ECMO PAL VV: using deep neural networks for survival prognostication in venovenous extracorporeal membrane oxygenation

BACKGROUND: Prognostication for venovenous extracorporeal membrane oxygenation (ECMO) outcomes is crucial for risk-adjusting centre performance. This study aimed to leverage a large, multicentre, international database to develop and evaluate AI-driven models for predicting survival to hospital discharge of adult patients receiving venovenous ECMO. The model was called ECMO PAL VV (ECMO – Predictive Algorithm for VV). METHODS: Training and temporal validation data were sourced from the Extracorporeal Life Support Organization Registry (ELSO), 39,501 patients across 660 hospitals. Deep neural networks were trained on all adult patients receiving VV ECMO between 2017 and 2023 (N = 35,182) to predict survival to hospital discharge. Temporal validation was performed on registry data cases from 2024 (N = 4,318). Model predictions were compared against published venovenous ECMO outcomes scores using the validation cohort. RESULTS: Internal training yielded an accuracy of 79% and an area under the receiver operating characteristic curve (AUC) of 0.87. Temporal validation revealed a drop in accuracy to 73% with an AUC of 0.78, primarily due to a reduction in sensitivity to mortality prediction (71% to 57%). ECMO PAL VV outperformed published venovenous ECMO scores, which had accuracies of 65% (RESP) and 60% (Lazzeri score) for predictions on the validation data. CONCLUSIONS: ECMO PAL VV demonstrated strong accuracy on contemporary international registry data (73%) with strong sensitivity (81%) and precision (77%) to predict survival to hospital discharge, outperforming existing published scores. ECMO PAL VV has the potential to improve risk adjustment and enable data-driven healthcare.