Research The association between lesion location and functional outcome after ischemic stroke Nawaf Yassi1, Leonid Churilov2, Bruce C.V. Campbell1, Gagan Sharma3, Roland Bammer4, Patricia M. Desmond3, Mark W. Parsons5, Gregory W. Albers4, Geoffrey A. Donnan2, and Stephen M. Davis1 on behalf of the EPITHET, and DEFUSE Investigators Background Infarct location has a critical effect on patient outcome after ischemic stroke, but the study of its role independent of overall lesion volume is challenging. We performed a retrospective, hypothesis-generating study of the effect of infarct location on three-month functional outcome in a pooled analysis of the EPITHET and DEFUSE studies. Methods Posttreatment MRI diffusion lesions were manually segmented and transformed into standard-space. A novel composite brain atlas derived from three standard brain atlases and encompassing 132 cortical and sub-cortical structures was used to segment the transformed lesion into different brain regions, and calculate the percentage of each region infarcted. Classification and Regression Tree (CART) analysis was performed to determine the important regions in each hemisphere associated with nonfavorable outcome at day 90 (modified Rankin score [mRS] > 1). Results Overall, 152 patients (82 left hemisphere) were included. Median diffusion lesion volume was 37·0 ml, and median baseline National Institutes of Health Stroke Score was 13. In the left hemisphere, the strongest determinants of nonfavorable outcome were infarction of the uncinate fasciculus, followed by precuneus, angular gyrus and total diffusion lesion volume. In the right hemisphere, the strongest determinants of nonfavorable outcome were infarction of the parietal lobe followed by the putamen. Conclusions Assessment of infarct location using CART demonstrates regional characteristics associated with poor outcome. Prognostically important locations include limbic, default-mode and language areas in the left hemisphere, and visuospatial and motor regions in the right hemisphere. Key words: brain atlas, infarct location, MRI, prognosis, recovery, stroke
Correspondence: Nawaf Yassi, Melbourne Brain Centre @ Royal Melbourne Hospital, Grattan St, Parkville, Vic. 3050, Australia. E-mail: [email protected] 1 Department of Neurology, Melbourne Brain Centre @ Royal Melbourne Hospital, University of Melbourne, Parkville, Vic., Australia 2 Florey Institute of Neuroscience and Mental Health, University of Melbourne, Parkville, Vic., Australia 3 Departement of Radiology, Royal Melbourne Hospital, University of Melbourne, Parkville, Vic., Australia 4 Department of Neurology and Neurological Sciences and Stanford Stroke Center, Stanford University Medical Center, Stanford, CA, USA 5 Priority Research Centre for Translational Neuroscience and Mental Health, University of Newcastle and Hunter Medical Research Institute, Newcastle, NSW, Australia Received: 24 November 2014; Accepted: 25 February 2015; Published online 4 June 2015 Conflict of interest: None declared. DOI: 10.1111/ijs.12537
1270
Vol 10, December 2015, 1270–1276
Introduction The impact of infarct location on stroke severity as well as recovery and functional outcome is intuitive to the stroke clinician. However, its systematic incorporation into the considerations of clinical trials or clinical practice represents a unique challenge given the complexity and heterogeneity of neuronal network organization as well as additional factors such as hemispheric dominance, premorbid functional level, and concurrent nonneurological comorbidity. Most imaging-based clinical trials focus on the alternate and more simplistic variable of infarct volume. Although infarct volume has been shown to be strongly associated with functional outcome (1), it is rarely considered in isolation of other factors in clinical practice, where the additional assessment of lesion location is critical in determining the likely outcome. We aimed to use pooled data from the Echoplanar Imaging Thrombolytic Evaluation Trial (EPITHET) (2) and the Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution (DEFUSE) (3) Study in a retrospective, hypothesisgenerating investigation of the impact of infarct location on functional outcome, independent of overall lesion volume.
Methods EPITHET was a phase-II prospective randomized, doubleblinded, placebo-controlled trial testing alteplase vs. placebo in patients with hemispheric ischemic stroke presenting three- to six-hours after onset. Patients underwent serial 1·5T magnetic resonance imaging (MRI) at baseline, day 3–5, and day 90. DEFUSE was a prospective, open-label cohort study of patients with hemispheric ischemic stroke treated with alteplase 3-6h from onset. MRI (1·5T) was performed at baseline, three- to six-hours post-alteplase and at day 30. Given that ischemic core on baseline imaging can evolve variably depending on timing and degree of reperfusion, and in order to minimize the potential confounding effects of this variable recanalization, infarct volume was defined on posttreatment MRI diffusion-weighted imaging (DWI), performed at day 3–5 in EPITHET and three- to six-hours posttreatment in DEFUSE. Of 101 patients in EPITHET, one patient withdrew consent, seven patients died prior to day 3–5 imaging, five had inadequate day 3–5 imaging data for analysis, and one patient was lost to follow-up prior to day 90. Thus, 87 patients were included in the analysis. Of 74 patients enrolled in DEFUSE, seven patients had no identifiable lesion on posttreatment DWI, one patient did not have posttreatment DWI, and one patient had severely motion © 2015 World Stroke Organization
Research
N. Yassi et al. degraded imaging. Thus, 65 patients were included in the analysis (a total of 152 in the pooled analysis). Posttreatment DWI images were manually segmented by a stroke neurologist based on maximum visual extent on the b = 1000 diffusion image with visual comparison with apparent diffusion coefficient maps to exclude T2 shine through. The diffusion images were then transformed with 12 degrees-of-freedom into standard-space using the Centre for Functional MRI of the Brain (FMRIB) Linear Image Registration Tool (FLIRT, version 6·0, Oxford, UK) (4). The resultant transformation matrices were subsequently applied to the DWI-lesion masks to transform these into standard-space using a nearest-neighbor transformation, thus creating binary DWI-lesion masks in standard-space. The results were visually inspected to ensure accurate transformation. In a random sample comprising 20% of the overall cohort (30 patients), a second stroke neurologist separately performed manual lesion masking on the DWI. Interrater reliability was assessed using the intraclass correlation coefficient and further validated using Lin’s concordance coefficient and reduced major axis regression analysis. A novel composite brain atlas comprising a total of 132 structures was derived by combination of three separate atlases (Appendices S1, S2) bundled within FMRIB’s Software Library (FSL, version 5.0.6, FMRIB, Oxford, UK). The Montreal Neurological Institute (MNI)-152 atlas (Copyright 1993–2009 Louis Collins, McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University) is derived from 152 T1-weighted MRI scans from healthy young adults, averaged together after nonlinear registration into a common coordinate space. It contains nine left-hemispheric and nine right-hemispheric structures including the cerebellar hemispheres, which were excluded from the composite atlas. The Harvard-Oxford Cortical Atlas (refer to acknowledgements) covers 48 left-hemispheric cortical structures and 48 right-hemispheric cortical structures derived from T1-weighted images of 37 healthy subjects registered to MNI-152 space. The Johns Hopkins University White-Matter Tractography Atlas identifies 20 white-matter tracts by averaging the results of deterministic tractography on 28 normal subjects (5). It includes nine left-hemispheric tracts and nine right-hemispheric tracts, as well as the forceps major and minor. The forceps major and minor were included as whole structures in each hemisphere’s atlas given that these structures cross the midline. Therefore, the entire number of regions overall was 132, with each hemisphere consisting of 67 regions. Each subject’s transformed DWI lesion was segmented using the composite brain atlas, and the percentage of each of the atlas regions infarcted was determined as volume of infarction within the region divided by the total volume of the region multiplied by 100. An example of the registration process in a single patient is shown in Appendix S3. Classification and Regression Tree (CART) analysis, specifically classification, was used to assess the important brain regions in each hemisphere and infarct percentage thresholds within these regions associated with nonfavorable outcome at day 90 [modified Rankin Scale (mRS) > 1]. CART analysis was performed in the Salford Predictive Modeler Software Suite version 7 (Salford Systems, San Diego, CA, USA). The variables included as inputs in © 2015 World Stroke Organization
the model were the percentage infarction within each of the 67 atlas regions from the relevant hemisphere as well as total infarct volume given its known prognostic effect. Further CART background and methodology, as well as results of secondary analyses with different inputs and outcomes are provided in Appendix S4. Briefly, CART is a data-driven binary partitioning statistical method that starts with the total sample and, in a stepwise manner, splits the sample into two sub-samples that are more homogenous with respect to a defined outcome (6). The input variable that achieves the most effective split is dichotomized at an optimal threshold determined by the automated analysis to maximize the separation between the resulting groups (nodes). The model also searches for any potential surrogate splitters at each node, which are variables that split the node in a very similar way to the primary splitter. Once further splitting is not possible, pruning of the tree takes place in order to minimize over-fitting and to obtain a final tree. This involves a 10-fold internal cross-validation method, whereby the data are randomly divided into 10 groups with nine used to build the model (training) and one used to validate (testing). During pruning, CART starts at the bottom of the full tree and sequentially prunes nodes that result in the least decrement in the performance of the tree in the testing set, until an optimal tree is reached. The overall process of training and testing is repeated 10-fold, and the results are combined to produce the final optimal performing tree. CART also computes a variable importance score ranging from 0 to 100 and indicating the importance of each input in the model. The score takes into account the role of variables as either splitters or surrogate splitters and is scaled relative to the highest scoring variable (hence it is unit-less). Variables that do not appear in the final tree may still appear as important variables if they act as surrogate splitters.
Results Overall, 152 patients were included in the analysis. Seventy-six patients (50%) were female, and 87 (57%) were from EPITHET. Eighty-two patients (54%) had left-hemispheric infarcts. Alteplase was administered to 106 patients (69·7%). There were no significant differences between patients with left- and righthemispheric infarcts with regard to baseline characteristics, stroke severity, treatment allocation, or outcome. Median diffusion lesion volume overall was 37·0 ml (interquartile range 15·2– 96·6 ml), and this was also comparable between patients with leftand right-hemispheric infarcts. In univariate logistic regression, there was a strong association between lesion volume (transformed to the 5th-root to approximate normality) and nonfavorable outcome (odds ratio [OR] 4·9 per ml0·2[transformed to the 5th root], 95% confidence interval (CI) 2·3–10·5, P < 0·0001). In the random sample of 30 patients who had lesion masking performed by two different operators, the intraclass correlation coefficient for DWI lesion volume was 0·97, Lin’s concordance correlation coefficient 0·97, reduced major axis slope 0·98 and intercept −0·6 ml indicating the absence of either a fixed or a proportional bias and excellent agreement between raters. Vol 10, December 2015, 1270–1276
1271
Research Left hemisphere CART analysis for the primary outcome of day-90 mRS > 1 in patients with left-hemispheric stroke identified the uncinate fasciculus, precuneus cortex, angular gyrus, and total lesion volume as the inputs most strongly associated with outcome in the final tree (Fig. 1). Although the impact of lesion volume may initially appear to be counterintuitive (patients with >5·2 ml appear to have more favorable outcomes than patients with 1 compared with 69·5% overall. Therefore, infarcts occurring in regions common to this sub-group were by default occurring in ‘less critical’ regions than the preceding steps and therefore conferred a better prognosis. The relative importance score for inputs with nonzero importance in the left-hemisphere analysis is shown in Fig. 2. Overall, the rate of successful classification was 87·7% in the learning set and 67·1% in the testing set. For correct classification of nonfavorable outcome in the testing set, the model had a sensitivity of 70%, specificity of 60%, positive predictive value (PPV) of 80% and negative predictive value (NPV) of 47%. Qualitative review of the final model also demonstrates clear sub-groups of patients with contrasting prognosis (black arrows). One of these sub-groups contains 40 patients with a 90% proportion of nonfavorable outcome compared with another with 25 patients with only a 20% proportion of nonfavorable outcome (OR 36·0, 95% CI 8·7–149·5, P < 0·0001). Right hemisphere For right-hemispheric infarcts, inputs most strongly associated with mRS > 1 were infarction in the parietal lobe and the putamen (Fig. 3). Overall, the rate of successful classification was 82·9% in the learning set and 72·9% in the testing set. The variable
N. Yassi et al. importance for inputs with nonzero importance in the righthemisphere analysis is shown in Fig. 4. For the prediction of nonfavorable outcome in the testing set, the model had a sensitivity of 77%, specificity of 64%, PPV of 82% and NPV of 56%. As in the left-hemispheric analysis, the model also identified specific sub-groups with quite distinct prognosis (black arrows). In one of these groups, 32% of patients had a nonfavorable outcome, compared with 92% of patients in the other group (OR 24·6, 95% CI 5·9–102·0, P < 0·0001). Results of CART analysis for alternative mRS dichotomies, and with the inclusion of age as an input, are presented in Appendix 4. Although age appeared to be important in some of these analyses, it was absent from the final tree in the majority of cases. We also ran the analysis with the patient’s original trial (EPITHET vs. DEFUSE) as a categorical input, in order to explore any major heterogeneity given the differences in imaging time-points and other potential differences between the trials. This did not result in any changes to the CART results or variable importance results at any of the specified primary or secondary mRS dichotomies. Table 1 shows differences in major characteristics between patients with favorable and nonfavorable outcomes.
Discussion We have demonstrated that the assessment of infarct location can provide additional information to lesion volume, by identifying topographical characteristics of infarction that are associated with functional outcome. It is well recognized that baseline infarct volume is strongly correlated with outcome in ischemic stroke. However, patients with relatively small infarct volumes in highly ‘eloquent’ areas may have poor outcomes based on lesion location, and this may be an equally important consideration in prognostication and interpretation of the patient’s response to a
Fig. 1 Classification tree results for patients with left-hemispheric strokes. Gy, gyrus, Vol, Volume.
1272
Vol 10, December 2015, 1270–1276
© 2015 World Stroke Organization
Research
N. Yassi et al.
Fig. 2 Variable importance for left hemisphere analysis (nonzero results).
Fig. 3 Classification tree results for patients with right-hemispheric strokes. Lb, Lobe.
reperfusion therapy. In situations where multiple variable interactions exist, tree-based statistical methods such as CART have some advantages over methods such as logistic regression as they allow for the exploration of data for potentially complex, nonlinear interactions between variables, which is arguably a more closely related methodology to clinical decision-making processes. CART is also better suited to analyses of datasets with a © 2015 World Stroke Organization
large number of potentially nonparametric variables, whereas logistic regression may not be ideal in these situations (7). In the left hemisphere, the primary areas of importance identified in the tree were the uncinate fasciculus, precuneus, and angular gyrus. The uncinate fasciculus connects the orbitofrontal and anterior temporal cortex and is traditionally understood to be a limbic structure with proposed functions including episodic Vol 10, December 2015, 1270–1276
1273
Research
N. Yassi et al.
Fig. 4 Variable importance for right-hemisphere analysis (nonzero results). Table 1 Baseline characteristics and stroke characteristics in patients with favorable and nonfavorable outcomes
Left hemisphere Age, years (mean, SD) (median, IQR) Baseline NIHSS (median, IQR) Thrombolysis (n, %) DWI volume, ml (median, IQR) Right hemisphere Age (mean, SD) (median, IQR) Baseline NIHSS (median, IQR) Thrombolysis (n, %) DWI volume (median, IQR)
mRS 0–1
mRS 2–6
P-value*
67 (13) 70 (65·5–75·5) 8 (7–11) 19 (76) 17·5 (10·8–38·1)
72 (14) 75 (63–83) 16 (11–20) 41 (72) 47·4 (19·3–111·9)
69 (16) 73 (63–79·5) 10 (8–14) 17 (77) 17·3 (8·2–39·7)
74 (13) 77·5 (65–84·75) 13 (9–17) 29 (60) 75·9 (24·1–159·1)
0·163
P < 0·001 0·702 0·008 0·146
0·062 0·168 P < 0·001
SD, standard deviation; NIHSS, National Institutes of Health Stroke Scale; IQR, interquartile range; n, number; DWI, diffusion-weighted imaging. *P-value quoted independent samples t-test (age), Wilcoxon–Mann Whitney test (NIHSS and DWI Volume), and Chi-square test (Thrombolysis).
memory, language (left-hemisphere), and social and emotional processing (8,9). The precuneus is a medial parietal structure that has extensive cortical and sub-cortical interconnections and has been implicated in a range of higher order functions including consciousness, visuospatial imagery, episodic memory retrieval, and self-processing (10). It has been identified as one of the primary brain areas demonstrating task-independent decreases in activation on functional imaging and is one of the most metabolically active brain regions during rest, being one of the principal components of the default-mode network (11).
1274
Vol 10, December 2015, 1270–1276
The left angular gyrus, along with Wernicke’s area (part of the superior temporal gyrus), the supramarginal gyrus, and parts of the middle temporal gyrus, form the posterior language area. It has important roles in language comprehension and arithmetic retrieval (12,13). In the right hemisphere, we identified the parietal lobe and putamen as key areas associated with nonfavorable outcome. The right parietal lobe is a complex structure with a rich array of functions including the primary somatosensory cortex (postcentral gyrus), as well as areas critical for somatosensory association © 2015 World Stroke Organization
N. Yassi et al. and visuospatial function. The putamen is part of the striatum with important cortical and sub-cortical connections that predominantly mediate motor function (12). These findings emphasize the complexity and multiplicity of clinical deficits that can be attributable to stroke. Thus, we would suggest that while the consideration of infarct volume is helpful, particularly in the context of the commonly used statistical methods in this field, it does ignore complex interactions present within the data that provide a more complete and refined understanding of the results. The strength of the present analysis lies in the ability of CART to identify such complex interactions within the dataset, which could be overlooked using other statistical techniques. Although less commonly used in medicine, there are previous examples of its implementation including in the stroke field, for example in exploring important determinants of outcome in the National Institute of Neurological Disorders and Stroke rt-PA Stroke Trial (14,15). Importantly, the interpretation of such models as we have presented relies heavily upon a clear understanding of the outcome measure of interest. Although the mRS is rapid, simple, validated, and widely used in the field (16), it does have limitations. For example, it is insensitive in distinguishing disability attributable to nonstroke comorbidities such as cardiac or musculoskeletal conditions. In addition, it cannot account for potential variations in patient functional outcomes based on variable access to support services due to cultural or socioeconomic factors. Nonetheless, the widespread use of the mRS makes it a suitable endpoint in our analysis. We chose the cut-off of mRS > 1 as it is a commonly used end-point in large stroke trials. On the other hand, review of the results for alternative mRS dichotomies (Appendix S4) demonstrates that different regions are associated with outcome when the end-point of interest is changed. Interestingly, the cortico-spinal tract appears to be more important in left-hemispheric stroke patients with mRS > 2, and even more so in left-hemispheric stroke patients with mRS > 4. This stands to reason given that this latter category includes patients with the most severe motor disability. On the other hand, in the right hemisphere, white-matter tracts such as the superior longitudinal fasciculus appeared to be associated with outcome at higher mRS dichotomies. Importantly, this method does not negate the importance of other prognostic variables such as age, NIHSS, or overall lesion volume. Rather it allows the identification of associations between lesion location and outcome, which are otherwise difficult to determine systematically. There are several potential limitations to this study. First, the number of patients in the analysis is relatively modest and hence a validation of the method using a larger dataset would be useful. However, the internal cross-validation incorporated in CART should mitigate this limitation. The overall classification success rates in the validation set were also relatively modest, and thus this specific model may not be generally applicable to all stroke patients. However, the outcome of those patients who were classified into highly homogenous nodes could be predicted with a reasonable degree of confidence using the model, and the identification of these sub-groups of © 2015 World Stroke Organization
Research patients is an advantage of this method. Additionally, the overall accuracy and applicability of the model may be improved using a larger dataset. Unfortunately, neither trial routinely collected data on handedness that may have impacted hemispheric dominance and the disability produced by specific lesions. However, although righthemispheric dominance increases in patients who are left-handed or ambidextrous, left-hemispheric dominance is still the most common pattern in these populations (17). We therefore do not anticipate that the lack of data on handedness would have significantly altered the results. The different imaging time-points in the two trials may be viewed as a potential limitation. We chose the posttreatment DWI scan for this proof-of-principle analysis in order to account as much possible for the possibility of variable degrees of reperfusion after treatment (or in the natural history in patients receiving placebo). Partial temporary diffusion lesion reversal has been described in a small number of DEFUSE participants (18). However, this is unlikely to significantly alter the topography of the lesion. The day-3 imaging in EPITHET patients coincided with maximal infarct edema (19) and may have caused some registration error in this cohort. Importantly, however, CART results were not affected at any of the mRS dichotomies by the inclusion of original trial as an input, which suggests that the potential heterogeneity between the trials did not significantly impact the location associations. The robustness of the imaging transformation and registration in this study has been examined visually and appears to be acceptable without serious misregistration. However, transformation of low-resolution DWI images in older patients acquired at 1·5T is invariably prone to some degree of misregistration given that DWI does not provide a high degree of structural resolution and that the MNI-space was derived in young normal individuals. On the other hand, the segmentation of acute infarction on alternative sequences such as T1-weighted images or Fluid Attenuated Inversion Recovery is difficult and would have also added uncertainty to the registration process. The use of a standard-space template atlas allows the method to be widely applicable outside of the present study. Finally, it is clear that despite the large number of regions included in the atlas, there are areas of brain-tissue that were not covered, and further coverage may have improved the strength of the model. On the other hand, we feel that adopting an extremely granular approach with a larger number of regions (or potentially single voxels) would detract from the clinically intuitive nature of this method. In this exploratory analysis, we have demonstrated that CART analysis of stroke outcome can identify important relationships within data that are not overtly obvious using other statistical methods and that are clinically plausible. In the left-hemisphere, we have identified that infarction in regions associated with limbic and language function, as well as the default-mode network, is associated with nonfavorable outcome. In the righthemisphere, the importance of motor and somatosensory association areas was observed. Secondary analysis with alternative levels of disability outcome demonstrated that different brain Vol 10, December 2015, 1270–1276
1275
Research regions appear to be important when the target outcome is adjusted. This type of decision-tree analysis may allow more specific prediction of patient outcomes based on imaging characteristics and may play a role in the design of clinical trials by allowing a more informed consideration of the likely outcome and hence a reasonable interpretation of treatment effect. Furthermore, a comprehensive consideration of infarct location at baseline using similar methodology, with inclusion of other parameters such as penumbral location and the location of reperfused penumbral tissue after treatment, may allow improved prediction of reperfusion response and hence inform treatment decisions. For this to be feasible, application of the method to more rapidly accessible modalities (e.g. perfusion CT) should be explored, and validation in larger imaging datasets should be performed.
Acknowledgements EPITHET was supported by the National Health and Medical Research Council of Australia, National Stroke Foundation, and National Heart Foundation of Australia. DEFUSE was funded by NIH grants RO1-NS39325-R01NS39325/NS/NINDS-NIH-HHS/United States, K24-NS044848K24-NS044848/NS/NINDS-NIH-HHS/United States, and K23NS051372-K23-NS051372/NS/NINDS-NIH-HHS/United States. The Harvard Oxford Cortical Atlas is bundled within FSL and developed by David Kennedy and Christian Haselgrove, Centre for Morphometric Analysis, Harvard; Bruce Fischl, Martinos Center for Biomedical Imaging, Massachusetts General Hospital; Janis Breeze and Jean Frazier, Child and Adolescent Neuropsychiatric Research Program, Cambridge Health Alliance; Larry Seidman and Jill Goldstein, Department of Psychiatry, Harvard Medical School; Barry Kosofsky, Weill Cornell Medical Center.
Author contributions Drs Yassi and Churilov – Study concept and design, analysis and interpretation, drafting manuscript. All other authors provided critical revisions of the manuscript for important intellectual content.
References 1 Thijs VN, Lansberg MG, Beaulieu C, Marks MP, Moseley ME, Albers GW. Is early ischemic lesion volume on diffusion-weighted imaging an independent predictor of stroke outcome? Stroke 2000; 31:2597– 602. 2 Davis SM, Donnan GA, Parsons MW et al. Effects of alteplase beyond 3 h after stroke in the Echoplanar Imaging Thrombolytic Evaluation Trial (EPITHET). Lancet Neurol 2008; 7:299–309. 3 Albers GW, Thijs VN, Wechsler L et al. Magnetic resonance imaging profiles predict clinical response to early reperfusion. Ann Neurol 2006; 60:508–17.
1276
Vol 10, December 2015, 1270–1276
N. Yassi et al. 4 Jenkinson M, Bannister P, Brady M, Smith S. Improved optimization for the robust and accurate linear registration and motion correction of brain images. Neuroimage 2002; 17:825–41. 5 Hua K, Zhang J, Wakana S et al. Tract probability maps in stereotaxic spaces: analyses of white matter anatomy and tract-specific quantification. Neuroimage 2008; 39:336–47. 6 James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning with Applications in R. 1 ed. New York, Springer Science+Business Media, 2013:303. 7 Strobl C, Malley J, Tutz G. An introduction to recursive partitioning: rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychol Methods 2009; 14:323–48. 8 Papagno C, Miracapillo C, Casarotti A et al. What is the role of the uncinate fasciculus? Brain 2011; 134:405–14. 9 Von Der Heide RJ, Skipper LM, Klobusicky E, Olson IR. Dissecting the uncinate fasciculus. Brain 2013; 136:1692–707. 10 Cavanna AE, Trimble MR. The precuneus. Brain 2006; 129:564–83. 11 Gusnard DA, Raichle ME. Searching for a baseline: functional imaging and the resting human brain. Nat Rev Neurosci 2001; 2:685–94. 12 Nolte J. The Human Brain: An Introduction to Its Functional Anatomy, 5th edn. St. Louis, Missouri, USA, Mosby, Inc., 2002:60. 13 Binder JR, Frost JA, Hammeke TA, Cox RW, Rao SM, Prieto T. Human brain language areas identified by functional magnetic resonance imaging. The Journal of Neuroscience 1997; 17:353–62. 14 Broderick JP, Lu M, Kothari R et al. Finding the most powerful measures of the effectiveness of tissue plasminogen activator in the NINDS tPA stroke trial. Stroke 2000; 31:2335–41. 15 Frankel M, Morgenstern L, Kwiatkowski T et al. Predicting prognosis after stroke A placebo group analysis from the National Institute of Neurological Disorders and Stroke rt-PA Stroke Trial. Neurology 2000; 55:952–9. 16 Banks JL, Marotta CA. Outcomes validity and reliability of the modified Rankin scale: implications for stroke clinical trials. Stroke 2007; 38:1091–6. 17 Knecht S, Dräger B, Deppe M et al. Handedness and hemispheric language dominance in healthy humans. Brain 2000; 123:2512–8. 18 Campbell BC, Purushotham A, Christensen S et al. The infarct core is well represented by the acute diffusion lesion. J Cereb Blood Flow Metab 2012; 32:50–6. 19 Lansberg MG, O’Brien MW, Tong DC, Moseley ME, Albers GW. Evolution of cerebral infarct volume assessed by diffusion-weighted magnetic resonance imaging. Arch Neurol 2001; 58:613–7.
Supporting information Additional Supporting Information may be found in the online version of this article at the publisher’s web-site: Appendix S1. Standard brain atlases used to derive composite atlas. Appendix S2. Brain structures included in composite atlas. Appendix S3. Example registration in a single patient. Appendix S4. Further methods and results for alternate dichotomies. Appendix S5. The EPITHET and DEFUSE Investigators.
© 2015 World Stroke Organization
Correspondence: Nawaf Yassi, Melbourne Brain Centre @ Royal Melbourne Hospital, Grattan St, Parkville, Vic. 3050, Australia. E-mail: [email protected] 1 Department of Neurology, Melbourne Brain Centre @ Royal Melbourne Hospital, University of Melbourne, Parkville, Vic., Australia 2 Florey Institute of Neuroscience and Mental Health, University of Melbourne, Parkville, Vic., Australia 3 Departement of Radiology, Royal Melbourne Hospital, University of Melbourne, Parkville, Vic., Australia 4 Department of Neurology and Neurological Sciences and Stanford Stroke Center, Stanford University Medical Center, Stanford, CA, USA 5 Priority Research Centre for Translational Neuroscience and Mental Health, University of Newcastle and Hunter Medical Research Institute, Newcastle, NSW, Australia Received: 24 November 2014; Accepted: 25 February 2015; Published online 4 June 2015 Conflict of interest: None declared. DOI: 10.1111/ijs.12537
1270
Vol 10, December 2015, 1270–1276
Introduction The impact of infarct location on stroke severity as well as recovery and functional outcome is intuitive to the stroke clinician. However, its systematic incorporation into the considerations of clinical trials or clinical practice represents a unique challenge given the complexity and heterogeneity of neuronal network organization as well as additional factors such as hemispheric dominance, premorbid functional level, and concurrent nonneurological comorbidity. Most imaging-based clinical trials focus on the alternate and more simplistic variable of infarct volume. Although infarct volume has been shown to be strongly associated with functional outcome (1), it is rarely considered in isolation of other factors in clinical practice, where the additional assessment of lesion location is critical in determining the likely outcome. We aimed to use pooled data from the Echoplanar Imaging Thrombolytic Evaluation Trial (EPITHET) (2) and the Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution (DEFUSE) (3) Study in a retrospective, hypothesisgenerating investigation of the impact of infarct location on functional outcome, independent of overall lesion volume.
Methods EPITHET was a phase-II prospective randomized, doubleblinded, placebo-controlled trial testing alteplase vs. placebo in patients with hemispheric ischemic stroke presenting three- to six-hours after onset. Patients underwent serial 1·5T magnetic resonance imaging (MRI) at baseline, day 3–5, and day 90. DEFUSE was a prospective, open-label cohort study of patients with hemispheric ischemic stroke treated with alteplase 3-6h from onset. MRI (1·5T) was performed at baseline, three- to six-hours post-alteplase and at day 30. Given that ischemic core on baseline imaging can evolve variably depending on timing and degree of reperfusion, and in order to minimize the potential confounding effects of this variable recanalization, infarct volume was defined on posttreatment MRI diffusion-weighted imaging (DWI), performed at day 3–5 in EPITHET and three- to six-hours posttreatment in DEFUSE. Of 101 patients in EPITHET, one patient withdrew consent, seven patients died prior to day 3–5 imaging, five had inadequate day 3–5 imaging data for analysis, and one patient was lost to follow-up prior to day 90. Thus, 87 patients were included in the analysis. Of 74 patients enrolled in DEFUSE, seven patients had no identifiable lesion on posttreatment DWI, one patient did not have posttreatment DWI, and one patient had severely motion © 2015 World Stroke Organization
Research
N. Yassi et al. degraded imaging. Thus, 65 patients were included in the analysis (a total of 152 in the pooled analysis). Posttreatment DWI images were manually segmented by a stroke neurologist based on maximum visual extent on the b = 1000 diffusion image with visual comparison with apparent diffusion coefficient maps to exclude T2 shine through. The diffusion images were then transformed with 12 degrees-of-freedom into standard-space using the Centre for Functional MRI of the Brain (FMRIB) Linear Image Registration Tool (FLIRT, version 6·0, Oxford, UK) (4). The resultant transformation matrices were subsequently applied to the DWI-lesion masks to transform these into standard-space using a nearest-neighbor transformation, thus creating binary DWI-lesion masks in standard-space. The results were visually inspected to ensure accurate transformation. In a random sample comprising 20% of the overall cohort (30 patients), a second stroke neurologist separately performed manual lesion masking on the DWI. Interrater reliability was assessed using the intraclass correlation coefficient and further validated using Lin’s concordance coefficient and reduced major axis regression analysis. A novel composite brain atlas comprising a total of 132 structures was derived by combination of three separate atlases (Appendices S1, S2) bundled within FMRIB’s Software Library (FSL, version 5.0.6, FMRIB, Oxford, UK). The Montreal Neurological Institute (MNI)-152 atlas (Copyright 1993–2009 Louis Collins, McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University) is derived from 152 T1-weighted MRI scans from healthy young adults, averaged together after nonlinear registration into a common coordinate space. It contains nine left-hemispheric and nine right-hemispheric structures including the cerebellar hemispheres, which were excluded from the composite atlas. The Harvard-Oxford Cortical Atlas (refer to acknowledgements) covers 48 left-hemispheric cortical structures and 48 right-hemispheric cortical structures derived from T1-weighted images of 37 healthy subjects registered to MNI-152 space. The Johns Hopkins University White-Matter Tractography Atlas identifies 20 white-matter tracts by averaging the results of deterministic tractography on 28 normal subjects (5). It includes nine left-hemispheric tracts and nine right-hemispheric tracts, as well as the forceps major and minor. The forceps major and minor were included as whole structures in each hemisphere’s atlas given that these structures cross the midline. Therefore, the entire number of regions overall was 132, with each hemisphere consisting of 67 regions. Each subject’s transformed DWI lesion was segmented using the composite brain atlas, and the percentage of each of the atlas regions infarcted was determined as volume of infarction within the region divided by the total volume of the region multiplied by 100. An example of the registration process in a single patient is shown in Appendix S3. Classification and Regression Tree (CART) analysis, specifically classification, was used to assess the important brain regions in each hemisphere and infarct percentage thresholds within these regions associated with nonfavorable outcome at day 90 [modified Rankin Scale (mRS) > 1]. CART analysis was performed in the Salford Predictive Modeler Software Suite version 7 (Salford Systems, San Diego, CA, USA). The variables included as inputs in © 2015 World Stroke Organization
the model were the percentage infarction within each of the 67 atlas regions from the relevant hemisphere as well as total infarct volume given its known prognostic effect. Further CART background and methodology, as well as results of secondary analyses with different inputs and outcomes are provided in Appendix S4. Briefly, CART is a data-driven binary partitioning statistical method that starts with the total sample and, in a stepwise manner, splits the sample into two sub-samples that are more homogenous with respect to a defined outcome (6). The input variable that achieves the most effective split is dichotomized at an optimal threshold determined by the automated analysis to maximize the separation between the resulting groups (nodes). The model also searches for any potential surrogate splitters at each node, which are variables that split the node in a very similar way to the primary splitter. Once further splitting is not possible, pruning of the tree takes place in order to minimize over-fitting and to obtain a final tree. This involves a 10-fold internal cross-validation method, whereby the data are randomly divided into 10 groups with nine used to build the model (training) and one used to validate (testing). During pruning, CART starts at the bottom of the full tree and sequentially prunes nodes that result in the least decrement in the performance of the tree in the testing set, until an optimal tree is reached. The overall process of training and testing is repeated 10-fold, and the results are combined to produce the final optimal performing tree. CART also computes a variable importance score ranging from 0 to 100 and indicating the importance of each input in the model. The score takes into account the role of variables as either splitters or surrogate splitters and is scaled relative to the highest scoring variable (hence it is unit-less). Variables that do not appear in the final tree may still appear as important variables if they act as surrogate splitters.
Results Overall, 152 patients were included in the analysis. Seventy-six patients (50%) were female, and 87 (57%) were from EPITHET. Eighty-two patients (54%) had left-hemispheric infarcts. Alteplase was administered to 106 patients (69·7%). There were no significant differences between patients with left- and righthemispheric infarcts with regard to baseline characteristics, stroke severity, treatment allocation, or outcome. Median diffusion lesion volume overall was 37·0 ml (interquartile range 15·2– 96·6 ml), and this was also comparable between patients with leftand right-hemispheric infarcts. In univariate logistic regression, there was a strong association between lesion volume (transformed to the 5th-root to approximate normality) and nonfavorable outcome (odds ratio [OR] 4·9 per ml0·2[transformed to the 5th root], 95% confidence interval (CI) 2·3–10·5, P < 0·0001). In the random sample of 30 patients who had lesion masking performed by two different operators, the intraclass correlation coefficient for DWI lesion volume was 0·97, Lin’s concordance correlation coefficient 0·97, reduced major axis slope 0·98 and intercept −0·6 ml indicating the absence of either a fixed or a proportional bias and excellent agreement between raters. Vol 10, December 2015, 1270–1276
1271
Research Left hemisphere CART analysis for the primary outcome of day-90 mRS > 1 in patients with left-hemispheric stroke identified the uncinate fasciculus, precuneus cortex, angular gyrus, and total lesion volume as the inputs most strongly associated with outcome in the final tree (Fig. 1). Although the impact of lesion volume may initially appear to be counterintuitive (patients with >5·2 ml appear to have more favorable outcomes than patients with 1 compared with 69·5% overall. Therefore, infarcts occurring in regions common to this sub-group were by default occurring in ‘less critical’ regions than the preceding steps and therefore conferred a better prognosis. The relative importance score for inputs with nonzero importance in the left-hemisphere analysis is shown in Fig. 2. Overall, the rate of successful classification was 87·7% in the learning set and 67·1% in the testing set. For correct classification of nonfavorable outcome in the testing set, the model had a sensitivity of 70%, specificity of 60%, positive predictive value (PPV) of 80% and negative predictive value (NPV) of 47%. Qualitative review of the final model also demonstrates clear sub-groups of patients with contrasting prognosis (black arrows). One of these sub-groups contains 40 patients with a 90% proportion of nonfavorable outcome compared with another with 25 patients with only a 20% proportion of nonfavorable outcome (OR 36·0, 95% CI 8·7–149·5, P < 0·0001). Right hemisphere For right-hemispheric infarcts, inputs most strongly associated with mRS > 1 were infarction in the parietal lobe and the putamen (Fig. 3). Overall, the rate of successful classification was 82·9% in the learning set and 72·9% in the testing set. The variable
N. Yassi et al. importance for inputs with nonzero importance in the righthemisphere analysis is shown in Fig. 4. For the prediction of nonfavorable outcome in the testing set, the model had a sensitivity of 77%, specificity of 64%, PPV of 82% and NPV of 56%. As in the left-hemispheric analysis, the model also identified specific sub-groups with quite distinct prognosis (black arrows). In one of these groups, 32% of patients had a nonfavorable outcome, compared with 92% of patients in the other group (OR 24·6, 95% CI 5·9–102·0, P < 0·0001). Results of CART analysis for alternative mRS dichotomies, and with the inclusion of age as an input, are presented in Appendix 4. Although age appeared to be important in some of these analyses, it was absent from the final tree in the majority of cases. We also ran the analysis with the patient’s original trial (EPITHET vs. DEFUSE) as a categorical input, in order to explore any major heterogeneity given the differences in imaging time-points and other potential differences between the trials. This did not result in any changes to the CART results or variable importance results at any of the specified primary or secondary mRS dichotomies. Table 1 shows differences in major characteristics between patients with favorable and nonfavorable outcomes.
Discussion We have demonstrated that the assessment of infarct location can provide additional information to lesion volume, by identifying topographical characteristics of infarction that are associated with functional outcome. It is well recognized that baseline infarct volume is strongly correlated with outcome in ischemic stroke. However, patients with relatively small infarct volumes in highly ‘eloquent’ areas may have poor outcomes based on lesion location, and this may be an equally important consideration in prognostication and interpretation of the patient’s response to a
Fig. 1 Classification tree results for patients with left-hemispheric strokes. Gy, gyrus, Vol, Volume.
1272
Vol 10, December 2015, 1270–1276
© 2015 World Stroke Organization
Research
N. Yassi et al.
Fig. 2 Variable importance for left hemisphere analysis (nonzero results).
Fig. 3 Classification tree results for patients with right-hemispheric strokes. Lb, Lobe.
reperfusion therapy. In situations where multiple variable interactions exist, tree-based statistical methods such as CART have some advantages over methods such as logistic regression as they allow for the exploration of data for potentially complex, nonlinear interactions between variables, which is arguably a more closely related methodology to clinical decision-making processes. CART is also better suited to analyses of datasets with a © 2015 World Stroke Organization
large number of potentially nonparametric variables, whereas logistic regression may not be ideal in these situations (7). In the left hemisphere, the primary areas of importance identified in the tree were the uncinate fasciculus, precuneus, and angular gyrus. The uncinate fasciculus connects the orbitofrontal and anterior temporal cortex and is traditionally understood to be a limbic structure with proposed functions including episodic Vol 10, December 2015, 1270–1276
1273
Research
N. Yassi et al.
Fig. 4 Variable importance for right-hemisphere analysis (nonzero results). Table 1 Baseline characteristics and stroke characteristics in patients with favorable and nonfavorable outcomes
Left hemisphere Age, years (mean, SD) (median, IQR) Baseline NIHSS (median, IQR) Thrombolysis (n, %) DWI volume, ml (median, IQR) Right hemisphere Age (mean, SD) (median, IQR) Baseline NIHSS (median, IQR) Thrombolysis (n, %) DWI volume (median, IQR)
mRS 0–1
mRS 2–6
P-value*
67 (13) 70 (65·5–75·5) 8 (7–11) 19 (76) 17·5 (10·8–38·1)
72 (14) 75 (63–83) 16 (11–20) 41 (72) 47·4 (19·3–111·9)
69 (16) 73 (63–79·5) 10 (8–14) 17 (77) 17·3 (8·2–39·7)
74 (13) 77·5 (65–84·75) 13 (9–17) 29 (60) 75·9 (24·1–159·1)
0·163
P < 0·001 0·702 0·008 0·146
0·062 0·168 P < 0·001
SD, standard deviation; NIHSS, National Institutes of Health Stroke Scale; IQR, interquartile range; n, number; DWI, diffusion-weighted imaging. *P-value quoted independent samples t-test (age), Wilcoxon–Mann Whitney test (NIHSS and DWI Volume), and Chi-square test (Thrombolysis).
memory, language (left-hemisphere), and social and emotional processing (8,9). The precuneus is a medial parietal structure that has extensive cortical and sub-cortical interconnections and has been implicated in a range of higher order functions including consciousness, visuospatial imagery, episodic memory retrieval, and self-processing (10). It has been identified as one of the primary brain areas demonstrating task-independent decreases in activation on functional imaging and is one of the most metabolically active brain regions during rest, being one of the principal components of the default-mode network (11).
1274
Vol 10, December 2015, 1270–1276
The left angular gyrus, along with Wernicke’s area (part of the superior temporal gyrus), the supramarginal gyrus, and parts of the middle temporal gyrus, form the posterior language area. It has important roles in language comprehension and arithmetic retrieval (12,13). In the right hemisphere, we identified the parietal lobe and putamen as key areas associated with nonfavorable outcome. The right parietal lobe is a complex structure with a rich array of functions including the primary somatosensory cortex (postcentral gyrus), as well as areas critical for somatosensory association © 2015 World Stroke Organization
N. Yassi et al. and visuospatial function. The putamen is part of the striatum with important cortical and sub-cortical connections that predominantly mediate motor function (12). These findings emphasize the complexity and multiplicity of clinical deficits that can be attributable to stroke. Thus, we would suggest that while the consideration of infarct volume is helpful, particularly in the context of the commonly used statistical methods in this field, it does ignore complex interactions present within the data that provide a more complete and refined understanding of the results. The strength of the present analysis lies in the ability of CART to identify such complex interactions within the dataset, which could be overlooked using other statistical techniques. Although less commonly used in medicine, there are previous examples of its implementation including in the stroke field, for example in exploring important determinants of outcome in the National Institute of Neurological Disorders and Stroke rt-PA Stroke Trial (14,15). Importantly, the interpretation of such models as we have presented relies heavily upon a clear understanding of the outcome measure of interest. Although the mRS is rapid, simple, validated, and widely used in the field (16), it does have limitations. For example, it is insensitive in distinguishing disability attributable to nonstroke comorbidities such as cardiac or musculoskeletal conditions. In addition, it cannot account for potential variations in patient functional outcomes based on variable access to support services due to cultural or socioeconomic factors. Nonetheless, the widespread use of the mRS makes it a suitable endpoint in our analysis. We chose the cut-off of mRS > 1 as it is a commonly used end-point in large stroke trials. On the other hand, review of the results for alternative mRS dichotomies (Appendix S4) demonstrates that different regions are associated with outcome when the end-point of interest is changed. Interestingly, the cortico-spinal tract appears to be more important in left-hemispheric stroke patients with mRS > 2, and even more so in left-hemispheric stroke patients with mRS > 4. This stands to reason given that this latter category includes patients with the most severe motor disability. On the other hand, in the right hemisphere, white-matter tracts such as the superior longitudinal fasciculus appeared to be associated with outcome at higher mRS dichotomies. Importantly, this method does not negate the importance of other prognostic variables such as age, NIHSS, or overall lesion volume. Rather it allows the identification of associations between lesion location and outcome, which are otherwise difficult to determine systematically. There are several potential limitations to this study. First, the number of patients in the analysis is relatively modest and hence a validation of the method using a larger dataset would be useful. However, the internal cross-validation incorporated in CART should mitigate this limitation. The overall classification success rates in the validation set were also relatively modest, and thus this specific model may not be generally applicable to all stroke patients. However, the outcome of those patients who were classified into highly homogenous nodes could be predicted with a reasonable degree of confidence using the model, and the identification of these sub-groups of © 2015 World Stroke Organization
Research patients is an advantage of this method. Additionally, the overall accuracy and applicability of the model may be improved using a larger dataset. Unfortunately, neither trial routinely collected data on handedness that may have impacted hemispheric dominance and the disability produced by specific lesions. However, although righthemispheric dominance increases in patients who are left-handed or ambidextrous, left-hemispheric dominance is still the most common pattern in these populations (17). We therefore do not anticipate that the lack of data on handedness would have significantly altered the results. The different imaging time-points in the two trials may be viewed as a potential limitation. We chose the posttreatment DWI scan for this proof-of-principle analysis in order to account as much possible for the possibility of variable degrees of reperfusion after treatment (or in the natural history in patients receiving placebo). Partial temporary diffusion lesion reversal has been described in a small number of DEFUSE participants (18). However, this is unlikely to significantly alter the topography of the lesion. The day-3 imaging in EPITHET patients coincided with maximal infarct edema (19) and may have caused some registration error in this cohort. Importantly, however, CART results were not affected at any of the mRS dichotomies by the inclusion of original trial as an input, which suggests that the potential heterogeneity between the trials did not significantly impact the location associations. The robustness of the imaging transformation and registration in this study has been examined visually and appears to be acceptable without serious misregistration. However, transformation of low-resolution DWI images in older patients acquired at 1·5T is invariably prone to some degree of misregistration given that DWI does not provide a high degree of structural resolution and that the MNI-space was derived in young normal individuals. On the other hand, the segmentation of acute infarction on alternative sequences such as T1-weighted images or Fluid Attenuated Inversion Recovery is difficult and would have also added uncertainty to the registration process. The use of a standard-space template atlas allows the method to be widely applicable outside of the present study. Finally, it is clear that despite the large number of regions included in the atlas, there are areas of brain-tissue that were not covered, and further coverage may have improved the strength of the model. On the other hand, we feel that adopting an extremely granular approach with a larger number of regions (or potentially single voxels) would detract from the clinically intuitive nature of this method. In this exploratory analysis, we have demonstrated that CART analysis of stroke outcome can identify important relationships within data that are not overtly obvious using other statistical methods and that are clinically plausible. In the left-hemisphere, we have identified that infarction in regions associated with limbic and language function, as well as the default-mode network, is associated with nonfavorable outcome. In the righthemisphere, the importance of motor and somatosensory association areas was observed. Secondary analysis with alternative levels of disability outcome demonstrated that different brain Vol 10, December 2015, 1270–1276
1275
Research regions appear to be important when the target outcome is adjusted. This type of decision-tree analysis may allow more specific prediction of patient outcomes based on imaging characteristics and may play a role in the design of clinical trials by allowing a more informed consideration of the likely outcome and hence a reasonable interpretation of treatment effect. Furthermore, a comprehensive consideration of infarct location at baseline using similar methodology, with inclusion of other parameters such as penumbral location and the location of reperfused penumbral tissue after treatment, may allow improved prediction of reperfusion response and hence inform treatment decisions. For this to be feasible, application of the method to more rapidly accessible modalities (e.g. perfusion CT) should be explored, and validation in larger imaging datasets should be performed.
Acknowledgements EPITHET was supported by the National Health and Medical Research Council of Australia, National Stroke Foundation, and National Heart Foundation of Australia. DEFUSE was funded by NIH grants RO1-NS39325-R01NS39325/NS/NINDS-NIH-HHS/United States, K24-NS044848K24-NS044848/NS/NINDS-NIH-HHS/United States, and K23NS051372-K23-NS051372/NS/NINDS-NIH-HHS/United States. The Harvard Oxford Cortical Atlas is bundled within FSL and developed by David Kennedy and Christian Haselgrove, Centre for Morphometric Analysis, Harvard; Bruce Fischl, Martinos Center for Biomedical Imaging, Massachusetts General Hospital; Janis Breeze and Jean Frazier, Child and Adolescent Neuropsychiatric Research Program, Cambridge Health Alliance; Larry Seidman and Jill Goldstein, Department of Psychiatry, Harvard Medical School; Barry Kosofsky, Weill Cornell Medical Center.
Author contributions Drs Yassi and Churilov – Study concept and design, analysis and interpretation, drafting manuscript. All other authors provided critical revisions of the manuscript for important intellectual content.
References 1 Thijs VN, Lansberg MG, Beaulieu C, Marks MP, Moseley ME, Albers GW. Is early ischemic lesion volume on diffusion-weighted imaging an independent predictor of stroke outcome? Stroke 2000; 31:2597– 602. 2 Davis SM, Donnan GA, Parsons MW et al. Effects of alteplase beyond 3 h after stroke in the Echoplanar Imaging Thrombolytic Evaluation Trial (EPITHET). Lancet Neurol 2008; 7:299–309. 3 Albers GW, Thijs VN, Wechsler L et al. Magnetic resonance imaging profiles predict clinical response to early reperfusion. Ann Neurol 2006; 60:508–17.
1276
Vol 10, December 2015, 1270–1276
N. Yassi et al. 4 Jenkinson M, Bannister P, Brady M, Smith S. Improved optimization for the robust and accurate linear registration and motion correction of brain images. Neuroimage 2002; 17:825–41. 5 Hua K, Zhang J, Wakana S et al. Tract probability maps in stereotaxic spaces: analyses of white matter anatomy and tract-specific quantification. Neuroimage 2008; 39:336–47. 6 James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning with Applications in R. 1 ed. New York, Springer Science+Business Media, 2013:303. 7 Strobl C, Malley J, Tutz G. An introduction to recursive partitioning: rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychol Methods 2009; 14:323–48. 8 Papagno C, Miracapillo C, Casarotti A et al. What is the role of the uncinate fasciculus? Brain 2011; 134:405–14. 9 Von Der Heide RJ, Skipper LM, Klobusicky E, Olson IR. Dissecting the uncinate fasciculus. Brain 2013; 136:1692–707. 10 Cavanna AE, Trimble MR. The precuneus. Brain 2006; 129:564–83. 11 Gusnard DA, Raichle ME. Searching for a baseline: functional imaging and the resting human brain. Nat Rev Neurosci 2001; 2:685–94. 12 Nolte J. The Human Brain: An Introduction to Its Functional Anatomy, 5th edn. St. Louis, Missouri, USA, Mosby, Inc., 2002:60. 13 Binder JR, Frost JA, Hammeke TA, Cox RW, Rao SM, Prieto T. Human brain language areas identified by functional magnetic resonance imaging. The Journal of Neuroscience 1997; 17:353–62. 14 Broderick JP, Lu M, Kothari R et al. Finding the most powerful measures of the effectiveness of tissue plasminogen activator in the NINDS tPA stroke trial. Stroke 2000; 31:2335–41. 15 Frankel M, Morgenstern L, Kwiatkowski T et al. Predicting prognosis after stroke A placebo group analysis from the National Institute of Neurological Disorders and Stroke rt-PA Stroke Trial. Neurology 2000; 55:952–9. 16 Banks JL, Marotta CA. Outcomes validity and reliability of the modified Rankin scale: implications for stroke clinical trials. Stroke 2007; 38:1091–6. 17 Knecht S, Dräger B, Deppe M et al. Handedness and hemispheric language dominance in healthy humans. Brain 2000; 123:2512–8. 18 Campbell BC, Purushotham A, Christensen S et al. The infarct core is well represented by the acute diffusion lesion. J Cereb Blood Flow Metab 2012; 32:50–6. 19 Lansberg MG, O’Brien MW, Tong DC, Moseley ME, Albers GW. Evolution of cerebral infarct volume assessed by diffusion-weighted magnetic resonance imaging. Arch Neurol 2001; 58:613–7.
Supporting information Additional Supporting Information may be found in the online version of this article at the publisher’s web-site: Appendix S1. Standard brain atlases used to derive composite atlas. Appendix S2. Brain structures included in composite atlas. Appendix S3. Example registration in a single patient. Appendix S4. Further methods and results for alternate dichotomies. Appendix S5. The EPITHET and DEFUSE Investigators.
© 2015 World Stroke Organization
