Machine Learning Prediction of Pediatric In-Hospital Survival Before Extracorporeal Membrane Oxygenation Cannulation.

Identifying suitable candidates for extracorporeal membrane oxygenation (ECMO) is still challenging. Our aim is to leverage machine learning (ML) to predict survival and identify critical variables influencing outcomes in pediatric patients requiring venovenous ECMO (VV-ECMO). This retrospective study used the Extracorporeal Life Support Organization (ELSO) registry to develop conventional ML algorithms and a transfer learning approach pretrained on the Multiparameter Intelligent Monitoring in Intensive Care IV (MIMIC-IV) database. The study included 4,169 pediatric patients (aged < 19 years). Model performance was assessed through internal and external validation using independent 2024 data for external testing. Overall survival to discharge was 73.2%. The transfer learning model achieved the highest predictive performance on external validation (accuracy: 0.73). It demonstrated robust results for survivors (recall: 0.92, F1-score: 0.83), although mortality prediction was significantly lower (F1-score: 0.29) due to outcome imbalance. Respiratory rate, SaO2, SpO2, and patient height were the most influential predictors across models. Transfer learning demonstrates strong predictive capacity for survival in pediatric VV-ECMO. Although significant outcome imbalance currently hinders mortality prediction, this methodology identifies key clinical variables. Future research should focus on ML techniques resilient to imbalanced outcomes to develop reliable clinical decision-support tools.