ORIGINAL RESEARCH I Comparative Effectiveness of CABG and PCI Contex We used propensity score matching to create the final Although randomized, controlled trials have shown small analysis cohort. We used all of the baseline characteristics reductions in mortality with coronary artery bypass graft listed in the Table as predictors of receiving CABG or PCI (CABG) surgery versus percutaneous coronary intervention as well as a history of ventricular tachycardia, ventricular (PCI) for coronary revascularization, the restrictive enroll fibrillation, other arrhythmias, implantable cardioverter- ment may limit generalizability to patients in real-world defibrillator. valvular heart disease. stroke, transient isch- emic attack, intracranial hemorrhage, other cerebrovascular Contribution disease, fluid and electrolyte abnormalities, gastrointestinal bleed anemla,F ascular disease, hypoth This comparative effectiveness study using data from roidism, chronic liver disease, AIDS, systemic cancer, obe- Medicare recipients found that CABG was associated with sity, dementia, depression, psychosis, alcohol use, or drug a small mortality benefit versus PCI and certain patient characteristics modified the association. The expected use. We matched each patient who received a multivessel survival advantage varies widely among individual PCI with one who received a multivessel CABG by using a Patients with diabetes, a history of smoking, peripheral greedy algorithm(9)that first matched propensity scores at arterial disease, and heart failure had a particular benefit; 7 digits, then at 6 digits, and so forth, down to a 2-digi those without these factors had slightly better survival match (that is, agreement at the 0.01 level of propensity ith PCI score). We also required that patients be matched by year of index procedure, diabetes status, and age within 1 year. Clinical details of individual patients could not be assessed In the primary analysis, we compared the propensity score-matched cohort of CABG and PCI patients over the first 5-year follow-up by using Cox proportional hazards Individual clinical variables are associated with differences models. We tested the treatment-by-covariate interaction in the benefits of cabg versus pcl with key baseline factors and reported the interaction P value. We repeated the interaction test after additional ad The editors justment for the baseline covariates in the Table. The modifiers of potential treatment effect analyzed included factors tested previously(8)in the pooled analysis of ran were identified from the 20% random sample of Part a domized trials(age, sex, diabetes, hypertension, hyperlipid Comorbid conditions were defined by using diagnosis emia, smoking, unstable angina, prior myocardial infarc- and procedure codes(available on request) found in Part tion [MI], heart failure, and peripheral arterial disease),as A and Part b data(a 5% random sample from 1992 to well as prespecified additional factors of interest that were 1997 and a 20% random sample from 1998 to 2008). We ot available in the previous analysis (race, chronic kidney excluded patients with procedures done before 1992 be disease, cerebrovascular disease, and atrial fibrillation cause of the unavailability of some necessary data during After testing the hypothesis that the comparative effec that period, but we looked back to 1986 to exclude pa- tiveness of CABG and PCI varied according to prespecified tients with prior CABG or PCI during that period. This baseline patient characteristics, we estimated the difference study was approved by the Stanford University Institu- in length of survival between these treatments as a function tional Review board We identified patients by using procedure codes for of patient characteristics. We calculated the life-years multivessel CABG (International Classification of Diseases added by CABG compared with PCI during 5 years of by using rtional hazards model. te 9th Revision, Clinical Modification [ICD-9-CM], codes develop a practical but individualized prediction of survival multivessel PCI(before October 2005: ICD-9-CM code benefit, we used 13 patient characteristics as predictors and 36.05: after October 2005: ICD-9.- CM code 00. 66 plus included selected interaction terms. For each 00.41,00.42, or 00.43 or Current Procedural Terminok- study population, we entered their baseline characteristics gy code 92981 or 92984). We excluded patients if they into the model to generate a predicted survival curve after ad single-vessel PCI or CABG, had concomitant cardiac treatment with CABG. We then numerically integrated procedures(such as valve replacement) at the time of this curve to estimate the life-years of survival over 5 years CABG or PCI, were of unknown race, or had end-stage We repeated this process with the treatment variable set to renal disease and were receiving long-term dialysis PCI. We then subtracted the life-years of survival predicted We defined comorbid conditions by using outpatient after PCI from those predicted after CABG to produce an and inpatient encounters in the year before the index pro- individualized estimate of the effect of CABG compared cedure. We considered a comorbid condition to be present with PCI for each patient in the study population. We if it was recorded as a primary or secondary diagnosis code used SAS, version 9.3(SAS Institute, Cary, North Caro- 21 May 2013 Annals of Internal Medicine Volume 158. Number 10 www.annals.orgwere identified from the 20% random sample of Part A data. Comorbid conditions were defined by using diagnosis and procedure codes (available on request) found in Part A and Part B data (a 5% random sample from 1992 to 1997 and a 20% random sample from 1998 to 2008). We excluded patients with procedures done before 1992 be￾cause of the unavailability of some necessary data during that period, but we looked back to 1986 to exclude pa￾tients with prior CABG or PCI during that period. This study was approved by the Stanford University Institu￾tional Review Board. We identified patients by using procedure codes for multivessel CABG (International Classification of Diseases, 9th Revision, Clinical Modification [ICD-9-CM], codes 36.12, 36.13, 36.14, 36.16, or 36.11 plus 36.15) and multivessel PCI (before October 2005: ICD-9-CM code 36.05; after October 2005: ICD-9-CM code 00.66 plus 00.41, 00.42, or 00.43 or Current Procedural Terminol￾ogy code 92981 or 92984). We excluded patients if they had single-vessel PCI or CABG, had concomitant cardiac procedures (such as valve replacement) at the time of CABG or PCI, were of unknown race, or had end-stage renal disease and were receiving long-term dialysis. We defined comorbid conditions by using outpatient and inpatient encounters in the year before the index pro￾cedure. We considered a comorbid condition to be present if it was recorded as a primary or secondary diagnosis code at either encounter. We used propensity score matching to create the final analysis cohort. We used all of the baseline characteristics listed in the Table as predictors of receiving CABG or PCI, as well as a history of ventricular tachycardia, ventricular fibrillation, other arrhythmias, implantable cardioverter￾defibrillator, valvular heart disease, stroke, transient isch￾emic attack, intracranial hemorrhage, other cerebrovascular disease, fluid and electrolyte abnormalities, gastrointestinal bleeding, anemia, pulmonary vascular disease, hypothy￾roidism, chronic liver disease, AIDS, systemic cancer, obe￾sity, dementia, depression, psychosis, alcohol use, or drug use. We matched each patient who received a multivessel PCI with one who received a multivessel CABG by using a greedy algorithm (9) that first matched propensity scores at 7 digits, then at 6 digits, and so forth, down to a 2-digit match (that is, agreement at the 0.01 level of propensity score). We also required that patients be matched by year of index procedure, diabetes status, and age within 1 year. In the primary analysis, we compared the propensity score–matched cohort of CABG and PCI patients over the first 5-year follow-up by using Cox proportional hazards models. We tested the treatment-by-covariate interaction with key baseline factors and reported the interaction P value. We repeated the interaction test after additional ad￾justment for the baseline covariates in the Table. The modifiers of potential treatment effect analyzed included factors tested previously (8) in the pooled analysis of ran￾domized trials (age, sex, diabetes, hypertension, hyperlipid￾emia, smoking, unstable angina, prior myocardial infarc￾tion [MI], heart failure, and peripheral arterial disease), as well as prespecified additional factors of interest that were not available in the previous analysis (race, chronic kidney disease, cerebrovascular disease, and atrial fibrillation). After testing the hypothesis that the comparative effec￾tiveness of CABG and PCI varied according to prespecified baseline patient characteristics, we estimated the difference in length of survival between these treatments as a function of patient characteristics. We calculated the life-years added by CABG compared with PCI during 5 years of follow-up by using a Cox proportional hazards model. To develop a practical but individualized prediction of survival benefit, we used 13 patient characteristics as predictors and included selected interaction terms. For each patient in the study population, we entered their baseline characteristics into the model to generate a predicted survival curve after treatment with CABG. We then numerically integrated this curve to estimate the life-years of survival over 5 years. We repeated this process with the treatment variable set to PCI. We then subtracted the life-years of survival predicted after PCI from those predicted after CABG to produce an individualized estimate of the effect of CABG compared with PCI for each patient in the study population. We used SAS, version 9.3 (SAS Institute, Cary, North Caro￾lina), to perform all statistical analyses. Context Although randomized, controlled trials have shown small reductions in mortality with coronary artery bypass graft (CABG) surgery versus percutaneous coronary intervention (PCI) for coronary revascularization, the restrictive enroll￾ment may limit generalizability to patients in real-world practice. Contribution This comparative effectiveness study using data from Medicare recipients found that CABG was associated with a small mortality benefit versus PCI and certain patient characteristics modified the association. The expected survival advantage varies widely among individuals. Patients with diabetes, a history of smoking, peripheral arterial disease, and heart failure had a particular benefit; those without these factors had slightly better survival with PCI. Caution Clinical details of individual patients could not be assessed. Implication Individual clinical variables are associated with differences in the benefits of CABG versus PCI. —The Editors Original Research Comparative Effectiveness of CABG and PCI 728 21 May 2013 Annals of Internal Medicine Volume 158 • Number 10 www.annals.org