Risk Score for Predicting Long-Term Mortality After Coronary Artery Bypass Graft SurgeryClinical Perspective
Background—No simplified bedside risk scores have been created to predict long-term mortality after coronary artery bypass graft surgery.
Methods and Results—The New York State Cardiac Surgery Reporting System was used to identify 8597 patients who underwent isolated coronary artery bypass graft surgery in July through December 2000. The National Death Index was used to ascertain patients' vital statuses through December 31, 2007. A Cox proportional hazards model was fit to predict death after CABG surgery using preprocedural risk factors. Then, points were assigned to significant predictors of death on the basis of the values of their regression coefficients. For each possible point total, the predicted risks of death at years 1, 3, 5, and 7 were calculated. It was found that the 7-year mortality rate was 24.2 in the study population. Significant predictors of death included age, body mass index, ejection fraction, unstable hemodynamic state or shock, left main coronary artery disease, cerebrovascular disease, peripheral arterial disease, congestive heart failure, malignant ventricular arrhythmia, chronic obstructive pulmonary disease, diabetes mellitus, renal failure, and history of open heart surgery. The points assigned to these risk factors ranged from 1 to 7; possible point totals for each patient ranged from 0 to 28. The observed and predicted risks of death at years 1, 3, 5, and 7 across patient groups stratified by point totals were highly correlated.
Conclusion—The simplified risk score accurately predicted the risk of mortality after coronary artery bypass graft surgery and can be used for informed consent and as an aid in determining treatment choice.
Statistical models have been developed to predict short-term1–8 and long-term9–11 mortality after coronary artery bypass graft (CABG) surgery. In addition, simplified bedside friendly risk scores have been created to overcome the complexity of computing predicted risk of death with statistical models in clinical settings.1,3,6–8 With the aid of such risk scores, predicted risks of death can be obtained by summing the weights of baseline risk factors to obtain the point total for the risk score and identifying the predicted risk of mortality for that point total. This information can be used by clinicians and patients in choosing treatment for the management of severe coronary artery disease. In the past, simple risk scores have been created mainly to predict procedural mortality after CABG surgery,1,3,6–8 but there is also a need for risk scores that predict long-term mortality.
Editorial see p 2409
Clinical Perspective on p 2430
Our group has previously developed a risk score that predicts the risk of in-hospital mortality after CABG surgery using the data of the New York State Cardiac Surgery Reporting System (CSRS).6 Built on this previous work, the present study develops a risk score for predicting the long-term (up to 7 years) mortality after CABG surgery.
The New York State CSRS was the main database used in this study. The CSRS was created in 1988 to establish a statewide registry for major cardiac surgery. The CSRS registry records all major cardiac procedures performed in nonfederal hospitals in the state. The database contains numerous variables, including patient demographics, preprocedural risk factors, procedure-related information, postprocedural complications, and discharge status. The completeness and accuracy of the data in the CSRS are maintained by rigorous data auditing that includes matching between the CSRS data and the statewide hospital discharge data to check the completeness of case reporting and reviewing samples of medical records in selected hospitals to check the accuracy of risk factors. Because documentation is required for coding patient risk factors, a risk factor without documentation to support its presence is coded as absent. An exception is that ejection fraction is allowed to be coded as missing when it was not evaluated before the procedure, and missing values were treated as a separate category of ejection fraction in this study.
A second database, the National Death Index, was used to ascertain patients' vital statuses after discharge by matching to the CSRS on patients' social security numbers. The National Death Index, maintained by the National Center for Health Statistics, compiles all death records in the United States.
Study Population and End Point
The study population consisted of 8597 patients who underwent isolated CABG surgery in 33 hospitals in New York State and were discharged between July and December 2000. The end point of interest was mortality after surgery in the follow-up period. Patient vital status was followed up through December 31, 2007, through the use of the National Death Index.
First, the bivariate relationship between each patient risk factor and long-term mortality was examined with Kaplan-Meier survival analyses. Patient risk factors examined were demographics; body surface area; body mass index (BMI); left main coronary disease (stenosis ≥50%); number of diseased coronary arteries (stenosis ≥70% in the left anterior descending artery, left circumflex artery, and right coronary artery); ejection fraction; history of myocardial infarction; hemodynamic state; the presence of comorbidities, including cerebrovascular disease, peripheral arterial disease, left ventricular hypertrophy, congestive heart failure, malignant ventricular arrhythmia, chronic obstructive pulmonary disease, extensively calcified aorta, diabetes mellitus, and renal failure; and history of open heart surgery and percutaneous coronary intervention. Continuous variables such as age, body surface area, BMI, and ejection fraction were divided into clinically meaningful categories. Significant (P<0.05) risk factors were identified and were used as candidate variables in the development of a multivariate Cox proportional hazards model.
Next, the study population was randomly divided into 2 groups. The data from 1 group (derivation group) of patients were used to derive a Cox proportional hazards model, which was then validated in the other group of patients (validation group). With the use of the derivation data, a Cox proportional hazards model was fit through backward selection to identify significant predictors (P<0.05). To account for the clustering of patients within hospitals, robust sandwich estimators of the standard errors of regression coefficients were obtained in the Cox proportional hazards model.12 Age as a continuous variable and age splines were used to select the best function of age in the model. Other continuous variables such as body surface area, BMI, and ejection fraction were treated as categorical variables in model fitting. Interactions between age and other significant predictors were believed to be the most clinically meaningful 2-way interactions to investigate, but none were found to be statistically significant and therefore were not included in the model.
The performance of the derivation model was then evaluated in the validation patient sample. The derivation model was used to predict each patient's survival at years 1, 3, 5, and 7 after surgery. To evaluate how well the derivation model predicted survival in the validation sample, C statistics were calculated to assess discrimination,13 and the observed and predicted mortality rates of 10 groups of patients of equal sizes grouped by the predicted risk of death were compared to assess model calibration.14,15
Then, the final Cox proportional hazards model was fit by including all of the significant variables identified in the derivation model using data from the entire study population. Because the effects of the significant risk factors in the Cox proportional hazards model could be time dependent, although in general the proportional hazards assumptions were satisfied for patient risk factors in this study, a separate set of Cox proportional hazards models was fit for 3 time periods: within 30 days, 31 days to 1 year, and >1 year after surgery. These 3 time periods were chosen on the basis of clinical significance and the similarity of regression coefficients when >3 models were developed. The C statistics measuring the discrimination of these 3 period-specific Cox proportional hazards models were 0.779, 0.774, 0.773, and 0.783 over 1, 3, 5, and 7 years of follow-up, respectively, and were similar to the respective C statistics of 0.773, 0.772, 0.773, and 0.782 for the final single Cox proportional hazards model. Therefore, the simple model was used to develop the risk score in the following steps.
Based on the final Cox proportional hazards model, a simplified risk score predicting long-term mortality after CABG surgery was developed by applying the method described by Sullivan and colleagues.16 First, a constant for the risk score system was determined as the increase in risk of death associated with a 5-year increase in age, and its value was expressed as the regression coefficient of age multiplied by 5 (0.0566×5=0.2830). Next, points for each category of every risk factor (age was divided into intervals) were obtained by dividing its respective regression coefficient in the final Cox proportional hazards model by the constant (0.2830) and then rounding to the nearest nonzero integer. The risk score assigned to each patient is the sum of the points assigned to each of the patient's risk factors. Next, the predicted risk of death for each possible point total at 1, 3, 5, and 7 years was computed by first calculating the survival rate, S0(t), at the mean values of the risk factors at each time point (t) of years 1, 3, 5, and 7. Then, the predicted risk of death for each point total at each time point (t) was calculated as 1−S0(t)exp[0.2830×(point total)−∑βixi]¯, in which ∑ βixi¯ is the sum of the products of the regression coefficient and the mean values for each risk factor in the Cox proportional hazards model.
To assess how well the risk score predicts the risks of death at years 1, 3, 5, and 7, the agreements between the predicted mortality rates estimated from the risk score and the respective observed mortality rates in those years were evaluated. The patient populations were divided into 10 groups based on the distribution of point totals of risk score and clinical significance of predicted mortality. Then, for each group at each time point, the average predicted mortality based on the risk score and the 95% confidence interval for the observed mortality were calculated. Good agreement between a predicted risk of death and its corresponding observed risk of death was defined as the predicted risk falling within the 95% confidence interval of the observed risk.
All statistical analyses were conducted in SAS version 9.1 (SAS Institute, Inc, Cary, NC).
Among the 8597 CABG patients, 2156 deaths occurred through the end of 2007. Figure 1 shows that the respective 1-, 3-, 5-, and 7-year mortality rates were 6.2%, 11.2%, 17.6%, and 24.2%.
Table 1 shows that higher risk of mortality was associated with older age, female sex, non-Hispanic black race, small body surface area, extreme BMI values, left main coronary disease, multivessel disease, low values of ejection fraction, history of myocardial infarction, unstable hemodynamic state/shock, the presence of comorbidities (cerebrovascular disease, peripheral arterial disease, left ventricular hypertrophy, congestive heart failure, malignant ventricular arrhythmia, chronic obstructive pulmonary disease, extensively calcified aorta, diabetes mellitus, hepatic failure, and renal failure), and history of open heart surgery or percutaneous coronary intervention before the current admission.
Table 2 presents the Cox proportional hazards model that identified the independent predictors of death after CABG surgery. The 13 independent risk factors were older age, BMI <25 or ≥40 kg/m2, lower ejection fractions, unstable hemodynamic state/shock, left main coronary disease, a few comorbidities (cerebrovascular disease, peripheral arterial disease, congestive heart failure, malignant ventricular arrhythmia, chronic obstructive pulmonary disease, diabetes mellitus, and renal failure), and history of open heart surgery. Age was represented as a continuous variable (number of years above 50) in the model, and each 1-year increase in age above 50 years was associated with a 6% increase of risk of death (adjusted hazard ratio = 1.06; P<0.001). The other risk factors were represented as categorical variables. Table 2 shows that renal failure requiring dialysis was associated with the highest relative risk of death (adjusted hazard ratio=5.53; P<0.001).
The C statistics, which measure the discrimination of the Cox proportional hazards model in the validation sample, were 0.768, 0.769, 0.771, and 0.783 for mortality at 1, 3, 5, and 7 years of follow-up, respectively. In addition, Figure 2A through 2D shows good agreement between the observed and predicted mortality rates at 1, 3, 5, and 7 years for the 10 equal groups of patients categorized by the predicted risk of death.
Table 3 presents the integer points assigned to each risk factor in the Cox proportional hazards model. The points assigned to the reference categories of all risk factors were 0. The range of points was from 1 for a number of risk factors (eg, age 51–59 years) to 7 for age ≥80 years. The possible range of point totals for individual patients ranged from 0 to 28. The largest observed point total in this study was 21.
For each point total, the predicted risks of death at 1, 3, 5, and 7 years after the index procedure are presented in Table 4. The predicted risk of death at 1 year after the procedure ranged from 0.87% for a point total of 0 to >98% for a point total of ≥22. For the 7-year predicted risk of death, the range was from 4.35% for a point total of 0 to >99% for a point total of ≥17. Tables 3 and 4 can be used together to obtain the corresponding predicted risks of death at different time intervals for each point total.
Figure 3A through 3D shows good agreements between the observed mortality rates and predicted mortality rates at years 1, 3, 5, and 7 during follow-up across the 10 groups of patients categorized by point totals of risk score. All but 1 of the 40 predicted risks of deaths were within the corresponding 95% confidence intervals of the observed mortality rates. The exception was the predicted risk for a point total of 7 at year 7, and the predicted risk (28.57%) was just outside the 95% confidence interval (29.66%–35.62%) for the respective observed mortality rate.
In this study, we fit a Cox proportional hazards model that identified the risk factors for long-term mortality of patients who underwent isolated CABG surgery. Using the Cox model, we developed and evaluated a risk score that can be used to estimate the risk of long-term mortality after isolated CABG surgery.
In general, the predictors for long-term mortality identified in the Cox proportional hazards model in this study are consistent with those identified in other studies.9–11 Age, extreme BMI, lower ejection fraction values, left main coronary disease, and a few comorbidities such as cerebrovascular disease, peripheral arterial disease, congestive heart failure, malignant ventricular arrhythmia, chronic obstructive pulmonary disease, diabetes mellitus, renal failure, and history of open heart surgery have also been found to be significant predictors in other studies.9–11 The impact of these risk factors on long-term mortality was monotone except for BMI, which shows a U-shaped relationship in which BMI of 25.0 to 39.9 kg/m2 is related to the lowest risk and lower and higher BMIs are associated with higher risk. A similar U-shaped relationship between BMI and mortality after CABG surgery has been reported in other studies.17,18 In addition, consistent with other studies,9–11 sex was not an independent predictor of long-term mortality in this study, although female sex is often reported to be associated with a higher risk of in-hospital and short-term mortality.6,19,19 In addition, no significant interaction between sex and BMI on the risk of long-term morality was observed in this study.
Using the Cox proportional hazards model, we developed a risk score that complements our risk score for in-hospital mortality6 by predicting mortality up to 7 years after procedure for CABG patients who have survived the first 30 days after the procedure. This new risk score can be used together with the New York State CSRS CABG surgery risk score for in-hospital mortality6 or other similar risk scores for short-term mortality.1,3,7,8 For example, on the basis of a patient's risk factors before isolated CABG surgery, including age, sex, hemodynamic state, ejection fraction, history of myocardial infarction, and previous open heart surgery, and a number of comorbidities such as chronic obstructive pulmonary disease, extensively calcified aorta, peripheral arterial disease, and renal failure, the CSRS CABG surgery risk score for in-hospital mortality can be used to predict the patient's short-term procedural risk.6 In addition, the risk score for long-term mortality developed in this study can be used to predict the risks of death at 1, 3, 5, and 7 years after CABG surgery.
Because there are possible variations in long-term mortality after CABG surgery between patient populations, we recommend recalibrating the predicted risks of death in this study in Table 4 to reflect the difference in mortality rates between the patient population of interest and our study population. First, the Cox proportional hazards model can be recalibrated with the approach described by D'Agostino and colleagues20 with the regression coefficients of the risk factors presented in Table 2, the prevalence of those risk factors, and the average survival rate in the new patient population. Then, the predicted risk of death for each point total in the new study population can be obtained with the method described in the Methods section of this article.
Our study has a few strengths. First, it used a large statewide population-based registry for CABG surgery; therefore, the generalizability of the study is likely to be high. Second, because we were able to match the patients to the National Death Index, the chance of loss to follow-up was minimized.
There are also a few limitations of the study. First, it was developed from data on patients who underwent isolated CABG surgery a decade ago. The quality of CABG surgery and follow-up care has improved over time, and outcomes after CABG surgery have also improved. However, our data are still the latest available data on long-term mortality. Second, the accuracy of this risk score should be tested in other patient populations and in other time periods. Third, the predictions from our statistical models are based on all-cause mortality, which is predicted on the basis of available risk factors before CABG surgery. Some risk factors for longer-term mortality in general (eg, cancer, smoking status) were not available. In addition, the values of some risk factors (eg, BMI) may change over time, but only the baseline values were available. Therefore, the accuracy of the Cox proportional models is limited. Nonetheless, many noncardiovascular risk factors (eg, cerebrovascular disease, chronic obstructive pulmonary disease, diabetes mellitus, renal failure, and hepatic failure) were available. At 7 years of follow-up, the C statistic, which describes the capability of model to differentiate high and low risks of death, was 0.782, which is a reasonably high value. Finally, the Cox proportional hazards model used to develop the risk score estimates the overall impact of significant risk factors on mortality during the entire 7-year follow-up period and does not estimate the impact of risk factors during subintervals of time within the 7 years. However, the simpler model has C statistics (0.773, 0.772, 0.773, and 0.782 for 1-, 3-, 5-, and 7- year mortality, respectively) similar to those of the period-specific models (0.779, 0.774, 0.773, and 0.783 for 1-, 3-, 5-, and 7- year mortality, respectively), indicating that these 2 sets of models have similar abilities for discriminating between high- and low-risk patients.
We created a risk score predicting long-term morality after isolated CABG surgery. We anticipate that this new risk score, like our CSRS CABG risk score and other risk scores for in-hospital mortality,1,3,6–8 can become a handy risk-stratification tool that can be used by clinicians and patients in the choice of treatment for severe coronary disease.
Sources of Funding
This work was supported by National Institutes of Health grant RC1HL099122. The views expressed are those of the authors and do not necessarily reflect those of the New York State Department of Health.
We thank the New York State Cardiac Advisory Committee for its encouragement and support of this study. We also thank Kimberly Cozzens, Rosemary Lombardo, and the cardiac catheterization laboratories and cardiac surgery programs of the participating hospitals for their tireless efforts to ensure the timeliness, completeness, and accuracy of the registry data.
- Received July 14, 2011.
- Accepted March 19, 2012.
- © 2012 American Heart Association, Inc.
- Nashef SAM,
- Rogues F,
- Michel P,
- Gauducheau E,
- Lemeshow S,
- Salamon R
- Ivanov J,
- Tu JV,
- Naylor CD
- Geissler HJ,
- Holzl P,
- Marohl S,
- Kuhn-Regnier F,
- Mehlhorn U,
- Sudkamp M,
- de Vivie ER
- Parsonnet V,
- Dean D,
- Bernstein AD
- MacKenzie TA,
- Malenka DJ,
- Olmstead EM,
- Piper WD,
- Langner C,
- Ross CS,
- O'Connor GT
- Hosmer DW,
- Lemeshow S
- Hosmer DW,
- Lemeshow S,
- May S
- Jin R,
- Grunkemeier GL,
- Furnary AP,
- Handy JR Jr.
Bedside friendly risk scores that predict in-hospital or 30-day mortality after coronary artery bypass graft (CABG) surgery are currently available. However, there is a need for a risk score that predicts long-term mortality after CABG surgery. In this study, we developed a Cox proportional hazards model that predicts mortality up to 7 years after CABG surgery using patient baseline risk factors. From this model, we created a simplified risk score that can accurately predict long-term mortality after CABG surgery. This risk score can be conveniently used to estimate long-term mortality after CABG surgery for a patient for whom CABG surgery is indicated; such information can also be useful while considering CABG surgery as an option to treat coronary artery disease.