(883) Predicting Mortality in Pediatric Chronic Lung Disease Requiring ECMO Using a Machine Learning Model

Introduction: Children with a history of bronchopulmonary dysplasia (BPD) are hospitalized at a higher rate for respiratory disease. Despite higher hospitalization rates, few studies have examined these patients and their outcomes when they require ECMO support. We hypothesize that a machine learning model using clinical, physiologic, and laboratory data can predict the risk of mortality in children with BPD and/or pulmonary hypertension (PHTN) that required ECMO support.

Methods: Pediatric patients with an ICD-9 or ICD-10 code consistent with BPD or PHTN from the years 2012-2020 were obtained from the Extracorporeal Life Support Organization (ELSO) international registry. The machine learning models were developed using a combination of demographic data, pre-ECMO data, and 24-hour post-ECMO data. Missing data were handled through multiple imputation, followed by a split into a training cohort (70%) and a testing cohort (30%). Four types of machine learning models were derived: neural network, extreme gradient boosting, random forest, and generalized linear model. Hyperparameters were optimized using cross-validation. Variables were ranked by importance, and the top variables were used in the final model. Model performance was then evaluated and validated using the internal testing cohort.

Results: Of the 1,489 patients included in this study, 490 patients (33%) died during their hospital stay after ECMO initiation. The extreme gradient boosting model performed the best out of all machine learning model types. The top 10 variables were age, race, ECMO mode, ECMO duration, PO2 at 24 hours post-ECMO, history of congenital diaphragmatic hernia, and pre-ECMO dopamine use, pre-ECMO PF ratio, vasopressor/inotropic support, pH, neuromuscular blockage. The final model contained 30 variables and had a receiver operating characteristic curve of 0.72 (95% confidence interval 0.67-0.78).

Conclusions: Machine learning models may aid clinicians in predicting mortality for children with BPD who require ECMO support regardless of the reason for requiring ECMO. Future research should focus on external validation and improving model performance with feature engineering.