Development of a Prognostic Tool for the Stratification of Cardiovascular Risk in Patients With Ischemic Stroke

Overview

The availability of several high-cost strategies for the prevention of cardiovascular morbidity and mortality in patients with established cardiovascular disease highlights the necessity of reliable risk stratification of these patients. Several such prognostic models are available for patients with coronary artery disease; however, for patients with ischemic stroke, the available risk stratification schemes are very few and have several limitations. This study aims to develop a prognostication tool to stratify the risk of cardiovascular outcomes in patients with ischemic stroke. The development of a well-designed prognostication tool for the stratification of cardiovascular risk in patients with ischemic stroke may assist to the identification of the highest-risk patients and hence, provide useful information to clinicians and authoritative bodies when prioritizing high-cost strategies for secondary stroke prevention.

Background and rationale Patients with established cardiovascular disease are at very high risk for recurrent cardiovascular events and mortality1. Nevertheless, within this very high risk group, there is significant variation of the underlying risk with some patients being at the extreme edge of the spectrum2,3. The identification of these patients is of utmost importance as it may have implications for management strategies such as prioritization of high-cost strategies like PCSK9 inhibitors and aggressive treatment of modifiable risk factors like arterial hypertension and dyslipidemia. Refined risk stratification may also guide treatment decisions in situations where the balance between the expected benefit and the risk of serious adverse events is borderline like in patients with high bleeding risk who need aggressive antithrombotic treatment, or patients with intracranial bleeding and an indication for antithrombotic treatment4. In addition, it may allow identify those patients who may benefit more from an intensive follow-up schedule. Finally, improved risk stratification may have a positive impact on the motivation of the patient to adhere to secondary prevention strategies. Identification of patients at greater risk of secondary vascular events after ischaemic stroke is challenging because stroke is an etiologically heterogeneous syndrome which may be caused by a diverse set of pathophysiologically discrete diseases like atrial fibrillation (AF), small vessel disease, atherosclerosis and others5. The CHA2DS2VASc score has been shown to predict long-term stroke outcomes in patients with ischaemic stroke, both with and without AF6-8. The Essen Stroke Risk score (ESRS) was derived from patients with ischaemic stroke in the CAPRIE trial and was shown to stratify the 1-year risk of stroke recurrence or major vascular events9. However, the discriminatory performance of both scores in patients with ischemic stroke was modest (c-statistic approximately 0.55 for 1-year stroke recurrence and cardiovascular events) and further refinements are required for clinical application10. Recently, a risk stratification tool was developed among placebo-treated patients with stable ischemic heart disease and previous myocardial infarction (MI) in the TRA2°P-TIMI50 trial11. This score is an integer-based scheme which consists of 9 easily assessed clinical parameters (age, diabetes mellitus,hypertension, smoking, peripheral arterial disease, previous stroke, previous coronary bypass grafting, heart failure and renal dysfunction) and showed a strong graded relationship with the rate of the composite outcome of cardiovascular death, MI and ischaemic stroke, as well as its individual components11. Stroke and ischaemic heart disease share many risk factors and the INTERHEART and INTERSTROKE studies have shown that the 9 or 10 common cardiovascular risk factors account for \>90% of MI or stroke12-14. In this context, several risk stratification models have been introduced to predict the overall cardiovascular risk (rather than its components like myocardial infarction or stroke), mainly in the general population at the primary care level15-18. In this context, it could be hypothesized that the prognostic performance of the TRA2°P score in patients with previous MI can be extended also to patients with ischemic stroke. However, the TRA2°P score performed less accurately in our cohort of ischemic stroke patients compared to the cohort of patients with previous MI in the original publication, with the c-statistics being 0.57 and 0.67 respectively (unpublished data). It becomes evident that the currently available schemes to predict the overall vascular risk in patients with ischemic stroke do not offer a reliable prognosis which could be incorporated in management decisions. Objective \& study implications The objective of the study is to develop a prognostication tool for the stratification of the risk of major adverse cardiovascular events (MACE) in patients with ischemic stroke regardless of the underlying etiology or pathophysiologic mechanism. MACE will be defined as a composite of nonfatal stroke, nonfatal myocardial infarction, and cardiovascular death during the follow-up of the patient. We will assess the time-to-event since the index stroke. In addition, we will also assess multiple events, i.e events occurring after the first outcome event. Stroke will be defined as an acute episode of neurological dysfunction caused by focal or global brain vascular injury and includes ischemic stroke, hemorrhagic stroke, and undetermined stroke. This includes fatal and non-fatal strokes. In case signs and symptoms resolve \<24 hours, stroke requires neuroimaging evidence of acute brain ischemia (i.e. Transient Ischemic Attack with positive neuroimaging). Myocardial infarction will be defined as evidence of myocardial necrosis in a clinical setting consistent with acute myocardial ischemia. The diagnosis of MI requires the combination of evidence of myocardial necrosis (either changes in cardiac biomarkers or post-mortem pathological findings) and supporting information derived from the clinical presentation, electrocardiographic changes, or the results of myocardial or coronary artery imaging. Cardiovascular death includes death due to stroke, myocardial infarction, heart failure or cardiogenic shock, sudden death or any other death due to other cardiovascular causes. In addition, death due to hemorrhage will be included. We will assess the performance (e.g. its sensitivity, specificity, accuracy, positive predictive value and negative predictive value) of different cut-off values of the score match requirements for specific clinical settings. The development of a well-designed prognostication tool for the stratification of cardiovascular risk in patients with ischemic stroke may assist to the identification of the highest-risk patients and hence, provide useful information to clinicians and authoritative bodies when prioritizing high-cost strategies for secondary stroke prevention like PCSK9 inhibitors. The generalizability of the prognostic tool will depend on the representativeness of the population included in the database; given that the analysis will be performed in all patients with ischemic stroke regardless of the underlying pathophysiologic mechanism, generalizability of the score is expected to be wide . Study design \& study population This will be a retrospective analysis in the Athens Stroke Registry, which is a prospective registry of all patients with acute first-ever ischemic stroke admitted between 1993 and 2010 within 24 hours after stroke onset and followed up for up to 10 years. An extended set of parameters is prospectively registered for each patient including demographics, medical history, vascular risk factors, previous treatment, stroke severity at admission, laboratory results, imaging data, in-hospital treatment and medication at discharge. Patients are followed up prospectively at the outpatient clinic at 1, 3 and 6 months after hospital discharge and yearly thereafter for up to 10 years or until death. For those patients who are unable to attend the outpatient clinic, follow-up was assessed over a telephone interview with the patient or proxies, or at the patient's residence by medical personnel. The outcomes assessed are cardiovascular and all-cause mortality, myocardial infarction, stroke recurrence and a composite cardiovascular event consisting of myocardial infarction, angina pectoris, acute heart failure, sudden cardiac death, ischaemic stroke recurrence and aortic aneurysm rupture. Death and its causes are assessed from death certificates, patients' hospital records and information from general practitioners or family physicians. The Athens Stroke Registry has supported many research projects with high-quality publications in high-profile journals, some of them may be found here. We expect that the dataset will include all eligible patients, i.e. approximately 3500 patients with ischemic stroke. The dataset will lock the day before the initiation of the study. Access to the data registered in the Athens Stroke Registry will be sought by the responsible parties. Inclusion criteria All patients with acute ischemic stroke registered in the Athens Stroke registry will be included in the analysis regardless of the underlying etiology or pathophysiologic mechanism. Exclusion criteria Patients with intracranial haemorrhage or transient ischemic attack. Primary outcome A well-validated prognostication tool for the stratification of the risk of major adverse cardiovascular events in patients with ischemic stroke regardless of the underlying etiology or the pathophysiologic mechanism of the index stroke. Study duration and description of steps The study is expected to be completed within 18 months after its initiation. Treatments This is a retrospective chart review analysis and as such, no treatment will be provided to study participants. Methodology \& Data Analysis The dataset will lock the day before the initiation of the study. Summaries of patient parameters and outcomes using appropriate descriptive statistics will be provided for all study variables including demographic and baseline characteristics. Mean, median, standard deviation, IQR, minimum, and maximum will be used to summarize continuous variables. Counts and percentages will be used to summarize categorical variables. Design and development of the algorithm We will develop the prognostic tool using two research methodologies: a) classical statistical analysis based on regression approach, and b) machine learning (ML). The overall predictive ability of the score will be measured via the area under the receiver-operating characteristic curve (AUC-ROC) generated by plotting sensitivity vs 1 - specificity. In addition, we will assess the performance (e.g. its sensitivity, specificity, accuracy, positive predictive value and negative predictive value) of different cut-off values of the score match requirements for specific clinical settings. Associations will be presented as hazard ratios (HR) with their corresponding 95% confidence intervals (95% CI). With regard to the two analytical methodologies which will be followed: * Classical statistical analysis based on regression We will perform multivariate stepwise regression with forward selection of covariates including demographics, medical history, vascular risk factors, previous treatment, stroke severity at admission, laboratory results, imaging data, in-hospital treatment and medication at discharge. For the multivariate analyses, the level of significance will be set at 5%. The log-odds of the final model will be used to define the coefficients in the proposed score. * Machine learning In addition to classical statistical data analysis, also state-of-the-art Machine Learning (ML) predictive algorithms will be applied to develop a prognostic system to predict the primary outcome. Recent advances in ML have greatly helped to accelerate the progress of scientific areas such as brain-computer interfaces, computer vision, natural language processing and understanding, sentiment analysis, time series forecasting, autonomous driving, fraud detection, etc. The incorporation of ML into clinical medicine holds promise for better analysis and understanding of the data. It also holds the keys to unlocking real-time clinical decision support. Prediction is not new to medicine, but recently proposed ML algorithms can substantially improve health care delivery. In this study, we will experiment with a range of ML approaches (e.g. traditional and Convolutional Neural Networks (Deep Learning), Support Vector Machines (SVMs)\] to build a robust prognostic system, capable to generalize to new and unknown inputs. Validation * Internal Validation Internal validation will be performed using bootstrapping and cross validation. Bootstrapping will assess the predictive ability of the model by creating copies of the datasets and recalculating AUC on these copies. Cross-validation will split the dataset in two parts (60%-40%), fits a model to one part (training dataset), and assesses its predictive ability using the other part (validation dataset). * Validation between the two analytical methods The approach of developing the algorithm using two different analytical approached (classical statistical analysis with regression and machine learning) will allow for an indirect method of internal validation. * External validation The developed algorithm will be externally validated in the LASTRO registry. The LASTRO registry is the ongoing, prospective registry of all patients with acute ischemic stroke admitted in the Department of Internal Medicine of the University of Thessaly at the Larissa University Hospital in Larissa, Greece. The registry was initiated in 2014 and is maintained by Prof. George Ntaios (the chief investigator of the Investigator-Initiated Study described in this document). The covariates registered in the LASTRO registry are grossly similar to the covariates registered in the Athens Stroke Registry, which will facilitate the external validation of the developed algorithm. In addition, we will seek to externally validate the developed algorithm in other external datasets, if feasible.