Abstract
Objective: To develop a simple prognostic model to predict outcome at 1 month after acute basilar artery occlusion (BAO) with readily available predictors.
Methods: The Basilar Artery International Cooperation Study (BASICS) is a prospective, observational, international registry of consecutive patients who presented with an acute symptomatic and radiologically confirmed BAO. We considered predictors available at hospital admission in multivariable logistic regression models to predict poor outcome (modified Rankin Scale [mRS] score 4–5 or death) at 1 month. We used receiver operator characteristic curves to assess the discriminatory performance of the models.
Results: Of the 619 patients, 429 (69%) had a poor outcome at 1 month: 74 (12%) had a mRS score of 4, 115 (19%) had a mRS score of 5, and 240 (39%) had died. The main predictors of poor outcome were older age, absence of hyperlipidemia, presence of prodromal minor stroke, higher NIH Stroke Scale (NIHSS) score, and longer time to treatment. A prognostic model that combined demographic data and stroke risk factors had an area under the receiver operating characteristic curve (AUC) of 0.64. This performance improved by including findings from the neurologic examination (AUC 0.79) and CT imaging (AUC 0.80). A risk chart showed predictions of poor outcome at 1 month varying from 25 to 96%.
Conclusion: Poor outcome after BAO can be reliably predicted by a simple model that includes older age, absence of hyperlipidemia, presence of prodromal minor stroke, higher NIHSS score, and longer time to treatment.
GLOSSARY
- AUC=
- area under the receiver operating characteristic curve;
- BAO=
- basilar artery occlusion;
- BASICS=
- Basilar Artery International Cooperation Study;
- mRS=
- modified Rankin Scale;
- NIHSS=
- NIH Stroke Scale
Posterior circulation stroke accounts for about 20% of all ischemic strokes. Unlike the anterior circulation, the posterior circulation depends on one main artery. The basilar artery supplies most of the brainstem and the occipital lobes and part of the cerebellum and thalami. The clinical presentation of basilar artery occlusion (BAO) is highly variable, ranging from TIAs or minor stroke to rapidly progressive brainstem dysfunction or coma at onset. Despite recent advances in the treatment of acute stroke, the rate of death or disability is almost 80%.1,–,3
Predicting outcome after BAO may be helpful to determine which therapeutic interventions should be given and to inform patients and their families about prognosis, but accurate prognostic models with admission data are not available. We describe a basic model that includes easily accessible clinical variables and additional models that include findings from the neurologic examination (NIH Stroke Scale [NIHSS]) and from CT imaging.
METHODS
Study population.
The Basilar Artery International Cooperation Study (BASICS) is a prospective, observational, international registry of consecutive patients aged 18 years or older who presented with clinical features of an acute symptomatic and radiologically confirmed BAO. Detailed characteristics of this registry have been described elsewhere.4,5 In summary, 619 patients with BAO from 48 centers in Europe (41), South America (3), North America (2), Australia (1), and the Middle East (1) were included in the registry from November 2002 to October 2007. Patients were eligible for entry in the registry if they presented with symptoms or signs attributable to disruption of the posterior circulation and had a BAO confirmed by CT angiography, magnetic resonance angiography, or conventional contrast angiography. BAO was defined as complete obstruction of flow in the proximal, middle, or distal portion of the basilar artery. The choice of treatment was left to the discretion of the treating physician.
Standard protocol approvals, registrations, and patient consents.
The BASICS protocol was approved by the ethics committee of the University Medical Center Utrecht (Utrecht, the Netherlands). The requirement for additional local ethics approval differed between participating countries and was obtained if required. Verbal or written informed consent was obtained from the patient or patient's representative, as required by national and local guidelines.
Data collection.
Detailed data were recorded in a Web-based data entry form that included information on demographics, stroke risk factors, stroke severity assessed by the NIHSS score, estimated time of BAO, and CT imaging findings before treatment. Slightly hypodense lesions with vague borders were classified as early ischemic changes, and lesions that were clearly hypodense and with sharp borders were classified as old infarctions. Estimated time of BAO was the time of onset of symptoms consistent with the clinical diagnosis of an acute BAO, as described by the patient or witness, or, if unknown, the time that the patient was last seen without such symptoms. TIA or minor stroke in the hours or days before the index event was not counted as the time of the occlusion but was recorded under the prodromal phase. For example, the estimated time of occlusion for a patient who was admitted with an acute minor cerebellar stroke but who developed a severe deficit the next day that was consistent with the clinical diagnosis of an acute BAO was recorded as the time of the onset of the severe deficit.
Outcome.
The primary outcome measure was poor outcome at 1 month. In view of the high risk of death and disability in patients with BAO, poor outcome was defined as a modified Rankin Scale (mRS) score of 4 or 5 (severe disability) or death.6 The mRS score at 1 month was assessed in person during admission or as an outpatient or through a telephone interview with the patient or caregiver. All patients had complete follow-up data.
Model development.
We considered predictors that could be determined easily and readily within the first few hours after BAO. For consecutive models, we presented groups of variables for inclusion according to the order in which information generally becomes available in clinical practice. Three prognostic models were defined: model 1 was based on demographics and stroke risk factors; model 2 had information on neurologic examination (NIHSS) added; and model 3 had imaging data from noncontrast CT and CT angiography added. Restricted cubic spline functions and graphs were used to determine whether continuous variables could be analyzed as linear terms or required transformation.7,8
Missing values of patient characteristics were imputed as determined by the correlation between patient characteristics with missing values with the other variables by means of single regression imputation.9 Logistic regression analysis was performed with poor outcome as the outcome variable. Candidate predictors were considered for entrance into multivariable regression models irrespective of their univariable association with poor outcome.8 All candidate predictors were included in a multivariable logistic regression model and were step by step excluded if the likelihood ratio test had p > 0.15. Interaction terms between predictors were examined with likelihood ratio tests, but none was of sufficient relevance to extend the models beyond the main effects for each predictor.
Model performance.
We evaluated both calibration and discrimination of the 3 models.10 The discriminative performance, i.e., the extent to which the prognostic models enable discrimination between patients with and without poor outcome, was described by area under the receiver operating characteristic curve (AUC). The AUC varies between 0.5 (a noninformative model) and 1.0 (a perfect model). The predictive accuracy of the prognostic models, i.e., the agreement of observed outcomes with predicted risk, was assessed by the Hosmer-Lemeshow test and graphically with a calibration plot.
Model validation.
Prognostic models derived from multivariable regression analysis are known to overestimate regression coefficients. This results in too extreme predictions when applied in new patients.7,10 Therefore, we internally validated our model with bootstrapping techniques, where in each bootstrap sample the entire modeling process was repeated.10 This resulted in a shrinkage factor for the regression coefficients.7 The bootstrap procedure was also used to estimate the AUC corrected for overoptimism. The corrected AUC may be considered an estimate of discriminative ability expected in future similar patients.
Application in clinical practice.
Based on the independent predictors from the model with the highest discriminatory value, a risk chart was developed to display the risks for poor outcome in patients with or without these predictors. Data were analyzed with SPSS for Windows (version 15.0; SPSS Inc., Chicago, IL) and R (version 2.10.1; http://www.r-project.org) with help of the libraries Hmisc and Design of Harrell.11
RESULTS
Study population.
Admission characteristics of the 619 patients are presented in table 1. Mean age at admission was 64 years (range, 19–95 years). Of the patients, 429 (69%) had a poor outcome at 1 month: 74 (12%) had a mRS score of 4, 115 (19%) had a mRS score of 5, and 240 (39%) had died.
Baseline characteristics of the 619 patients with acute basilar artery occlusion
Prognostic models.
The multivariable models for prediction of poor outcome at 1 month after BAO are presented in table 2. Predictors of poor outcome were older age, absence of hyperlipidemia, presence of prodromal minor stroke, higher NIHSS score, longer time to treatment, proximal and middle vs distal location of the BAO, and presence of old posterior circulation stroke or early ischemic changes on CT. A poor outcome occurred in particular for those with higher NIHSS scores and older age. Type of treatment and CT imaging findings had no additive contribution to the prediction of poor outcome.
Multivariable predictors of poor outcome at 1 month after basilar artery occlusion in a series of 3 prognostic models of increasing complexity based on admission characteristicsa
Model performance.
The discriminatory ability of the models became larger with increasing complexity. The AUC corrected for overoptimism was 0.64 for the model restricted to demographics and stroke risk factors only, 0.79 for that also using information about stroke severity (NIHSS) and time to treatment, and 0.80 for the model using a combination of demographics, stroke risk factors, and findings from the neurologic examination and from CT imaging. Calibration of the 3 models was good (Hosmer-Lemeshow tests, p > 0.30) (figure e-1 on the Neurology® Web site at www.neurology.org).
Clinical application.
Because the second model had almost the same discriminatory performance as the third model, we developed a risk chart on the basis of the second model. Figure 1 systematically displays the predicted 1-month risk of poor outcome for each combination of the 5 predictors from model 2. These risks range from 25% for a patient younger than 60 years with hyperlipidemia, a low NIHSS score, and no prodromal minor stroke to 96% for a patient older than 60 years with a high NIHSS score and prodromal minor stroke. The AUC corrected for overoptimism of the risk chart was 0.75 (95% confidence interval 0.71–0.79).
NIHSS = NIH Stroke Scale.
DISCUSSION
We described the development of a series of prognostic models of increasing complexity, based on admission characteristics, to predict the risk of poor outcome at 1 month in patients with an acute symptomatic BAO. The largest amount of prognostic information could be obtained with a set of 5 predictors: age, stroke severity, time to treatment, presence of prodromal minor stroke, and hyperlipidemia.
To the best of our knowledge, this is the first study that identified independent predictors for poor outcome after acute BAO and combined these clinical variables into a simple risk chart. Obviously, risk charts will be more reliable if they include predictors that are already well-established risk factors for poor outcome after stroke. Older age, time to treatment, and stroke severity are such risk factors.2,12 Patients with hyperlipidemia had a lower risk of poor outcome after BAO. This was an unexpected finding but might be explained by potential benefits of treatment with statins.13 However, we do not have data on statin use in our patients. The underlying mechanism of this association is uncertain and deserves further study, although prognostic factors do not necessarily need to have a causal association with the outcome.
We found limited or no added value of CT imaging findings and type of treatment for the prediction of poor outcome. This does not mean that treatment and CT imaging findings are not important; instead, these variables have limited added value for the prediction of poor outcome when data on demographics, stroke risk factors, and clinical stroke severity are already accounted for. Hence, we consider it unlikely that our prediction model has been influenced importantly by withdrawal of life-sustaining interventions. The limited role of type of treatment as a predictor of outcome is probably caused by the lack of a clearly superior treatment strategy in patients with BAO.5
An important strength of our study lies in the large number of patients from which the models were derived. Second, we believe our dataset is representative of the current practice in dedicated stroke centers around the world for patients with an acute symptomatic BAO. Standard neurologic and functional scores and risk factor data were systematically collected from all sites. Furthermore, all measurements were obtained in routine clinical practice. Therefore, the predictors in our model are well defined, easily measured clinical variables already available at admission.
Our study has certain limitations. First, we might not have captured all variables related to outcome. In the design of our data entry form, however, we tried to take into account all factors that could affect outcome. Second, there were some missing values in our database. Regression imputation was used to predict missing values with information from other predictors. Both theoretical and empirical support is growing for the use of imputation methods instead of traditional complete case analysis.14 Third, some misclassification may have occurred in classification of poor vs good outcome. However, the mRS is widely used and has the advantage that it suffers little from such potential bias. Furthermore, our models are based on clinical outcome at 1 month, which is too early a time point to capture a patient's full neurologic recovery from an acute symptomatic BAO. Some patients with an acute BAO have been described to experience remarkably good long-term functional outcomes despite an initial locked-in state and extensive brainstem infarction.15 Thus, it is likely that some patients labeled as having a poor outcome at 1 month would have shifted to having a good outcome if studied at a later time point. The most important drawback of this study is the lack of validation of the risk chart in another population than that in which it was derived (external validation). Although bootstrapping techniques were applied to shrink regression coefficients to correct for overoptimism (internal validation), there might be an overestimation of the true performance. Future studies are needed to confirm the validity of our risk chart.
Prognostic models are particularly useful for a more efficient design of randomized controlled trials. For example, we can exclude those with a very good or very poor prognosis. In addition, these models can be useful for stratification and covariate adjustment of a treatment effect in clinical trials. The proposed risks may also guide clinicians in their initial assessment of the prognosis of a patient who presents with an acute symptomatic BAO. We note, however, that prognostic models can only augment, not replace, clinical judgment.
Our study shows that poor 1-month outcome after BAO can be reliably predicted by a simple model that includes older age, absence of hyperlipidemia, presence of prodromal minor stroke, higher NIHSS score, and longer time to treatment. The model predictions may support clinical practice and research, including the design and analysis of randomized controlled trials.
AUTHOR CONTRIBUTIONS
Dr. Schonewille had the idea for the registry, developed the Internet database, and encouraged international colleagues to contribute data to the registry. Dr. Greving, Dr. Schonewille, Dr. Kappelle, and Dr. Algra designed the study. Dr. Greving did the statistical analyses. Dr. Schonewille and Dr. Greving wrote the first draft of the manuscript and drafted tables and figures.
DISCLOSURE
Dr. Greving reports no disclosures. Dr. Schonewille receives research support from The Netherlands Heart Foundation. Dr. Wijman reports no disclosures. Dr. Michel serves on scientific advisory boards for Bayer Schering Pharma and Boehringer Ingelheim; has received funding for travel or speaker honoraria from Bayer Schering Pharma and Lundbeck, Inc.; serves on the editorial board for the International Journal of Stroke; serves as a consultant for Servier and Lundbeck, Inc.; and receives research support from Lundbeck, Inc., the Swiss National Science Foundation, and the Swiss Cardiology Foundation. Dr. Kappelle has served on a scientific advisory board and received funding for travel and speaker honoraria from for Boehringer Ingelheim; serves on the editorial board of Cerebrovascular Diseases; and receives research support from The Netherlands Heart Foundation and The Dutch Brain Foundation. Dr. Algra reports no disclosures.
ACKNOWLEDGMENT
The authors thank the BASICS study group and all health professionals from participating centers for their contribution to the BASICS registry (see Lancet Neurol 2009;8:724–730).
Footnotes
-
Coinvestigators of the BASICS Study Group are listed on the Neurology® Web site at www.neurology.org.
-
Study funding: This study was funded by an unconditional grant from the Brain Foundation of the Netherlands (grant 2010(2).01). The development of the BASICS registry was supported by the Department of Neurology, University Medical Center Utrecht, Utrecht, the Netherlands.
Supplemental data at www.neurology.org
- Received August 12, 2011.
- Accepted November 17, 2011.
- Copyright © 2012 by AAN Enterprises, Inc.
REFERENCES
Letters: Rapid online correspondence
- Respective prediction models for respective treatment managements
- Re: Respective prediction models for respective treatment managements
REQUIREMENTS
You must ensure that your Disclosures have been updated within the previous six months. Please go to our Submission Site to add or update your Disclosure information.
Your co-authors must send a completed Publishing Agreement Form to Neurology Staff (not necessary for the lead/corresponding author as the form below will suffice) before you upload your comment.
If you are responding to a comment that was written about an article you originally authored:
You (and co-authors) do not need to fill out forms or check disclosures as author forms are still valid
and apply to letter.
Submission specifications:
- Submissions must be < 200 words with < 5 references. Reference 1 must be the article on which you are commenting.
- Submissions should not have more than 5 authors. (Exception: original author replies can include all original authors of the article)
- Submit only on articles published within 6 months of issue date.
- Do not be redundant. Read any comments already posted on the article prior to submission.
- Submitted comments are subject to editing and editor review prior to posting.
You May Also be Interested in
Cutaneous α-Synuclein Signatures in Patients With Multiple System Atrophy and Parkinson Disease
Dr. Rizwan S. Akhtar and Dr. Sarah Brooker
► Watch
Topics Discussed
Alert Me
Recommended articles
Neutrophil count predicts poor outcome despite recanalization after endovascular therapy
Frequency of early rapid improvement in stroke severity during interfacility transfer
MRI-based thrombolytic therapy in patients with acute ischemic stroke presenting with a low NIHSS
Patient outcomes from symptomatic intracerebral hemorrhage after stroke thrombolysis