2.1

Data source

We utilized data from the 2004–2012 National Inpatient Sample (NIS), collected by the Agency for Healthcare Research and Quality Healthcare Cost and Utilization Project (AHRQ HCUP), which is the largest all-payer inpatient publicly available database in the United States . The NIS provides annual information on approximately 8 million inpatient stays from about 1000 hospitals, and estimates a 20% stratified sample from a sampling frame that comprises 90% of US acute care hospital admissions. International Classification of Diseases, Ninth Revision (ICD-9) codes were utilized to ascertain CKD patient population, defined as any discharge record with a CKD stage 3, stage 4, stage 5, or end stage renal disease (ESRD) diagnoses (ICD-9 diagnosis codes 585.3, 585.4, 585.5, or 585.6, respectively). Carotid revascularization-related procedures (ICD-9 procedure codes: CAS 00.61 & 00.63; CEA 38.12) were identified. The Clinical Classifications Software (CCS) codes for ICD-9 were utilized to detect associated diagnoses during hospitalization of these patients. CCS is a diagnosis and procedure categorization scheme that collapses multitudes of associated ICD-9 codes into a smaller number of clinically meaningful classes that have been standardized and extensively previously employed in multiple analyses of diagnoses and procedures .

2.2

Study population

We identified within the NIS database a total of 5331 patients discharged after CEA and 1290 after CAS at U.S. hospitals between 2004 and 2012 based on ICD-9 procedure codes. The sample population was then divided into patients with stage 3 and 4 CKD (glomerular filtration rate = 15–59 mL/min/1.73 m ² ) and those patients with stage 5 CKD/ESRD (glomerular filtration rate < 15 mL/min/1.73 m ² or on dialysis). Patients were not included in analyses if they had overlapping renal impairment diagnoses/coding of stage 3, 4, or 5 CKD or ESRD. Additionally, patients were removed from analyses if they underwent both CEA and CAS during hospital admission.

Cohorts were further refined to better delineate those patients with post-operative stroke complication from those presenting with primary stroke. By way of how the NIS database is designed (primary diagnosis variables and secondary diagnosis variables) we removed any patient with primary diagnosis of stroke (ICD-9 code 433.01, 433.11, 433.21, 433.31, 433.81, 433.91, 434.01, 434.1, 434.10, 434.11, or 434.91). However, if a patient had a secondary diagnosis of iatrogenic cerebrovascular infarction or hemorrhage (ICD code 997.02) then the patient was not removed from analyses. The latter was done in order to account for inherent limitations of an ICD-9 coded database, where coders may list the most severe complication of a hospitalization as the primary diagnosis regardless if it was not the primary cause of the hospitalization. This does not presume that the cohorts studied were completely asymptomatic, but instead ensures that our data were not skewed by patients with primary stroke presentation. The final analyses were performed on a cohort sample sizes of 3299 undergoing CEA and 652 undergoing CAS in patients with stage 3 and 4 CKD, and 1630 undergoing CEA and 511 undergoing CAS among patients with stage 5 CKD/ESRD.

2.3

Patient characteristics and outcome measures

All patient clinical characteristics were obtained from the NIS. Demographic and medical history data extracted included age, gender, race, and associated medical comorbidities. CCS coding was utilized to define underlying coronary artery disease (CCS code 101), peripheral vascular disease (CCS code 114), chronic obstructive pulmonary disease and bronchiectasis (CCS code 127), diabetes mellitus (CCS codes 49 and 50), hypertension (CCS codes 98 and 99), and hyperlipidemia (CCS code 53) rates in our sample.

The primary outcome of interest was MACCE, defined as a composite of in-hospital death, acute myocardial infarction (CCS code 100) and stroke (ICD-9 code 433.01, 433.11, 433.21, 433.31, 433.81, 433.91, 434.01, 434.1, 434.10, 434.11, 434.91, or 997.02). Other outcomes of interest surveyed included the individual endpoints, gastrointestinal hemorrhage (CCS code 153), heart failure (CCS code 108), cardiac dysrhythmias (CCS code 106), and cardiac arrest and ventricular fibrillation (CCS code 107).

2.4

Statistical analysis

Univariate analyses were used to compare demographics, medical comorbidities, and outcomes of patients in each year. Data were summarized by descriptive statistics. Categorical variables were presented as percentages and were compared with the chi-square test. Continuous variables were presented as means with standard deviations and were compared using student’s t-test. Multivariable logistic regression analyses were performed to determine clinical characteristics independently associated with MACCE among both cohorts of stage 3 and 4 CKD and stage 5 CKD/ESRD patients. A separate multivariable logistic regression was performed to examine whether treatment type (CAS or CEA) was independently associated with stroke among both renal impairment groups also. Regression model variables included all factors that had a p-value <0.1 in the univariate models.

2.5

Propensity score matching analysis

Patients with stage 3 and 4 CKD and patients with stage 5 CKD/ESRD were analyzed separately and independently. Propensity score (PS) matching analysis was used to estimate the adjusted marginal (population average) differences in MACCE rate between patients who underwent CAS and CEA . All available patient characteristics were used to estimate PS according to a logistic regression model and the treatment type (CAS or CEA) was the response variable. These characteristics included age as a categorical variable, race, gender, region, insurance type, HCUP Emergency Department service indicator, elective versus non-elective admission, teaching hospital, patient location, transferred status, surgery year, median household income national quartile, old myocardial infarction, transient ischemic attack, peripheral vascular disease, and other specific comorbidities. Propensity score matching was used to select patients from CAS and CEA groups who were matched in PS (i.e., in terms of probability of getting CAS). A 1:1 matching algorithm without replacement was used, where all CAS patients were matched to the closest CEA ones within a range of 0.20 standard deviations of the logit of the estimated PS . In our study, a number of pairs were successfully matched. The success of the PS matching was assessed by checking standardized differences between groups (CAS vs. CEA) before and after matching, (i.e., the absolute differences in sample means divided by an estimate of the pooled standard deviation of the variable) . If the standardized differences were less than 0.2, the differences in the matched samples were considered as minimal. Based on matched samples, McNemar’s tests were carried out for MACCE outcome. Sensitivity analysis for PS matching was carried out to determine the potential impact of unmeasured confounding variables on the significance of the observed treatment effect. The range of significant levels was in the context of McNemar’s test .

In PS model, some covariates had missing values in the dataset. To make full use of the dataset, multiple imputation procedure with Markov Chain Monte Carlo (MCMC) method was also applied to impute the missing values of the covariates before performing PS matching analysis. Variables used to impute the missing values included all the patients’ covariates used in the PS matching as stated above, all comorbidities, treatment type, and MACCE. These imputed values for categorical variables were truncated by 0.5 to create binary values (yes/no). These imputed data sets were then analyzed by PS matching and McNemar’s test with complete data and the results from these analyses were combined using Robin’s rule .

STATA 12 (StataCorp LP, College Station, TX) was used for data analysis and a two-tailed p-value of less than or equal to 0.05 was regarded as statistically significant. SAS 9.3 (SAS Institute Inc., Cary, NC) was used for PS analysis and the significance level was set at p-value < 0.05.