Development Of Risk Prediction Model For Heart Valve Surgery: A Multi-center Study

【Objective】This study analyses the clinical character of patients who underwent heart valvesurgery in four cardiac surgery centers in China from Jan.1st,2006to Dec.31,2008retrospectively. The performance of EuroSCORE Ⅱ, STS2008risk model andSinoSCORE for these patients was evaluated. A new adult heart valve operation riskprediction model for Chinese was established upon the data of these patients. Then therisk prediction model was validated in each of the four cardiac center.【Methods】1． Current situation of Chinese heart valve surgery-a mulitycenter analysis1． Patients who underwent heart valve surgery in four cardiac surgicalcenters between Jan.1,2006and Dec.31,2008were included in this study.323patients whose age below18years and79patients who had missing value of keyinformation were excluded. The key information of patients included as surgery types,sex, operative procedure and status of discharge. Patients underwent valve surgerycombined with CABG, arotic surgery, congenital heart surgery were not excluded andresulted a final study group of11774patients.2． The patients were divided into3and4groups according to the years ofoperation and institutions, respectively. Patients’ clinical characters were comparedamong groups accordingly.2. Validation of cardiac valve surgery risk models in our patients set:1． The patient set were the same as that of part one. The endpoint event wasdefined as hospital death. Cardiac surgery risk models listed blew were selected toevaluate their performance in this patient set: EuroSCOREⅡ, STS2008, SinoSCORE.2． Value of risk factors for every model was obtained for each patient includedin the study. The predicted operative mortality for each patient was calculated by thelogistic regression formular of risk models.3． Model discrimination was evaluated by c-index and model calibration wasevaluated by Hosmer-Lemeshow good-of-fit test. Observed/Expected mortality (O/E ratio) was calculated to determine the prediction accuracy of each model.3. Establishment of risk prediction model and risk score for patients underwentadult heart valve surgery:1． The patient set were the same as that of part one. Potential risk factors forhospital mortality were selected. The endpoint event of the study was hospitalmortality.2． Univaries logistic regression was performed for all potential risk factors attheir appropriate coding format to reveal their relationship with hospital mortality.Examination for complete separation and multi-collinearity were also performed.3． The patients set was splited randomly into development set (70%) andvalidation set (30%), and then the risk model was developed using a logisticregression model with backward selection upon the development set. Risk modelderived from the development set was tested in the validation set.4． Risk score was established upon the logistic regression model. The list ofpredicted mortality to risk score was calculated. Then, this addictive risk model wasvalidated in both development set and validation set.4. Validation of cardiac valve surgical risk prediction model in patients of eachcardiac surgical center.1. patient set in this part of research was as the patient set of first part. Hospitaldeath was defined as the endpoint.2. Validation of Logistic risk model as well as the additive model established inprevious study, STS2008, EuroSCORE Ⅱ and SinoSCORE were perform forpatients of each cardiac center.【Results】1. Current situation of Chinese heart valve surgery-a mulitycenter analysis:1． Comparation of patients’ clinical character between operation years. There isan annual increase in case of valve surgery. There is no significant difference aboutweight and height distribution according to the operative year. Asymptomatic patientsand patients with NYHA class Ⅰ/Ⅱ increased by year. Smoking, hypertension, hyperlipidemia also increased by year in the patients set. Preoperative treatment withdopamine, digitalis, and diuretics also became more popular in recent years.Distribution of surgical procedures was relatively constant except for the annualdecreasment of triscupid repairment. The useage of post-operative adrenaline,norepinephrine and phenylephrine were decreased but the useage of phosphodiesteraswas increase. Percentage of patients who had blood transfusion increased but theaverage amount of blood per patients decreased. The hospital mortality, ventilationtime, time of intensive care unit also decreased by year.2． Comparation between cardiac surgery centers. Gender composition of fourinstitutions was similar and there are more yong patients in Center A. Average heightand weight of patient were significant lower in Center D than those of other Centers.Proportion of asymptomatic patients was higher in Center A than that of other Centers.Mitral stenosis, mitral insufficiency and aortic insufficiency consiste the majority ofvalve lesions in all the four Centers while the composition of valve lesion differs fromeach other. Hospital mortality was high in Center C and low in Center A while theother two was in the middle position. There are differences between institutions forincidence of other operative complications.2. Validation of published cardiac valve surgery risk models with our patientsset.1． Model calibration. There are237death in the whole patients set whichresulted a mortality of2%. Expected mortality by the EuroSCOREⅡ, STS2008,SinoSCORE was2.63%,1.27%,2.83%, respectively. Observed mortality/Expectedmortality (O/E ratio) of EuroSCOREⅡ, STS2008, SinoSCORE was0.77,1.58,0.71,respectively. All three cardiac surgery risk models showed poor performance withpatients set of this study. EuroSCOREⅡ and SinoSCORE overestimated hospitalmortality and STS2008underestimated hospital mortality. With theHosmer-Lemeshow good-of-fit test, EuroSCORE Ⅱ had the best performance amongthree models while the STS2008had the worst performance for the whole patients set.2． Model discrimination. C-index of EuroSCOREⅡ, STS2008, SinoSCORE was0.715,0.685,0.752, respectively. C-index of EuroSCOREⅡand SinoSCOREexceeds0.7means that EuroSCOREⅡand SinoSCORE had better discriminate powerthan STS2008had.3． Single valve surgery vs. multiple valve surgery. SinoSCORE showed goodmodel discrimination for either Single valve surgery (C-index0.761) or multiplevalve surgery (C-index0.746). EuroSCOREⅡand STS2008risk models were ofbetter discrimination for single valve surgery (C-index0.739,0.721) than that formultiple valve surgery (C-index0.708,0.660). Hosmer-Lemeshow good-of-fit testdemonstrated that only STS2008risk model for the single valve surgery group havegood calibration (P>0.05).3. Establishment of risk prediction model for adult heart valve surgery:1． The final risk prediction model included16risk factors: age (OR:1.21),female (OR:1.38), body surface area (OR:1.78), smoking (OR:1.54), hypertension(OR:1.64), renal failure (OR:6.59), previous heart valve surgery (OR:3.86), leftventricular ejection fraction (OR:1.54), severe pulmonary hypertension (OR:7.27),Severe tricuspid regurgitation (OR:1.98); NYHA class Ⅲ (OR:3.01); NYHA class Ⅳ(OR:7.04), preoperative critical state (OR:4.82), number of conary artery disease(OR:1.56); multiple valve surgery (OR:1.96), combined CABG (OR:7.98).2． The new developed heart valve surgery risk model showed good calibrationand discriminative power for the development set, validation set and the wholepatients set, with the Hosmer–Lemeshow good-of-fit test’s P value of0.46,0.16and0.63, respectively. C-index of the risk model in development set, validation set and thewhole patients set was0.827,0.763and0.810, respectively. The risk model showgood calibration and discrimination power in the whole study population. Thevalidation study of the risk model in the patient set of each cardiac surgery center alsodemonstrated that the model have excellent prediction ability.3． Based upon the logistic regression model, an additive model was established.Risk score of the additive model listed blow: age (above65years):0.5score per year;femal:1score; body surface area<1.5m2:1.5score, smoking:1score, hypertension: 1.5score, renal failure:5score, previous heart valve surgery:2score, left ventricularejection fraction<40%:4score, severe pulmonary hypertension:5.5score, severtriscupid insuffiency:2score; NYHA class Ⅲ:3score; NYHA class Ⅳ:5.5score,preoperative critical state:4score, number of conary artery disease:1score per vessel,multiple valve surgery:2score, combined CABG:6score. For bedside useage,corresponding predicted mortality to the full range of risk score was also calculated.The additive model’s prediction accuracy was confirmed by validation on the wholepatients set, with O/E ratio of1.11, C-index of0.808and Hosmer–Lemeshowgood-of-fit test’s P value of0.28. The validation of this addictive model for thedevelopment set, validation set and every cardiac surgery center demonstrated thatthe additive model have excellent prediction ability.4. Validation of cardiac valve surgical risk prediction model in patients of eachcardiac surgical center.1. Both the logistic risk model and additive risk model have excellent predictionpower for patients of each cardiac surgical center.2. The new established valve surgical risk model was more suitable forestimating operative mortality after valve surgery than STS2008,EuroSCOREⅡand SinoSCORE.【Conclusion】1. This project analyses the clinical character of patients who underwent heartvalve surgery in four cardiac surgery centers from Jan.1st,2006to Dec.31,2008retrospectively. Then the annual changes and hospital level difference of patients’clinical character were identified.2. EuroSCOREⅡ, STS2008, SinoSCORE were validated by the patients setfrom four Chinese cardiac surgical center. These three cardiac risk models weredemonstrated of poor calibration and reasonable discrimination. STS2008risk modelwas the only risk model suitable for single valve surgery with good calibration andreasonable discrimination.3. A new adult cardiac risk prediction model and corresponding addictive risk model was developed based on our patients set. This risk model was developed for alladult cardiac valve surgery in China. The inner validation of this risk model showedexcellent performance.4. Validation study of the new established risk prediction model was performedfor patients of each cardiac center. The risk predition model showed high accuracy inestimating the operative mortialy, and it is better than the STS2008, EuroSCOREⅡand SinoSCORE.