025 * NATIONAL ADMINISTRATIVE DATA PRODUCES A VERY ACCURATE RISK PREDICTION MODEL FOR SHORT- AND LONG-TERM MORTALITY FOLLOWING CARDIAC SURGERY

Abstract

Objectives: Various models exist to predict short-term risk-adjusted outcomes after cardiac surgery. Many models are calibrated using administrative data and are subject to controversy. This study compares predictive performance for short- and long-term mortality of a procedure-specific risk prediction model based on administrative NHS Health Episode Statistics (HES) data for England against the re-calibrated logistic EuroSCORE. Methods: Procedure-specific models were built on administrative data and tested using an independent dataset sampled at a later period. Models are applicable to patients having isolated CABG, isolated valve or combined CABG and valve surgery. Outcomes at hospital discharge and 1 year are linked with national death registrations (ONS). Comparisons between the models are conducted on a local cohort of patients between 2010 and 2013 using C-statistic for performance and Hosmer-Lemeshow for calibration across different risk strata. Results: A total of 84,791 patients in England and 1174 locally were identified. HES models have higher or comparable predictive performance compared to the re-calibrated logistic EuroSCORE (Table 1). Model calibration assessed via Hosmer-Lemeshow procedure also demonstrates good performance. C-statistics test for risk prediction models C-statistics test for risk prediction models Conclusion: The national administrative dataset has produced very accurate short- and long-term mortality prediction after cardiac surgery. Future iterations should explore further statistical methods and focus on developing procedure-specific algorithms.