A machine learning approach to predict mortality and neonatal persistent pulmonary hypertension in newborns with congenital diaphragmatic hernia. A retrospective observational cohort study

Abstract

Congenital diaphragmatic hernia (CDH) has high morbidity and mortality rates. This study aimed to develop a machine learning (ML) algorithm to predict outcomes based on prenatal and early postnatal data. This retrospective observational cohort study involved infants with left-sided CDH, born from 2012 to 2020. We analyzed clinical and imaging data using three classification algorithms: XGBoost, Support Vector Machine, and K-Nearest Neighbors. Medical records of 165 pregnant women with CDH fetal diagnosis were reviewed. According to inclusion criteria, 50 infants with isolated left-sided CDH were enrolled. The mean o/eLHR was 37.32%, and the average gestational age at delivery was 36.5 weeks. Among these infants, 26 (52%) had severe persistent neonatal pulmonary hypertension (PPHN), while 24 (48%) had moderate or mild form; 37 survived (74%), and 13 did not (26%). The XGBoost model achieved 88% accuracy and 95% sensitivity for predicting mortality using ten features and 82% accuracy for PPHN severity with 14 features. The area under the ROC curve was 0.87 for mortality and 0.82 for PPHN severity.

Conclusion: ML models show promise in predicting CDH outcomes and supporting clinical decisions. Future research should focus on more extensive studies to refine these algorithms and improve care management.

Clinical trial registration: NCT04609163.

What is known: • Congenital diaphragmatic hernia (CDH) is a serious condition characterized by high morbidity and mortality rates, making it critical to predict neonatal outcomes for effective clinical management accurately. • Traditional prenatal diagnostic methods often struggle to predict complications such as Neonatal Persistent Pulmonary Hypertension (PPHN) in CDH, highlighting the need for innovative predictive approaches.

What is new: • Machine learning (ML) models, particularly XGBoost, have been shown to accurately forecast mortality and the severity of PPHN in infants with CDH based on prenatal and early postnatal clinical and imaging data. • ML-based predictive models can enhance prenatal counseling, optimize birth planning, and tailor postnatal care for patients with CDH, enabling real-time risk assessment and adaptive management strategies.

Keywords: Congenital diaphragmatic hernia; Deep learning; Machine learning; Mortality; Neonatal persistent pulmonary hypertension; Newborn.

Fig. 1

The PRISMA flow diagram of the study

Fig. 2

Mortality prediction. Left, Receiver Operating Characteristic (ROC) curve; right, the Precision-Recall (P-R) curve for mortality predictions, obtained with the XGBoost classifier on prenatal clinical variables. No shape features extracted from magnetic resonance imaging (MRI) were used, as required by the feature-selection procedure

Fig. 3

Neonatal persistent pulmonary hypertension (PPHN) prediction. Left, Receiver Operating Characteristic (ROC) curve; right, the Precision-Recall (P-R) curve for PPHN predictions with the XGBoost classifier on prenatal clinical variables and shape features extracted from magnetic resonance imaging (MRI)