Supervised Machine Learning Models For Prediction Of Weaning Success In Veno-arterial Extracorporeal Membrane Oxygenation Based On Continuous Hemodynamic Parameters

Veno-arterial extracorporeal membrane oxygenation (VA ECMO) provides temporary cardiac support during cardiogenic shock but is resource-intensive and associated with significant mortality. This study aims to develop a machine learning model to predict weaning success using continuous hemodynamic parameters early in the VA ECMO treatment. The parameters used to predict weaning success include heart rate, pulse pressure, mean arterial pressure, central venous pressure, vasoactive inotropic score, lactate levels and ECMO flow from the first three days of the ECMO run. A total of 108 ECMO runs were included. Features were extracted by calculating the slope, standard deviation, and mean for each day and for the full three-day period. Machine learning models, specifically random forest, KNN, logistic regression, gradient boosting, and support vector machine, were developed and compared based on RMSE, MAE, and R². KNN and random forest models showed the best results, with RMSE of 0.48 and 0.48, MAE of 0.45 and 0.42,