The Use of Blood Biomarkers to Predict Poor Outcome After Acute Transient Ischemic Attack or Ischemic Stroke
Background and Purpose—The prediction of death or disability (“poor outcome”) after stroke by validated clinical models might be improved by the addition of blood biomarker measurements. We investigated whether such measurements improved the classification of patients into 4 categories of predicted risk of poor outcome: very high, intermediate high, intermediate low, and very low.
Methods—We prospectively recruited symptomatic patients within 24 hours of ischemic cerebrovascular events. We measured clinical prognostic variables in each patient. We drew blood soon after admission and measured markers of inflammation, thrombosis, cardiac strain, and cerebral damage. We assessed poor outcome at 3 months with the modified Rankin Scale and recovery of symptoms at 24 hours. We measured the association between blood marker levels and poor outcome after adjustment for stroke severity and age with multivariate logistic regression. Where these associations were statistically significant, we calculated the net reclassification index.
Results—We recruited 270 patients with acute ischemic cerebrovascular events. At 3 months, 112 patients had a poor outcome. After adjustment for stroke severity and age, only interleukin-6 and N-terminal pro-brain natriuretic peptide were significantly associated with poor outcome. The addition of either interleukin-6 or N-terminal pro-brain natriuretic peptide to National Institutes of Health Stroke Scale and age did not improve the prediction of a poor outcome.
Conclusions—Neither interleukin-6 nor N-terminal pro-brain natriuretic peptide had sufficient predictive power to be of clinical use to predict poor outcome after stroke. The search for better markers to improve the classification of patients across clinically relevant boundaries of predicted probabilities of outcome events needs to continue.
After ischemic stroke, the early prediction of death or disability (“poor outcome”) is of great interest. Statistical models, constructed with clinical variables such as age or neurological impairment, make similar predictions of poor outcome to experienced stroke physicians.1 Biomarkers of the processes active in ischemic stroke might add predictive power to these simple statistical models based on bedside clinical examination.
Many studies have examined the association between blood marker levels and poor outcome after stroke.2,3 However, the associations seen in group data, unless very strong, do not often lead to better predictions of outcome in individuals.4 To be useful, a given marker should at least improve on the predictions from validated prognostic variables.
We hypothesized that the addition of biomarker levels to the National Institutes of Health Stroke Scale (NIHSS) and age5 would improve the prediction of poor outcome (or recovery at 24 hours) in patients with ongoing symptoms due to cerebral ischemia who presented to the hospital soon after the onset. We decided to measure improvement in prediction with a novel statistic, the net reclassification index, which gives a clinically meaningful measure of improvement in prediction.6
We examined biomarkers of pathophysiological processes that were plausibly associated with poor outcome after stroke after a systematic review of the available literature.2
The study was approved by the Scotland A Research Ethics Committee (REC Reference No. 06/MRE00/119). All patients or their relatives provided written informed consent.
Between March 2007 and February 2009 we included consecutive patients with ischemic stroke or transient ischemic attack who presented in normal working hours with symptoms of <24 hours' duration that were still present at the time of initial assessment. Because of laboratory availability, patients were only eligible if they could be recruited within normal working hours. All patients were assessed at the time of presentation when they still had ongoing symptoms of their acute “brain attack.” A study neurologist assessed each patient in the acute stage and recorded prognostic factors for poor outcome after stroke (age, stroke severity, symptoms in the previous week consistent with infection, medical history, atrial fibrillation, NIHSS); medications at admission; electrocardiographic findings; and time when last seen well. We defined prior cognitive impairment as a history of cognitive problems before the onset of symptoms that were sufficient to impair activities of daily living, elicited by the study neurologist.
For each patient, a panel of experts agreed on a final diagnosis of confirmed ischemic stroke, transient ischemic attack, or not a vascular event after considering the presentation, CT or MRI brain imaging, and clinical course blinded to the results of blood marker levels. We defined an ischemic stroke as a clinically definite stroke in a patient whose brain imaging showed either positive evidence of a relevant ischemic lesion or was normal and excluded intracranial hemorrhage and stroke mimics and where the symptoms lasted for >24 hours. We used a similar definition for transient ischemic attack, although symptoms had to last for <24 hours.
We drew blood from each patient before infusion of any thrombolytic therapy into 2 2.5-mL ethylenediaminetetraacetic acid tubes and an 8-mL tube containing clot activator and gel for serum separation. Samples were transferred on water ice and centrifuged at 3000 revolutions per minute for 10 minutes and the supernatant stored at −80°C.
In serum and plasma blinded to clinical information, clinical and research laboratories measured as markers of inflammation: adiponectin, C-reactive protein, intracellular adhesion molecule 1, interleukin-6 (IL-6), IL-10, matrix metalloproteinase 9, tissue necrosis factor α, von Willebrand Factor, and white cell count; thrombosis: d-dimer, fibrinogen, and tissue-type plasminogen activator; cardiac strain: N-terminal pro B-type natriuretic peptide (NT pro-BNP) and troponin T; and neural and glial damage: tau, S100B, and creatinine, and glucose (see Supplementary Material for detailed methods; http://stroke.ahajournals.org).
At 24 hours, we determined whether the presenting symptoms had resolved completely (to differentiate transient ischemic attack from ischemic stroke). We ascertained vital status by contacting the patient's general practitioner at 3 months after onset and sent surviving patients a short questionnaire based on the modified Rankin Scale (mRS) by mail. If a patient failed to return an interpretable questionnaire by mail, we interviewed them in a structured way by telephone to measure the mRS. We dichotomized a patient's outcome into “poor” if he or she was dependent on others for activities of daily living (mRS scores 3, 4, and 5) or dead and “good” if he or she was independent in activities of daily living 3 months after stroke onset (mRS 0, 1, and 2). We determined the cause of death by inspection of the hospital or general practitioner records.
We examined the association between biomarker levels and delay to admission with a series of correlation analyses. We made adjustment for NIHSS at the time of presentation with multiple linear regression analysis.
We examined the association between clinical variables and blood markers with poor outcome at 3 months with a series of univariate logistic regression analyses. We constructed 2 logistic regression models to adjust for potential confounders. First, in a simple model, we adjusted the association between blood markers with poor outcome for the baseline NIHSS and age. Then in a second series of more complex models, we adjusted not only for NIHSS, age, and premorbid disability, but also for the potentially confounding effects of infection and statins on the associations with inflammatory markers and for the potentially confounding effects of current or previous atrial fibrillation, cardiac failure, or previous cardiac vascular disease on the associations with cardiac markers. We measured the crude association of blood markers with recovery at 24 hours and adjusted these associations for NIHSS and age.
The Additional Predictive Value of Blood Markers Over NIHSS and Age
We assessed if the addition of biomarkers significantly associated with poor outcome made better predictions of poor outcome than a model built based on NIHSS and age alone.5 We made the most conservative estimate of the improvement in prediction after the addition of blood markers by allowing the association of each clinical covariate with poor outcome to vary. We assessed changes in goodness of fit (Bayes information criterion), calibration (Hosmer Lemeshow χ2), and discrimination (area under receiver operator characteristic curves) after the addition of biomarkers to the baseline clinical model. We calculated the net reclassification index6 across the clinically relevant thresholds of predicted risk: 0.1 (“very low risk”), 0.5 (“intermediate”), and 0.9 (“very high risk”). We prespecified thresholds of <10% and >90% for predicted probability of poor outcome because we believe that one would need to be very certain of a good or poor outcome before avoiding treatments such as thrombolysis or selecting patients for palliative care only. The net reclassification index examines whether the addition of a biomarker moves those with poor outcome to higher risk categories more often than lower risk categories and those with a good outcome to lower risk categories more often than higher risk categories. Because Bonferroni adjustments for over 5 comparisons are likely to be too conservative,7 we considered a 2-tailed P<0.01 to be statistically significant. We analyzed the data with Stata 11.
We recruited 270 patients who presented with ongoing symptoms due to cerebral ischemia. In 40 patients, the symptoms resolved completely within 24 hours, and in 230, they were persistent (Figure 1). At 3 months, a mRS score was available for 268 (99%) patients and vital status for all patients. At 3 months, 31 (11%) patients had died and 112 (42%) had a poor outcome. The cause of death was the effects of the original stroke (21), cancer (3), extracranial hemorrhage (2), cardiac arrhythmia (1), myocardial infarction (1), ischemic colitis (1), heart failure (1), and was unavailable in 1. Blood marker data were missing for some patients, although each patient had blood drawn. Data appeared to be missing at random. The clinical characteristics of the study participants are summarized in Table 1 and the levels of blood markers are shown in an online Data Supplement.
Clinical Features and Outcome
The risk of poor outcome doubled per decade of patient age (OR, 2.03; 95% CI, 1.58–2.62) and per 3-U increase in the NIHSS (OR, 1.98; 95% CI, 1.63–2.39; Table 1). A patient with poor outcome was significantly (P<0.01) more likely to have atrial fibrillation, prior cognitive impairment, or have been admitted more quickly to the hospital. Treatment with statins before stroke was associated (P=0.02) with poor outcome, but not symptoms of infection at admission (P=0.14). Seven patients were treated with intravenous recombinant tissue-type plasminogen activator.
Blood Markers and Delay to Admission
The median delay to blood draw from symptom onset was 7.1 hours (interquartile range, 3.2–20.1 hours). Although there was a significant correlation (P<0.01) between increasing loge delay to blood taking and lower levels of loge d-dimer (r=−0.21), loge IL-6 (r=−0.18), and loge tissue-type plasminogen activator (r=−0.13), after making adjustment in multivariate linear regression for NIHSS at admission, these associations were no longer statistically significant.
Blood Markers and Poor Outcome
The relationship between blood marker levels and poor outcome was approximately log linear for all markers except NT pro-BNP and S100B in which a loge transformation was log linear (Figure 2). Although higher levels of most of the markers were positively associated with poor outcome, after adjustment for age and the baseline NIHSS, only the associations with IL-6 (75th:25th centile OR, 2.05; 95% CI,1.23–3.41) and NT pro-BNP (OR, 1.26; 95% CI, 1.24–4.12) reached statistical significance.
Further adjustment of the association (Supplemental Table II) between IL-6 with poor outcome for the prescription of statin medication before admission and symptoms of infection before stroke attenuated this relationship by a small amount (OR, 1.86; 95% CI, 1.10–3.18). Adjustment of the relationship between loge NT pro-BNP with poor outcome for a prior diagnosis of cardiac failure, atrial fibrillation at the time of admission and prior cardiac vascular disease, also weakened the association to a small degree (OR, 2.09; 95% CI, 1.11–3.93). For other markers, further adjustment made little material difference to the direction, strength, or statistical significance of the associations.
There were no important differences in the magnitude or statistical significance of the associations between any of the blood markers and poor outcome when the analyses were repeated in the subgroup of 190 patients with positive imaging findings.
There was an association of higher levels of NT-pro BNP with continuing symptoms at 24 hours, which remained after adjustment for NIHSS and age (OR, 2.72; 95% CI, 1.25–5.96). The levels of other markers were not significantly associated with clinical outcome at 24 hours (Supplemental Figure I).
The Addition of Blood Markers to NIHSS and Age
We tested the addition of those markers associated significantly with poor outcome to a baseline model constructed with the covariates NIHSS and age (Table 2). The baseline model was well calibrated and showed good discrimination in this cohort.
The addition of NT pro-BNP or IL-6 to the baseline model significantly improved the goodness of fit but made little additional difference to either the discrimination (measured by area under the receiver operating characteristic curve) or the calibration of the models. Although a few patients moved across clinically relevant boundaries of predicted probability of poor outcome, these numbers were small and not statistically or clinically significant. The full models are summarized in Supplemental Table III.
Higher levels of NT pro-BNP and IL-6 were strongly associated with poor outcome 3 months after stroke or transient ischemic attack after adjustment for age and neurological impairment. Despite this strong association, the addition of either NT pro-BNP or IL-6 to NIHSS and age did not improve the classification of patients to an important degree into “very high risk” or “very low risk” of poor outcome 3 months after symptom onset. These findings were not affected by delay to admission, statin prescription, or the presence of symptoms of prior infection, which was a concern in our previous study of IL-6 and poor outcome.8 It is therefore unlikely that doctors will find predictive instruments that use the measurement of IL-6 or NT pro-BNP clinically more useful than age and NIHSS to predict short-term functional outcome in patients with acute stroke. In addition, only NT pro-BNP was associated with failure to recover by 24 hours after symptom onset, although there are wide limits of uncertainty about the estimates of other markers, because the number of patients with complete recovery at 24 hours was small.
Our finding of the association between NT pro-BNP and poor outcome after cerebrovascular disease is consistent with several published studies.9–11 Plausible explanations for the association are: (1) NT-pro BNP is a measure of early cardiac dysfunction after stroke, which is associated with more severe stroke and worse outcome; (2) NT-pro BNP is associated with cardioembolic strokes, which tend to have a worse outcome that other causes of ischemic stroke; or (3) NT pro-BNP is released from the brain after cerebral ischemia. However, despite the association of BNP with many potential causes of poor outcome after stroke, it does not seem to have a role in clinical prediction of poor outcome.
The significant association of higher levels of adiponectin with poor outcome after adjustment for potentially confounding variables is also of interest, although this did not reach our chosen threshold of statistical significance for this study (P=0.02). In previous studies of the association of adiponectin with death after stroke, lower levels of adiponectin were associated with worse outcome12; our finding suggested that this may not be the case.
To be useful for predicting outcome in clinical practice, markers must either outperform established clinical models when measured alone or improve predictions when added to these models. The net reclassification index6 is a measure of the number of patients who are better classified across clinically meaningful thresholds after the addition of a new biomarker. Because there is no widely agreed definition of clinically important boundaries for the prediction of poor outcome after stroke, in this study, we chose boundaries after discussion with experienced stroke physicians. However, further work is still needed to determine the boundaries that are most important to patients and less experienced physicians.
This study did not aim to elucidate the causal role of particular molecules or the underlying physiological pathways. There are many plausible hypotheses about how higher levels of each marker might cause a poor outcome after cerebral ischemia. However, even in a perfect observational study, free of the influence of selection or information bias, causality is not the only factor that might explain the observed associations. Other explanations include reverse causality, the choice of confounders used for statistical adjustment, and imperfect biomarkers.
We recruited patients with a wide range of severity of baseline neurological impairments and time points of study entry up to 24 hours after symptom onset. A study of a large group of patients with a uniform baseline stroke severity or at a sooner time would have more power to detect important improvements in prediction with blood markers in patients presenting at early time points or with particularly severe (or mild) strokes. However, then the results would be less generalizable.
We measured the mRS outcome by a postal questionnaire and telephone interview, similar to large trials of stroke treatments,13 although we note that there are problems with most methods of mRS measurement.14 It is possible that with this method of outcome measurement, our results were affected by a nondifferential information bias, which would tend to weaken the observed associations.
This study met almost all the methodological criteria for a reliable study of prognosis.15 We recruited 270 patients, of whom 112 had a poor outcome at the end of the study. Because approximately 10 outcome events per covariate has been suggested as a “rule of thumb” to ensure logistic regression models have sufficient power, we could examine approximately 10 covariates.16 Although some patients had missing blood marker levels, these were random so unlikely to have led to an important selection bias. Furthermore, the study was larger than most of the previous studies of blood biomarkers and stroke outcome. By including patients with ‘clinically probable” as well as patients with “clinically definite” stroke, we avoided the selection biases associated with including only patients with imaging positive strokes, and therefore our cohort is more likely to be representative of patients with stroke in routine clinical practice.
We have confirmed the associations between blood markers and poor outcome with effects of a similar magnitude to previous studies. Although these blood markers were therefore of potential clinical value, it is sobering to note that none of the very plausible biomarkers measured in this study added a clinically useful degree of prediction of poor outcome after stroke to that provided by the measurement of the simple clinical variables, the NIHSS and age. Future studies of blood markers for predicting outcome after stroke will need to demonstrate that the candidate marker: (1) has a statistically independent association with outcome; (2) adds statistically significant predictive value to a clinical model; and (3) improves the classification of patients across relevant boundaries of predicted probabilities of poor outcome that have clinical use.
Sources of Funding
W.W. was supported the Chief Scientist's Office (CAF/06/30) and is now funded by a UK Medical Research Council Clinician Scientist Fellowship (G0902303). J.W. was supported by the Scottish Funding Council through the Scottish Imaging Network, a Platform for Scientific Excellence Collaboration. The study was funded by The Chief Scientist's Office and a National Health Service Lothian endowment, Research and Education in Medical Neurological Disorders.
We are indebted to the Wellcome Trust Clinical Research Facility, many research fellows, and the patients. The imaging was conducted in the SFC Brain Imaging Research Centre at the University of Edinburgh, a SINAPSE center.
Ralph L. Sacco, MD, MS, was the Guest Editor for this paper.
The online-only Data Supplement is available at http://stroke.ahajournals.org/lookup/suppl/doi:10.1161/STROKEAHA.111.634089/-/DC1.
- Received July 26, 2011.
- Accepted August 2, 2011.
- © 2012 American Heart Association, Inc.
- Counsell C,
- Dennis M,
- McDowall M
- Whiteley W,
- Chong WL,
- Sengupta A,
- Sandercock P
- Pepe MS,
- Janes H,
- Longton G,
- Leisenring W,
- Newcomb P
- Konig IR,
- Ziegler A,
- Bluhmki E,
- Hacke W,
- Bath PMW,
- Sacco RL,
- et al.
- Bland JM,
- Altman DG
- Makikallio AM,
- Makikallio TH,
- Korpelainen JT,
- Vuolteenaho O,
- Tapanainen JM,
- Ylitalo K,
- et al.
- Sharma JC,
- Ananda K,
- Ross I,
- Hill R,
- Vassallo M
- Efstathiou SP,
- Tsioulos DI,
- Tsiakou AG,
- Gratsias YE,
- Pefanis AV,
- Mountokalakis TD
- Quinn TJ,
- Dawson J,
- Walters MR,
- Lees KR