An Australian risk prediction model for determining early mortality following aortic valve replacement

Abstract

Objective: To develop a multivariable logistic risk model for predicting early mortality following aortic valve replacement (AVR) in adults, and to compare its performance against existing AVR-dedicated models. Methods: Prospectively collected data from the Australasian Society of Cardiac and Thoracic Surgeons (ASCTS) database project were used. Thirty-five preoperative variables from AVR literature were considered for analysis by chi-square method and multiple logistic regression. Using the bootstrap re-sampling technique for variable selection, five plausible models were identified. Based on models' calibration, discrimination and predictive capacity during n-fold validation, a final model, the AVR-Score, was chosen. An additive score, derived from the final model, was also validated externally in a consecutive cohort. The performance of AVR-dedicated risk models from the North West Quality Improvement Program (NWQIP) and the Northern New England Cardiovascular Study (NNE) groups were also assessed using the receiver operating characteristic (ROC) curve and the Hosmer-Lemeshow (H-L) chi-square test. Results: Between July 2001 and June 2008, a total of 3544 AVR procedures were performed. Early mortality was 4.15%. The AVR-Score contained the following predictors: age, New York Heart Association class, left main disease, infective endocarditis, cerebrovascular disease, renal dysfunction, previous cardiac surgery and estimated ejection fraction. Our final model (AVR-Score) obtained an average area under ROC curve of 0.78 (95% confidence interval (CI): 0.76, 0.80) and an H-L p-value of 0.41 (p> 0.05) during internal validation, indicating good discrimination and calibration capacity. External validation of the additive score on a consecutive cohort of 1268 procedures produced an ROC of 0.73 (0.62, 0.84) and an H-L p-value of 0.48 (p> 0.05). The NWQIP and NNE risk models achieved acceptable discrimination of ROC of 0.77 (0.73, 0.81). However, both models obtained H-L p-values of 0.002 (p< 0.05), indicating a poor fit in our cohort. Conclusion: Existing AVR-dedicated risk models were deemed inappropriate for risk prediction in the Australian population. A preoperative risk model was developed using prospective data from a contemporary AVR cohort.