This double-masked cohort study was conducted in the NICU to assess NRDS severity in neonates using LUS and other clinical indicators. The entire research process was conducted within the NICU to ensure legality and ethical integrity. NICU clinical management adhered to local guidelines, with the study not directly influencing clinical practices. Researchers ensured neonates received appropriate treatment and care while collecting clinical data and LUS scores for subsequent analysis. Researchers used the STROBE checklist during manuscript preparation to ensure the study’s reliability and scientific rigor.

This study involved 336 pediatric patients admitted to the NICU at Yantaishan Hospital from December 2020 to July 2023. Neonates were selected based on strict inclusion criteria to ensure they met the diagnostic standards for NRDS. From the original cohort, 106 patients were excluded for reasons including an unclear diagnosis, withdrawal from treatment, congenital anomalies, tuberculosis, or congestive heart failure. As a result, 230 neonates remained in the study, with 119 identified as having NRDS and 111 categorized as N-NRDS. The diagnostic procedures were carried out using a Philips CX50 portable ultrasound machine equipped with a linear array probe, operating at a frequency range of 8 to 12 MHz. The study protocol adhered to medical ethical standards and received approval from the Medical Ethics Committee of Yantai Mountain Hospital (approval number: 20220001), with informed consent obtained from the patients’ families. All data collection and processing followed relevant guidelines and standards.

The inclusion criteria for this study were grounded in the Berlin criteria, encompassing clinical symptoms, arterial blood gas analysis, and chest X-ray results. A diagnosis of NRDS was confirmed by evaluating clinical signs, arterial blood gas measurements, and radiographic evidence. Eligible neonates were those born within 6 to 12 h before disease onset, showing severe respiratory distress, cyanosis, expiratory grunting, inspiratory retractions, and related symptoms. Arterial blood gas analysis required evidence of hypoxemia, defined as an arterial oxygen partial pressure below 60 mmHg, or hypercapnia, indicated by a carbon dioxide partial pressure exceeding 50 mmHg. Chest X-rays showed decreased lung transparency, increased lung markings, evenly distributed granulation and reticular opacities, blurred cardiac and diaphragmatic outlines, bronchial aeration, and signs of severe pneumonia in both lung fields. Inclusion required that the neonate have a complete clinical medical record.

The exclusion criteria for this study include neonates with congenital disorders such as complex respiratory malformations, congenital heart disease, congenital mental disorders, or chromosomal abnormalities. Neonates showing signs of cardiogenic pulmonary edema or persistent pulmonary hypertension. Neonates with severe intracranial hemorrhage, sepsis, severe hypoxic–ischemic encephalopathy, severe sepsis, septic shock, or other serious illnesses and multi-system complications. Neonates diagnosed with pulmonary hemorrhage, neonatal wet lung, pneumothorax, meconium aspiration syndrome, neonatal infectious pneumonia, or other respiratory disorders. Additionally, neonates were excluded if factors led to increased measurement errors or significantly compromised ultrasound image quality.

Lung ultrasound score has diverse applications in clinical assessment (19). This study employed a 12-region scoring system, dividing the pediatric chest into 12 zones, with each lung region systematically scanned and scored based on specific ultrasound images, leading to the calculation of the LUS score. The scoring criteria outlined in Table 1 provide a standardized approach for the quantitative evaluation of pulmonary diseases and their severity. Continuous, smooth A-lines or the presence of fewer than three isolated B-lines are assigned a score of 0, indicating normal lung status. A score of 1 is given when three or more scattered, distinct B-lines reflect a mild abnormality. A score of 2 is assigned for diffusely confluent B-lines appearing in a waterfall pattern, indicating moderate abnormality. Lung consolidation, with a score of 3, indicates severe abnormality. The total LUS score ranges from 0 to 36, as depicted in Figure 1, which presents representative images from our institution’s assessment of NRDS scores between 0 and 3 (20).

The OI is a critical measure for assessing the severity of NRDS, calculated as the ratio of arterial oxygen partial pressure (PaO₂) to the fraction of inspired oxygen (FiO₂) (21). The formula for its calculation is OI = PaO₂/FiO₂. Typically, the normal range for the oxygenation index is between 400 and 500 mm Hg; lower scores suggest worsening pulmonary function and oxygenation capacity in the neonate, thereby indicating the severity of the condition. RI is crucial for assessing children’s respiratory function. It is calculated as the ratio of the alveolar-arterial oxygen pressure difference (A-aDO₂) to PaO₂. The formula for its calculation is RI = A-aDO₂/PaO₂. RI between 0.8 and 1.2 is considered normal, indicating good respiratory function, regular lung activity, and sufficient gas exchange. An RI between 1.2 and 1.6 suggests mild abnormalities, potentially indicating mild dyspnea due to respiratory muscle fatigue or upper respiratory tract infection. An RI of 1.6 to 2.0 indicates moderate abnormalities, usually signifying moderate dyspnea and pulmonary function decline, possibly due to conditions like chronic obstructive pulmonary disease. An RI exceeding 2.0, particularly above 3.0, indicates severe abnormalities, often reflecting significant respiratory dysfunction, potentially due to severe pulmonary diseases or respiratory acidosis, leading to substantial dyspnea and critically impaired lung function.

The SOFA score includes six systems: respiratory, coagulation, liver, circulatory, central nervous, and renal (22). Each system is scored from 0 to 4, with the total SOFA score being the sum of these scores. Table 2 shows that when evaluating the SOFA score in children with NRDS, special attention should be given to each system’s score, as a higher total score indicates more severe organ dysfunction. According to the criteria, the SOFA score typically ranges from 0 to 24.

Figure 2 shows the flowchart of this study. This study addresses a binary classification problem ideally suited for supervised learning in machine learning. Supervised learning techniques encompass a variety of algorithms such as linear and LR, SVM, Naive Bayes, extreme gradient boosting, decision trees, and RF. In this research, we selected LR, RF, NN, and SVM as the training models commonly used in medical treatment.

Data sets with statistically significant differences were selected to construct machine learning metrics. Subsequently, 161 cases (70%) were randomly chosen from the entire sample as the training set. The remaining 69 cases (30%) were used as the test set. The training set data underwent five-fold cross-validation to achieve the optimal training model. The advantage of cross-validation is that it uses all data for both training and validation sets, with each subset independently validated, providing a more robust reflection of the training model. Finally, an independent external validation set was selected, consisting of 44 NRDS and 43 N-NRDS patients, to test and evaluate the model’s accuracy.

Data entry was performed using SPSS software. Continuous variables that followed a normal or approximately normal distribution were expressed as mean ± standard deviation ( x ¯ ±s), with group comparisons made using the independent samples t-test. For continuous variables not following a normal distribution, data were reported as median (interquartile range), and group differences were assessed using non-parametric rank-sum tests. All data analyses and the construction of machine learning models were performed using R software (version 4.0.3). All models were built using the Caret package, with the importance of predictive variables assessed through the vamp function within Caret. In this study, a p-value of <0.05 was considered statistically significant.

Table 3 presents the baseline characteristics of the two groups. In the NRDS cohort, 47.90% of the participants were male, and the median gestational age was 32 (27,36) weeks. The mean birth weight was 2013 ± 622 g, and 60.50% of deliveries were by cesarean section. In the N-NRDS group, the proportion of males was 60.30%, the median gestational age was 35 (30,39) weeks, the mean birth weight was 2,291 ± 741 g, and cesarean deliveries accounted for 54.05% of the total deliveries.

The LUS and clinical score variations across different patient groups were compared, as shown in Table 4. The NRDS group exhibited higher LUS, SOFA, and RI values than the N-NRDS group, while OI values were lower (p < 0.01).

Patients with NRDS were categorized based on the severity of their pulmonary disease into mild, moderate, and severe groups. Table 5 displays LUS levels and clinical scores corresponding to various degrees of NRDS, revealing significant statistical differences among the groups (p < 0.01). The severe NRDS group showed significantly elevated LUS, SOFA, and RI scores relative to the mild and moderate groups, whereas the OI was notably lower.

To develop the NRDS predictive model, four commonly used machine learning algorithms were employed: LR, RF, NN, and SVM. Based on this foundation, ensemble and cascading methods were used to combine the individual classification models, and the performance of the integrated classifier was subsequently evaluated.

We visualized the top eight variables according to their weights. The length of the bars is directly proportional to the importance of each variable. The results indicate that the model’s key variables fall into two groups, with the first group—LUS, OI, SOFA, and RI—significantly impacting predictive performance. LUS is the most influential variable in the LR, RF, and NN models. LUS is the most influential variable for the LR, RF, and NN models. The second group consists of cesarean birth, gestational age, weight, and gender, which have a relatively minor impact on prediction; detailed information is provided in Figure 3.

This research utilizes five key performance metrics, the AUC, accuracy, sensitivity, specificity, and the F1 score, to assess the effectiveness of four machine learning algorithms, as detailed in Table 6. ROC curves were plotted using models constructed from the training dataset (Figure 4), with test set AUC values as follows: LR (AUC = 0.841, 95% CI: 0.712–0.881), RF (AUC = 0.894, 95% CI: 0.811–0.976), NN (AUC = 0.828, 95% CI: 0.701–0.935), and SVM (AUC = 0.726, 95% CI: 0.623–0.887). The RF model demonstrated the highest performance of the four models, with an AUC of 0.894, accuracy of 0.808, and sensitivity of 0.706. In contrast, the AUC values for the LR, NN, and SVM models were notably lower. Given the superior performance of the RF model, it was selected as the final model for this study.

This study used calibration curves to evaluate the accuracy of the model. Figure 5 depicts the calibration curves for the training, internal validation, and external validation sets. The results show that the nomogram’s predictive performance across all sets is closely aligned with ideal performance.