Leukoaraiosis Attenuates Diagnostic Accuracy of Large-Vessel Occlusion Scales

BACKGROUND AND PURPOSE: Prehospital stroke scales may help identify patients likely to have large-vessel occlusion to facilitate rapid triage to thrombectomy-capable stroke centers. Scale misclassification may result in inaccurate decisions and possible harm. Pre-existing leukoaraiosis has been shown to attenuate the association between deficit type and stroke severity. We sought to determine whether leukoaraiosis affects the predictive ability of 5 commonly used large-vessel occlusion scales. MATERIALS AND METHODS: We retrospectively analyzed 274 consecutive patients with stroke with available brain MR imaging and vessel imaging. We used the following large-vessel occlusion scales: the 3-Item Stroke Scale; Field Assessment Stroke Triage for Emergency Destination; Rapid Arterial Occlusion Evaluation; Vision, Aphasia, Neglect score; and Cincinnati Prehospital Stroke Severity Scale. For diagnostic scale accuracy, we assessed sensitivity, specificity, positive predictive value, negative predictive value, and κ. Multivariable logistic regression was used to determine the predictive ability of the scales after adjustment for leukoaraiosis and potential confounders. RESULTS: In unadjusted analyses, all scales predicted the presence of large-vessel occlusion (n = 46, P < .01 each), though diagnostic accuracy was attenuated among patients with moderate-to-severe leukoaraiosis. After adjustment, the Field Assessment Stroke Triage for Emergency Destination (OR = 3.2; 95% CI, 1.1–9.5; P = .033) and Rapid Arterial Occlusion Evaluation (OR = 3.7; 95% CI, 1.3–10.8; P = .015), but not the 3-Item Stroke Scale (OR = 5.4; 95% CI, 0.86–33.9; P = .073), Vision, Aphasia, Neglect score (OR = 2.5; 95% CI, 0.8–7.2), and Cincinnati Prehospital Stroke Severity Scale (OR = 2.8; 95% CI, 1.0–8.0), predicted large-vessel occlusion. CONCLUSIONS: The diagnostic accuracy of the tested large-vessel occlusion scales was attenuated in the presence of moderate-to-severe leukoaraiosis. This information that may aid the design of future studies that require large-vessel occlusion scale screening of patients who are likely to have concomitant leukoaraiosis.

may increase the risk for misclassification to improve prediction accuracy.
Pre-existing leukoaraiosis is identified in most patients with stroke on brain MR imaging as white matter hyperintensities. 14 Prior studies have shown that leukoaraiosis alters the classic association between the NIHSS deficit and hemispheric lateralization, relates to greater-than-expected NIHSS scores in relation to the acute infarct extent, and worsens outcome after EST. [14][15][16][17] Because most available LVO scales are based on the NIHSS, 6-10 we sought to determine whether leukoaraiosis severity affects the diagnostic accuracy of several previously developed prehospital stroke scales. [6][7][8][9][10] We hypothesized that the diagnostic scale performance is reduced among patients with stroke with moderate-to-severe pre-existing leukoaraiosis. Clarifying this issue may aid clinical study design and choice of LVO scales to more reliably identify patients with stroke with LVO in the prehospital setting when neuroimaging data are unavailable, and it is expected that a substantial proportion of screened patients will have advanced leukoaraiosis.

Study Cohort
We retrospectively analyzed 274 consecutive patients with acute ischemic stroke shown on brain MR imaging who were included in the University of Massachusetts Memorial Medical Center Stroke Registry between January 2013 and January 2014. 15,18 Patient demographics, laboratory data, comorbidities, preadmission medications, and stroke pathogenesis (using the Causative Classification System for Ischemic Stroke as previously described 15,18 ) after completion of diagnostic evaluation were collected on all patients. NIHSS scores were assessed at the time of presentation by members of the stroke team certified in the NIHSS. Only patients with complete details on all NIHSS subcategories were included (On-line Fig 1). All patients underwent head CT and either CTA (n ϭ 249) or MRA (n ϭ 25) at admission. Ten patients underwent both CTA and MRA. To reliably determine the white matter lesion burden (leukoaraiosis), we only included patients with available brain MR imaging. Our investigation was approved by our Institutional Review Board (No. H00006964), and a Health Insurance Portability and Accountability Act waiver of informed consent was granted. We adhere to the Strengthening the Reporting of Observational Studies in Epidemiology (www.strobe-statement.org) and Standards for Reporting of Diagnostic Accuracy Studies (www.stard-statement. org) guidelines.

Neuroimaging Protocol
All MR imaging and MRA sequences were acquired on a 1.5T whole-body MR imaging scanner (Signa HD; GE Healthcare, Milwaukee, Wisconsin) between 1 and 7 days after stroke. Brain MR imaging sequences included T1, T2, FLAIR, and DWI. DWI was performed with echo-planar imaging with a TR of 8000 ms, TE of 102 ms, FOV of 22 ϫ 22 cm, image matrix of 128 ϫ 128, slice thickness of 5 mm with a 1-mm interslice gap, and b-values of 0 and 1000 s/mm 2 . FLAIR was performed with a TR of 9002 ms, TE of 143 ms, FOV of 22 ϫ 22 cm, image matrix of 256 ϫ 224, and slice thickness of 6 mm with a 1-mm interslice gap. All MRA was performed with TOF-echo-spoiled gradient-echo pulse sequences. Head MRA was performed with a TR of 25 ms, flip angle of 20°, FOV of 20 cm, matrix of 256 ϫ 224, and slice thickness of 1.4 mm. Neck MRA performed with contrast used a TE of 1.8, flip angle of 45°, FOV of 33, slice thickness of 1 mm, and matrix of 28.4 ϫ 22.4 cm. Patients received 20 mL of gadobenate dimeglumine (MultiHance; Bracco Diagnostics, Princeton, New Jersey). Neck MRA was performed without contrast with a TR of 24 ms, TE of 5.1 ms, flip angle of 60°, FOV of 20, slice thickness of 1.5 mm, and matrix of 512 ϫ 160.
All CT sequences were obtained on a 64 -detector row scanner (Brilliance; Philips Healthcare, Best, the Netherlands). 17,19 CT was performed in a nonhelical mode at 120 kV(peak) and 200 mA, with data reconstruction at 5-mm axial sections. CTA was performed using a 64 ϫ 0.625 mm detector configuration with a pitch of 0.673 from the arch of the aorta to the vertex using 120 kV(p), 300 mA, and 0.5-second rotation time. Patients received 60 -80 mL of iopamidol (Isovue 370; Bracco Diagnostics) in the antecubital vein at a rate of 4 mL/s through a power injector, followed by 40 mL of saline. 3D orthogonal MIP images were created in 3 planes.

Leukoaraiosis Grading
In all patients, leukoaraiosis was defined on MR imaging as supratentorial white matter FLAIR hyperintensity lesions according to the Standards for Reporting Vascular Changes on Neuroimaging 20 criteria and graded according to the Fazekas scale as previously described in detail. 14,15 The total Fazekas scale score was calculated by adding the periventricular and subcortical scores. 14 We have previously shown substantial interrater reliability with an intraclass correlation coefficient of 0.969 (95% CI, 0.943-0.983). 14 In addition, we dichotomized the degree of leukoaraiosis according to the median Fazekas scale score to 0 -2 (n ϭ 145, absent-to-mild leukoaraiosis) versus 3-6 (n ϭ 129, moderate-tosevere leukoaraiosis) for statistical purposes.

Definition of Large-Vessel Occlusion
Large-vessel occlusion was defined on the admission CTA or MRA as the presence of an apparent occlusion of the intracranial internal carotid artery and proximal middle cerebral artery (M1). The included prehospital stroke scales were not designed to assess the presence of occlusion in the distal MCA (M2), basilar, and intracranial vertebral arteries. Therefore, occlusion in these vessels was considered non-LVO for this study, and only considered in the exploratory analyses. All images were reviewed by a neuroradiologist as part of the routine clinical work-up. The site of the LVO was abstracted from the radiologic report and by cross-validating with the original scans by one of the authors (Y.M.), who was masked to clinical data. Discrepant interpretations were resolved by consensus after review by a second neurologist (N.H.).

Prehospital Stroke Scales
To determine the impact of pre-existing leukoaraiosis on LVO prediction, we examined 5 previously published scales that could be reliably reconstructed by abstraction from the admission NIHSS score sheet and the detailed neurologic examination documented in our medical records. We used the following addi-tional prespecified rules for derivation of the scores because we previously found that severe leukoaraiosis substantially attenuates the classic hemispheric lateralization of the NIHSS deficit 15 : If both sides were weak, the weaker side determined the sign scored; and if both sides were equally weak, the presence of the sign was scored.

Statistics
Unless otherwise stated, continuous variables are reported as mean Ϯ SD or median (25th-75th percentile). Categoric variables are reported as proportions. Between-group comparisons for continuous and ordinal variables were performed with the Mann-Whitney U test. Categoric variables were compared using the 2 test or Fisher exact test as appropriate. To determine the diagnostic accuracy of the LVO scales, we calculated the area under receiver operating curves (C statistics, ordinal scales) as well as sensitivity, specificity, positive predictive value, negative predictive value, and (dichotomized scales each) with corresponding 95% CIs.
We created multivariable logistic regression models to determine whether the tested LVO scale cutoffs predicted large-artery occlusion (dependent variable) independent of leukoaraiosis (as determined by the Fazekas scale score). All models were adjusted for factors known to be associated with leukoaraiosis, including age, sex, and preadmission mRS and a history of hypertension, stroke/TIA, and atrial fibrillation. In addition, models were adjusted for the Fazekas scale score ϫ LVO Scale Interaction to determine whether the leukoaraiosis burden differentially modified the predictive ability of the tested scales. Variables were sequentially removed (likelihood ratio) from the models at a significance level of 0.1 to avoid model overfitting. Collinearity diagnostics were performed (and its presence rejected) for all multivariable regression models. The Hosmer-Lemeshow goodness-of-fit statistic was used to assess model fit.
Two-sided significance tests were used throughout; unless stated otherwise, a 2-sided P Ͻ .05 was considered statistically significant. All statistical analyses were performed with SPSS, Version 22 (IBM, Armonk, New York).

RESULTS
Of 502 patients with ischemic stroke, 274 fulfilled the study criteria and were included in the analysis (On-line Fig 1).
Baseline characteristics of included patients stratified by the presence of LVO are summarized in Table 1. Overall, 46 patients had a proximal LVO (17%), and 228 (83%) patients had no LVO. Baseline characteristics of the included patients were well-balanced except for a higher NIHSS (P Ͻ .001) and Fazekas scale score (P ϭ .009) among patients with LVO. The relative distribu-  tion of the NIHSS scores and types of arterial occlusion in the included patients are summarized in the Figure. On-line Fig 2 shows the percentage of LVO cases that would have gone unidentified across different LVO score thresholds. On average, more cases would have gone unnoticed in patients with moderate-to-severe leukoaraiosis compared with those with only absent-to-mild leukoaraiosis. Table 2 summarizes the diagnostic accuracy of the established LVO scale cutoffs both in the entire cohort and when stratified by leukoaraiosis. Overall, the diagnostic accuracy was modest in the entire cohort and further attenuated among patients with moderate-to-severe leukoaraiosis.
We calculated the area under receiver operating curve to examine the diagnostic performance of the 5 tested prehospital LVO scales for the presence of a proximal LVO as an ordinal variable in the entire cohort as well as stratified by leukoaraiosis burden (On-line Fig 3A-C). All scales predicted LVO in the entire cohort (P Ͻ .01 each, Table 3) and in patients with absent-to-mild leukoaraiosis (P Ͻ .001 each). However, none predicted LVO among patients with moderate-to-severe leukoaraiosis (P Ͼ .05 each). Results did not meaningfully change when the definition of LVO was liberalized to include MCA-M2, vertebral artery, and basilar artery occlusions (On-line Fig 3D-F). Last, results were similar in exploratory analyses excluding patients with small subcortical infarcts as well as those with posterior circulation infarcts (not shown).
Finally, on multivariable logistic regression with adjustment for leukoaraiosis and potential confounders, we found that despite an overall similar direction of association of all scales, only the FAST-ED and RACE scales predicted LVO (Table 3). In all models, leukoaraiosis severity was independently associated with the presence of LVO (P Ͻ .05, each). No Leukoaraiosis ϫ LVO Scale Interaction was found in the examined models.

DISCUSSION
The goal of our study was to determine whether the diagnostic accuracy of previously validated prehospital LVO stroke scales is affected by the presence of leukoaraiosis-knowledge that may aid investigators in choosing scales that allow more reliable patient screening in the prehospital setting. Using a hospital-based acute ischemic stroke cohort, we found that the diagnostic accuracy of several LVO scales was attenuated among patients with severe leukoaraiosis. After we adjusted for leukoaraiosis severity and potential confounders, only 2 of the 5 tested scales (FAST-ED and RACE) predicted LVO.
Prehospital stroke scales have been developed to aid stroke recognition and prediction of LVO by non-stroke specialists when neuroimaging information is unavailable. Our results may aid the clinical study design when prehospital patient screening involves LVO scales, and it is expected that a substantial proportion of patients are likely to have leukoaraiosis. If confirmed, our data suggest that in such a setting, scales like the FAST-ED and RACE may be advantageous because they appeared less impervious to leukoaraiosis-related misclassification. Furthermore, LVO scales are increasingly used to render decisions regarding immediate patient transfer to EST-capable hospitals when LVO is suspected, 13,21-23 including in the setting of telemedicine consultations. 24 In this setting, our results may be impactful when a plain head CT is reviewed (and provides information on leukoaraiosis severity) but no CTA was obtained due to lack of resources or immedi-

FIGURE.
NIHSS and LVO distribution in the tested cohort. A, Relative frequency of NIHSS scores as stratified by the presence-versus-absence of large-artery occlusion. B, Distribution of NIHSS scores according to large-artery occlusion status. There was no difference in the distribution of NIHSS scores (bins) between subjects with absent-to-mild versus moderate-to-severe leukoaraiosis (P ϭ .319, 2 test).
ate availability or because repeat CTA is not desirable to avoid repeat imaging at the EST-capable hub hospital. 25,26 Leukoaraiosis has been repeatedly shown to confer an increased risk of poor outcome in patients with LVO, including those undergoing EST. 17,19 Arguably, the presence of severe leukoaraiosis could thus prompt CTA at an outside hospital (versus transfer without CTA) to minimize the cost associated with unnecessary transfers. 27 However, to date, no study has shown that patients with severe leukoaraiosis do not benefit from EST. Accordingly, our data should not direct medical decision-making. Nevertheless, one could envision that the presence of specific, yet-to-be-defined, clinical patient characteristics trigger "prescreening" of patients with CTA to aid transfer decisions. Such factors may be very old age, prestroke functional status, and cognitive impairment, which have been associated with leukoaraiosis and limited benefit from EST. 28 The strengths of our study relate to inclusion of consecutive patients with available MR imaging, imaging assessment by experienced examiners, and adjustment for important factors associated with leukoaraiosis and stroke severity. Limitations relate to the retrospective and observational nature of the study, as well as the inclusion of a moderately sized single-center cohort. In this respect, the overall diagnostic accuracy of the tested scales was only modest in our cohort. This result may be partially explained by our hospital-based study, which only included patients with proved stroke and who had available vessel imaging because the diagnostic accuracy is sensitive to the characteristics (disease prevalence and spectrum) of the tested population. Nevertheless, our results are in line with observations from other hospital-based studies suggesting generalizability of our results. 29,30 Last, due to our study design, we only examined LVO stroke scales based on the NIHSS. Accordingly, our results may not translate to other scales and should be interpreted cautiously.

CONCLUSIONS
Our results indicate that the presence of moderate-to-severe leukoaraiosis attenuates the diagnostic accuracy of available LVO scales and that the RACE and FAST-ED scales appeared more resilient to the confounding effects of leukoaraiosis. If confirmed in a larger cohort, this information may aid the design of studies that require prehospital LVO scale screening of patients likely to have concomitant leukoaraiosis.   .015 a Each model was adjusted for leukoaraiosis burden (as determined by the Fazekas scale score) as well as Leukoaraiosis ϫ Stroke Scale Interaction. When additionally entered into the model, the prestroke modified Rankin Scale, age, sex, hypertension, history of stroke/transient ischemic attack, and atrial fibrillation were not retained in the final step of the analysis. In all analyses, leukoaraiosis and stroke scales were entered as dichotomized variables. When all analyses were repeated by entering the Fazekas scale score and stroke scale scores as ordinal variables, the VAN (P ϭ .029), CPSSS (P ϭ .024), FAST-ED (P ϭ .003), RACE (P ϭ .002), and 3I/SS (P ϭ .006) scores predicted LVO independent of the leukoaraiosis burden (not shown). There was no Leukoaraiosis ϫ Stroke Scale Interaction in any of the examined models (P Ͼ .05, each).