Enhancing Survival Prediction After Venoarterial Extracorporeal Membrane Oxygenation Using Machine Learning

In-hospital mortality after venoarterial extracorporeal membrane oxygenation (VA-ECMO) remains high. This study compared the performance of the Survival after Venoarterial ECMO (SAVE) score with machine learning (ML) models incorporating rich electronic medical record data to evaluate survival for patients on VA-ECMO support. We retrospectively reviewed adults undergoing VA-ECMO (2016-2022) at a single tertiary care center. The CatBoost algorithm was trained using leave-one-out cross-validation (LOOCV) on 74 extracted vital signs, laboratory values, and ventilator settings. Shapley Additive Explanations (SHAP) was used to identify key predictive features for logistic regression. Overall, 194 VA-ECMO patients (median age = 58 years, 36.6% female) were included, with 133 (69%) experiencing mortality. The SAVE score was compared to two predictive models: a pre-ECMO model (≤ 24 hours before cannulation) and an on-ECMO model (including up to the first 48 hours of ECMO). The LOOCV area under the receiver-operator characteristics curves (AUC) for the SAVE score, pre-ECMO, and on-ECMO models was 0.73, 0.77, and 0.83, respectively. Logistic regressions using ML-identified variables showed stepwise AUC improvements: 0.82 (pre-ECMO), 0.86 (on-ECMO), and 0.89 (combined). A novel, interpretable ML model predicted survival for VA-ECMO patients with accuracy comparable to the SAVE score. Incorporating on-ECMO variables significantly increased predictive performance and revealed novel variables associated with survival.