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

Area of Science:

• Pediatric critical care medicine
• Biomedical informatics
• Machine learning applications in healthcare

Background:

• Identifying optimal candidates for extracorporeal membrane oxygenation (ECMO) remains a clinical challenge.
• Predicting survival in pediatric patients requiring venovenous ECMO (VV-ECMO) is crucial for resource allocation and treatment strategies.

Purpose of the Study:

• To utilize machine learning (ML) for predicting survival in pediatric VV-ECMO patients.
• To identify critical variables impacting outcomes in this patient cohort.
• To evaluate a transfer learning approach for enhanced predictive performance.

Main Methods:

• Retrospective analysis of the Extracorporeal Life Support Organization (ELSO) registry.
• Development of conventional ML algorithms and a transfer learning model pretrained on MIMIC-IV data.
• Inclusion of 4,169 pediatric patients (< 19 years) with internal and external validation using 2024 data.

Main Results:

• Overall survival to discharge was 73.2%.
• The transfer learning model achieved the highest external validation accuracy (0.73), with strong recall (0.92) and F1-score (0.83) for survivors.
• Mortality prediction was limited by outcome imbalance (F1-score: 0.29).
• Key predictors identified include respiratory rate, SaO2, SpO2, and patient height.

Conclusions:

• Transfer learning demonstrates significant predictive power for survival in pediatric VV-ECMO.
• Outcome imbalance currently impedes accurate mortality prediction, necessitating further research into robust ML techniques.
• The study successfully identified critical clinical variables influencing VV-ECMO outcomes in children.