Predicting Survival for Veno-Arterial ECMO Using Conditional Inference Trees-A Multicenter Study

Area of Science:

• Cardiology
• Critical Care Medicine
• Medical Informatics

Background:

• Veno-arterial extracorporeal membrane oxygenation (VA-ECMO) therapy, despite increased use, has high mortality rates.
• Accurate, timely survival predictions are crucial for clinical decision-making before initiating VA-ECMO.

Purpose of the Study:

• To develop and validate a user-friendly prognostic model for predicting in-hospital mortality in VA-ECMO patients.
• To assess the performance of machine learning models using conditional inference trees for VA-ECMO survival prediction.

Main Methods:

• A multicenter retrospective study involving 837 patients (2007-2019).
• Development and validation of prognostic models using conditional inference trees with small and comprehensive variable sets.
• Model performance evaluated using Area Under the Curve (AUC), Brier score, and error rate.

Main Results:

• Models demonstrated moderate predictive accuracy in derivation cohorts (AUC 0.70-0.71).
• Error rates were comparable between small (35.79%) and comprehensive (35.35%) datasets.
• External validation showed AUCs of 0.60 (small tree) and 0.63 (comprehensive tree), with significant variability between validation sets.

Conclusions:

• Conditional inference trees can enhance clinical decision-making for VA-ECMO patients.
• These models offer a degree of accuracy in mortality prediction and prognostic stratification.
• Readily available variables are sufficient for developing effective prognostic tools.