This study was a prospective cohort study. Patients were newborns enrolled in diagnosed cases of NRDS meeting criteria from January 2022 to January 2025 at The First Affiliated Hospital of Anhui Medical University. Inclusion criteria: (1) compliance with the European Consensus Guidelines on the Management of Respiratory Distress Syndrome: 2022 Update [14], regarding clinical symptoms and imaging data; (2) single fetus; (3) gestational age $\le$37 weeks and birth weight 2500 g. Exclusion criteria: (1) congenital developmental anomalies, inherited metabolic and chromosomal disorders; (2) severe intrauterine infections (e.g., sepsis); and (3) severe asphyxia (Apgar score $\le$3 and lasting for 10 minutes), as well as abandonment of treatment, or missing data $>$20%. The data were then obtained from the records related to newborns who met the above inclusion and exclusion criteria, and 163 subjects were finally included. The study was approved by the Ethics Committee of The First Affiliated Hospital of Anhui Medical University (Ethical approval number: No. 2021AMU-0532), and the guardians of the newborns voluntarily signed an informed consent form.

Maternal information: age, gestational comorbidities (e.g., gestational diabetes mellitus or hypertension), prenatal hormone use, and mode of delivery (eutocia or cesarean section) were recorded.

Neonatal information: gestational age, birth weight, gender, Apgar scores at 1 and 5 minutes (1-min Apgar and 5-min Apgar), initial respiratory support (non-invasive continuous positive airway pressure [CPAP] or mechanical ventilation, initial respiratory support equipment including Weikang V60 (Shenzhen Weikang Biological Technology Co., Ltd., Shenzhen, Guangdong, China) non-invasive ventilator 25A (Delgado Medical Systems, Frankfurt, Germany) for non-invasive CPAP treatment, Delgado Evita XL invasive ventilator for mechanical ventilation treatment), and chest X-ray classification (grade I–IV based on the severity of NRDS). Chest X-ray examination was performed using Philips DigitalDiagnost DR chest X-ray machine (C90, Philips Healthcare, Amsterdam, The Netherlands) [14].

Radial artery or dorsalis pedis artery (preferred arterial catheter) was selected and sterilized with 75% alcohol twice. Blood samples (0.3–0.5 mL) were collected by puncture using a heparinized syringe, gently mixed for anticoagulation, and sent for testing within 15 min. Testing was performed using the Siemens Rapidpoint 500 blood gas analyzer (Siemens Healthineers, Erlangen, Germany). Blood collection time, site and newborn status were recorded to avoid hemolysis or air bubble interference. The device was calibrated daily, with abnormal results (e.g., pH 7.0) requiring retesting for confirmation.

Blood gas levels were dynamically monitored at several time points: within 6 h of birth (baseline), and at 12, 24, and 48 h after birth, as well as during any condition worsening, like a sudden decrease in the oxygenation index or the need to adjust the ventilator parameters. In terms of monitoring indicators, pH, PaO₂, PaCO₂, HCO₃^–, base excess (BE), and Lac were monitored. The derived indicators, i.e., OI (FiO₂ $\times$ mean airway pressure $\times$ 100/PaO₂) and PaO₂/FiO₂ ratio were also measured.

The endpoint metrics consisted of primary and secondary outcomes. The primary outcomes were a poor prognostic composite endpoint that encompassed death in the first week after birth, pneumothorax or mediastinal emphysema requiring closed chest drainage, and severe respiratory failure requiring high-frequency oscillatory ventilation support for $>$24 h high-frequency oscillatory ventilation equipment utilized was the Stephanie BabyLog 8000+ ventilator (Stephan Medizintechnik GmbH, Gackenbach, Germany). The secondary endpoints included changes in dynamic blood gases within 24 h and within 24–48 h of birth.

The required sample size (N) must satisfy the following criteria: limiting overfitting: N $\ge$ k/(R² $\times$ 0.05); accurate variable estimation of the number of events: E $\ge$ 10 $\times$ k. Here, k represents the expected number of predictor variables included in the model, R² denotes the expected model explanatory power (set at 0.20), and E is the required number of events.

Based on preliminary studies and literature, k = 4 was set. Substituting into the formula yields a required event count of 40. This study ultimately enrolled 163 pediatric patients, with 30 events. While the total sample size did not fully meet the theoretical calculation, it reached a common scale for this type of research.

Statistical analyses were performed using SPSS 25.0 software (IBM Corp., Armonk, NY, USA). To evaluate the normality of the data, Shapiro-Wilk tests were conducted. Continuous variables conforming to normal distribution were expressed as mean $\pm$ standard deviation (SD), and differences between groups were compared by independent samples t-test. Non-normally distributed data were described by median [M (Q1, Q3)], and Mann-Whitney U-test was applied for comparison between groups; and categorical variables were expressed as frequency (percentage) [n (%)] and analyzed for between-group differences by chi-square test. The magnitude of change in key blood gas indices (e.g., pH, PaO₂, PaCO₂, etc.) was assessed by calculating the difference between the 24-hour value and the 6-hour baseline value and the difference between the 48-hour value and the 24-hour value. Using two robust feature selection methods—Elastic Net regression and the Boruta algorithm—predictors were screened from all clinical baseline characteristics and dynamic blood gas parameters. Ultimately, four core dynamic blood gas parameters were selected for inclusion in the final model. Subsequently, a multivariable Cox proportional hazards regression model was established. The time-dependent receiver operating characteristic (ROC) curve (7 days) of the predictive model was plotted for evaluation. To assess model calibration, 1000-sample Bootstrap internal validation was performed, and a calibrated calibration curve was plotted. Finally, decision curve analysis (DCA) was conducted to evaluate the clinical net benefit of the model at different risk thresholds. All statistical tests were completed using specialized software with two-sided test level $\alpha$ = 0.05, and p 0.05 was considered statistically different.

This study included 163 newborns with NRDS who were categorized into a good prognosis group (n = 133) and a poor prognosis group (n = 30) based on the composite endpoints. As for maternal baseline characteristics, there were no significant differences between the two groups in terms of age (p = 0.076) and prenatal hormone use (p = 0.659) and cesarean section rate (p = 0.322), but there was a significantly higher incidence of pregnancy comorbidities in the poor prognosis group (50.0% vs. 26.3%, p = 0.011). Based on neonatal characteristics, gestational age (p = 0.019), birth weight (p 0.001), and 5-min Apgar (p = 0.034) were significantly lower in the poor prognosis group than in the good prognosis group, whereas the differences in gender (p = 0.583), 1-min Apgar (p = 0.056). Although the proportion of chest radiographs graded $\ge$II was higher in the poor group (40.0% vs. 28.6%), it did not reach a significant level (p = 0.220). Notably, there was a significant difference in the mode of initial respiratory support between the two groups: 53.3% of the poor prognosis group directly required mechanical ventilation, whereas only 27.1% of the good group did so. Accordingly 72.9% of the good group initiated treatment with CPAP, which was significantly higher than the 46.7% in the poor group (p = 0.005). These findings suggest that maternal pregnancy comorbidities, low gestational age, low birth weight and initial need for mechanical ventilation may be strongly associated with poor prognosis in newborns with NRDS (Table 1).

Neonatal blood gas indices in the poor and good prognosis groups were dynamically monitored (Table 2), and significant differences were found between the two groups in some parameters (Table 2). The poor prognosis group showed a triad of evolution of blood gas indicators in comparison with the good prognosis group. A significant acidosis was observed in the early stage (6 h) (pH 7.21 $\pm$ 0.10 vs. 7.28 $\pm$ 0.07), and OI_24–48 h continued to climb (24 h 27.35 $\pm$ 3.20 vs. 24.03 $\pm$ 4.20, 48 h 24.10 $\pm$ 3.01 vs. 19.88 $\pm$ 4.33) accompanied by a decrease in oxygenation efficiency (PaO₂/FiO₂ 243.00 $\pm$ 45.76 vs. 265.35 $\pm$ 38.16) with a significant delay in Lac clearance (24 h 2.86 $\pm$ 1.17 vs 2.15 $\pm$ 0.93, 48 h median 2.60 vs 1.60 mmol/L). Although the trends of persistent low HCO₃^–, insufficient PaCO₂ clearance and BE recovery were not statistically significant, the persistent elevation of 48 h OI and impaired Lac metabolism constituted the most critical early warning indicators, suggesting the existence of a vicious pathological cycle of “early acidosis-progressive oxygenation-delayed metabolic clearance” in poorly prognosticated newborns.

Subsequently, the dynamic characteristics of key blood gas indicators in newborns in the poor prognosis group were compared with those in the good prognosis group, and significant differences were found between the two groups in the magnitude of changes in several indicators (Table 3). In terms of acid-base balance, the increase in pH was greater in the poor prognosis group during the first 24 h after birth (p = 0.016), whereas the change in pH tended to slow down during the period of 24–48 h (p 0.001); HCO₃^– showed a more pronounced and sustained increase in the poor prognosis group, which was significantly better than the poor prognosis group at 24 h (p 0.001) and 24–48 h (p = 0.013); PaCO₂ changes, on the other hand, showed a more significant decrease in the poor prognosis group at 24 h (p = 0.004), but not as much improvement as in the good prognosis group at 24–48 h (p 0.001), whereas changes in BE were not statistically different between the two groups. Regarding oxygenation and metabolic indices, OI worsened more significantly in the poor prognosis group at 24 h (p 0.001) and improved less at 24–48 h (p 0.001). Oxygenation efficiency (PaO₂/FiO₂) was more significantly elevated at 24–48 h in the good prognosis group (p 0.001). In terms of Lac clearance, the clearance efficiency was significantly lower in the poor prognosis group than in the good prognosis group both at 24 h (p 0.001) and 24–48 h (p 0.001). Changes in PaO₂ showed a more pronounced rise in the good prognosis group only at 24 h (p 0.001), with no significant difference for the rest of the period.

Elastic Net analysis (Supplementary Table 1) identified five non-zero coefficient variables from all candidate variables under the optimal regularization parameter ($\lambda$ = 0.175894834354564, determined via 10-fold cross-validation; Supplementary Fig. 1). The Boruta algorithm (Supplementary Table 2, Supplementary Fig. 2) further validated the importance of these features. This algorithm explicitly confirmed four variables: $△$OI_24 h, $△$PaCO₂_24–48 h, $△$PaO₂_24 h, and Lac_48 h. Among these, $△$OI_24 h demonstrated significantly higher importance (Mean Importance = 33.1547790) than other variables, establishing its core predictive role. Conversely, $△$HCO₃^–_24 h was rejected by the algorithm, suggesting its information may be covered by other variables. Both methods jointly identified four core dynamic blood gas parameters—$△$OI_24 h, $△$PaCO₂_24–48 h, $△$PaO₂_24 h, and Lac_48 h—for constructing the final model.

Incorporating the above four variables into a multivariate Cox regression model yielded the results shown in Table 4: $△$OI_24 h and Lac_48 h were independent risk factors for adverse outcomes within 7 days in children with NRDS. However, $△$PaO₂_24 h (hazards ratio (HR) = 0.82, 95% confidence interval (CI): 0.66–1.02, p = 0.075) and $△$PaCO₂_24–48 h (HR = 1.22, 95% CI: 0.92–1.62, p = 0.165) did not reach statistical significance (Table 4).

As shown in the time-dependent ROC curve in Fig. 1, the model achieved an area under the curve (AUC) of 0.96 (95% CI: 0.92–1.00) for predicting 7-day adverse outcomes, indicating good discriminatory ability between poor and good outcomes. The adjusted curve obtained through 1000-time internal validation via Bootstrap resampling (Fig. 2A) demonstrates high consistency between predicted and observed risks, indicating excellent predictive accuracy with no significant bias after adjustment. DCA analysis indicated that when the risk threshold exceeded 0.15, the clinical net benefit derived from risk stratification using this predictive model consistently surpassed that of both intervention and no intervention strategies (Fig. 2B). This demonstrates the model’s strong clinical utility in supporting decision-making by clinicians.

Fig. 1.

Time-dependent ROC curve of the predictive model (7 days). AUC, area under the curve.

Fig. 2.

Internal validation of the model. (A) Calibration curve from internal validation via 1000 Bootstrap resamples. (B) DCA evaluating the model’s clinical net benefit at different risk thresholds. DCA, decision curve analysis.

This study has several limitations that warrant caution in interpreting the findings. It is a single-center study with a limited sample size (particularly in the poor prognosis group, n = 30), which may have restricted the power of certain statistical tests and increased the risk of model overfitting. The findings require further validation in prospective, multicenter, large-scale cohort studies. Although the model adjusted for several important variables, unmeasured confounders may still exist. These include detailed ventilator settings strategies, fluid management practices, patent ductus arteriosus and its hemodynamic impact, as well as the specific timing and frequency of surfactant administration—all of which could influence outcomes. The primary outcome of this study was defined as a composite endpoint within 7 days postnatal. These dynamic indicators were not evaluated for their association with longer-term outcomes such as bronchopulmonary dysplasia. Future studies may extend follow-up periods. Based on this, a standardized dynamic monitoring pathway was established using the 6/24/48 h blood gas monitoring points and incorporating $△$OI_24 h, Lac_48 h into the routine NRDS monitoring indicators. Clinicians can rapidly assess risk using simple calculation formulas without complex equipment, making it suitable for implementation. Additionally, this approach could facilitate the clinical integration of dynamic functional monitoring as an auxiliary criterion for NRDS risk stratification, providing quantitative evidence for developing personalized treatment plans and ultimately improving pediatric patient outcomes. However, it must be emphasized that this study currently offers only preliminary proof-of-concept for this concept. The aforementioned translational application prospects are entirely contingent upon validation of this model in subsequent larger-scale, prospective studies. At the present stage, neither $\mathrm{\Delta }$OI_24 h nor Lac_48 h should be used as independent decision-making criteria, and clinical physicians must interpret these metrics holistically in conjunction with the specific circumstances of each pediatric patient.

This study has several limitations that warrant caution in interpreting conclusions and generalizing findings. It is a single-center observational study with a limited sample size, particularly in the poor prognosis group (n = 30). This limits statistical power, and the generalizability of results requires validation. While the primary composite outcome encompassed severe events, follow-up duration was short (within 7 days). Therefore, the findings cannot infer associations between these dynamic indicators and critical long-term neonatal outcomes (e.g., bronchopulmonary dysplasia, neurodevelopmental prognosis), which are equally vital for assessing the overall burden of NRDS. Although we conducted internal validation using Bootstrap and demonstrated good performance, this remains an internal validation. Whether the model maintains the same predictive efficacy in other populations or medical centers with different clinical practice standards must be confirmed through rigorous external validation. Future studies should involve multicenter collaboration to establish an external validation cohort of $\ge$500 cases. This cohort should validate the predictive accuracy of $△$OI_24 h, Lac_48 h, and model stability, and compare their clinical utility with traditional scoring systems. Follow-up should extend to 40 weeks’ corrected gestational age or 1–2 years, incorporating long-term outcomes such as bronchopulmonary dysplasia, Bayley scales, and pulmonary function tests and assessing the predictive value of core indicators for long-term prognosis.