Missing Value Imputation Improves Mortality Risk Prediction Following Cardiac Surgery: An Investigation of an Australian Patient Cohort

Background The aim of this study was to evaluate the impact of missing values on the prediction performance of the model predicting 30-day mortality following cardiac surgery as an example. Methods Information from 83,309 eligible patients, who underwent cardiac surgery, recorded in the Australia and New Zealand Society of Cardiac and Thoracic Surgeons (ANZSCTS) database registry between 2001 and 2014, was used. An existing 30-day mortality risk prediction model developed from ANZSCTS database was re-estimated using the complete cases (CC) analysis and using multiple imputation (MI) analysis. Agreement between the risks generated by the CC and MI analysis approaches was assessed by the Bland-Altman method. Performances of the two models were compared. Results One or more missing predictor variables were present in 15.8% of the patients in the dataset. The Bland-Altman plot demonstrated significant disagreement between the risk scores (p<0.0001) generated by MI and CC analysis approaches and showed a trend of increasing disagreement for patients with higher risk of mortality. Compared to CC analysis, MI analysis resulted in an average of 8.5% decrease in standard error, a measure of uncertainty. The MI model provided better prediction of mortality risk (observed: 2.69%; MI: 2.63% versus CC: 2.37%, P<0.001). Conclusion ‘Multiple imputation’ of missing values improved the 30-day mortality risk prediction following cardiac surgery.