An In-hospital Mortality Risk Model for Patients Undergoing Coronary Artery Bypass Grafting in China

Journal Pre-proof An In-hospital Mortality Risk Model for Patients Undergoing Coronary Artery Bypass Grafting in China Zhan Hu, MD, PhD, Sipeng Chen, MS, Junzhe Du, MD, Dachuan Gu, MD, Yun Wang, PhD, Shengshou Hu, MD, PhD, Zhe Zheng, MD, PhD PII:

S0003-4975(19)31404-3

DOI:

https://doi.org/10.1016/j.athoracsur.2019.08.020

Reference:

ATS 33042

To appear in:

The Annals of Thoracic Surgery

Received Date: 6 December 2018 Revised Date:

20 June 2019

Accepted Date: 8 August 2019

Please cite this article as: Hu Z, Chen S, Du J, Gu D, Wang Y, Hu S, Zheng Z, An In-hospital Mortality Risk Model for Patients Undergoing Coronary Artery Bypass Grafting in China, The Annals of Thoracic Surgery (2019), doi: https://doi.org/10.1016/j.athoracsur.2019.08.020. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 by The Society of Thoracic Surgeons

An In-hospital Mortality Risk Model for Patients Undergoing Coronary Artery Bypass Grafting in China Running head: CABG mortality risk model in China

Zhan Hu, MD, PhD1; Sipeng Chen, MS1; Junzhe Du, MD,1; Dachuan Gu, MD1; Yun Wang, PhD2; Shengshou Hu, MD, PhD1; Zhe Zheng, MD, PhD1

1. Department of Cardiovascular Surgery, Fuwai Hospital, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, People’s Republic of China. 2. Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, US. Center for Outcomes Research and Evaluation, Yale New Haven Health, CT, US.

Classifications: coronary artery bypass grafting, database, risk modeling

Word count: 4981

Corresponding Author: Zhe Zheng, MD, PhD, Department of Cardiovascular Surgery, National Clinical Research Center of Cardiovascular Diseases, Fuwai Hospital, National Center for Cardiovascular 1

Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College. No. 167, Beilishi Rd, Xicheng District, Beijing, People’s Republic of China 100037. E-mail [email protected]

2

ABSTRACT Background: To meet the demand of increasing surgical volume and changing of patient’s risk profiles of coronary artery bypass grafting in China, we develop a new risk model that predicts in-hospital mortality. Methods: The analysis included patients who underwent coronary artery bypass grafting between January 2013 and December 2016 at 87 hospitals in the Chinese Cardiac Surgery Registry. Patients in years 2013-2015 were randomly divided into training (n=31297, 75%) and test (n=10432, 25%) samples; 2016 patients (n=15047) comprised the validation sample. Demographic and clinical risk factors were identified

. Harrell’s C-statistic was used to

evaluate model discrimination, and the Hosmer-Lemeshow goodness-of-fit test was used to assess calibration. Results: The 56776 patients had a mean age of 61.8 (standard deviation 8.8) years, and 24.6% were female. Overall, in-hospital mortality was 2.1%. The final model included 21 risk factors represented by 16 unique variables. The model achieved good discrimination, with a C-statistic of 0.79 (95% CI 0.77-0.80) in the training sample, 0.79 (95% CI 0.76-0.82) in the test sample, and0.78 (95% CI 0.76-0.81) in the validation sample. Model calibration was good according to the Hosmer-Lemeshow test (P>0.05 in the three samples). When compared with the SinoSCORE and EuroSCORE II, the model had better discrimination and calibration.

3

Conclusions: We developed and evaluated a model with 16 risk factors that predicted inhospital mortality risk after coronary artery bypass grafting in China. This updated model may help surgeons and hospitals better identify high-risk patient.

Word count: 240

4

Ischemic heart disease is the leading cause of death worldwide (1), and coronary artery bypass grafting (CABG) is one of most common and reliable procedures to treat this disease. The procedure, however, still carries an in-hospital mortality rate of 1%-2% in China and the US (2). To limit adverse postoperative outcomes, clinical risk prediction models were introduced into cardiac surgical practice more than 30 years ago (3). These models have been used to risk stratify patients, facilitate informed consent, and aid clinical decision making, outcome evaluation, and quality improvement. Several models have been developed for the prediction of early mortality after CABG. The European System for Cardiac Operative Risk Evaluation (EuroSCORE) (4, 5) and the Society of Thoracic Surgeons risk evaluation system (6) are the most widely accepted of these tools and have been recommended in guidelines for CABG surgical risk assessment (7). The first risk model predicting in-hospital mortality among Chinese patients undergoing CABG, the SinoSCORE, was developed a decade ago (8). When compared with other well-accepted risk models, the SinoSCORE exhibited better performance in the prediction of mortality after cardiac surgery in the Chinese population (8, 9). CABG admissions have declined substantially in the US over the past decade (10), but they continue to rise in China (2, 11). Advances in cardiac surgical care and the Chinese government’s efforts to improve population health have helped to decrease surgical mortality over time. A recent study showed that the SinoSCORE risk model may now overestimate surgical mortality and poorly discriminate between patients (12). Consequently, a new risk 5

model is required to meet the demand of increasing surgical volume and to improve the usefulness of risk modeling in contemporary cardiac surgical practice in China. Established in 2013, the Chinese Cardiac Surgery Registry (CCSR) is a nationwide multicenter registry that provides a platform for monitoring adult cardiac surgical care in mainland China (13). Compared with the database used to develop the SinoSCORE, the CCSR includes a broader range of participating sites and higher quality data. The purpose of this study was to identify risk factors and develop a risk model that predicts CABG postoperative mortality using this new national database.

PATIENTS AND METHODS Data Source and Study Sample The CCSR database includes data from consecutive patients undergoing cardiac surgery at 87 participating centers located in nearly all provinces and directly controlled municipalities in China(13). Preoperative risk factors and in-hospital mortality are recorded for each patient. Each year, 5%-10% of these records are randomly selected from each participating site for audit by the National Coordinating Center. During this audit, medical record data are independently abstracted by two qualified physicians and compared with data submitted by the sites. Agreement was 95% for mandatory variables in our study. We limited the sample for our study to the 56775 unique patients aged 18 to 90 years who underwent CABG, including combined surgery, between January 2013 and December 2016. ，

We randomly divided the 41729 patients in years 2013 to 2015 into two subsamples according to a 3:1 ratio; the training sample included 31297 patients, and the test sample included 10432 patients (Figure 1). We selected the 15047 patients in 2016 for the validation sample. The training sample was used to select risk factors and develop the risk model, and the test and validation samples were used to evaluate and validate the risk model. The rationale for the 2016 validation sample was to provide an additional opportunity to evaluate performance stability and assess performance in current clinical practice.

Potential Risk Factors Candidate risk factors for the model included patient demographic characteristics (age, sex, insurance coverage), medical history and comorbidities (body mass index [BMI], smoking, diabetes mellitus, hypertension, chronic obstructive pulmonary disease [COPD], extracardiac arteriopathy, prior cerebrovascular accident, prior myocardial infarction, angina, atrial fibrillation, prior percutaneous coronary intervention [PCI]), hospital evaluation and workup (New York Heart Association [NYHA] class, left ventricular ejection fraction [LVEF], renal function [creatinine clearance], critical preoperative state[demonstrating any one of the following

preoperatively and during the same hospital admission as surgery: cardiogenic shock; cardiopulmonary resuscitation; ventricular fibrillation or flutter; or intra-aortic balloon pump implantation]), and procedure-related factors (prior cardiac surgery, urgency of surgery, types of combined procedures) (see Supplemental eTable 1 in the supplement for detailed variable 7

definitions). BMI was calculated as weight (kg)/height2 (m2), and creatinine clearance was calculated using the Cockcroft-Gault formula. These candidate variables were chosen because they were clinically meaningful, supported by existing literature, and occurred at a frequency of more than 1%. Variables with missing values were imputed using multiple imputation with 10 imputations(14). The final imputed value was an average of the 10 imputations. Rates of missing ranged from 0% (0) to 3.3% (1896) (Supplemental eTable 2 in the supplement).

Outcome The outcome for the prediction model was in-hospital mortality, defined as all-cause in-hospital death after CABG.

Statistical Analysis Model Development and Validation A stepwise logistic regression model with an entrance threshold of P=0.3 and an exit threshold of P=0.2 was used to select potential risk factors in the training sample. The bootstrap simulation procedure, a computer-based resampling statistical method designed to assign measures of accuracy, was employed in conjunction with the logistic regression procedure. The simulations were carried out 2000 times for the training sample, with each iteration of the simulation using a new sample generated by drawing n observations with replacement from the original sample to fit the regression model and yield a set of variables that were statistically 8

significantly associated with the outcome. A variable that was statistically significant at least 75% of the time was considered a robust risk factor for predicting the outcome and retained in the final model. To construct the final model, we refit the logistic model to the training sample, without stepwise elimination, using the variables chosen by the bootstrapping approach. We calculated Harrell’s C-statistic to evaluate the discriminatory performance of the model and used the Hosmer-Lemeshow goodness-of-fit test to assess model calibration. Additionally, the original Chinese CABG risk model (SinoSCORE) and the EuroSCORE II model were used to compare predictive performance in all 3 samples.

Risk Score To facilitate the use of our risk model in clinical practice, we assigned the independent predictors in the final model weighted point scores proportional to their β regression coefficient values. We calculated risk score points for each variable by dividing the variable’s coefficient by the sum of all coefficients, multiplying by 100, and rounding to the nearest integer. The risk score for each patient was calculated by summing the points for each factor. Patients were stratified into three risk groups based on the risk score distribution: low (<10th percentile), average (10th-90th percentile), and high (>90th percentile). All analyses were performed using SAS version 9.4 (SAS Institute Inc. Cary, NC). The Fuwai Hospital institutional review board approved the study and waived the requirement for informed consent. The study followed the Transparent Reporting of a Multivariable Prediction Model for 9

Individual Prognosis or Diagnosis (TRIPOD) guideline for developing and evaluating a prediction model; each of the 22 items of the statement was addressed.

RESULTS Study Sample There were 56776 patients included in the study from 87 participating centers in 32 provinces across China. Patients had a mean age of 61.8 (standard deviation 8.8) years, and 75.4% were male. The mean BMI was 25.0±3.2 kg/m2, and common comorbidities included history of hypertension (60.1%), insulin-dependent diabetes mellitus (9.0%), myocardial infarction (19.0%), prior PCI (7.7%), extracardiac arteriopathy (3.6%), and preoperative atrial fibrillation (3.0%). Approximately 26.1% of patients had left main disease, and 61.8% had three-vessel disease. The patient characteristic profiles were similar across the training, test, and validation samples (Table 1). In-hospital mortality in the aggregate cohort was 2.1% (95% CI 2.0%-2.2%). For the training, test, and validation samples, in-hospital mortality rates were 2.0% (95% confidence interval [CI] 1.8%-2.1%; 622 of 31297 patients), 2.1% (95% CI 1.8%-2.4%; 218 of 10432 patients), and 2.2% (95% CI 2.0%-2.5%; 337 of 15047 patients), respectively (P=0.20).

Model Development and Validation

10

We identified 24 candidate preoperative variables for inclusion in the multivariable analysis with bootstrap samples (Table 1). The results of the 2000 random repeated samples are summarized in Figure 2. There were 21 risk factors, comprising 16 unique variables, identified as significant in at least 75% of the samples that were chosen for the final multivariable logistic regression model: demographic characteristics (age 60-69 years, age ≥70 years, female), medical history and comorbidities (BMI <18.5 kg/m2, myocardial infarction in the past 21 days, angina, prior PCI, prior cerebrovascular accident, COPD), hospital evaluation and workup (NYHA class III, NYHA class IV, LVEF <35%, LVEF ≥35% to <45%, LVEF ≥45% to <55%, creatinine clearance ≥50 mL/min to <80 mL/min, creatinine clearance <50 mL/min or on dialysis, critical preoperative state), and procedure-related factors (prior cardiac surgery, non-elective, combined valve surgery, combined surgery except valve surgery) (Supplemental eFigure 1). The estimated β coefficients and odds ratios for the risk factors in the final risk model are reported in Table 2. The final risk model exhibited good discrimination and calibration in the training data. The Cstatistic was 0.79 (95% CI, 0.77-0.80; Figure 3), and the mean observed in-hospital mortality ranged from 0.3% in the lowest predicted decile to 8.4% in the highest predicted decile, a range of 8.1% (Figure 4). The P value of the Hosmer-Lemeshow was 0.30, and observed in-hospital mortality coincide with predicted in-mortality by deciles (Figure 4), suggesting the model was fitted well (Table 3). The risk model achieved comparable performance in the test and validation sample. For test and validation sample, respectively, the C-statistic was 0.79 (95%CI, 0.7611

0.82) and 0.78 (95%CI, 0.76-0.81), Hosmer-Lemeshow P value was 0.13 and 0.29, the observed in-hospital mortality ranged from 0.5% and 0.2% in the lowest predicted decile to 9.4% and 10.2% in the highest predicted decile. The observed and predicted in-hospital mortality also coincide well by deciles(Figure 4). In consideration of huge variation in surgical volumes among centers, we also refit a mixed model with hospitals as a random effects for the 16 selected variables (Supplemental eTable 3). Although two variables (angina and prior PCI) were no long statistically significant, the model performance was similar. When applied to the training, test and validation sample, the original SinoSCORE model and EuroSCORE II model obtained relatively worse predicting performance than our risk model. The results of C-statistic and Hosmer—Lemeshow P-value are shown in Table 3 and area under the ROC curve was shown in Figure 3.

Risk Score The risk points for each factor ranged from 1(creatinine clearance ≥50 mL/min to <80 mL/min) to 10 (prior cardiac surgery) (Table 2). Prior cardiac surgery, NYHA class IV, ejection fraction <35%, combined valve surgery, and non-elective surgery were the top five factors with odds ratios greater than 6. The mean (SD) risk scores for the training, test, and validation samples were 10.4 (6.2), 10.4 (6.3), and 10.7 (6.5), respectively. The proportion of patients assigned to the low-risk (score 0-3), average-risk (score 4-18), and high-risk (score ≥19) groups and the corresponding probabilities of in-hospital mortality for these groups (Figure 5 and Supplemental 12

eTable 4) were similar in the training, test, and validation samples. The predicted probabilities of in-hospital for CABG patients in training sample ranged from less than 1% for a risk score of 0, to 35% for the risk scores of 33 and higher (Table 4 and Supplemental eFigure 2)

COMMENT In this study, we identified 21 risk factors that predicted in-hospital mortality for patients undergoing CABG surgery in China. Data on these factors are easily collected and readily available prior to surgery, allowing our risk model and its corresponding risk score to stratify patients into low, average, and high-risk groups for surgical mortality prior to CABG. Risk stratification is not only important to better inform patients and doctors about procedure risk and aid clinical decision making, but also it is crucial for facilitating clinical research projects, establishing benchmarks, and improving cardiac surgical care. Enrolling patients undergoing combined procedures (about 15% of all CABG patients) in our study sample, our risk model was not confined to isolated CABG, and could also be applied to predict mortality in combined surgery. The CCSR database was established in 2013 for risk assessment, outcome evaluation and quality improvement of adult cardiac operations in China. We chose the data elements based on preliminary works(including the Original SinoSCORE model) and with reference to wellestablished cardiac surgical database, and continued to modify and update the data elements based on feedbacks. This registry established a nationally representative network of 87 13

participating centers with adult cardiac surgery volume greater than 100 patients per year. Data quality audits and feedback reporting to the participating sites are routinely performed to ensure the timeliness, completeness, and accuracy of data submitted to the CCSR (13). Our risk model was developed using this updated and reliable database with improved data quality and representativeness. Therefore, our model should better reflect current cardiac surgery practice in China. In candidate predictor variables selection process, we used an approach similar approach that was similar to STS risk model. Potential predictor variables were screened for their overall frequency, coding concerns, clinically value and presence on existing CABG risk model. In final variables selection process, we used bootstrap variable selection in this study, rather than the traditional single backward stepwise elimination method. The advantages of bootstrap resampling have previously been described (15, 16) , and they include an improved ability to assess the strength and stability of risk factors and improved selection of independent predictors and elimination of noise variables. The high calibration and discrimination power of the model in our test and validation samples suggest we identified true independent risk factors for postoperative in-hospital mortality. Most of the predictors included in our model are consistent with existing literatures on cardiac surgery risk (5, 6, 17). We updated the original SinoSCORE model with seven new risk factors, including myocardial infarction in the past 21 days, combined cardiac surgery other than valve surgery, angina, prior cardiac surgery, prior PCI, prior cerebrovascular accident, and female 14

sex. This increase in model predictors is likely due, in part, to the increased sample size in our study and differences in variable selection methods. We also updated the SinoSCORE by using creatinine clearance, rather than serum creatinine, in our analysis. Creatinine clearance is a better indicator of renal dysfunction and better predictor of mortality (18). A similar modification was also made when updating the EuroSCORE II model(5). The further classification of renal dysfunction (creatinine clearance ≥50 mL/min to <80 mL/min, <50 mL/min or on dialysis) may also have contributed to the better discriminatory power of the new model. Preoperative atrial fibrillation

diabetes

disease vessel

and peripheral vascular disease are indicators in STS

risk model(6) that were not select in our final model. All these variables appearing in less than 75% of the random bootstrap samples (Figure 2). This could be explained by that these variables may not be a stable independent predictor of early mortality in our cohort. Unlike STS risk model or EuroSCORE II model, body mass index, a useful measurement of body habitus, was used as candidate variable. And BMI< 18.5 kg/m2 was remained in our final risk model, suggesting that underweight or undernourished could be a potential predictor of surgical mortality. By contrast, the original SinoSCORE model was developed using 2007-2008 data and may not reflect current conditions. When originally developed, the SinoSCORE model achieved a Cstatistic of 0.78 in its own 2007-2008 validation sample. In our 2016 validation sample, the Cstatistic for the SinoSCORE model was 0.75, which was lower than our new risk model (0.77). In all three samples, the Hosmer-Lemeshow test was statistically significant (P<0.05), indicating 15

that the SinoSCORE model may not be adequately calibrated for the contemporary patient cohort. Improvements in cardiac surgical practice quality and changes in surgical candidate profiles in mainland China over the last decade likely contributed to changes in the predictors of surgical mortality over time. The modifications we made to variable definitions and the addition of seven new predictors to the risk model also could have contributed to the improved performance of our model. Our risk model also achieved better calibration and discrimination results in our cohort than did the EuroSCORE II model (Table 3), which was developed as a global standard for cardiac surgical risk calculations. This indicates that the patients undergoing CABG in China may have their own specific features, and our risk model may serve as a better assessment tool for the evaluation of CABG mortality in the Chinese population. Because no risk factors in our model were specific to China, and cardiac surgical practice in China is more similar to the practices of Asian countries and developing nations than to Western countries, our risk model has potential utility for other Asian or developing countries. Validation in additional studies and regions is needed.

Limitations Our study has several limitations. First, our primary outcome was in-hospital mortality after CABG. Although 30-day mortality is often used as an endpoint for risk models in Western countries, it is not readily available in the CCSR database, and collecting such information could 1，

significantly increase registry costs with unclear benefit. Additionally, the length of postoperative hospital stay is longer in China than in the United States (2), and death outside the hospital within 30 days of surgery is uncommon in China because of the lack of post-acute care services and limited insurance coverage for out-of-hospital care. Second, the new model was not validated using external data resources. The CCSR database is currently the only available resource with high quality data on CABG patients in China. For this reason, we considered the 2016 data separately in our study and used these data to create the validation sample. Our new risk model developed using 2013-2015 data achieved similar satisfactory results in the validation process with the latest data in the CCSR database. Third, our risk model was procedure specific. CABG only accounts for 40%-50% of the adult cardiac surgery volume in China (13). Future work on other cardiac surgery risk models or non-procedure-specific risk models is needed. Fourth, our model may miss some risk variables that were not included in our database used to develop the model. Future improvement in our model is necessary.

Conclusion In conclusion, our 21-risk factor logistic regression model for prediction of in-hospital mortality after CABG had good prediction performance and was able to identify high-risk patients before operation. The use of the CCSR database and the bootstrap variable selection method strengthened the risk model. This risk prediction model provides a basis for improved assessment of CABG mortality risk in the Chinese population. 17

REFERENCES 1.

Lozano R, Naghavi M, Foreman K et al. Global and regional mortality from 235 causes

of death for 20 age groups in 1990 and 2010: A systematic analysis for the global burden of disease study 2010. Lancet 2012;380(9859):2095-2128. 2.

Zheng Z, Zhang H, Yuan X et al. Comparing outcomes of coronary artery bypass

grafting among large teaching and urban hospitals in china and the united states. Circulation Cardiovascular quality and outcomes 2017;10(6). 3.

Paiement B, Pelletier C, Dyrda I et al. A simple classification of the risk in cardiac

surgery. Can Anaesth Soc J 1983;30(1):61-68. 4.

Nashef SA, Roques F, Michel P, Gauducheau E, Lemeshow S, Salamon R. European

system for cardiac operative risk evaluation (euroscore). Eur J Cardiothorac Surg 1999;16(1):913. 5.

Nashef SA, Roques F, Sharples LD et al. Euroscore ii. Eur J Cardiothorac Surg

2012;41(4):734-744; discussion 744-735. 6.

Shahian DM, O'Brien SM, Filardo G et al. The society of thoracic surgeons 2008 cardiac

surgery risk models: Part 1--coronary artery bypass grafting surgery. Ann Thorac Surg 2009;88(1 Suppl):S2-22. 7.

Neumann F, Sousa-Uva M, Ahlsson A et al. 2018 esc/eacts guidelines on myocardial

revascularization. Eur Heart J 2018.

18

8.

Zheng Z, Zhang L, Li X, Hu S, Chinese CRS. Sinoscore: A logistically derived additive

prediction model for post-coronary artery bypass grafting in-hospital mortality in a chinese population. Front Med 2013;7(4):477-485. 9.

Ma X, Wang Y, Shan L et al. Validation of sinoscore for isolated cabg operation in east

china. Sci Rep 2017;7(1):16806. 10.

Culler S, Kugelmass A, Brown P, Reynolds M, Simon A. Trends in coronary

revascularization procedures among medicare beneficiaries between 2008 and 2012. Circulation 2015;131(4):362-370; discussion 370. 11.

Diseases CNCfC. Report on cardiovascular diseases in china 2016 (in chinese). .

Encyclopedia of China Publishing House;2016. p. 162-167. 12.

Bai Y, Wang L, Guo Z et al. Performance of euroscore ii and sinoscore in chinese

patients undergoing coronary artery bypass grafting. Interact Cardiovasc Thorac Surg 2016;23(5):733-739. 13.

Rao C, Zhang H, Gao H et al. The chinese cardiac surgery registry: Design and data

audit. Ann Thorac Surg 2016;101(4):1514-1520. 14.

Sterne JA, White IR, Carlin JB et al. Multiple imputation for missing data in

epidemiological and clinical research: Potential and pitfalls. BMJ (Clinical research ed) 2009;338:b2393. 15.

Ariyaratne TV, Billah B, Yap CH et al. An australian risk prediction model for determining

early mortality following aortic valve replacement. Eur J Cardiothorac Surg 2011;39(6):815-821. 19

16.

Blackstone EH. Breaking down barriers: Helpful breakthrough statistical methods you

need to understand better. J Thorac Cardiov Sur 2001;122(3):430-439. 17.

Billah B, Reid CM, Shardey GC, Smith JA. A preoperative risk prediction model for 30-

day mortality following cardiac surgery in an australian cohort. Eur J Cardiothorac Surg 2010;37(5):1086-1092. 18.

Walter J, Mortasawi A, Arnrich B et al. Creatinine clearance versus serum creatinine as

a risk factor in cardiac surgery. BMC Surg 2003;3:4.

20

Table 1. Patient characteristics for the training, test, and validation samples No. (%) Overall

Training

Test

Validation

(2013-2016)

(2013-2015)

(2013-2015)

(2016)

(N=56776)

(N=31297)

(N=10432)

(N=15047)

<60

20600 (36.3)

11638 (37.2)

3816 (36.6)

5146 (34.2)

60-69

25429 (44.8)

13900 (44.4)

4646 (44.5)

6883 (45.7)

≥70

10747(19.0)

5759 (18.4)

1970 (18.9)

3018 (20.1)

Female

13974 (24.6)

7717 (24.7)

2607 (25.0)

3650 (24.3)

No health insurance

10511 (18.5)

4867 (15.6)

1624 (15.6)

4020 (26.7)

918 (1.6)

508 (1.6)

176 (1.7)

234 (1.6)

≥18.5 to <24

21171 (37.3)

11731 (37.5)

3921 (37.6)

5519 (36.7)

≥24

34687 (61.1)

19058 (60.9)

6335 (60.7)

9294 (61.8)

5130 (9.0)

2889 (9.2)

927 (8.9)

1314 (8.7)

Hypertension

34147 (60.1)

18651 (59.6)

6265 (60.1)

9231 (61.3)

Smoking

31503 (55.5)

17145 (54.8)

5669 (54.3)

8689 (57.7)

COPD

722 (1.3)

382 (1.2)

147 (1.4)

193 (1.3)

Extracardiac arteriopathy

2064 (3.6)

1166 (3.7)

414 (4.0)

484 (3.2)

Prior cerebrovascular accident

4179 (7.4)

2237 (7.2)

728 (7.0)

1214 (8.1)

I or II

34904 (61.5)

19779 (63.2)

6667(64.0)

8458(56.2)

III

19949 (35.1)

10500 (33.6)

3439 (33.0)

6010 (39.9)

Variable

Age, y

Body mass index, kg/m2 <18.5

Insulin-dependent diabetes mellitus

NYHA class

21

19231922 IV

1018 (3.3)

326 (3.1)

579 (3.8)

1724 (3.0)

931 (3.0)

306 (2.9)

429 (2.9)

None

46007 (81.0)

24669 (78.8)

8304 (79.6)

13034 (86.6)

<21 d

1979 (3.5)

1219 (3.9)

359 (3.4)

401 (2.7)

≥21 d

8790 (15.5)

5409 (17.3)

1769 (17.0)

1612 (10.7)

41361 (72.8)

24091 (77.0)

8081 (77.5)

9189 (61.1)

4395 (7.7)

2533 (8.1)

775 (7.4)

1087 (7.2)

Left main disease

14838 (26.1)

8369 (26.7)

2823 (27.1)

3646 (24.2)

Three-vessel disease

35079 (61.8)

19587 (62.6)

6606 (63.3)

8886 (59.1)

<35

738 (1.3)

386 (1.2)

171 141(1.35)

211 (1.4)

≥35 to <45

3614 (6.4)

1995 (6.4)

601 (5.8)

1018 (6.8)

≥45 to <55

9198 (16.2)

5220 (16.7)

1751 (16.8)

2227 (14.8)

≥55

43226 (76.1)

23696 (75.7)

7939 (76.1)

11591 (77.0)

Prior cardiac surgery

668 (1.2)

367 (1.2)

147 (1.4)

154 (1.0)

Non-elective surgery

1307 (2.3)

771 (2.5)

264 (2.5)

272 (1.8)

Combined valve surgery

5635 (9.9)

3109 (9.9)

1033 (9.9)

1493 (9.9)

Combined surgery except valve

3669 (6.5)

1639 (5.2)

551 (5.3)

1479 (9.8)

4653(8.2)

2164(6.9)

762 (7.3)

1727 (11.5)

CC ≥50 to <80

21615 (38.1)

11961 (38.2)

3914 (37.5)

5740 (38.1)

CC ≥80

30508 (53.7)

17172 (54.9)

5756 (55.2)

7580 (50.4)

662 (1.2)

375 (1.2)

133 (1.3)

154 (1.0)

(3.4) Preoperative atrial fibrillation Prior myocardial infarction

Angina Prior PCI

Left ventricle ejection fraction, %

Renal function, mL/min CC <50 or on dialysis

Critical preoperative state

22

CC, creatinine clearance; COPD, chronic obstructive pulmonary disease; NYHA, New York Heart Association; PCI, percutaneous coronary intervention.

23

Table 2. Final risk model for in-hospital mortality based on training sample Regression Risk Factor

Score P value

OR (95% CI)

Coefficient

Points

Age, y 60-69

0.3763

<0.001

1.46 (1.17-1.81)

3

≥70

0.7393

<0.001

2.09 (1.63-2.69)

5

0.3970

0.062

1.49 (0.98-2.26)

3

III

0.6149

<0.001

1.85 (1.54-2.22)

4

IV

1.2478

<0.001

3.48 (2.62-4.62)

9

Female

0.2533

0.006

1.29 (1.07-1.54)

2

Prior myocardial infarction <21 d

0.5025

0.002

1.65 (1.21-2.25)

4

Critical preoperative state

0.7309

<0.001

2.08 (1.39-3.11)

5

CC ≥50 to <80

0.1787

0.079

1.20 (0.98-1.46)

1

CC<50 or on dialysis

0.8216

<0.001

2.27 (1.75-2.95)

6

COPD

0.5051

0.033

1.66 (1.04-2.64)

4

Angina

0.4807

<0.001

1.62 (1.30-2.02)

3

<35

1.3029

<0.001

3.68 (2.45-5.53)

9

≥35 to <45

0.7301

<0.001

2.08 (1.61-2.67)

5

≥45 to <55

0.5250

<0.001

1.69 (1.38-2.07)

4

Prior cardiac surgery

1.4751

<0.001

4.37 (2.99-6.39)

10

Non-elective surgery

1.0411

<0.001

2.83 (2.06-3.89)

7

Body mass index <18.5 kg/m2 NYHA class

Renal function, mL/min

Left ventricle ejection fraction, %

24

Combined valve surgery

0.9686

<0.001

2.63 (2.14-3.24)

7

Combined surgery except valve

0.7131

<0.001

2.04 (1.59-2.62)

5

Prior cerebrovascular accident

0.3135

0.016

1.37 (1.06-1.77)

2

Prior PCI

0.3348

0.010

1.40 (1.08-1.80)

2

CC, creatinine clearance; CI, confidence interval; COPD, chronic obstructive pulmonary disease; NYHA, New York Heart Association; OR, odds ratio; PCI, percutaneous coronary intervention.

25

Table 3. C-statistics and Hosmer-Lemeshow P values for risk models applied to the training, test, and validation samples New risk model

Original SinoSCORE

EuroSCORE II

C-statistic

H-L

C-statistic

H-L

C-statistic

H-L

(95% CI)

P value

(95% CI)

P value

(95% CI)

P value

0.79 (0.77-0.80)

0.30

0.75 (0.73-0.77)

<0.01

Sample

Training

0.77 (0.75-

(2013-2015)

<0.01 0.79)

Test

0.77 (0.730.79 (0.76-0.82)

0.13

0.77 (0.74-0.81)

<0.01

(2013-2015)

<0.01 0.80) 0.76 (0.74-

Validation 0.78 (0.76-0.81)

0.29

0.75 (0.72-0.78)

(2016)

<0.01

<0.01 0.79)

CI, confidence interval; H-L, Hosmer-Lemeshow.

2，

Table 4. Predicted risk of in-hospital mortality associated with individual risk scores in training sample Total Risk Score Predicted Risk (%) 0 <1 1 <1 2 <1 3 <1 4 <1 5 1 6 1 7 1 8 1 9 1 10 1 11 1 12 2 13 2 14 2 15 2 16 3 17 3 18 4 19 4 20 5 21 5 22 6 23 7 24 8 25 9 26 10 27 12 28 13 29 15 30 17 31 19 32 21 33+(1.0% of all patients) 35

27

FIGURE LEGENDS Figure 1. Flow chart of data and final sample division.

Figure 2. Frequency of each candidate variable selected in the 2000 bootstrap samples. CC, creatinine clearance; COPD, chronic obstructive pulmonary disease; NYHA, New York Heart Association; LVEF, left ventricle ejection fraction; PCI, percutaneous coronary intervention; BMI, Body mass index; MI, myocardial infarction.

Figure 3. Area under the receiver operating characteristic curve for risk models applied to the training, test, and validation samples.

Figure 4. Observed vs predicted in-hospital mortality by deciles in the training (left panel), test (middle panel), and validation (right panel) sample for the risk model.

Figure 5. Risk stratification by risk scores. For the training, test, and validation samples, respectively, the high-risk group includes 10.1%, 10.3%, and 11.8% of the patients; the average-risk group includes 77.5%, 77.4%, and 75.7% of the patients; and the low-risk group includes 12.3%, 12.3%, and 12.5% of the patients.

28