Related Experiment Videos
Predicting survival in extremely preterm infants: A multicenter machine learning study from Spain
Paula Sol Ventura1, Juan Leva-Bueno2, Angélica Atehortúa2
1Barcelona Institute for Global Health (ISGlobal), Barcelona, Spain; Neonatology Unit, Paediatric Department, Hospital Germans Trias i Pujol, Universitat Autónoma de Barcelona, Badalona, Spain; Paediatric Department, Hospital Universitario Arnau de Vilanova Lleida, Lleida, Spain.
Background And Objective:
Accurately predicting outcomes in extremely preterm infants remains a major challenge in neonatology. Traditional scores such as the Clinical Risk Index for Babies (CRIB I) and its update (CRIB II) offer valuable initial assessments but have limited predictive power. We developed interpretable machine learning (ML) models to predict survival to hospital discharge using data from 8080 extremely preterm infants in the Spanish Neonatal Network SEN1500 born at 22-26 weeks' gestation between 2004 and 2019.
Methods:
All actively managed infants with documented survival outcomes were included. Multiple ML classifiers were trained to predict survival to hospital discharge; a CRIB-based binary classifier served as a benchmark. Discrimination was evaluated using the area under the receiver operating characteristic curve (ROC AUC) and the area under the precision-recall curve (PR AUC) with nested cross-validation. Probabilities were calibrated using Platt scaling, and interpretability was assessed with SHAP (Shapley Additive Explanations). Unsupervised clustering of SHAP values via UMAP identified subgroups with shared survival profiles.
Results:
XGBoost achieved the best performance (PR AUC: 0.81; ROC AUC: 0.77), outperforming classifiers based on CRIB I and CRIB II (PR AUC: 0.80 and 0.73; ROC AUC: 0.74 and 0.68, respectively). Key predictive factors included gestational age, birth weight, admission temperature, nasal continuous positive airway pressure (nCPAP), 5-minute Apgar score, and antenatal corticosteroid exposure. SHAP-UMAP analysis revealed two phenotypes with survival rates of 76% and 52%, respectively.
Conclusions:
Interpretable ML models improve survival prediction over CRIB I and CRIB II while preserving clinical transparency. By capturing nonlinear relationships among routinely collected variables, these models enable earlier identification of infants with high and low survival probabilities and may support targeted interventions. Prospective external validation and integration into electronic health records with real-time interpretability are key next steps for clinical translation.
