The findings address a difficult period in neonatal medicine. Extremely premature infants have immature lungs, limited surfactant production and impaired fluid clearance, and some develop severe respiratory injury that evolves into BPD. Death and BPD are distinct conditions, but the researchers combined them as one endpoint because death before 36 weeks’ postmenstrual age prevents a later assessment of BPD. The composite outcome consisted of 23 surviving infants with grade 2–3 BPD and 22 deaths, including 17 before 36 weeks’ postmenstrual age. The study population was highly selected: infants intubated within their first 60 minutes, those born outside the hospital, infants with congenital anomalies and several with missing data were excluded. The findings therefore describe a specific single-center population of spontaneously breathing, inborn extremely preterm infants rather than all premature newborns.

Unlike many prediction tools based on a small number of variables measured at fixed clinical checkpoints, the study condensed nearly a full day of continuous monitoring into 17 prespecified predictors. Respiratory data were sampled once per minute from 60 minutes to 24 hours after birth and included inspired oxygen fraction, or FiO₂; oxygen saturation measured by pulse oximetry; ventilation mode; mean airway pressure; and transcutaneous carbon dioxide. The researchers calculated the Oxygen Saturation Index, or OSI, from mean airway pressure, FiO₂ and SpO₂ as a non-invasive measure of hypoxic respiratory failure. FiO₂, SpO₂ and airway pressure were smoothed with a 10-minute moving average, and maximum values for FiO₂, OSI and transcutaneous carbon dioxide were defined as the highest levels sustained for at least 30 minutes. Although the source signals had minute-level resolution, the model itself received only one summary statistic for each respiratory measure rather than full trajectories, variability indices or slopes.

Chest radiographs supplied a second type of early information. The admission image, acquired at approximately 30 to 60 minutes of life while infants received non-invasive support, was converted into an objective lung-density score. Measurements from three pulmonary regions were normalized against soft-tissue and darkest references on the same image, reducing variation caused by exposure and acquisition conditions. In a reliability substudy, three trained operators measured 30 randomly selected radiographs twice while blinded to one another and their previous readings. The single-measure intraclass correlation coefficient was 0.855, with a 95% bootstrap confidence interval of 0.759–0.906, and model-based agreement limits were ±0.067 within operators and ±0.099 between operators. The score was therefore reasonably reproducible under routine conditions, although two-dimensional overlap, patient positioning and differences in image acquisition could still affect it.

The analytical dataset was divided by stratified random allocation into 115 training infants and 77 held-out test infants. Hyperparameters were selected through 10-fold cross-validation within the training set, and the test cohort remained excluded from training and tuning. Because the training set contained only 27 composite events across 21 encoded predictor columns—about 1.3 events per variable—the researchers applied strong regularization: elastic-net penalties for logistic regression, constrained feature sampling for random forest and structural constraints for XGBoost. Class imbalance was addressed with synthetic oversampling inside training folds for logistic regression and random forest, and with a weighted-loss parameter for XGBoost. They also used 2,000 bootstrap resamples to estimate 95% confidence intervals for test-set performance. After evaluation, the models were retrained on all 192 infants with the selected settings, but that final retraining did not alter the reported test results.

Random forest was chosen as the headline model not because it had the highest AUROC, but because it produced the most reliable probability estimates. Its test-set AUROC of 0.831 had a 95% confidence interval of 0.707–0.933, compared with 0.847 for logistic regression and 0.847 for XGBoost. Random forest also had the highest area under the precision-recall curve, 0.598, and the lowest Brier score, 0.131. Its calibration slope of 1.016 and intercept of −0.035 were close to ideal, while its expected calibration error was 0.053. XGBoost discriminated as well but produced overconfident probabilities, with a calibration slope of 0.805 and expected calibration error of 0.091. The retuned logistic regression was severely miscalibrated, with a slope of 11.3 and expected calibration error of 0.314, so the researchers retained it as a transparent comparator rather than a probability tool. Decision-curve analysis indicated that random forest and XGBoost offered greater net benefit than treating everyone at the same threshold, while random forest had the highest net benefit at thresholds of 25% and 30%.

SHAP, or SHapley Additive exPlanations, was used to show how individual variables contributed to each model prediction. In the random forest, chest X-ray density, gestational age and maximum FiO₂ had the three largest mean absolute SHAP values, 0.067, 0.063 and 0.043, respectively; maximum OSI followed at 0.039. Across all three algorithms, gestational age, maximum FiO₂ and chest X-ray density had the strongest combined rankings, with mean cross-model ranks of 1.3, 2.3 and 3.3. The dependence plots revealed several non-linear patterns: risk contributions shifted around 28 weeks’ gestation, radiographic density contributions rose above a normalized score of roughly 0.70, and FiO₂ showed a marked step between 30% and 35%. Birthweight z-score had a U-shaped association, with the lowest model contribution between −1 and +0.5. These patterns describe how the fitted algorithms used the data; they are not causal thresholds or clinical cutoffs.

The researchers also compared the full models with baselines restricted to gestational age and birth weight. Adding respiratory and radiographic information increased AUROC by 0.073 for random forest and 0.097 for XGBoost, supporting the incremental value of the broader data set in this cohort. Individual SHAP decompositions could distinguish contributions from constitutional factors, such as gestational age and birthweight, from early respiratory and radiographic markers, such as sustained oxygen requirement, OSI and lung density. The authors presented this as a way to make individual risk estimates more transparent, potentially helping clinicians distinguish modifiable respiratory signals from non-modifiable vulnerability. They explicitly cautioned that high SHAP contributions identify biomarkers or markers of severity and management, not interventions proven to prevent BPD. Randomized evaluation would be needed to establish whether algorithm-guided care improves outcomes.

Several features of the study limit how the results should be interpreted. The sample was small, drawn from one center and restricted to a selected group of spontaneously breathing infants, so performance may change under different respiratory-support or oxygenation practices. The training set’s very low events-per-variable ratio left the models vulnerable to overfitting despite regularization, and no formal sample-size calculation was performed. The composite endpoint cannot distinguish the separate biological pathways leading to death and BPD; an exploratory analysis of the validation set was too small and imprecise to support outcome-specific models. Transcutaneous carbon dioxide was missing in 23% of cases, and maximum FiO₂ and OSI were missing in 5%, although sensitivity analyses for the latter two were conducted. Radiographs were not systematically timed relative to less-invasive surfactant administration, and respiratory summaries may still reflect treatment trajectories as well as underlying disease severity.

The most defensible conclusion is therefore that day-one data can improve early risk stratification for a defined group of extremely preterm infants. The internally tested random forest combined clinical history, objective radiographic density and sustained respiratory measurements to estimate the risk of grade 2–3 BPD or death with calibration that was close to ideal in the held-out cohort. The work also illustrates why interpretability matters in neonatal prediction: it makes clear which measurements shaped an individual estimate and which apparent effect patterns are nonlinear features learned by the model. Yet the tool remains investigational. Independent multicenter validation, standardized imaging and monitoring protocols, calibration monitoring and prospective comparison with standard care are needed before its predictions can safely guide treatment. Even with promising internal results, the study does not show that using the model improves survival, reduces BPD or changes clinical decisions for the better.

Subject of Research: Prediction of severe bronchopulmonary dysplasia or death in extremely preterm infants using day-one respiratory, radiographic and perinatal data

Article Title: Early prediction of moderate-to-severe BPD or death in preterm infants using minute-level respiratory monitoring and explainable machine learning

Article References: Louifi, T., Delange, B., Leroux, M., Pladys, P., & Beuchée, A. (2026). Early prediction of moderate-to-severe BPD or death in preterm infants using minute-level respiratory monitoring and explainable machine learning. Pediatric Research. https://doi.org/10.1038/s41390-026-05503-5

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05503-5

Keywords: preterm infants, bronchopulmonary dysplasia, machine learning, respiratory monitoring, chest radiography, neonatal intensive care, risk prediction, explainable AI, random forest, Early, prediction, moderate-to-severe