Lyden P, Claesson L, Havstad S, Ashwood T, Lu M. Factor Analysis of the National Institutes of Health Stroke Scale in Patients With Large Strokes. Arch Neurol. 2004;61(11):1677-1680. doi:10.1001/archneur.61.11.1677
The National Institutes of Health Stroke Scale (NIHSS) was created to detect treatment-related differences in clinical trials and was designed to measure right- and left-sided cerebral hemispheric function.
To validate the original design in patients with very large strokes.
A previously published factor structure was fit to the data. Then, a new analysis was conducted to explore the underlying structure of the scale in this population. Finally, NIHSS scores and infarction volumes were compared.
The Clomethiazole for Acute Stroke Study–Ischemic, conducted in academic and community hospitals.
Individuals with acute stroke seen within 12 hours of onset. Of 1191 records available, 98% had complete NIHSS scores.
Main Outcome Measure
Goodness-of-fit statistic (Bentler) for each factor solution.
Two factors were found underlying the NIHSS, corresponding to the left and right hemispheres (goodness of fit = 0.97), using the previously published factor analysis. The new exploratory analysis also suggested 2 factors representing left and right brain function. The median (range) NIHSS scores were 15 (5-25) for right brain strokes and 19 (6-32) for left brain strokes (P<.001). The median (range) infarction volumes were 56.2 mL (0.1-381.5 mL) for right brain strokes and 37.8 mL (0.2-255.1 mL) for left brain strokes (P<.001). The correlation coefficient between NIHSS score and lesion volume was 0.37 (P<.001).
The underlying structure of the NIHSS conforms to cerebral hemispheric lateralization, confirming previous findings in a new population of large hemispheric strokes. Left- brain strokes score 4 points higher on the NIHSS than right brain strokes of larger volume.
The underlying structure of the National Institutes of Health Stroke Scale (NIHSS) contained 2 underlying factors, corresponding to the cerebral hemispheres, in a previous population of mild-to-moderate strokes, thus validating the original, empirical design. In another validation, the NIHSS correlated with infarction volume, but in one study, patients with left brain strokes tended to score, on average, 4 points higher than those with right brain strokes, thus challenging the criterion validity of the scale.1- 3 The scale correlated with alternative measures of neurologic outcome, such as activities of daily living scales, and other deficit scales, further suggesting validity.4 It remains unknown whether the same 2 hemisphere factors are found in other study populations, notably, patients with large hemispheric strokes, who are now the preferred targets in neuroprotection trials.5
We proposed to confirm the previous findings in a new group of patients with larger hemispheric infarctions. The Clomethiazole for Acute Stroke Study–Ischemic (CLASS-I) tested clomethiazole therapy in patients with a major ischemic stroke (resembling the total anterior circulation syndrome [TACS])6 with a combination of limb weakness, higher cortical dysfunction, and visual field deficits within 12 hours of symptom onset.7 The NIHSS score was obtained before treatment with any study material. Clomethiazole therapy did not improve outcomes in patients with major ischemic stroke, but CLASS-I remains the largest collection of well-studied TACS-like patients, and thus the study group afforded a unique population for studying the NIHSS.8
The CLASS-I methods and results were published previously.7,8 The CLASS-I investigators underwent NIHSS annual certification via videotapes.9,10 Each patient was scored by a certified investigator at initial stroke evaluation (baseline). Brain computed tomography was performed 30 days later.
We first tested a published factor structure—derived from the National Institute of Neurological Disorders and Stroke tPA [tissue plasminogen activator] Stroke Study2—with the present data. Next, we conducted a new exploratory analysis, followed by a confirmatory analysis using published methods.2 Half of the data were randomly chosen for the initial exploratory factor analysis assuming a 2- or 4-factor structure based on previous experience3; the remaining data were used for a confirmatory factor analysis. We explored 15 NIHSS items using PROC FACTOR in SAS, with Harris-Kaiser rotation assuming correlations among the factors.11 We conducted structure modeling using PROC CALIS in SAS.11,12 We calculated the percentage of direct contribution (eg, communality) of each NIHSS item to the derived factor structure. We also calculated the goodness of fit, the Bentler Comparative Fit Index, for the selected factor structure.12- 14 The Comparative Fit Index ranges from 0 to 1 and is viewed as the percentage of variation in the observed measure (the scale items) explained by a given structure (such as R2 in the regression model), where a value greater than 0.90 indicates an excellent fit and validation of the given factor structure.12,13 Finally, we validated the factor structure using the remaining half of the data.
Infarction volumes were computed using semiquantitative image analysis data from 30-day computed tomographic brain scans.7 We report NIHSS scores and infarction volumes for left and right brain strokes, and the Wilcoxon rank sum test was used to test the difference. We also calculated Spearman correlation coefficients between NIHSS scores and infarction volumes.
The NIHSS version used in this study is given in Table 1. Of 1198 CLASS-I patients, 1169 completed the baseline NIHSS assessment; in the remainder, scores were either incomplete or missing. The demographic data are provided in a previous study.8 The median baseline NIHSS score was 17 (range, 4-32). The median NIHSS score for right brain strokes was 15 (range, 5-25) and for left brain strokes was 19 (range, 6-32) (P<.001). These severe scores reflect the fact that the CLASS-I required a severe syndrome, similar to TACS, for inclusion.6
In the preliminary analysis (n = 1169), the previously derived factor structure fit CLASS-I data well, with an excellent goodness of fit (Comparative Fit Index = 0.97). Consciousness (item 1a) and ataxia (item 7) loaded poorly.
The 2-factor exploratory analysis using half the data at random (n = 586) yielded a goodness of fit of 0.89 (data not shown). The exploratory 4-factor solution is given in Table 2 and shows an excellent goodness of fit (Comparative Fit Index = 0.97). The factor structure is identical to published analyses using less severe strokes: factor 1 seems to represent left hemisphere cortical function, and factor 2 generally represents right hemisphere cortical function. Gaze (item 2) and visual fields (item 3) load on the right hemisphere owing to the confounding effects of aphasia (less cooperation with testing) and neglect (more likely to exhibit an apparent visual field cut). Similarly, the facial and sensory items (items 4 and 8, respectively) load on the right hemisphere because aphasia interferes with the testing. Factors 3 and 4 represent left and right hemisphere motor functions, respectively. The factor loadings are remarkably similar to the previous result, although we retained facial weakness—owing to its greater loading—in this analysis and not in the previous one.2 In the present study, dysarthria loaded with factor 1 (left hemisphere cortical function), but it previously loaded with factor 3 (left hemisphere motor function). This finding likely reflects the paucity of lacunar strokes in the present study. Using the remaining half of the data (n = 583), we confirmed that the 2-factor solution was identical to the exploratory analysis and again showed an excellent goodness of fit (data not shown). We then confirmed the 4-factor solution (Table 3).
Infarction volumes (from 30-day computed tomography) were different between left and right brain lesions. The median right brain infarction volume was 56.2 mL (range, 0.1-381.5 mL), and the median left brain infarction volume was 37.8 mL (range, 0.2-255.1 mL) (P<.001). Combining both sides, the correlation coefficient between infarction volume and baseline NIHSS score was 0.37 (P<.001).
Using 1191 patients with very large infarctions, we explored the NIHSS to confirm the underlying scale structure. First, we imposed a factor structure—from a previous work2—on the present data and confirmed that the NIHSS represents the left and right hemispheres independently (goodness of fit = 0.97). This finding validates that the NIHSS measures cerebral function in a new population of very large strokes. Some items are known to serve the user poorly: poor scale items that do not function as intended are said to “load” or correlate poorly with the overall scale. As in a previous work,2 ataxia (item 7) and consciousness (item 1a) loaded poorly, so these 2 items were dropped from further analysis. The facial palsy item (item 4) loaded well, unlike in the previous experiment, and was retained; this probably reflects the greater severity of the strokes examined in the CLASS-I.
Next, a de novo exploratory analysis, using half of the data chosen at random, confirmed the previous results. The 2 hemisphere factors resolved into 2 subfactors, 1 representing cortical function and the other representing motor function, which resulted in better goodness of fit (Table 2 and Table 3). Brainstem deficits do not appear in any factor because such patients were excluded from the CLASS-I.
In the previous analysis of the National Institute of Neurological Disorders and Stroke tPA data,2 the factor solution was robust over time after stroke: using different examiners up to 3 months after stroke, the same factors were found. These repeated confirmations suggest that the 4-factor solution is likely to be found in future study populations. Furthermore, the previous results15 showed that the scale reported true differences between treatment groups. Taken together, these findings confirm the utility of the scale as an outcome measure, its most important function in large clinical trials of putative therapies. That trial included all stroke subtypes, including lacunes. In the present trial, only patients with large, nonlacunar strokes were included, and, thus, the present results are a true, independent, repeated confirmation of the factor structure underlying the NIHSS.
In no study has the ataxia item loaded on (correlated with) any factor in the NIHSS. Furthermore, this item is known to be highly unreliable in many trials, but not in the Trial of ORG 10172 in Acute Stroke Treatment.9,10 Whether ataxia should remain in the NIHSS depends on the perceived utility of the scale for brainstem strokes. The modified NIHSS dropped the ataxia item given the poor reliability and the expectation that few patients with brainstem events would be enrolled in clinical trials.2,16,17 For routine clinical use (eg, in a stroke unit), the ataxia item might be retained to allow for coverage of all possible stroke presentations.4,9
Factor analysis characterizes a large set of independent variables in terms of a few underlying variables, called factors. Such derived factors usually have meaningful clinical interpretations that are best measured using multi-item scales. If a single scale item correlates well with one of the underlying factors, it is said to load on that factor. By studying the factor loadings—interpreted as correlation coefficients—one can determine how well the factors explain the data. Factor analysis uses regression methods; therefore, results can vary in different data sets. It is critical to confirm the factor analysis results using a new data set. Therefore, we sought to confirm the factor analysis derived from the National Institute of Neurological Disorders and Stroke tPA Stroke Study using the data set from the CLASS-I. The CLASS-I included only patients with very large hemispheric strokes and, hence, offered an excellent opportunity to confirm the previous NIHSS factor analysis.
Correspondence: Patrick D. Lyden, MD, Stroke Center (8466), Third Floor, OPC, Suite 3, 200 W Arbor Dr, San Diego, CA 92103-8466.
Accepted for publication: May 13, 2004.
Author Contributions:Study concept and design: Lyden, Claesson, Ashwood, and Lu. Acquisition of data: Lyden, Claesson, and Ashwood. Analysis and interpretation of data: Claesson, Havstad, and Lu. Drafting of the manuscript: Lyden, Havstad, and Lu. Critical revision of the manuscript for important intellectual content: Claesson and Ashwood. Statistical analysis: Claesson, Havstad, and Lu. Administrative, technical, and material support: Ashwood. Study supervision: Lu.
Funding/Support: AstraZeneca LLP, the Research Department of the Department of Veterans Affairs, and grant P50 NS044148 from the National Institute of Neurological Disorders and Stroke.