Fig. 5

Machine learning models predicting complications based on common risk factors, and by adding a combination of Ang-2 and body impedance electrical analysis-derived measurements to the risk factors. A shows ROC-AUC of machine learning (ML) algorithms to predict acute kidney injury (AKI), low oxygenation index (P/F-ratio), dependence on vasoactive drugs, mortality, and postoperative dependence on ECMO from standard risk factors (grey curve). The red curves represent an augmented feature set, adding the phenotype of capillary leak to the standard risk factors (with Ang-2 and body impedance electrical analysis-derived measurements), thus showing a significant improvement in predicting the respective postoperative complication (blue dotted lines: random selection). B shows different cross-validation methods and their respective ROC-AUC to predict complications derived from ML. Models could maintain their robust performance in terms of ROC-AUC regardless of the splitting approaches used in various cross-validation strategies. Color from dark to light (counter-clockwise) in each segment: repeated train-test splitting validation, standard cross-validation, stratified cross-validation, leave-one-patient-out cross-validation, and leave-one-surgery-out cross-validation