphasia assessments to underlying cognitive processes and their neural substrates. Neurorehabil Neural Repair 2017; 31: 442–450.

36.

Fama

Hayward

Snider

et al.

Subjective experience of inner speech in aphasia: preliminary behavioral relationships and neural correlates. Brain Lang 2017; 164: 32–42.

37.

Griffis

Nenert

Allendorfer

et al.

Damage to white matter bottlenecks contributes to language impairments after left hemispheric stroke. Neuroimage Clin 2017; 14: 552–565.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

1.50 MB