Page 19 - ANZCP Gazette May 2023

Patients and Methods
Nine cardiac surgical centers in Australia and New Zealand prospectively collected data using the ANZCPR as previously described (3). Intraoperative CPB and physiological data were collected every 20-60 seconds during CPB. Pressure artefacts and periods of partial CPB were automatically excluded from analysis programmatically as part of the routine data collection process at each center. Institutional Ethics Review Board approval was obtained at each participating center, and this study was approved by the Southern Adelaide Clinical Human Research Ethics Committee (386.15), the ANZCPR Steering Committee and registered on the Southern Adelaide Local Health Network Quality Register Library (2327). Centers collected data on 34,687 adult patients undergoing cardiac surgery using CPB between January 2011 and December 2020. Data collected prior to 2011 was excluded to provide all ANZCTS variables required for the analysis. Patients were excluded if they were missing 30-day mortality data (1018), underwent circulatory arrest (932), had a minimum nasopharyngeal temperature <25oC (729), or were missing intraoperative electronic perfusion (either blood flow, blood pressure or blood gas) data (1863). In total, 8.3% of the patients were excluded from the analysis because of missing observations. Imputation using mean replacement of 20 chained iterations was performed for residual missing values (<1% of dataset). A total of 30,145 patients were included in the study. The ANZCPR is listed in the Australian Commission on Safety and Quality in Health Care register of clinical registries (https://www.safetyandquality.gov.au/publications- and-resources/australian-register-clinical-registries). Clinical data definitions were based on the ANZSCTS registry (4). Complete ANZCPR variable definitions are available (http://www.anzcpr.org).
Statistical software package Stata (version 15) was used for the analyses. The a-priori selection of risk factors was made by including variables reported by Billah et al (1). CPB parameter selection was determined by evaluating each variables association with 30-day mortality. Variables were selected with a significant association (p<0.05) including; intraoperative red blood cell transfusion (pre, during or post CPB), minimum and maximum blood glucose, minimum nasopharyngeal temperature, minimum oxygen delivery, and minimum arterial pCO2, and minute duration variables including oxygenator arterial outlet temperature >36.5oC, CPB, aortic cross clamp, arterial flow <1.6 l/min/m2, arterial flow < 1.8 l/min/m2, venous oxygen saturation < 60%, mean arterial pressure <50mmHg and mean arterial pressure <40mmHg. The linearity of the relationship between continuous CPB variables and 30-day mortality was assessed using LOWESS plots. CPB duration and minimum oxygen delivery were converted to quintiles due to non-linearity. Subsequent model development
was performed according to the method described by Billah et al (1). Bootstrap methods along with automated variable selection procedures were used to develop a parsimonious model using multiple logistic regression (5). The data were randomly divided into two sets: model creation set (n = 15,073 50% of the total patients) and model validation set (n = 15,072). In all, 1000 bootstrap samples (each of size of 15,073) were selected from the model creation set and a multiple logistic regression model was developed for each of the 1000 samples. The number of times each variable was identified as significant in 1000 bootstraps was recorded and then ranked. In each bootstrap, a p-value of 0.05 or less was regarded as significant in the variable selection. Then we developed seven multiple logistic regression models with the variables selected in at least 100%, 90%, 80%, 70%, 60%, 50% and all variables of bootstrap samples. In the creation set, the Bayesian Information Criteria (BIC) (6) and the prediction mean square error (MSE) were calculated for each of these models and the final model was selected based on the value obtained for the BIC and MSE. The prediction performance of the selected model was assessed by calculating average ROC, Hosmer—Lemeshow p-value and MSE from a multifold (100) validation on the validation set. Multicollinearity was assessed by the variance inflation factor.
A base comparative model was created using the variables reported by Billah et al (1). To assess the diagnostic accuracy of each model, we calculated the area under the receiver- operating characteristic curve (AUROC) in the validation dataset. Similarity of diagnostic accuracy for 30-day mortality of the models was determined by using the chi- square test to compare the 2 AUROC curves. The extent of over and underestimation associated with the model was graphically described using calibration plots.
Results
In total, 30,145 patients were included in the study (15,073; creation dataset, 15,072; validation). Patient preoperative and intraoperative characteristics were similar between training and validation cohorts (Table 1). Isolated CABG accounted for 50% of procedures, isolated valve repair or replacement 20%, valve repair or replacement + CABG 11%, and other procedures 19%. Most of the procedures were elective (70%) in patients <70 years of age (60%) with normal left ventricular function (82%). Predominant risk factors included hypercholesterolemia (60%) and body mass index >25 kg/m2 (70%). Overall, 735 (2.4%) patients died within 30 days of surgery with similar rates of 30-day mortality between training and validation datasets (2.3% vs 2.6% respectively). Rates of other major adverse outcomes were also similar (Table 2).
MAY 2023 | www.anzcp.org 16