531 - Predicting the Likelihood of In-hospital Survival of Pediatric Patients Prior to ECMO Cannulation Utilizing Artificial Intelligence (AI)

Publication Number: 1510.531

Ayaka Tsutsumi, SSM Cardinal Glennon Children's Medical Center, St. Louis, MO, United States; Chiara Camerota, Saint Louis University, St. Louis, MO, United States; Wazma Ali, Saint Louis University, Saint Louis, MO, United States; Flavio Esposito, Saint Louis University, Saint Louis, MO, United States; Erik C. Madsen, SSM Cardinal Glennon Children's Medical Center, Saint Louis, MO, United States; Richard Herman, SSM Cardinal Glennon Children's Medical Center, St Louis, MO, United States; Priya Reddy, Saint Louis University School of Medicine, St. Louis, MO, United States; Jessica Qiu, Saint Louis University School of Medicine, Saint Louis, MO, United States; Ruizhi Huang, Saint Louis University, Cedar Park, TX, United States; Mary B. Taylor, Children's of Mississippi, Jackson, MS, United States; Takanari Ikeyama, Aichi Children's Health and Medical Center, Obu, Aichi, Japan; Shin Miyata, SSM Cardinal Glennon Children's Medical Center, St Louis, MO, United States

Background: Patient selection for Venovenous Extracorporeal Membrane Oxygenation (VV-ECMO) remains challenging, often relying solely on the clinician's experience.

Objective: To leverage machine learning (ML) to predict the likelihood of survival and in-hospital death, and to reveal the critical variables influencing prediction in pediatric patients requiring VV-ECMO.

Design/Methods: This was a retrospective study that involved the development and evaluation of both conventional machine learning algorithms (including Logistic Regression, Random Forest, XGBoost, Support Vector Machine, Neural Network, and Naïve Bayes) and a transfer learning approach. Data for completed ECMO runs spanning from 2018 to 2024 were retrieved from the Extracorporeal Life Support Organization (ELSO) registry. The study participants included pediatric patients (age < 19 years) who underwent completed VV-ECMO runs. The transfer learning algorithms learn from the Multiparameter Intelligent Monitoring in Intensive Care IV (MIMIC-IV) database to generalize knowledge on the ELSO dataset. Model performance was assessed using internal and external validation (using independent 2024 data). Accuracy, recall (sensitivity), precision (positive predictive value, or PPV), and the F1-score were the metrics utilized. The F1-score was calculated as: 2×(PPV×Sensitivity)/(PPV+Sensitivity).

Results: Of the 4169 participants, 3052 (73.2%) survived to discharge (Figure 1). The ability of all models to predict survival consistently outperformed the prediction of in-hospital death (Table 1). Among the models assessed, the transfer learning model (MIMIC autoencoder) achieved the highest predictive performance upon external validation, demonstrating an overall accuracy of 0.73. This model showed robust results for the survivor class, yielding a recall of 0.92 and an F1-score of 0.83. Conversely, the external validation performance for in-hospital death was significantly lower (F1-score 0.29). Variables consistently ranked as most influential across the best-performing models included Respiratory Rate, SaO2, SpO2, and patient height. G-computation analysis identified low SpO2, low SaO2, and short patient height as the most negatively impactful factors on survival.

Conclusion(s): The transfer learning model demonstrated robust capacity in predicting survival among the pediatric VV-ECMO population. However, due to significant data imbalance, the predictive capacity for mortality remained a challenge. Further investigations using advanced ML methodologies resilient against imbalanced outcomes are warranted to transition these findings into a reliable clinical tool.

ELSO Entry Selection


Comparative Analysis of Model Performance
*w/o: without, Ext.: External, Val.: Validation, SVM: Support Vector Machine, MIMIC: Multiparameter Intelligent Monitoring in Intensive Care