Early prediction of moderate-to-severe BPD or death in preterm infants using minute-level respiratory monitoring and explainable machine learning.
Authors
Affiliations (4)
Affiliations (4)
- Neonatology Department, CHU Rennes, Rennes, France.
- University of Rennes, CHU Rennes, Inserm, LTSI-UMR 1099, Rennes, France.
- Neonatology Department, CHU Rennes, Rennes, France. [email protected].
- University of Rennes, CHU Rennes, Inserm, LTSI-UMR 1099, Rennes, France. [email protected].
Abstract
Early identification of extremely preterm infants at risk of moderate-to-severe bronchopulmonary dysplasia (BPD) or death remains challenging. We assessed whether minute-by-minute respiratory data from the first 24 h improves early prediction of this outcome using explainable machine learning. Retrospective cohort of 192 infants <30 weeks' gestation (2019-2024). Logistic regression, random forest, and XGBoost models were trained on 17 predictors and summarized 24-h respiratory data. Evaluation used AUROC, AUPRC, Brier score, calibration, and decision curves, with SHAP for interpretability. Forty-five infants (23.4%) reached the composite endpoint. All models achieved comparable discrimination (AUROC 0.831-0.847). Random forest was selected as the headline model based on superior calibration (slope 1.016; ECE 0.053), lowest Brier score (0.131), and highest net benefit at decision thresholds ≥20%. Chest X-ray density, gestational age, and maximum FiO₂ were the top SHAP predictors in the headline model; gestational age achieved the highest cross-model rank. SHAP identified a gestational age threshold at 28 weeks, an FiO₂ risk step at 30-35%, and a U-shaped birthweight z-score effect. Combining early radiographic and high-frequency respiratory data with explainable machine learning enables well-calibrated BPD risk stratification on day 0-1. SHAP decomposition supports personalized care by distinguishing modifiable from non-modifiable risks. A machine learning model integrating respiratory monitoring, radiography, and perinatal variables from the first 24 h of life predicts moderate-to-severe BPD or death in very preterm infants with an AUROC of 0.831 and well-calibrated risk estimates (calibration slope 1.016). Unlike existing tools relying on data from day 7 or later, this day-1 approach significantly improves early prognosis beyond gestational age and birthweight alone. Providing net clinical benefit across a 10-30% threshold, it enables individualized, actionable risk stratification at the bedside, while clinical trajectories can still be modified.