Abstract 13293: Predicting Successful ECMO Decannulation - A Novel Machine Learning Approach

Introduction: Extra Corporeal Membrane Oxygenation (ECMO) has emerged as a crucial life support intervention, yet the decision-making process for safe decannulation remains challenging due to a paucity of data.

Methods: 199 patients who underwent venoarterial ECMO for cardiogenic shock between 2015 and 2021 were identified. Demographic, hemodynamic, laboratory, and echocardiographic data obtained within 24 hours of decannulation were collected. Successful decannulation was defined as survival without relapse to mechanical circulatory support or heart transplant within 30 days. The dataset was randomly split into 80/20 train-validation splits, and four machine learning models were employed. 5-fold cross-validation and error analysis were performed. Feature importance was derived from the best performing model.

Results: 103 patients were successfully weaned off ECMO. Table 1 provides patient characteristics, which were used as features in the models. Among the models tested, Random Forest (RF) exhibited the highest performance (accuracy 85.0%, precision 90.0%, recall 85.7%, AUROC 0.94). Figure 1 depicts the performance for each model as well as the ranked feature importance and Receiver Operating Curve for the RF model. The most important features were mixed venous oxygen saturation, systolic blood pressure, ECMO flow rate and pulmonary diastolic blood pressure.

Conclusion: This study demonstrates the successful application of machine learning in predicting ECMO weaning outcomes. Future research will add features and patients to enhance model performance with the goal of developing a clinical tool to assist physicians caring for ECMO patients.