3D Vector Component Analysis of the Modified National Institutes of Health Neurological Stroke Scale
To the Editor:
The National Institutes of Health Stroke Scale (NIHSS) is the most widely used clinical instrument to measure the consequences of stroke for prognosis, follow-up, and assessment of therapeutic interventions. The neurological impairment associated with stroke reflects global and/or focal changes in brain vascular territories, eliciting syndromes in 3 major domains, including consciousness/orientation, sensorimotor functions, and speech/verbal communication. For a neurological scale to possess high validity, it should reflect changes in the different domains. First, the component items of the scale must represent specific domains without redundancy (ie, content validity), and most importantly the scale must have a high ability to detect meaningful change in clinical status over time (ie, convergent validity, a measure of responsiveness). Recently, Lyden et al1 proposed a modified NIHSS (mNIHSS). Their goal was to increase both the reliability of the scale and its capacity to reflect more meaningful clinical information. However, the approach adopted by the authors could not examine responsiveness with respect to predicting within-individual changes. Recently, we have demonstrated the responsiveness of another scale, the United Form for Neurological Scoring of Hemispheric Stroke With Motor Impairment (UNSS),2 using the 3D vector component analysis to overcome some of the limitations of using cumulative scores, which may represent widely different abnormalities in different domains. This work presents a method to represent the magnitude and direction of change in the various domains of the new mNIHSS in comparison with the original NIHSS with the aim of examining within-individual changes.
We present findings in 18 patients (14 men and 4 women, aged 63 ±1.7 years [mean±SE]), who suffered a stroke, as defined by WHO,3 of the middle cerebral artery and underwent a standardized neurological assessment with use of the NIHSS on admission and at discharge, after a mean (±SE) duration of hospitalization of 23±3.8 days. The scores on the mNIHSS were derived simply by analyzing the 11 items proposed for the mNIHSS from the 15 items of the original NIHSS. The scores of the items of the NIHSS and mNIHSS reflecting global severity (consciousness/orientation) (item 1), sensorimotor functions (items 2 to 8, 11), and speech/verbal communication (items 9 to 10) were summed for each domain and for each patient. For the NIHSS, the mean±SE values for consciousness/orientation on admission and at discharge were 1.78±0.58 and 0.17±0.12, respectively (P=0.008, t test). The sensorimotor functions domain on admission and at discharge was 10.5±1 and 4±0.89, respectively (P=0.0000001). The speech/verbal communication scores on admission and discharge were 1.78±0.39 and 0.72±0.24, respectively (P=0.0008). For the mNIHSS, the mean±SE values for consciousness/orientation on admission and at discharge were 1.39±0.44 and 0.17±0.12, respectively (P=0.008). The sensorimotor functions domain on admission and at discharge was 7.3±0.8 and 3±0.67, respectively (P=0.0000009). The speech/verbal communication scores on admission and discharge were 1±0.25 and 0.34±0.14, respectively (P=0.004).
The individual cases were plotted on a 3D axis scatterplot that shows the relationship between 3 variables with a statistical software package (Statistica, StatSoft, Inc). The backdrop of the 3D axis scatterplot is not a base plane and grid as with the standard 3D scatterplot,2 but a horizontal plane drawn at the middle of the range of the vertical (z) variable with a single pivoting axis through its center. The variables were assigned to the axes such that the variable that is most likely to discriminate between the patterns of relations among the other 2 was specified as z (the vertical axis) and was assigned to the consciousness/orientation domain. Then, by interactively rotating the display, the levels of consciousness/orientation at which the pattern of relationship between x (sensorimotor functions) and y (speech/verbal communication) changes were identified. Points above the horizontal plane reflect the most critically ill patients with severely altered consciousness; those below were less impaired in consciousness/orientation. The 3D axis scatterplot shown in the Figure revealed the relationships between the 3 domains: each point in the plot represents 1 case on admission (filled circles) and at discharge (open circles). Both panels revealed that the strength of the relationship between the domains increased at discharge compared with admission. This could be used to evaluate the effectiveness of therapeutic interventions and comparison of various treatment protocols and procedures. For example, to examine the effect of treatment in this study cohort, the main outcome variable was set at the lower range of the 95% confidence interval of the admission scores in each domain. This range includes all cases with scores less than the lower range of the 95% CI of the admission scores in each domain. In other words, this range of scores defined the beneficial effect of interventions, or “best treatment effect.” The latter must appear distinct by any stroke scale and may provide a measure of the responsiveness common for all stroke scales rather than the use of decrements (in some scales, increments) in points from baseline scores used until now. For the NIHSS, the 95% CIs for consciousness/orientation, sensorimotor functions, and speech/verbal communication were 0.62 to 2.94, 8.5 to 12.5, and 0.99 to 2.56, respectively. There were 14 points (cases) within the range of best treatment effect, of which 11 cases were patients at discharge and 3 at admission. In other words, 11 cases were true treatment effect or true positives, while 3 cases were false-positives. The “cluster zone of best treatment effect” for the NIHSS can be seen in the circled area in panel A. For the mNIHSS, the 95% CIs for consciousness/orientation, sensorimotor functions, and speech/verbal communication were 0.52 to 2.26, 5.6 to 8.94, and 0.49 to 1.51, respectively. There were 11 points (cases), of which 10 were patients at discharge and 1 a case of mild partial stroke at admission. The cluster zone of best treatment effect for the mNIHSS could be seen in the circled area in panel B. Therefore, the predictive value can be calculated by assessing the proportion of the total of positive cases that are truly treatment related. The predictive value of the NIHSS was 11/14, or 78.6%. This means that the likelihood that a patient detected as having a best treatment effect by the NIHSS was a true positive effect of treatment was 78.6%. Similarly, the predictive value of the mNIHSS was 10/11, or 91%. This means that the likelihood that a patient detected as having a best treatment effect by the mNIHSS was a true positive effect of treatment was 91% in this cohort population of patients.
The NIHSS predicted that 11 cases (11/18, or 61.1%) benefited significantly from the treatment at discharge. However, 1 patient included in this cluster zone lies at the border of the cluster and thus had significant motor impairment. Conversely, mNIHSS predicted that only 10 cases (10/18, or 55.5%) benefited significantly from the treatment at discharge, with minimal impairment in all domains. This may suggest that the NIHSS may overestimate a putative drug effect, mainly due to its wide 95% confidence range, thus confirming the findings by Lyden et al1 that mNIHSS, because of its narrower confidence range, has improved power to detect treatment effect. Both the responsiveness and the predictive validity of the mNIHSS compared with the NIHSS were improved. It is obvious that the greater the proportion of cases in the cluster zone of the mNIHSS to the overall study cohort the more effective was the treatment regimen under investigation. The latter may provide a visually demonstrable index of drug efficacy in stroke clinical trials.
A few pitfalls of the present approach were the use of data from the NIHSS rather than prospective data collected directly from the mNIHSS. This may imply bias, as discussed by Lyden et al.1 The other pitfall is rather a technical problem: the overlap of multiple data points may create a potentially time-consuming effort when a large number of cases are analyzed. Software improvements of the 3D scatterplot module in the statistical package that could automatically track and separate overlapping points by a user-selected tolerance limit are required to enable the widespread use of this approach in major stroke clinical trials. In conclusion, the NIHSS and mNIHSS were shown as constructs with several dimensions. However, the mNIHSS has improved responsiveness. The mNIHSS is a welcome improvement, and if used in combination with the 3D vector component analysis may provide ready detection of drug efficacy in future stroke clinical trials.
Lyden PD, Lu M, Levine SR, Brott TG, Broderick J. A modified National Institutes of Health Stroke Scale for use in stroke. Stroke. 2001; 32: 1310–1317.
Njemanze PC, Chidi-Ebere A. Three-dimensional vector component analysis of the neurological stroke scales. Stroke. 1999; 30: 1731–1732.Letter.
World Health Organization. Proposal for the Multinational Monitoring of Trends and Determinants in Cardiovascular Disease (MONICA Project). Geneva, Switzerland: World Health Organization; 1983. WHO document MNC/82.1, Rev 1.