Page 21 - ANZCP Gazette May 2023

Table 2. Postoperative outcomes in the creation and validation datasets.
Creation dataset Validation dataset
p
n=15073 n=15072
30-day mortality 349 (2.3%) 386 (2.6%) 0.17
Stroke 240 (2%) 213 (1%) 0.2
Acute kidney injury (RIFLE class risk or greater) 2030 (14%) 2121 (14%) 0.12
Mechanical ventilation > 48 hrs 1112 (8%) 1200 (8%) 0.052
Return to operating theatre 1045 (7%) 1128 (7%) 0.064
Intra-aortic balloon pump 163 (1%) 159 (1%) 0.82
Dialysis 444 (3%) 459 (3%) 0.61
Postoperative stay (days) 7 (6, 10) 7 (6, 10) 0.41
Data presented as n (%) or median (IQR). RIFLE: risk, injury, failure, loss or end stage.
A total of 15 preoperative and 14 intraoperative variables were included as covariates for the 30-day mortality logistic regression model. In 1000 bootstraps of the regression model, the number of times each variable appeared as significant was recorded (Table 3). Variables that were identified as significant in all bootstrap samples included age (group), red blood cell transfusion and CPB time (quintiles). Variables selected in at least 90% of the samples included ejection fraction estimate, preoperative dialysis and urgency of procedure. CPB mean arterial pressure <50 mmHg (minutes) was identified in 83%, CPB minimum oxygen delivery (quintiles) in 75%. CPB cardiac index <1.6 l/min/m2 (minutes) and respiratory disease were selected in at least 60%. Cardiogenic shock, NYHA class and critical preoperative state were selected in at least 50%. Other variables were selected as independent predictors in less than 50%. Using the results in Table
3, we developed six plausible risk prediction models. The average ROC, Hosmer—Lemeshow p-value and prediction MSE for each candidate model are reported in Table 4. The model with variables selected in at least 60% of the samples was selected as our final model based on firstly the minimum MSE and then the lowest BIC. The average 100-fold cross validation ROC (0.8214, 95% CI: 0.8001—0.8425) and Hosmer-Lemeshow p-value (0.2882) show good discrimination and calibration. The validation comparison including CPB parameters had a ROC 0.8306, which was significantly improved in comparison to the base model of preoperative risk factors only (ROC 0.7833), p<0.001 (see Figure 1 for ROC curves). A calibration plot of the validation model (Figure 2) shows strong calibration (intercept 1.000, slope 1.000, calibration in the large; 0.000, AUC 0.831).).
Table 3. Number of times each candidate variable was selected in 1000 bootstrap samples drawn from model creation dataset.
Risk factor Frequency
%
Age group 1000 100
Red blood cell transfusion 1000 100
CPB time (quintiles) 1000 100
Ejection fraction estimate
997
99.7
Preoperative dialysis 982 98.2
Urgency of procedure
978
97.8
CPB mean arterial press <50 mmHg (min)
828
82.8
CPB minimum oxygen delivery (quintiles) 746 74.6
CPB cardiac index <1.6 l/min/m2 (minutes) 641 64.1
Respiratory disease 614 61.4
Cardiogenic shock 578 57.8
NYHA class
554
55.4
Critical preoperative state 521 52.1
Peripheral vascular disease 441 44.1
Previous cardiac surgery 358 35.8
CPB mean arterial pressure <40mmHg (min) 329 32.9
Procedure type 325 32.5
Aortic cross clamp time 320 32
Cerebrovascular disease 256 25.6
CPB arterial blood temp >36.5oC (min) 224 22.4
CPB minimum arterial pCO2 182 18.2
Female 176 17.6
CPB cardiac index <1.8 l/min/m2 (minutes) 169 16.9
CPB minimum nasopharyngeal temperature 137 13.7
Hypercholesterolaemia 106 10.6
Body mass index >25 kg/m2 97 9.7
CPB minimum blood glucose 94 9.4
CPB maximum blood glucose
93
9.3
CPB venous saturation <60% (minutes) 51 5.1
CPB; cardiopulmonary bypass, NYHA; New York Heart Association
MAY 2023 | www.anzcp.org 18