Derivation and Validation of a Novel Prognostic Scale (Modified–Stroke Subtype, Oxfordshire Community Stroke Project Classification, Age, and Prestroke Modified Rankin) to Predict Early Mortality in Acute Stroke

There are a growing number of prognostic models and scales designed to predict mortality or other outcomes after acute stroke.^1–8 Although many of the scales have favorable properties reported, few have been incorporated into routine clinical practice. Two common limitations in prognosis research are failure to modify scales with new covariates that may improve the performance and lack of external validation of scales in independent cohorts.⁹

The stroke subtype [ischemic or hemorrhagic], Oxfordshire Community Stroke Project classification, age, and prestroke modified Rankin (SOAR) score was recently proposed for early mortality prediction after an acute stroke.^10,11 The score may potentially be clinically useful because of its simplicity, with uncomplicated scoring rules and use of readily available data. Although the prognostic accuracy of SOAR score in the original description was reasonable,^10,11 there was scope to improve further. The major limitation of SOAR is the lack of consideration of acute stroke severity marker. The best predictive values possible are needed to avoid giving patients and family wrong and potentially damaging prognostic information. An initial acute stroke severity score, National Institutes of Health Stroke Scale (NIHSS) on admission, is well established as an independent predictor of outcome post stroke.^12–14 Furthermore, the usefulness of SOAR beyond inpatient stay is not examined previously.

In this study, we aimed to evaluate whether including a measure of initial stroke severity (NIHSS) to form a modified-SOAR (mSOAR) score would improve the prognostic accuracy and also assessed that this would be applicable ≤90 days post stroke in an independent data set.

Our findings suggest that the addition of an initial stroke severity scale (NIHSS) may significantly improve prognostic accuracy of the SOAR score for predicting early mortality post stroke. We think that our modified mSOAR score offers prognostic utility while remaining relatively easy to score. A criticism of previous studies that have suggested favorable properties of prognostic tools has been the lack of replication. We performed validation analyses using an independent and geographically distinct population. Prognostic utility of mSOAR was confirmed in this cohort.

Although several stroke prognosis scales are described, few have been adopted for widespread clinical use. An example of a prognostic tool that has translated into practice is the ABCD₂ (age, blood pressure, clinical features and duration of TIA, diabetes) risk score for transient ischemic attack.²² We think that mSOAR shares several features with ABCD₂ that should make it attractive to clinical teams. The input covariates of mSOAR are easily available; in fact, most of the features needed to score mSOAR are recorded as standard in national stroke audits. We recognize that premorbid mRS is an imperfect measure, but it remains the most common method of describing prestroke functioning.²³ Furthermore, we and others have previously shown that prestroke mRS is an independent predictor of stroke outcomes, such as mortality and length of stay.^24,25 Creating a total mSOAR score is also straightforward with only 5 variables requiring scoring and no need of external software or complicated arithmetic. Derivation and validation of mSOAR score were based on a heterogeneous population, and so the score should be applicable to all stroke syndromes.

Our study had a specific hypothesis around adding a marker of initial stroke severity to an existing prognostic tool. We recognized that stroke impairment measures, such as NIHSS, have prognostic value as a standalone assessment, and so intuitively adding these data to SOAR should improve properties. Our approach could be applied to any of the other stand alone or multivariate prognostic tools available in acute stroke. Indeed, future prognostic research in stroke requires further examination of which are the best predictors of various relevant outcomes, including mortality and functional outcome, both as standalone and as part of multivariable predictive tools. Perhaps, inclusion of other covariates especially more sophisticated or clinically relevant variables would improve the prognostic value. For example, the addition of neuroimaging finding to the ABCD₂ score²² (which resulted in ABCD₃-I score²⁶) improves the risk stratification after transient ischemic attack in secondary care settings.²⁶ Another example is the widely used and well-validated Acute Stroke Registry and Analysis of Lausanne (ASTRAL) score, which includes time delay from stroke onset to admission, specific clinical signs at presentation, and baseline glucose level.^8,27 It would also be informative to compare clinical utility of mSOAR against intuition of an experienced clinician. Future prognostic research perhaps should focus on using large composite datasets with comprehensive baseline data that would allow for such an analysis and help describe an optimal predictive covariate set for relevant outcomes.

We describe an optimal performing cut point for our scale; this is the point of most equitable trade-off between false positives and false negatives. In practice, the mSOAR score that has the greatest utility will vary according to the purpose for which the score is used. For example, if clinicians wish to use the prognostic information to inform discussions around ceilings of care, they may prefer a cut point on the scale that minimizes false positives.

A difficulty with mSOAR and indeed all stroke prognostic scales is how the clinician uses the tool. If decisions on pursuing or withholding treatment are to be based on a risk estimate then that estimate needs to be robust. Although mSOAR has favorable properties, that are comparable or better than many other prognostic tools, it is probably still not suitable as the sole basis for therapeutic decisions. Even at the optimal performing cut point, mSOAR will suggest early mortality for a substantial proportion who go on to survive past 90 days and equally will suggest survival for a proportion who have early mortality. Nonetheless, patients and families want early prognostic information, and clinicians have to make therapeutic decisions based on likely prognosis each day. We think that mSOAR offers some structure and evidence base to inform prognostic assessment.

We agree that the NRI seems high. The NRI works best for dichotomous data.^20,21 The majority of the reclassifications in our data were within the low-risk categories; hence, the practical importance of these reclassifications will be limited. Thus, we think that the NRI should be interpreted with caution and prefer the ROC as a better measure of the difference between the risk scores.

Strengths of our analysis include the large sample size, large number of outcomes, and real-world populations. We think that our results will have greater external validity than prognostic scores derived and validated exclusively using selected clinical trial participants. Furthermore, for the first time, we found that SOAR and mSOAR are useful ≤90 days post stroke mortality prediction. Despite the observable differences between the derivation and validation cohorts in terms of age and prestroke mRS and the follow-up time points (inhospital versus within 90 days), the mSOAR score performed similarly and consistently, which may indicate its generalizability.

A limitation of our derivation data set was substantial missing NIHSS data. Our analysis comparing those with and without NIHSS suggests that missing NIHSS data were more likely to be associated with markers of poor outcome (age and prestroke mRS) and higher in-hospital mortality, approximately doubled the mortality risk. The data we used for our derivation work were from a clinical registry. At the time of data collection, NIHSS assessment and recording in the registry were not mandatory. We acknowledge that those with NIHSS recorded may be systematically different from those without; however, the derivation cohort still has a spread of NIHSS and outcomes that are in keeping with data from other registries and so seem to have external validity. The effect of missing NIHSS could potentially bias the prognostic properties of mSOAR, but we are reassured that in our validation cohort (with good data capture for NIHSS), the properties of mSOAR were similar to the derivation data. We also hope that the effect of any misreporting of key variables will be modest given our relatively large and similar size data sets for derivation and validation and the internal and external quality control used within such national registries. The final cohorts are all old, and changes in stroke care may have affected early mortality and hence the properties of the tool. Of note, the derivation data were based on in-hospital mortality, and we were unable to ascertain their vital status at 90 -days. Nevertheless, examination of ASCNES data (drawn from the same ASHCN database) with 1-year follow-up data demonstrated that 81% of within 90 days were in-hospital deaths.

In conclusion, our results demonstrate the reliability of modified-SOAR score to predict early death in an acute stroke cohort. Adding NIHSS data to the original SOAR score, to create modified-SOAR score, improved the prognostic utility in both derivation and validation data sets. Modified SOAR may potentially help clinicians better predict early stroke mortality.

We thank all who were involved in the Anglia Stroke and Heart Clinical Network data. Dr Myint supervised the project. Drs Abdul-Rahim, Alder, and Clark conducted the analyses. Drs Abdul-Rahim and Quinn drafted the initial article. Drs Abdul-Rahim, Quinn, Clark, and Myint involved in reviewing and reporting of the work. All authors provided critical revision of the article for important intellectual content and approved the final version.