This was a single-center prospective observational cohort study conducted in a tertiary NICU. The study period was from January 1, 2024, to June 30, 2025. The protocol adhered to a noninterventional design; all ultrasound examinations were performed solely for research purposes and did not influence clinical management. The study was approved by the Institutional Review Board of Shijiazhuang No.4 Hospital, Hebei Province, China (Approval No. 20220085). Written informed consent was obtained from the legal guardian of each infant prior to participation.

Clinical management followed a standardized departmental protocol for preterm infants, which included predefined oxygen saturation targets (90–95%), standardized noninvasive ventilation settings, and explicit criteria for surfactant administration and intubation. This approach minimized the potential for variations in clinical practice to systematically bias the observation of the study outcome.

The primary outcome was the escalation of respiratory support within 72 hours after birth. Based on local historical data, the estimated incidence of this outcome was 40%. For the most complex predictive model (Model C), we anticipated a maximum of 8 candidate predictors. Following the recommendation of at least 10 outcome events per predictor variable, a minimum of 80 outcome events was required, corresponding to a total sample size of 200 infants. A sample size calculation based on comparisons of AUC further supported this estimate. Assuming an AUC of 0.75 for the intermediate model (Model B) and 0.83 for the full model (Model C), with a correlation of 0.60 between the receiver operating characteristic (ROC) curves, a sample size of 200 infants provided 80% power at a two-sided α of 0.05. Anticipating a 5% rate of attrition or missing data, we planned to enroll 210 infants over the 18-month study period.

All ultrasounds were performed at the bedside within ≤6 hours of birth using a portable ultrasound system. Infants were examined in the supine position in a thermoneutral environment. Prewarmed gel was used, and gentle swaddling was employed as needed to minimize agitation; no sedation was administered. Ultrasound examinations were conducted during the infant’s prevailing clinical condition. To reflect real-world bedside practice, no adjustments were made to existing respiratory support during the examination.

Data were collected prospectively using an electronic case report form with built-in range checks and dual-entry verification. All ultrasound cine loops and still images were stored in a deidentified archive. For the reproducibility subset, measurements were independently performed by two readers, and one reader repeated the measurements after a minimum interval of two weeks.

A dedicated quality control officer performed monthly audits on 10% of randomly selected cases. Prespecified criteria for excluding poor-quality measurements included incomplete cardiac cycles, excessive Doppler angle (>20°), and significant artifact obscuring anatomical boundaries.

During the study period, 287 preterm infants were screened for eligibility. Of these, 205 infants completed the full multiparametric ultrasound assessment within ≤6 hours after birth. The primary outcome (respiratory support escalation within 72 hours) could not be ascertained for 2 infants, resulting in a primary analysis cohort of 203 infants (70.7% of screened infants). The first 30 consecutively enrolled infants formed the prespecified reproducibility subset (Fig. 1).

Fig. 1.

Flow diagram of participant screening, enrollment, and inclusion in the different analyses phases. CHD, congenital heart disease; PDA, patent ductus arteriosus; PFO, patent foramen ovale.

The baseline characteristics of the cohort stratified by the primary outcome are presented in Table 1. Infants who required escalation of respiratory support (n = 82, 40.4%) had significantly lower median gestational age (29.8 vs. 31.7 weeks, p < 0.001) and birth weight (1294 vs. 1618 g, p = 0.002). They were also less likely to have received a complete course of antenatal corticosteroids (63.4% vs. 83.5%, p = 0.002). In addition, the escalation group had higher initial PaCO₂ levels (58 vs. 47 mmHg, p < 0.001), a higher prevalence of hemodynamically significant PDA (hsPDA) (28.0% vs. 13.2%, p = 0.011), and more frequent administration of postnatal caffeine (72.0% vs. 56.2%, p = 0.027). No significant differences were observed between groups regarding sex, 5-minute Apgar score, mode of delivery, or multiple gestation status.

Data on surfactant administration prior to the ≤6-hour ultrasound assessment were not systematically recorded and therefore could not be reported.

Baseline values for the multiparametric ultrasound measurements, obtained within 6 hours of birth, are summarized in Table 2. The median LUS was 11 (IQR, 8–13), the median diaphragmatic excursion was 4.7 (IQR, 3.9–5.6) mm, the median diaphragmatic thickening fraction was 27.6% (IQR, 21.3–33.8), and the median pulmonary artery AT/ET ratio was 0.31 (IQR, 0.27–0.36).

The reproducibility of the key ultrasound measurements was excellent. In the subset of 30 infants, all interobserver and intraobserver ICCs were ≥0.87, indicating good-to-excellent agreement (Table 3).

Overall data completeness was high. Minor missing data (<5%) occurred for the 5-minute Apgar score (n = 4) and antenatal corticosteroid status (n = 2), and these variables were handled via multiple imputation (Supplementary Table 1). PaCO₂ values were missing in 9 cases (4.4%); however, because PaCO₂ was not included as a predictor in the final models, these values were not imputed.

The clinical utility of the models was assessed using decision curve analysis (Fig. 3). Across a clinically relevant range of threshold probabilities (approximately 0.22 to 0.50), the full multiparametric model (Model C) provided a higher net benefit for clinical decision-making than both the simpler models (A and B) and the default strategies of “escalate all” or “escalate none”.

Fig. 3.

Decision curve analysis. Net benefit of each prediction model and default strategies across a range of threshold probabilities.

Exploratory outcomes are summarized in Table 6. The median time to first respiratory support escalation was 7.8 hours (IQR, 3.4–18.6). The rates of noninvasive ventilation failure, repeated surfactant administration within 7 days, and in-hospital mortality were 15.3%, 8.9%, and 3.0%, respectively.

In this prospective cohort study, we demonstrated that a multiparametric ultrasound protocol performed within 6 hours after birth, integrating structural (LUS), mechanical (diaphragmatic ultrasound), and circulatory (pulmonary artery Doppler) assessments, significantly improves the early prediction of respiratory support escalation in preterm infants. The full model (Model C), which incorporated all three sonographic dimensions alongside key clinical variables, achieved superior discriminatory performance (AUC: 0.88), excellent calibration, and provided a greater net clinical benefit across a range of decision thresholds compared with models based solely on clinical factors or on clinical variables combined with LUS alone. This performance, which remained robust after rigorous internal bootstrap validation, underscores that respiratory support escalation is a multifactorial event that is more accurately captured by a physiology-integrated approach than by maturity indices or lung aeration status alone [8].

Our findings are consistent with the evolving understanding of neonatal respiratory failure as a triad of impaired lung aeration, inefficient respiratory pump function, and altered pulmonary perfusion [4]. Although LUS effectively quantifies loss of lung aeration and interstitial fluid accumulation, which are key drivers of hypoxemia and surfactant deficiency [6,7], it does not fully assess respiratory muscle reserve or pulmonary vascular adaptation. Diaphragm ultrasound provided independent prognostic value, as reduced diaphragmatic excursion and thickening fraction reflect impaired capacity to generate adequate tidal volume, thereby increasing the risk of hypercarbia and apnea, particularly in infants with highly compliant chest walls [10,16]. Concurrently, a lower pulmonary artery AT/ET ratio, indicative of elevated pulmonary vascular resistance, was independently associated with an increased risk of respiratory support escalation. This parameter likely reflects compromised pulmonary perfusion and ventilation-perfusion mismatch, thereby contributing to oxygenation failure that may not be directly detected by LUS alone [12,17]. Integration of these physiological signals creates a more comprehensive bedside assessment of pulmonary immaturity, which likely explains why Model C outperformed models lacking these components.

The independent associations of traditional risk factors, such as lower gestational age and incomplete antenatal corticosteroid treatment, with escalation risk confirm the fundamental role of structural immaturity and surfactant deficiency in neonatal respiratory failure [20]. Our model does not replace these critical clinical variables but rather complements them with real-time, modifiable physiological data, thereby potentially refining risk stratification among infants of similar gestational ages.

Current evidence supporting the use of ultrasound for respiratory prognostication remains fragmented. Many studies have validated LUS for predicting surfactant need or failure of noninvasive ventilation; however, these investigations have primarily reported discriminative metrics without adequately evaluating calibration or clinical utility [14,15]. Similarly, studies examining diaphragmatic function [9] or pulmonary hemodynamics [13] have typically been conducted independently or at different postnatal time points, thereby preventing an integrated assessment [18]. Our study addresses these gaps by prospectively quantifying all three physiological dimensions within a standardized early postnatal window (≤6 h after birth) and by employing a comprehensive model evaluation framework, assessing discrimination, calibration, and decision-curve net benefit, as recommended for clinical prediction models [21].

Our choice of the six-zone LUS protocol warrants further justification. Although extended protocols (e.g., 10- or 12-zone approaches) have been described and may provide more detailed granular regional aeration data, we selected the six-zone method for several reasons specific to our early postnatal assessment window. First, this protocol has been well-validated in preterm populations for predicting surfactant need and noninvasive ventilation failure [6,7], thereby providing a robust evidence base. Second, the six-zone approach offers a balance between comprehensiveness and efficiency; in our experience, it can be completed within 3–5 minutes, which is critical when assessing clinically unstable infants during the first 6 hours after birth. Third, minimizing handling and examination time reduces infant agitation and motion artifact, which likely directly contributed to the high measurement reproducibility observed in our study (ICC ≥0.94). Finally, although more extensive protocols may offer incremental diagnostic information, our primary aim was to evaluate the additive value of a multiparametric assessment strategy (diaphragm + Doppler) beyond a practical and clinically applicable LUS baseline. Therefore, the six-zone protocol represents a pragmatic compromise between diagnostic detail and real-world feasibility in the time-sensitive NICU setting.

Our study employed trained sonologists to ensure measurement consistency during this validation phase. However, the protocol was designed for future clinical translation. Standardized LUS systems are increasingly being performed by clinicians, and a recent systematic review concluded that LUS is feasible when performed by appropriately trained neonatologists [22]. Formal training curricula and credentialing pathways for neonatal POCUS, which encompasses lung, cardiac, and procedural applications, have now been established in North America, as evidenced by a recent national position statement [23]. The excellent interobserver reproducibility (ICC ≥0.87) we observed further suggests that, with such structured competency-based training, reliable measurements are attainable by clinical NICU teams, supporting the potential for widespread application.

The high reproducibility of all ultrasound measurements (ICC ≥0.87) confirms the feasibility of this protocol in unsedated neonates and supports its potential for broader clinical implementation [24]. Furthermore, the observed clustering of respiratory support escalation events early after birth (median, 7.8 hours) validates the clinical rationale for performing assessments within the ≤6-hour assessment window, which aims to inform management decisions before significant therapeutic interventions.

This study has several limitations. First, as a single-center study, our findings require external validation across diverse clinical settings with varying patient demographics, levels of ultrasound expertise, and respiratory management protocols (e.g., surfactant use strategies) [25]. Second, while we employed a strict blinding protocol and predefined objective escalation criteria, the observational design cannot fully exclude the possibility of residual confounding or subtle clinician-related bias in treatment decisions. Third, diaphragm ultrasound metrics can be influenced by the mode and level of concurrent respiratory support. Our study design prioritized a pragmatic bedside assessment approach; therefore, ultrasound examinations were performed without altering the infant’s ongoing respiratory management. Consequently, our measurements reflect a combination of intrinsic muscle function and the effects of ongoing respiratory assistance rather than isolated muscle performance. However, this composite assessment likely represents the most clinically relevant context for a predictive tool intended for use in the dynamic NICU environment. Fourth, although the ≤6-hour window was designed to capture early postnatal physiology, some infants may have received stabilizing support (e.g., continuous positive airway pressure [CPAP]) before ultrasound evaluation. Consequently, the ultrasound may partially reflect a postintervention physiological phenotype. Fifth, we used conventional logistic regression without exploring nonlinear relationships or interaction effects among predictors. Future work could investigate more flexible machine learning methods (e.g., random forest models) to enhance predictive performance. Sixth, we did not apply formal multiple comparison corrections (e.g., Bonferroni) to the model performance metrics (e.g., AUC comparisons). Our analyses were based on prespecified sequential model comparisons rather than exploratory testing across unrelated hypotheses. Furthermore, internal validation via bootstrap resampling was performed to correct for overoptimism. The consistency of our findings across discrimination, calibration, and clinical utility metrics supports the robustness of the primary conclusion regarding the added value of the multiparametric model. Seventh, our primary outcome focused on short-term respiratory support escalation; therefore, the association between this model and longer-term outcomes, such as bronchopulmonary dysplasia, remains to be investigated.

Future research should prioritize external validation of the model and, if successful, progress to interventional trials to evaluate whether model-guided management strategies (e.g., targeting high-risk infants for enhanced monitoring or early selective therapy) improve clinical outcomes without increasing unnecessary interventions.