An Exploratory Study of a Prognostic Risk Warning Model for Neonatal Respiratory Distress Syndrome Based on Dynamic Monitoring of Blood Gas: A Prospective Cohort Study

Background: This study is a prospective cohort study. It aims to investigate the relationship between blood gas parameters and neonatal respiratory outcomes and to develop a prognostic prediction model. Methods: A total of 163 preterm newborns who satisfied the diagnostic criteria outlined in the European Guidelines for the Prevention and Treatment of NRDS-2010 from January 2022 to January 2025 were included. The baseline data of mothers and newborns were collected, and the blood gas parameters were dynamically monitored at 6, 12, 24, and 48 h after birth, including pH, oxygen partial pressure (PaO2), carbon dioxide partial pressure (PaCO2), lactic acid (Lac), and oxygenation index (OI), as well as PaO2/fraction of inspired oxygen (FiO2). Elastic net regression and the Boruta algorithm were used to screen predictive variables, and a multivariate Cox proportional hazards regression model was established. The performance of the model was evaluated using a time-dependent receiver operating characteristic (ROC) curve, Bootstrap resampling calibration curve, and decision curve analysis (DCA). Results: The poor prognosis group (n = 30) experienced a higher rate of maternal pregnancy comorbidities (50.0% vs. 26.3%; p = 0.011), had a smaller gestational age (29.4 weeks; p = 0.019), lower birth weight (1412.5 g; p < 0.001) and 5-minute Apgar score (p = 0.034), and a higher need for initial mechanical ventilation (53.3% vs. 27.1%; p = 0.005). Dynamic monitoring revealed significant acidosis in the early phase (6 hours), which remained at persistently low levels even at 48 hours. The OI progressively increased, oxygenation efficiency declined, and lactate clearance was markedly delayed. Elastic net regression (optimized λ = 0.1759 via 10-fold cross-validation) and Boruta algorithm screening identified core variables for inclusion in a multivariate Cox regression. Meanwhile, △OI_24 h (hazard ratio (HR) = 1.82, 95% confidence interval (CI) 1.51–2.21; p < 0.001) and Lac_48 h (HR = 1.95, 95% CI: 1.40–2.73; p < 0.001) were identified as independent risk factors. The model predicted a 7-day poor prognosis with an area under the ROC curve of 0.96 (95% CI 0.92–1.00). A 1000 Bootstrap validation model demonstrated high concordance between predicted and actual risks. The DCA indicated that the model provided a significant clinical net benefit compared to intervention or no intervention strategies when the risk threshold exceeded 0.15. Conclusions: △OI_24 h and Lac_48 h serve as core early warning indicators for poor prognosis in NRDS. The model was constructed using elastic net regression and the Boruta algorithm, demonstrating robust predictive performance and clinical utility, and providing a basis for early risk stratification.