Predictive Pre and Peroperative Factors for MODS-2 in Pediatric Cardiac Surgery

Overview

Pediatric cardiac surgery has a relatively high morbi-mortality. Despite great advances in surgical techniques, today the mortality rate is about 3% and morbidity is about 30-40%. Outcome has been related to demographic factors, like age; peroperative factors, like duration of cardiopulmonary bypass as well as postoperative factors like positive fluid balance. Willems et al defined a new score (MODS2), an outcome score combining either patient's death or a high postoperative morbidity. This morbidity is defined as minimum of 2 organ failures: either respiratory insufficiency, prolonged use of inotropic agents or renal insufficiency. The aim of this study is to identify pre and peroperative factors which are predictors of MODS2. Patients operated between 2008 and 2018 for pediatric cardiac surgery with cardiopulmonary bypass will be included. Variables extracted from our database will be: sex, ASA score, cyanotic cardiac pathology, redo surgery, RACH1 score, use of antifibrinolytic agents, aortic cross-clamping, deep hypothermic circulatory arrest, selective cerebral perfusion, red cell transfusion in the operating room, administration of fresh frozen plasma in the operating room, age, preoperative weight, weight difference between preop weight and weight at postop day 2, emergency surgery, duration of aortic cross clamping, duration of selective cerebral perfusion, duration of cardiopulmonary bypass, duration off deep hypothermic circulatory arrest, duration of surgery, minimal core temperature, cardiopulmonary priming volume, calculated hemodilution, use of red blood cells in the cardiopulmonary bypass priming, preoperative hemoglobin, preoperative hematocrit, preoperative platelet count, preop international normalized ratio, preop fibrinogen, preop creatinin, toal fluid balance, blood loss during surgery. A statistical analysis (see detailed description) will be used to establish a prediction model for MODS2. The variables describing best the MODS2 outcome will be retained.

Detailed statistical analysis: 5 multiple imputations via the mice R package will be performed, which is the most appropriate method for a risk model. We then take the mean of the imputed datasets in order to start the data mining models on one single dataset. All of the variables will be entered in the model. Before to run the data mining models, we will perform three transformations on the continuous variables: 1) standardization; 2) best normalisation via the bestNormalize R package and 3) taking the variable from its power 2 to its power 10. The dataset will be split into a training set (75% of the cases) and test set (25% of the cases). For all tested data mining models, we will use a 10-fold cross-validation method on the training set before applying the retained model on the test set. The following data mining models will be tested: 1) a regression tree, 2) a logistic regression (GLM), 3) a Neural Network (NN); 4) a Support Vector Machine (SVM); 5) a Random Forest (RF); 6) a Multivariate Adaptive Regression Spline model (MARS) and 7) a Non-Linear Support Vector Machine (SVM NL). The models will be drawn with the caret R package. The confusion matrix, reporting the sensibilities, specificities, accuracies will be drawn on the test set based on the models developed on the training set, and the calibration plot will be drawn for three model competitors. The R software (R Core Team, 2019), version 3.6.1. will be used to produce the results.