Real world effectiveness of warfarin among ischemic stroke patients with atrial fibrillation: observational analysis from Patient-Centered Research into Outcomes Stroke Patients Prefer and Effectiveness Research (PROSPER) study

Study design and patient involvement

Details of the design and conduct of the Patient-Centered Research into Outcomes Stroke Patients Prefer and Effectiveness Research (PROSPER) study have been previously described.17 Briefly, PROSPER is a Patient-Centered Outcomes Research Institute (PCORI) sponsored project designed to help patients, physicians, and other stakeholders to make informed decisions about stroke care and to improve outcomes through innovative comparative effectiveness research. The study was conceived and designed by the PROSPER team, a multidisciplinary group of researchers at the Duke Clinical Research Institute, Massachusetts General Hospital, University of California at Los Angeles, University of Texas Southwestern Medical Center, Brigham and Women’s Hospital, and University of Calgary, partnering with patient investigators and stakeholders representing the varied interests of patients, caregivers, physicians, nurses, healthcare systems, policy makers, governmental agencies, and professional societies.

To identify the most relevant research topics and meaningful outcomes, we worked with our patient co-investigators and conducted a series of focus groups with stroke survivors and administered a qualitative survey customized to stroke patients. Knowledge gained from these efforts was further reviewed by our patient co-investigators and stakeholder committee, who transformed these identified research needs into patient centered research questions to be studied, including the effectiveness of warfarin among ischemic stroke patients with atrial fibrillation.

The concept of patient involvement also translated to the study design and execution phases of the research. We provided basic educational materials and a research tutorial to help to encourage familiarity with research concepts and terminology. The pros and cons of different types of study design and research method were discussed with patient co-investigators, using laypersons’ language to facilitate common understanding and solicit their feedback. We then revisited our aims and revised the research plans to align them better with things that patients cared about the most. As results emerged, we reviewed the results with patient co-investigators to obtain their perspectives and feedback to ensure that we presented the findings in the most effective way beyond the research community to general populations. Through these iterative processes, PROSPER patient co-investigators made high value contributions to both the design and the implementation of the study, while we retained the rigors of the scientific work.

Data sources

PROSPER builds on the American Stroke Association/American Heart Association GWTG-Stroke program, which is an ongoing nationwide quality improvement initiative.18 19 Standardized data collection in GWTG-Stroke includes patients’ demographics, medical history, diagnostic testing, brain imaging, in-hospital treatment, discharge drugs, and inpatient outcomes. The eligibility of each admission is confirmed through chart review. The validity and reliability of data collection in the GWTG-Stroke database has been reported in previous research.20

As GWTG-Stroke is an inpatient registry, we linked GWTG-Stroke data to Medicare claims by matching on a series of indirect identifiers, including admission date, discharge date, patient’s age or date of birth, and sex, for the assessment of longitudinal outcomes. This linkage method, developed by our analytic team, has been successfully completed and validated using Medicare inpatient claims.21 Previous work has shown that patients in the linked GWTG-Stroke/CMS database are representative of the national Medicare ischemic stroke population.22 Quintiles Inc serves as the data collection (through its patient management tool) and coordination center for GWTG-Stroke. The Duke Clinical Research Institute serves as the data analysis center and has an agreement to analyze the aggregate de-identified data for research purposes.

Study population

Our analysis included Medicare fee for service beneficiaries admitted to GWTG-Stroke hospitals for acute ischemic stroke and discharged alive with documented persistent or paroxysmal atrial fibrillation or flutter between 1 January 2009 and 31 December 2011. Each patient had at least one year of follow-up after the index hospital discharge, using hospital claims data submitted to CMS through December 2012. Figure 1⇓ shows details of the study’s inclusion and exclusion criteria. Briefly, we excluded patients with documented contraindications for anticoagulation treatment, as well as those who received comfort measures only or were transferred out to other hospitals. To avoid prevalent user bias, we further excluded patients receiving chronic anticoagulation treatment before the index hospital admission.23 We abstracted the discharge drug treatment from the medical record of the index hospital admission. We divided patients into two groups according to discharge drug treatment: patients treated with warfarin versus those not treated with any oral anticoagulant at discharge. Patients who refused warfarin but had no contraindications were considered eligible for warfarin but did not receive the treatment (n=1162). Novel oral anticoagulants such as dabigatran and rivaroxaban were not recorded in GWTG-Stroke until October 2011. Therefore, we also excluded patients discharged on novel oral anticoagulants or other agents such as low molecular weight heparin or fondaparinux. As the National Institutes of Health Stroke Scale (NIHSS) score (a measure of neurologic deficits ranging from 0 to 42, with a higher score indicating greater severity of stroke) is a critical predictor of outcomes in acute ischemic stroke,24 25 our primary study population consisted of 12 552 patients with complete NIHSS data from 1487 GWTG-Stroke hospitals in the United States.

Fig 1 Study population from initial cohort, through exclusions, to final population. CMS=Centers for Medicare and Medicaid Services; GWTG-Stroke=Get With The Guidelines-Stroke; NIHSS=National Institutes of Health Stroke Scale

• Download figure
• Open in new tab
• Download powerpoint

Outcome measures

We conducted focus group and surveys of patients to identify outcomes most relevant and meaningful to patients who have survived stroke.17 The most frequent and important outcome identified was “alive at home, without recurrent stroke, and without being hospitalized for complications.”15 16 On the basis of the interpretation of patients’ responses and further refinement by patient co-investigators, we translated this concept into home time, which we defined as the total number of days alive and out of the hospital or a skilled nursing facility within two years after the index hospital discharge. We calculated home time by using Medicare inpatient care, inpatient rehabilitation facility, and skilled nursing home claims.26 Similarly, a major adverse cardiovascular event (MACE) is a composite measure of all cause mortality or cardiovascular readmission, reflecting patients’ desire of “being alive at home, without recurrent stroke, or being hospitalized for complications.” Secondary outcomes included all cause mortality, cardiovascular readmission, ischemic stroke readmission, and hemorrhagic stroke readmission.

Where applicable, we determined the date of death through the Medicare denominator files. We ascertained the date and cause of readmission from the hospital claims data submitted to CMS, with the international classification of diseases, ninth revision, clinical modification (ICD-9-CM) primary diagnosis code of 433.x1 or 434.x1 for ischemic stroke and 430 or 431 for hemorrhagic stroke and the Agency for Healthcare Research and Quality’s clinical classification software diagnosis category 96-118 for cardiovascular admission. The clinical classification software is a diagnosis and procedure categorization scheme developed by the Agency for Healthcare Research and Quality to identify disease specific or procedure specific hospital admission by grouping individual ICD-9-CM codes into clinically meaningful categories.27 The use of ICD-9-CM codes 433.x1, 434.x1, 430, and 431 has been proposed by the American Heart Association/American Stroke Association for defining stroke using ICD-9-CM codes and for CMS stroke core measures.28 29 30

Statistical analyses

Our pre-specified statistical analysis plan has been previously described.17 To summarize, we used means, medians, and percentages to describe the distribution of continuous and categorical variables between patients discharged on warfarin and those discharged not on any oral anticoagulants. We defined warfarin treatment as warfarin prescribed at the time of hospital discharge, as documented in GWTG-Stroke. We used standardized difference to compare baseline characteristics between patient cohorts. An absolute standardized difference greater than 10% indicates significant imbalance of a covariate, whereas a smaller value supports the assumption of balance between treatment groups.31

Unlike randomized clinical trials, the decision to treat in “real world” practice is often based on prognostic factors. Therefore, the effectiveness estimate for warfarin might be confounded as a result of treatment selection. We used a stabilized inverse probability weighting approach to control for potential bias. Inverse probability weighting is an extension of the propensity score method used to summarize the conditional probability of assignment for a treatment.32 33 Using this method, the weights are the inverse probability of assigning a treatment derived from a logistic model with warfarin treatment as the dependent variable and all observed patient level and hospital level characteristics as the independent variables. These variables included age; sex; race/ethnicity; history of previous stroke/transient ischemic attack, coronary artery disease/myocardial infarction, carotid stenosis, peripheral vascular disease, hypertension, dyslipidemia, diabetes, or smoking; emergency medical services transportation; NIHSS at presentation; hospital bed size; annual ischemic stroke volume; American Heart Association/American Stroke Association/Joint Commission certification as a primary stroke center; academic status; rural/urban location; and calendar year. Each patient was then weighted by the inverse of the estimated probability of the treatment received.

We then used inverse probability weighting regression models weighted with warfarin treatment as the independent variable to estimate the adjusted associations between warfarin treatment and each outcome estimated, controlling for other discharge drugs as covariates. Discharge drugs included antihypertensives, statins, and antiplatelets (aspirin, aspirin/dipyridamole, clopidogrel, aspirin/clopidogrel dual antiplatelet therapy, and other antiplatelet agents). Owing to the presence of over-dispersion in home time and the fact that not all patients had the same length of time in follow-up (for example, patients lost Medicare fee for service coverage or transitioned to Medicaid), we used a negative binomial regression with an offset for log of the proportion of follow-up time to model home time. We used separate Cox proportional hazards models for time to event data including MACE, all cause mortality, all cause readmission, cardiovascular readmission, ischemic stroke readmission, and hemorrhagic stroke readmission.

In addition to the general population of ischemic stroke patients with atrial fibrillation, we did these analyses in clinically relevant subgroups by age (65-80 and >80 years), sex, history of previous stroke/transient ischemic attack and coronary artery disease/myocardial infarction, and stroke severity as measured by NIHSS (mild 0-5, moderate 6-14, and severe 15-42). Our analyses also accounted for within hospital clustering by using random effects for negative binomial models and robust sandwich estimators for Cox proportional hazards models. In a sensitivity analysis, we included 7574 (37.6%) patients with missing NIHSS data and extended our study population to the entire study cohort (total n=20 126). We used a multiple imputation method to impute missing NIHSS values. The characteristics of the patient cohorts including those without NIHSS recorded and only patients with NIHSS recorded are shown in supplementary table A.

All analyses followed the intention to treat principle. All P values are two sided. Owing to multiple outcomes and subgroups assessed, we report 99% confidence intervals and considered P values less than 0.01 to be statistically significant. We used SAS version 9.3 for all statistical analyses.