Day-90 Acute Ischemic Stroke Outcomes Can Be Derived from Early Functional Activity LevelOvbiagele B. · Saver J.L.
Stroke Center and Department of Neurology, UCLA Medical Center, and Department of Neurology, Olive View – UCLA Medical Center, Los Angeles, Calif., USA Corresponding Author
Background: The time point generally recognized as most appropriate to assess final functional outcome after acute ischemic stroke is 3 months. However, identifying earlier reliable prognostic time points may allay patient anxiety about the recovery process, enable assignment of outcome in patients lost to follow-up, and provide earlier and more readily available options for clinical outcome assessment in adaptive design and proof of concept studies. We assessed whether day-7/10 functional outcome predicted day-90 functional outcome among acute ischemic stroke patients. Methods: The NINDS-tPA Study database was analyzed. Global disability was assessed using the modified Rankin Scale (mRS). Spearman correlation evaluated the association of day-7/10 versus day-90 mRS, and observed agreement was computed using the weighted κ agreement, both unadjusted and adjusted by multivariable ordinal logistic modeling for demographic and clinical variables known to influence stroke outcomes. Results: 581 subjects (93%) were alive at 7–10 days. There was strong correlation between mRS score at day 7/10 and day 90 (r = 0.81, p < 0.0001), and the weighted κ agreement was 0.82 (p < 0.0001). In multivariable analysis, day-7/10 day mRS was independently and strongly associated with day-90 mRS, while among other baseline variables, only baseline NIH Stroke Scale score (per unit increase) and a history of congestive heart failure (CHF) were significantly associated with a worse day-90 mRS. Conclusions: Global disability status 1 week after an index ischemic stroke strongly predicts final 3-month disability outcome. Functioning at 1 week, supplemented by initial stroke severity and CHF history information, may provide an early outcome guide useful for patient and family counseling.
© 2009 S. Karger AG, Basel
Compromise of vocational and self-care capacity (disability) after stroke occurrence is of major clinical relevance to stroke survivors and their caregivers, and residual disability after stroke exerts a substantial economic toll on society . The time course of stroke recovery must be considered when selecting the optimal time point for disability assessment for public policy, clinical trials, and clinical practice . Neural repair and functional recovery evolve over weeks and months, making early assessments liable to underestimate final outcome . However, competing causes of morbidity and mortality are more likely to arise as more time passes, interfering with direct assessment of the effects of the index stroke. Three months after ischemic stroke is the time point generally recognized as best balancing these contending forces to provide the most informative assessment of final functional state, and is the time frame most frequently chosen for primary outcome measurement in acute stroke trials [2,3,4].
Nonetheless, there is a need to understand the relationship between early assessments of outcome and final 3-month functional state. Patients and their caregivers are understandably worried early about the extent of recovery to be expected from the stroke. Following discharge from the acute care hospital, patients in a mobile society and fragmented medical system may be lost to follow-up, complicating efforts to assess the efficacy of new treatments in clinical trials and the effectiveness of established care in quality improvement programs.
Few studies have looked at how well functional outcomes around 7–10 days approximate the standard 3-month assessment. Being able to predict 3-month outcomes 1 week after the index stroke may alleviate patient/caregiver anxiety about the recovery process and facilitate home care planning, yield reliable early endpoints for dose exploration in adaptive-design dose optimization clinical trials, and enable reliable estimation of missing final outcome data among lost to follow-up patients in pivotal clinical trials.
The modified Rankin Scale (mRS) is a clinician-reported measure of global disability, and is a simple and time-efficient outcome index frequently used in large-scale multicenter trials [5,6,7]. The mRS defines 7 clinically discrete patient disability categories, including 6 levels of disability and 1 for death [5,6,7]. Several types of evidence attest to the validity and reliability of the mRS [8,9,10]. Our study objectives were to assess if and how well day-7/10 mRS functional outcome alone predicts the final day-90 mRS score, and to derive a formula that predicts day-90 mRS using a patient’s known day-7/10 mRS and additional baseline demographic/clinical variables.
Data from Trial 1 and Trial 2 of the NINDS-tPA Study were analyzed . The intermediate endpoint assessed was mRS at day 7/10, whereas mRS at day 90 was the final endpoint. Baseline demographic and clinical covariates to be examined were preselected based on prior studies of factors that influence outcomes after acute ischemic stroke [12,13,14,15,16,17,18,19,20,21]. Candidate predictive factors handled as continuous variables were age, admission serum glucose, admission temperature, admission systolic blood pressure, baseline National Institutes of Health Stroke Scale Score (NIHSSS), and time to thrombolytic treatment. Candidate predictive factors handled as dichotomized variables were: gender, race-ethnicity, presumed stroke subtype (small vessel, cardioembolic, large vessel, other), and history of the following: stroke or transient ischemic attack, hypertension, diabetes, myocardial infarction, congestive heart failure (CHF), smoking and premorbid use of aspirin.
Time to tPA was calculated by summing the time from onset of symptoms to admission and the time from admission to treatment. The time predictor was defined as the inverse of the time to tPA in the case of tPA patients. Therefore, for patients who had a short time to tPA, the inverse time variable would be large relative to patients who had a long time to tPA. In the case of placebo patients, the time predictor was assigned the value of 0 since time to tPA for these patients is infinity, and therefore the inverse of the time variable approaches 0.
We set out to derive a formula to predict how a patient with a given day-7/10 mRS level of 0, 1, 2, 3, 4, or 5 would fare on day 90. Those with day-7/10 mRS of 6 (those who were dead) were excluded from the analysis, since by definition they would still be dead at 90 days. For the bivariate analysis, the percent of 90-day mRS correctly predicted (accuracy) and the observed agreement were initially computed by cross-tabulating the day-7/10 mRS versus the day-90 mRS. The observed agreement is the proportion of the subjects who had exactly the same score at 7–10 days and 90 days. However, it does not take into account that the outcomes are ordinal; hence, a day-7/10 score that is only 1 point different from the day-90 score is better than a day-7–10 score that is more than 1 point different. As such, we also computed the Spearman correlation and the weighted ĸ agreement. The ĸ and correlation statistics correctly take the ordinality of the outcomes into account as measures of association. Furthermore, the weighted ĸ statistic measures the agreement after removing chance agreement, and takes into account that the mRS scores are ordinal by giving partial credit when the 2 mRS scores disagree slightly.
We determined that most of the covariates had no or very few missing values (<1%). There were only 4 covariates for which the amount of missing values was >1%. These were history of CHF (4.5% missing), history of myocardial infarction (4.6% missing), history of stroke or TIA (4.3% missing), and baseline temperature (7% missing). Among these 4 covariates, we found that only history of CHF was not missing at random as patients with a missing value for this covariate were significantly more likely to have a poor outcome after controlling for the other covariates. For the other 3 covariates, the missingness was unrelated to the outcome controlling for the other covariates, and hence they were viewed as missing at random. Since all but 1 of the covariates had either no or extremely few missing data, or were missing at random, imputation was chosen as the optimal solution for handling the missing data. Although history of CHF was not missing at random, there were only 26 patients (4.5%) with a missing value for this covariate, and hence any resulting bias was very unlikely to be considerable. Missing values were singly imputed using nearest neighbor (hot deck) imputation.
For the multivariable analysis, the same aforementioned statistics utilized in the bivariate analyses were computed under an ordinal logistic regression model evaluating the association of day-7/10 versus day-90 mRS controlling for all of the covariates prespecified here. We tested for interaction effects between treatment group versus day-7/10 mRS score and, upon finding no significant interaction effects, went ahead to combine the placebo and tPA groups in the analysis.
Of 624 subjects in the dataset, 581 (93%) were left for analysis after excluding 38 patients who were dead at 7–10 days, and an additional 5 patients who had missing values for day-7/10 mRS. The unadjusted analysis of the effect of rtPA (vs. placebo) on day-7/10 global disability in 621 subjects (99.5%) with available data is shown in table 1, and indicated a significantly higher incidence of good outcome in those who received tPA.
|Table 1. Day-7/10 mRS scores (tPA vs. placebo) among subjects in the NIH Stroke Trial (n = 619)|
Table 2 shows the relationship between actual versus predicted day-90 mRS using day-7/10 mRS as the predicted value. The cross-tabulation of day-7/10 mRS (predicted value) versus day-90 mRS (actual value) among 581 subjects showed that the average accuracy was 44% (55/90 + 46/125 + 11/59 + 17/84 + 51/103 + 91/120) while the observed agreement was 47% [(55 + 46 + 11 + 17 + 51 + 91)/581], which were both relatively low, although these statistics do not take into account that the mRS score is ordinal. When the most appropriate statistic for ordinal outcomes (weighted ĸ) was applied, the agreement was substantial: 0.793, standard error (SE) = 0.018, p < 0.0001. Furthermore, there was a strong positive correlation between the mRS score at days 7/10 versus day 90 (Spearman r = 0.82, p < 0.0001).
|Table 2. Relationship between actual versus predicted day-90 mRS using day-7/10 mRS as the predicted value (n = 581)|
After imputing the missing values, table 3 displays the association between actual versus predicted day-90 mRS using day-7/10 mRS as the predictor and adjusting for the covariates. The cross-tabulation of day-90 mRS (predicted value) versus day-90 mRS (actual value) among 581 subjects showed that the average accuracy was 48% (59/90 + 86/125 + 0/59 + 37/84 + 50/103 + 73/120) while the observed agreement was 50% [(59 + 86 + 0 + 37 + 50 + 73)/581], which were both relatively low, although these statistics do not take into account that the mRS score is ordinal. When the most appropriate statistic for ordinal outcomes (weighted ĸ) was applied, the agreement was substantial: 0.83, SE = 0.02, p < 0.0001. Furthermore, there was a strong positive correlation between the predicted day-90 mRS score versus actual day-90 mRS (Spearman r = 0.83, p < 0.0001).
|Table 3. Relationship between actual versus predicted day-90 mRS (using day-7/10 mRS as the predictor) and adjusting for covariates (n = 581)|
Table 4 shows the results of the multivariable ordinal logistic modeling (again after imputing missing values). The goodness of fit analysis indicated that this model fitted the data adequately (χ2 = 6.2, d.f. = 9, p = 0.720). Day-7/10 mRS was strongly associated with day-90 mRS after controlling for all the covariates. For example, the cumulative odds ratio of having a worse day-90 outcome was increased by a factor of 6.74 in those who had a day-7/10 mRS of 1 relative to day-7/10 mRS of 0. The odds ratios for the effect of 7/10 mRS on day-90 mRS became progressively larger with an increasing day-7/10 mRS, implying that increases in the day-7/10 mRS led to increasingly worse outcomes. Expected distribution of day-90 functional activity after acute ischemic stroke by level of mRS score at days 7–10 based on the multivariable model can be seen in table 5.
|Table 4. Ordinal logistic multivariable model for predicting day-90 mRS score among subjects who were alive at 7–10 days and who had no missing covariates (n = 470)|
|Table 5. Distribution of day-90 functional activity after acute ischemic stroke by level of mRS score at days 7–10|
Based on the multivariable analysis among the 581 subjects, the weighted ĸ statistic was 0.82 (SE = 0.019, p < 0.0001), and the Spearman correlation was 0.82 (p < 0.0001), both of which closely corresponded to the bivariate results, implying that the addition of all the predictors to the model did not improve the prediction by much. Correspondingly, the p values for the covariates were mostly not statistically significant (table 4) except for baseline NIHSS (per unit increase) and having a history of CHF, both of which were significantly associated with an increased cumulative odds of having a worse 90-day outcome.
Although it is generally agreed that outcome measurement within 2 weeks of acute ischemic stroke may be too early in the recovery process to detect maximal recovery , our analysis of the NINDS-tPA stroke trial suggests that the day-7/10 mRS can serve as a good proxy for day-90 mRS, when the latter time point cannot be practically obtained or when early outcome information is needed to drive adaptive clinical trial design algorithms.
The day-7/10 mRS and day-90 mRS in the NINDS-TPA dataset were strongly correlated in this patient cohort. Moreover, in the multivariable model predicting day-90 functional activity after ischemic stroke, there was a strong independent association of mRS level with day-90 outcome, and apart from initial NIHSSS and history of CHF, the addition of 13 established demographic and clinical prognosticators to the model including r-TPA treatment did not significantly improve the prediction of day-90 mRS. These findings would suggest that the effect of these known baseline predictor variables are already reflected in day-7/10 day mRS, which in turn affects day-90 mRS. Indeed, the percent difference in treatment effect (TPA vs. placebo) on good outcome at 7–10 days (mRS <2) was 15 points comparable to the 13 percent point difference between treatment arms on good outcome at 90 days .
There are several potential practical and research implications of our results. In regular clinical practice, at 7–10 days after stroke onset when the acute course has generally stabilized, patients, family and caregivers have a natural human and a practical planning interest in understanding what the final degree of disability will be 3 months hence. Our data demonstrate that day-7/10 mRS, initial NIHSSS values and CHF history provide clinicians with information that can be given to the recent stroke survivor and their caregivers as roughly approximating future disability. In this study, we derived a formula that can be readily applied to a stroke patient using these values to predict the day-90 functional activity status.
In the research arena, the close correlation between global disability at days 7–10 and at day 90 after acute ischemic stroke may permit the early outcome assessment to serve as a useful secondary endpoint in future stroke clinical trials to further confirm or disconfirm the efficacy of a given study drug, especially in small early stage trials to minimize challenges of attrition and in trials exploring candidate dose regimens using adaptive designs that require early feedback of treatment results into the selection of next tested dose tiers . Furthermore, even in large trials that use day-90 assessment as the primary endpoint, our data provide a useful method for dealing with the infrequent but not rare occurrence of patients being lost to follow-up after discharge from the acute care setting. The close correlation of day-7/10 and day-90 values supports the frequent practice in acute stroke clinical trials of using the technique of last observation carried forward to fill in missing mRS outcome data, and the predictive formula we derived provides an even more precise method for imputing missing final outcome data.
This study has some limitations. It was a secondary analysis of completed randomized trials, and so despite controlling for established stroke prognosticators, we cannot exclude the possibility that unmeasured confounding may explain some of our findings. Furthermore, for the multivariable analysis not all the data were complete at 7–10 days, although we handled this issue by the use of imputation. Also, it has been suggested that there may be interobserver variability in mRS assessment across regions ; however, the NINDS-tPA trials were conducted within 1 country with relatively few study sites. The NINDS-tPA trials were extremely well executed, but their sample sizes were relatively modest, and so it would be helpful to derive formulas for predicting final stroke recovery based on early assessment from much larger and more heterogeneous datasets. Finally, although the agreement and correlation indicated a very high association between the day-7/10 and day-90 mRS, it was not 100%, and this will need to be noted when presenting prognostic information to the patient or family mem- ber(s).
An early and reliable prognosis for final global disability in stroke patients is important for patient and family counseling and for outcome assessment in randomized clinical trials. We found that global disability assessed at 7–10 days after acute ischemic stroke may serve as a good proxy for final disability status, thereby providing an opportunity for earlier reliable prognostication among acute stroke patients and subjects enrolled in acute stroke clinical trials. Our results need to be confirmed independently.
The authors are grateful to Daniela Markovic, MS, and Jeffrey Gornbein, PhD, for expert statistical consultation and to the NINDS-tPA Study Investigators.nbein, PhD, for expert statistical consultation and to the NINDS-tPA Study Investigators.
Bruce Ovbiagele, MD, MS
Stroke Center and Department of Neurology, University of California at Los Angeles
710 Westwood Plaza
Los Angeles, CA 90095 (USA)
Tel. +1 310 794 6379, Fax +1 310 267 2063, E Mail email@example.com
Received: March 11, 2009
Accepted: August 12, 2009
Published online: November 5, 2009
Number of Print Pages : 7
Number of Figures : 0, Number of Tables : 5, Number of References : 23
Vol. 29, No. 1, Year 2010 (Cover Date: December 2009)
Journal Editor: Hennerici M.G. (Mannheim)
ISSN: 1015-9770 (Print), eISSN: 1421-9786 (Online)
For additional information: http://www.karger.com/CED