Predicting Postoperative Cardiac Events and Mortality for People with Kidney Failure Having Noncardiac Surgery

Key Points Three models developed specifically for people with kidney failure were externally evaluated in a distinct Canadian province. All three models performed well, with some improvement with the re-estimation of predictor coefficients. Models have the potential to improve clinical decision making, and future research should evaluate their clinical effect. Background Patients with kidney failure undergoing noncardiac surgery are at high risk of adverse cardiac events and mortality; however, existing perioperative risk prediction tools for these outcomes are not valid in these patients. Recently, three models were developed from a kidney failure cohort in Alberta, Canada. In this study, we evaluated these Alberta models in a kidney failure cohort that had surgery in Manitoba, Canada. Methods The cohort included adults from Manitoba, Canada (18 years or older), with preexisting kidney failure (eGFR <15 ml/min per 1.73 m 2 or receiving maintenance dialysis) undergoing noncardiac surgeries between 2007 and 2019. The primary outcome was a composite of acute myocardial infarction, cardiac arrest, ventricular arrhythmia, and all-cause mortality within 30 days. The three models included an increasing number of variables: demographics and surgical characteristics (model 1), comorbidities (model 2), and preoperative albumin and hemoglobin (model 3). Model performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC), calibration, and Brier score on Manitoba data. This was evaluated by applying Alberta model coefficients for all three models to predict outcomes on Manitoba data and also by re-estimating the Alberta model predictor coefficients using logistic regression on Manitoba data. Results We identified 12,082 surgeries performed in 4175 participants; 569 outcomes were observed (4.7%). All three models performed well with both approaches, with AUC-ROC ranging from 0.821 (model 1) to 0.874 (model 3) using the models with Alberta coefficients. Calibration slopes were 1.32, 1.40, and 1.24 for models 1, 2, and 3, respectively. On refitting, AUC-ROC ranged from 0.830 (model 1) to 0.861 (model 3). Calibration slopes approximated one across all the re-estimated models. Brier scores remained <0.1 across all original and re-estimated models. Conclusions Our external validation study confirmed that the kidney failure specific postoperative outcome models developed in Alberta, Canada, performed well in a geographically distinct Canadian population. Future research should explore the performance of these models in different settings and evaluate their clinical effect with prospective implementation.