Improving early risk stratification for bronchopulmonary Dysplasia in preterm infants using machine learning: development and external validation

Abstract

Background

Despite advances in perinatal care and improved survival of preterm infants, bronchopulmonary dysplasia (BPD) remains a major neonatal morbidity with long-term consequences. Accurate early risk prediction is therefore clinically important. Conventional models may be limited in capturing the complex and nonlinear interactions underlying BPD, whereas machine learning offers greater flexibility for high-dimensional clinical data. However, limited interpretability remains a barrier to clinical implementation.

Method

This retrospective study included 1,083 preterm infants with gestational age <32 weeks and birth weight <2,500 g admitted to the Affiliated Women’s Hospital of Jiangnan University from January 2011 to December 2025. Data were split 7:3 into training and validation sets. Seven models were evaluated, with hyperparameters optimized by 10-fold cross-validation using AUC. Missing data were handled by complete case analysis. Performance was assessed using AUC-ROC, Brier score, and decision curve analysis (DCA), while SHAP and partial dependence plots (PDP) were used for interpretation. The optimal model was externally validated in an independent cohort of 1,000 neonates.

Result

GBM showed the best internal performance, with an AUC of 0.880, accuracy of 0.793, sensitivity of 0.827, specificity of 0.759, and Brier score of 0.141. Birth weight, gestational age, and 1-minute Apgar score were the leading predictors. Neonatal anemia was strongly associated with BPD (aOR = 5.036, 95% CI: 3.35–7.57). PDP demonstrated a nonlinear association between birth weight and BPD risk, with a sharp increase between 1,250 and 1,500 g. DCA indicated favorable net clinical benefit across relevant threshold probabilities.

Conclusion

The GBM model provides interpretable, individualized early BPD risk estimates using routinely available clinical variables and may support identification of high-risk preterm infants. It is intended for risk stratification rather than treatment guidance. Prospective multicenter validation is required before routine clinical implementation.

KEY MESSAGES

• A gradient boosting machine (GBM) model incorporating routinely available clinical variables achieved high predictive accuracy (AUC 0.880) for early risk stratification of bronchopulmonary dysplasia in preterm infants and demonstrated strong discriminative performance in external validation.

• Gestational age, birth weight, and 1-minute Apgar score were the strongest predictors of BPD, with neonatal anemia emerging as a major independent risk factor (aOR 5.036, 95% CI: 3.35–7.57).

• The model provides interpretable, individualized risk estimates with favorable clinical utility across a range of threshold probabilities, offering a practical tool to support early risk stratification decisions in the neonatal intensive care unit.

1. Introduction

Bronchopulmonary dysplasia (BPD) remains a major challenge in neonatal intensive care. Although advances in neonatal management, including corticosteroid therapy, surfactant administration, and lung-protective ventilation, have substantially improved the survival of preterm infants, a corresponding reduction in the incidence of BPD has not been observed. The number of BPD cases keeps rising. Researchers refer to this trend as a ‘survivor paradox’. Since these highly fragile infants survive early birth, a larger group naturally faces the risk of chronic lung damage. In the past, physicians viewed BPD as a short-term breathing delay. Current clinical evidence tells a different story. We now understand that the disease triggers lasting heart, lung, and brain problems. These health struggles frequently continue well into adulthood [1,2]. Identifying at-risk infants early is an urgent need. Medical teams must spot these vulnerable patients right away to start prompt interventions.

Accurate risk assessment in BPD remains an elusive goal, largely due to the heterogeneous clinical landscape of the disease. Central to its development is the developmental epoch during which the primary noxious stimulus occurs, an interaction that defines the resultant lung pathology. Adopting a ‘multiple-hit’ perspective, we see BPD as an evolving process initiated by intrauterine sensitizing events that compromise the fetal baseline. The neonatal lung, already in a state of compromised equilibrium, must navigate a gauntlet of postnatal stressors; the cumulative burden of ventilatory support and infectious episodes often suffices to arrest the orchestrated progression of alveolarization and structural remodeling. The resulting biochemical cascade of inflammation and oxidative stress acts as a fundamental barrier to normal alveolar and microvascular development [3–5]. Managing this interplay of stressors presents a pervasive diagnostic dilemma, complicating the delivery of precise, individualized care in the NICU.

Traditional BPD prediction models are predominantly based on conventional regression methods and a limited number of predefined clinical variables. Although these models offer relatively straightforward interpretation, their performance may be constrained by linearity assumptions, limited capacity to capture complex interactions, and insufficient accommodation of high-dimensional and nonlinear clinical data. Moreover, most traditional models provide static risk estimates at predefined time points and therefore may not fully reflect the dynamic evolution of BPD. These limitations can restrict predictive accuracy, generalizability, and individualized risk stratification in heterogeneous preterm populations.

These structural flaws directly cause suboptimal predictive performance. The AUC metrics for these models typically plateau between 0.65 and 0.75. Clinical teams require much higher accuracy for bedside decision-making. Researchers have published numerous BPD prediction tools over the years. The persistent mathematical and clinical limitations keep almost all of them out of daily NICU workflows [6,7].

Modern intensive care units generate massive amounts of complex Electronic Health Record (EHR) data. Classical statistical methods fail to capture the high-dimensional, non-linear nature of this information. Clinical teams now track intricate patient details, including continuous respiratory waveforms and complex inflammatory profiles. Advanced machine learning (ML) architectures easily process these vast clinical datasets.

Our research group designed a new framework to resolve these exact barriers. We integrated SHapley Additive Explanations (SHAP) directly into our methodology. This analytical tool exposes the precise variables driving each risk assessment. Clinicians finally receive a transparent and practical system for patient risk stratification.

2. Methods and data collection

2.1. Research design and study subjects

This investigation utilized an extensive clinical repository spanning a fifteen-year window from January 2011 to December 2025 at the Affiliated Women’s Hospital of Jiangnan University. We conducted a retrospective cohort study to evaluate differences between BPD-diagnosed preterm neonates and unaffected controls. The study protocol was approved by the Institutional Ethics Committee and carried out in accordance with the Declaration of Helsinki (No. YLSL2025-094).

We selected the 5th day after birth as the prediction time point. All collected data were available prior to this observation time point. For multiple data points of the same variable, we selected the one closest to the observation time point for inclusion.

The inclusion criteria for the BPD group are as follows: 1) Gestational age at birth < 34 weeks; 2) Diagnosis of BPD meeting the 2001 NICHD criteria; 3) Birth weight < 2,500 g; 4) Complete medical records; the exclusion criteria are: 1) Presence of severe congenital malformations, chromosomal abnormalities, or inherited metabolic diseases; 2) Hospital stay of less than 28 days or death; 3) Incomplete clinical information. The control group consisted of infants without BPD. The inclusion criteria for this group were items 1), 3), and 4) from the BPD inclusion criteria, while the exclusion criteria were identical to those of the BPD group.

A 1:1 individual matched case-control design was employed, with birth month as the matching factor. Controls were selected using simple random sampling from the eligible pool, and random replacement was performed for any non-compliant subjects.

During the study period, a total of 1684 patients were enrolled, of whom 21 had incomplete data, 14 died, and 566 did not meet the inclusion criteria. Ultimately, 1083 patients were included in the final analysis. This cohort maintained a balanced distribution, containing 542 BPD cases and 541 controls. As illustrated in the flowchart, every one of these 1,083 cases was utilized in the final statistical analysis (Figure 1). The external validation cohort included neonates admitted to the NICU of the First Affiliated Hospital of Nanjing Medical University between January 2020 and December 2025. The sampling methods, inclusion and exclusion criteria were identical to those used in the modeling cohort. During this period, the NICU admitted approximately 8,870 neonates, among whom 500 developed BPD, yielding an incidence rate of 5.64%. Baseline characteristics are presented in Table S1.

Figure 1.

Incorporate into the flowchart.

2.2. Clinical features

Our research group extracted candidate predictors for BPD modeling directly from the institution’s EHR records. We selected these independent variables using a combination of daily clinical expertise and published medical research. A rigorous evaluation of the infants’ clinical profiles requires strict data organization. Our team separated the chosen variables into four distinct conceptual domains to structure the final analysis.

2.2.1. Perinatal and sociodemographic variables

We extracted a comprehensive set of perinatal predictors from the medical records. The dataset includes maternal baselines like age, parity, conception mode, and obstetric history. We tracked major pregnancy morbidities, specifically pregnancy-induced hypertension (PIH), gestational diabetes (GDM), hyperthyroidism, and cervical insufficiency. Our team also monitored critical intrauterine factors. These conditions involved placental or fetal anomalies, amniotic fluid issues, premature rupture of membranes (PROM), and intrauterine infections. We carefully documented the administration of antenatal corticosteroids (ACS). For the neonate, we recorded fundamental delivery metrics: gestational age (GA), birth weight (BW), sex, and delivery method. We finalized the clinical profile using immediate transition indicators. These indicators included fetal distress, 1- and 5-minute Apgar scores, and initial vital signs upon NICU admission.

2.2.2. Neonatal complications and morbidities

An accurate model must reflect the infant’s complete multisystemic health status. We systematically extracted data on major respiratory and cardiovascular challenges. Medical records specified cases of Respiratory Distress Syndrome (RDS) and Patent Ductus Arteriosus (PDA). Clinical teams also tracked other associated pulmonary conditions. Vulnerable neonates are often exposed to severe systemic and hematological stressors. We explicitly documented all episodes of sepsis and asphyxia. Gastrointestinal implications add another layer of clinical risk. Our database captures exact occurrences of internal hemorrhage and distension. This rigorous tracking establishes a clear health baseline for every patient.

2.2.3. Therapeutic interventions

We categorized therapeutic interventions to capture exact NICU care details. Our dataset includes complete infant medication records. We specifically tracked antibiotics, pulmonary surfactant (PS), and methylxanthines like caffeine and theophylline. We also documented advanced respiratory support strategies. These clinical records specify all mechanical ventilation modalities and tracheostomy requirements.

2.2.4. Laboratory and diagnostic investigations

To capture the multifaceted nature of neonatal health, the research protocol prioritized the documentation of acute cardiorespiratory status. Initial assessments included immediate blood gas analysis to measure PaO₂ and pH, along with bedside echocardiography to evaluate ductus arteriosus hemodynamics. Beyond cardiovascular monitoring, the clinical team actively tracked systemic inflammation. Specifically, they monitored C-reactive protein (CRP) levels, leukocyte counts, and neutrophil percentages to gauge the infant’s overall physiological stress. To complete the clinical picture, these parameters were substantiated by microbiological data, with blood and sputum culture results being meticulously extracted from the inpatient records during critical windows of care.

2.3. Diagnostic criteria

Neonatal complications were diagnosed according to established international criteria.

2.3.1. Bronchopulmonary dysplasia

The diagnosis of BPD required the use of supplemental oxygen for 28 days or more. We assessed disease severity at 36 weeks postmenstrual age (PMA) and classified it into three levels: mild (infant breathes room air), moderate (requires FiO₂ < 30%), and severe (requires FiO₂ ≥ 30% or positive pressure ventilation) [8].

2.3.2. Pulmonary hypertension

Pulmonary hypertension (PH) was diagnosed via echocardiography during the first 72 h of life. The diagnosis required a systolic pulmonary arterial pressure (sPAP) ≥ 35 mmHg [9].

2.3.3. Neonatal asphyxia

The diagnosis of neonatal asphyxia required an umbilical cord blood pH ≤7.15 and an Apgar score <8 at 1 or 5 min. For infants with a 5-minute Apgar score below 7, resuscitation was maintained as clinically indicated. A definitive diagnosis also required observed neurological dysfunction after excluding other causes [10].

2.3.4. Other complications

Neonatal anemia: Hemoglobin < 145 g/L within 5 days after birth.

Sepsis: Positive blood cultures for pathogenic bacteria.

PDA: Confirmed by echocardiography performed within 5 days after birth.

Surfactant use: Administration of pulmonary surfactant within 5 days after birth.

Respiratory intervention: Any form of respiratory support provided within 5 days after birth.

Other neonatal and maternal conditions were defined according to standard international guidelines and institutional protocols.

2.4. Statistical analysis and model development

2.4.1. Computational environment and data pre-processing

All statistical analyses were performed using R software (version 4.5.2; R Foundation for Statistical Computing, Vienna, Austria). Laboratory parameters with incomplete data were excluded from the analysis.

2.4.2. Descriptive statistics and hypothesis testing

The normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed data are presented as mean ± standard deviation (SD) and were compared using Student’s t-test. Non-normally distributed data are expressed as median [interquartile range, IQR] and compared using the Mann–Whitney U test.

Categorical variables are presented as counts and percentages [n (%)]. Group differences were compared using the Chi-square test or Fisher’s exact test, as appropriate. Yates’ continuity correction was applied when expected cell frequencies were less than 5. A two-sided p-value < 0.05 was considered statistically significant.

2.4.3. Feature engineering and model building

We developed the model using ten clinical predictors selected for their clinical relevance and statistical significance: GA, BW, neonatal anemia, RDS, PS administration, 1-minute Apgar score, postnatal resuscitation, PDA, ACS exposure, and sex.

We converted categorical variables into a binary format for the machine learning algorithms: sex (male = 1, female = 0), delivery mode (cesarean = 1, vaginal = 0), and early-onset sepsis (present = 1, absent = 0). We left all continuous variables unscaled to preserve their clinical interpretability.

In the modeling process, the dataset was split at a 7:3 ratio. Strict 10-fold cross-validation was performed within the training set (Table S3). We compared seven machine learning algorithms to identify the optimal predictive model: Gradient Boosting Machine (GBM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), and Extra-Trees.

2.4.4. Validation and performance evaluation

Medical algorithms require rigorous testing before hospital deployment. Our team first evaluated the internal discriminative capacity of the predictive framework. We measured the exact AUC, sensitivity, and specificity. We also established the model’s calibration using the Brier score. Pure statistical accuracy does not guarantee practical value. We applied DCA to confirm the actual clinical net benefit.

The inherent ‘black-box’ nature of high-dimensional models often precipitates a lack of trust among frontline clinicians, hindering widespread adoption. A key objective of this study was to make the model’s decision-making process transparent. By incorporating SHAP analysis, we successfully mapped the specific weights driving each risk assessment. Beyond internal transparency, an algorithm must perform well outside its original training environment. We confirmed this by evaluating the framework against an external cohort of 1,000 cases at Nanjing Medical University’s First Affiliated Hospital, proving that its predictive power holds up against differing institutional practices.

3. Results

3.1. Participant characteristics and variable selection

Univariate analysis identified several potential risk factors for BPD (Table 1). For the adjusted odds ratios, please refer to Table S4. Based on clinical relevance and statistical significance, we selected ten primary predictors for the model. These included GA, BW, neonatal anemia, RDS, PS administration, 1-minute Apgar score, postnatal resuscitation, PDA, ACS exposure, and sex. These variables were selected to ensure the development of a nuanced, multidimensional predictive framework capable of capturing the heterogeneous nature of BPD development.

Table 1.

Results of univariate analysis of each variable.

Characteristics/variables	Non-BPD group (n = 540)	BPD group (n = 543)	p-value	Statistic (Z/χ²)	OR (95% CI)
Maternal age (years)	30(27–33)	30(28–33)	0.016	2.402	1.03 (1.00–1.06)
Gravidity (n)	2(1–3)	2(1–3)	0.033	2.134	1.10 (1.01–1.21)
Parity (n)	1(1–2)	1(1–2)	0.398	0.845	0.92 (0.76–1.12)
GA (weeks)	31(30–32)	29(28–30)	<0.001	18.075	0.50 (0.46–0.55)
BW (g)	1425(1260–1600)	1170(1030–1365)	<0.001	14.037	1.00 (1.00-1.00)
Apgar score (1-min)	8(7–9)	7(6–8)	<0.001	11.138	0.69 (0.64–0.74)
Apgar score (5-min)	9(9–10)	8(8–9)	<0.001	11.201	0.57 (0.50–0.64)
Respiratory rate (bpm)	68(66–70)	68(66–70)	0.915	0.106	1.00 (0.99–1.00)
Systolic blood pressure (mmHg)	60(58–62)	57(52–60)	<0.001	12.621	0.85 (0.82–0.87)
Diastolic blood pressure (mmHg)	31(30–32)	29(28–31)	<0.001	10.587	0.80 (0.76–0.84)
WBC (×10^9/L)	8.51(6.45–12.2025)	9.3(6.705–14.4975)	0.004	2.907	1.03 (1.02–1.05)
Neutrophil percentage (%)	40.85(30.7–51.1)	46.7(31.9–61.5)	<0.001	5.27	1.02 (1.01–1.03)
PLT (×10^9/L)	231(185.75–277)	225(186–269.5)	0.343	0.948	1.00 (1.00-1.00)
HCT (%)	49.7(45.675–53.825)	48.59(44.9–52.2)	0.001	3.247	0.98 (0.96–0.99)
CRP (mg/L)	4(2–4)	4(4–4)	0.004	2.907	1.05 (1.02–1.09)
PaO2 (mmHg)	69.8(62–77)	68.9(60.15–77)	0.203	1.274	0.99 (0.98–1.00)
pH value	7.28(7.23–7.32575)	7.28(7.23–7.33)	0.549	0.599	0.59 (0.12–2.95)
Na+ (mmol/L)	137(135–139.65)	136.4(134.5–139)	0.023	2.28	0.98 (0.95–1.01)
K+ (mmol/L)	4.6(4.2–5.2)	4.67(4.2–5.405)	0.047	1.983	1.24 (1.08–1.43)
Creatinine (μmol/L)	66.2(48.45–83.5)	63.1(46.025–84.725)	0.396	0.849	1.00 (0.99–1.00)
BUN (mmol/L)	4.54(3.1275–6.7125)	4.76(2.935–7.765)	0.187	1.319	1.06 (1.02–1.10)
HR (bpm)	142(139.75–146)	144(140–150)	<0.001	4.669	1.03 (1.02–1.05)
Body temperature (°C)	36.5(36.5–36.6)	36.5(36.4–36.6)	0.027	2.211	1.01 (0.99-NA)
MPV (fL)	10(9.5–10.6)	10.1(9.7–10.7)	0.002	3.039	0.99 (0.95–1.00)
PMI (fL × 10^9/L)	2277.3(1843.38–2752.2)	2286.6(1891.15–-2665.95)	0.941	0.075	1.00 (1.00-1.00)
ART, n(%)	87(16.1%)	129(23.8%)	0.002	9.44	1.62 (1.20–2.20)
Adverse obstetric history, n(%)	205(38.0%)	252(46.4%)	0.006	7.576	1.42 (1.11–1.80)
ACS, n(%)	320(59.3%)	368(67.8%)	0.004	8.104	1.45 (1.13–1.86)
Complete course of antenatal corticosteroids, n(%)	266(49.3%)	290(53.4%)	0.192	1.702	1.18 (0.93–1.50)
GDM, n(%)	81(15.0%)	125(23.0%)	0.001	10.792	1.69 (1.25–2.32)
HDP, n(%)	119(22.1%)	88(16.2%)	0.017	5.654	0.68 (0.50–0.93)
Cervical incompetence, n(%)	25(4.6%)	43(7.9%)	0.035	4.435	1.77 (1.07–2.98)
Thyroid dysfunction, n(%)	30(5.6%)	28(5.2%)	0.876	0.025	0.92 (0.54–1.57)
Intrauterine infection, n(%)	21(3.9%)	29(5.3%)	0.320	0.987	1.39 (0.79–2.51)
Placental abnormalities, n(%)	108(20.0%)	81(14.9%)	0.034	4.509	0.70 (0.51–0.96)
Abnormal amniotic fluid, n(%)	136(25.2%)	185(34.1%)	0.002	9.827	1.54 (1.18–2.00)
Multiple pregnancy, n(%)	165(30.6%)	141(26.0%)	0.112	2.53	0.80 (0.61–1.04)
PROM, n(%)	165(30.6%)	210(38.7%)	0.006	7.529	1.43 (1.11–1.84)
Cesarean section, n(%)	326(60.4%)	271(49.9%)	0.001	11.561	0.65 (0.51–0.83)
Intrauterine distress, n(%)	46(8.5%)	21(3.9%)	0.002	9.306	0.43 (0.25–0.72)
Postnatal resuscitation, n(%)	193(35.7%)	322(59.3%)	<0.001	59.315	2.62 (2.05–3.35)
Male, n (%)	298(55.2%)	354(65.2%)	0.001	10.905	1.52 (1.19–1.94)
Postnatal surfactant administration, n(%)	230(42.7%)	418(77.0%)	<0.001	131.115	4.49 (3.46–5.86)
Caffeine/Theophylline therapy, n(%)	159(29.4%)	396(72.9%)	<0.001	203.166	6.46 (4.97–8.43)
Endotracheal intubation, n(%)	98(18.1%)	258(47.5%)	<0.001	104.481	4.08 (3.11–5.40)
Respiratory support, n(%)	510(94.4%)	538(99.1%)	<0.001	17.145	6.33 (2.66–18.68)
Antibiotic utilization, n(%)	524(97.0%)	538(99.3%)	0.013	6.205	4.11 (1.49–14.41)
Neonatal infection, n(%)	295(54.6%)	364(67.0%)	<0.001	16.973	1.69 (1.32–2.16)
PDA, n(%)	233(43.1%)	343(63.2%)	<0.001	42.78	2.26 (1.77–2.89)
Pulmonary hypertension, n(%)	17(3.1%)	20(3.7%)	0.747	0.104	1.18 (0.61–2.30)
Neonatal sepsis, n(%)	30(5.6%)	71(13.1%)	<0.001	17.227	2.56 (1.66–4.04)
Pulmonary hemorrhage, n(%)	5(0.9%)	27(5.0%)	0.000	14.121	5.61 (2.33–16.65)
Neonatal asphyxia, n(%)	152(28.2%)	259(47.7%)	<0.001	42.829	2.32 (1.81–2.99)
RDS, n (%)	244(45.2%)	462(85.1%)	<0.001	188.166	6.92 (5.20–9.30)
Neonatal anemia, n(%)	246(45.6%)	473(87.1%)	<0.001	207.649	8.08 (6.00–11.00)
Gastrointestinal bleeding, n(%)	67(12.4%)	68(12.5%)	1	0	1.01 (0.70–1.45)
Abdominal distension, n(%)	193(35.7%)	266(49.0%)	<0.001	18.916	1.73 (1.35–2.20)
Sputum culture, n(%)	8(1.5%)	19(3.5%)	0.053	3.742	2.41 (1.08–5.89)
Blood culture, n(%)	32(5.9%)	41(7.6%)	0.345	0.893	1.30 (0.81–2.10)

Note: Continuous variables are displayed as the median (interquartile range, IQR) or mean ± SD appropriately selected. Categorical variable is presented in terms of frequency and percentage(n%, n); P-values come from using the Mann-Whitney U test, a Student’s t-test, or χ² test.

GA: gestational age; BW: birth weight; ACS: antenatal corticosteroids; ART: assisted reproductive technology; PaO₂: partial pressure of arterial oxygen; PLT: platelet count; HT: hematocrit; BUN: blood urea nitrogen; HR: heart rate; MPV: mean platelet volume; PMI: Platelet Mass Index; GDM: gestational diabetes mellitus; PROM: preterm rupture of membranes; NRDS: neonatal respiratory distress syndrome; PDA: patent ductus arteriosus.

3.2. Maternal and antenatal risk factors

Maternal age and parity were both higher in the BPD group (p = 0.016 and p = 0.033, respectively). Mothers of infants with BPD had a higher prevalence of gestational diabetes $\mathrm{(}\text{23.0\% vs.15.0\%;}\mathrm{a}\mathit{\text{OR = 1.762, }}p\mathrm{<}\text{0.010}\mathrm{)}$ and were more likely to have a history of adverse obstetric outcomes (46.4% vs. 38.0%, p < 0.05). While ACS exposure was more common in the BPD cohort $\mathrm{(}\mathrm{a}\mathit{\text{OR = 1.600, }}p\mathrm{<}\text{0.014}\mathrm{)}$.

3.3. Perinatal and neonatal baseline characteristics

Marked disparities in physiological maturity were evident at birth, with the BPD cohort exhibiting a notably lower median gestational age compared to controls (29 vs. 31 weeks; p < 0.001). A corresponding deficit was observed in birth weight, where infants eventually diagnosed with BPD averaged 1,170 g, as opposed to 1,425 g in the unaffected group (p < 0.001). This baseline vulnerability was further corroborated by significantly lower Apgar scores at both the 1-minute and 5-minute intervals (p < 0.001). Collectively, these clinical markers underscore the compromised capacity for postnatal adaptation characteristic of the BPD population during the immediate transition after delivery.

Several perinatal complications were associated with a higher risk of BPD. Specifically, PROM $\mathrm{(}\mathrm{a}\mathit{\text{OR = 1.738, }}p\text{= 0.006}$, and cesarean section $\mathrm{(}\text{aOR = 1.625,}p\text{= 0.028}\mathrm{)}$ all increased the likelihood of BPD development. Abnormal amniotic fluid volume, cervical incompetence, and intrauterine distress had no statistically significant effect on the incidence of BPD.

3.4. Clinical interventions and neonatal complications

BPD-affected infants required significantly more intensive clinical support than the control group. Specifically, the use of PS was higher $\mathrm{(}\text{77.0\% vs.42.7\%; OR = 4.49,}p\mathrm{<}\text{0.001}\mathrm{)}$, as were the rates of mechanical ventilation $\mathrm{(}\text{99.1\% vs.94.4\%; OR = 6.33,}p\mathrm{<}\text{0.001}\mathrm{)}$ and endotracheal intubation $\mathrm{(}\mathit{\text{OR = 4.08, }}p\mathrm{<}\text{0.001}\mathrm{)}$. Additionally, infants in the BPD cohort were more likely to require postnatal resuscitation $\mathrm{(}\mathit{\text{OR = 2.62, }}p\mathrm{<}\text{0.001}\mathrm{)}$.

We found that certain respiratory and hematological conditions were highly prevalent in the BPD cohort. Specifically, neonatal anemia was the most significant independent predictor of BPD, affecting 87.1% of cases compared to 45.6% of controls $\mathrm{(}\mathrm{a}\mathit{\text{OR = 5.036, }}p\mathrm{<}\text{0.001}\mathrm{)}$. RDS $\mathrm{(}\mathrm{a}\mathit{\text{OR = 2.211, }}p\text{= 0.018}\mathrm{)}$ also demonstrated strong positive associations with BPD development.

PDA $\mathrm{(}\mathit{\text{OR = 1.647, }}p\text{= 0.006}\mathrm{)}$ and male sex $\mathrm{(}\mathit{\text{OR = 2.074, }}p\mathrm{<}\text{0.001}\mathrm{)}$ are also high-risk factors for BPD. Compared to females, male infants have a higher incidence of BPD.

3.5. Laboratory indicators and early hemodynamic profiles

BPD infants showed evidence of a stronger systemic inflammatory response at birth. Compared to the control group, the BPD cohort had significantly higher WBC counts (median 9.30 vs. 8.51 × 10⁹/L, p = 0.004) and neutrophil percentages (46.7% vs. 40.85%, p < 0.001). CRP levels were also significantly elevated in the BPD group $\mathrm{(}p\text{= 0.004}\mathrm{)}$.

Hemodynamic data revealed early cardiovascular instability in infants with BPD. Median systolic and diastolic blood pressures were significantly lower at 57 mmHg and 29 mmHg, respectively. Conversely, the median heart rate was significantly higher at 144 beats/min (all p < 0.001). These physiological differences form a distinct clinical profile that appears before BPD is formally diagnosed.

3.6. Machine learning model construction and comparative performance

We evaluated seven machine learning architectures – GBM, LR, ET, RF, SVM, KNN, and DT – to identify the best predictive model. Model stability was assessed through 10-fold cross-validation. The performance metrics for each model are summarized in Table 2, and their respective ROC curves are compared in Figure 2.

Table 2.

Comparison results among different models.

Model	AUC	AUC 95%CI	Accuracy	Sensitivity	Specificity	F1_Score	Brier_Score
GBM	0.88	(0.843–0.917)	0.793	0.827	0.759	0.8	0.1415
LR	0.876	(0.838–0.914)	0.796	0.815	0.778	0.8	0.1428
ET	0.875	(0.837–0.914)	0.793	0.821	0.765	0.799	0.1405
RF	0.872	(0.833–0.911)	0.796	0.821	0.772	0.801	0.1423
SVM	0.87	(0.830–0.910)	0.806	0.821	0.79	0.809	0.1427
KNN	0.867	(0.827–0.906)	0.809	0.858	0.759	0.818	0.145
DT	0.816	(0.772–0.861)	0.787	0.889	0.685	0.807	0.1602

GBM: Gradient Boosting Machine; LR: Logistic Regression; RF: Random Forest; SVM: Support Vector Machine; KNN: K-Nearest Neighbours; DT: Decision Tree and ExtraTrees.

Figure 2.

ROC curve comparison of various machine learning methods in predicting the severity of BPD. Figure shows the ROC curve of seven machine learning model predictions on the test dataset. GBM obtained the highest AUC of 0.880, which is more discriminating; therefore, it is chosen for further applications based on its stability in multi-folds CV.

Different algorithms address different clinical priorities. DT and KNN provided the highest sensitivities (0.889 and 0.858), making them suitable for early screening. Alternatively, SVM achieved a specificity of 0.790, which is ideal for accurately identifying low-risk neonates and avoiding false-positive diagnoses.

3.7. Calibration integrity and goodness-of-fit appraisal for the GBM framework

3.7.1. Visual and statistical calibration assessment

We assessed the GBM model’s calibration using visual plots and the Hosmer-Lemeshow test. The calibration plot demonstrated strong concordance between predicted and observed BPD cases. With a Brier score of 0.1415 and a Hosmer-Lemeshow p-value of 0.060, the model showed excellent goodness-of-fit with minimal estimation error (Figure 3).

Figure 3.

Curve of the GBM model. At a 45-degree angle are the predicted results; The solid blue line shows how well the GBM model performs compared to these predictions. Dotted lines are employed for observing the calibration condition. The Hosmer–Lemeshow test (p = 0.060) and Brier Score (0.1415) show excellent goodness-of-fit and calibration.

The calibration plot showed that the predicted probabilities from the GBM model closely followed the 45-degree diagonal reference line. This alignment suggests that the model provides accurate risk estimates without significant bias across the entire probability spectrum. These results indicate that the algorithm’s predictions effectively reflect the actual incidence of BPD observed in our study population.

3.7.2. Statistical validation and goodness-of-fit performance

Goodness-of-fit evaluations further confirmed the model’s reliability. For the internal validation set, the Hosmer-Lemeshow test yielded a χ² value of 14.947 (p = 0.060) (Table 3). Since the P-value was greater than 0.05, there was no significant difference between the predicted probabilities and actual clinical observations, indicating a good model fit.

Table 3.

Test of goodness of fit.

Hosmer–Lemeshow test	χ2	p-value
GBM (Internal validation set)	14.947	0.060
GBM (External validation set)	115.86	<0.001
LR (Internal validation set)	23.549	0.002
LR (External validation set)	82.487	<0.001

The GBM model’s Brier score was 0.1415 (Table 2). This metric reflects a high level of predictive accuracy with minimal error, confirming that the algorithm is reliable for practical application.

3.8. Clinical utility and interpretable feature importance

We assessed the clinical utility of the GBM model using DCA (Figure 4). The analysis demonstrated a positive net benefit across a wide range of threshold probabilities. These findings indicate that the model outperforms both ‘treat-all’ and ‘treat-none’ strategies, suggesting its value for improving risk stratification in the NICU.

Figure 4.

DCA for the optimal GBM model. The x-axis represents the threshold probability for BPD, and the y-axis indicates the net benefit. The GBM curve remains above both the ‘Intervene All’ and ‘Intervene None’ reference lines across a wide range of threshold probabilities, confirming the clinical utility of the model.

SHAP analysis was performed to visualize the contribution of each clinical feature to the model’s predictions. GA was identified as the strongest predictor of BPD, followed by BW and 1-minute Apgar scores (Figure 5).

Figure 5.

Feature importance ranking based on SHAP values. This figure presents the contribution of each variable to the model’s prediction. A longer bar indicates higher feature importance. BW was identified as the most critical predictor, followed by GA and Apgar score.

PDP analysis was performed to visualize the impact of BW on BPD risk. The plots revealed a non-linear relationship, with BPD risk increasing significantly as birth weight declined. This approach allows for a more detailed assessment of physiological vulnerability compared to traditional linear methods (Figure 6).

Figure 6.

PDP showing how BW affects the risk of BPD. PDP curve presents a non-linear association among birth-weight data and the likelihood of having BPD through marginalisation across all attributes besides these. A steep rise in the risk of BPD with decreasing birth weight.

Figure 7.

Calibration Curves for External Validation of LR Model and GBM Model. At a 45-degree angle are the predicted results; The solid blue line shows how well the GBM model performs compared to these predictions. Dotted lines are employed for observing the calibration condition.

3.9. External cohort validation and comparative model performance

We evaluated the model’s generalizability through external validation using an independent cohort of 1,000 neonates from the First Affiliated Hospital of Nanjing Medical University. This step was necessary to confirm the model’s performance in a different clinical setting before broader implementation.

In the external validation set, the LR, RF, and GBM models all showed strong discriminative performance, with AUCs exceeding 0.97. The LR model performed best, achieving an AUC of 0.977, an accuracy of 92.6%, a sensitivity of 91.0%, and a specificity of 94.2% (Table 4). Despite the high discriminative performance of both models, external calibration was imperfect, as indicated by the calibration plots and quantitative calibration metrics (Figure S7; Table S2).

Table 4.

Performance comparison of each model during external validation.

Model	AUC	Accuracy	Sensitivity	Specificity	F1_Score	Brier_Score
LR	0.977	0.926	0.91	0.942	0.924	0.0632
RF	0.973	0.912	0.894	0.93	0.91	0.0699
GBM	0.971	0.898	0.861	0.934	0.894	0.0755
ET	0.971	0.904	0.878	0.93	0.901	0.0785
SVM	0.97	0.919	0.906	0.932	0.918	0.0742
KNN	0.966	0.907	0.898	0.916	0.906	0.0696
DT	0.943	0.918	0.988	0.847	0.923	0.0932

GBM: Gradient Boosting Machine; LR: Logistic Regression; RF: Random Forest; SVM: Support Vector Machine; KNN: K-Nearest Neighbours; DT: Decision Tree; ET: Extra-Trees.

4. Discussion

Using a comprehensive 15-year longitudinal dataset, we developed and tested a GBM model to predict BPD in premature infants. The framework reached a validation AUC of 0.880, demonstrating a level of predictive accuracy that strongly aligns with current clinical standards [11]. Standard logistic regression relies on rigid mathematical assumptions, making it poorly equipped to handle the highly unpredictable and complex nature of a newborn’s physiology. In contrast, ensemble learning methods easily capture these non-linear clinical variables, making them a much better fit for this specific task. Because this framework successfully captures the intricate variables driving BPD, it serves as a highly practical tool for patient risk stratification. With these clear risk thresholds, NICU teams can confidently deliver targeted treatments to the most vulnerable infants.

Our analysis supports GA and BW as the most critical predictors of BPD (p < 0.001), in line with established research [12–14]. Preterm infants face significant lung immaturity, including alveolarization arrest and microvascular complications [15–18]. This structural fragility makes them vulnerable to postnatal stressors. In such cases, even standard life-support interventions can exacerbate lung injury, promoting the development of BPD.

Low Apgar scores reflect the physiological failure of extrauterine adaptation. These metrics capture acute hypoxia and systemic instability during the immediate transition after birth. Such early insults ignite inflammatory pathways and disrupt lung maturation [19,20]. Our ML framework leverages these markers to refine early risk identification. This precision allows clinical teams to implement targeted preventive care within the first hours of life.

Supporting earlier population-based observations, our model isolates neonatal anemia as a primary and independent driver of BPD [21,22]. This relationship is driven by a compromised oxygen-carrying potential, a condition that necessitates more aggressive ventilation. This intervention, while life-saving, subjects underdeveloped lung tissue to cumulative barotrauma. Crucially, besides mechanical support, allogeneic blood transfusions may act as a ‘secondary hit’, further damaging the lung and compounding the trauma from prolonged mechanical ventilation. The altered lung architecture seen in BPD can be driven by the interruption of normal pulmonary maturation. This disruption often occurs when premature infants are exposed to older stored blood components. Over time, these blood products amass pro-oxidative species and bioactive lipids, which can trigger systemic inflammatory cascades and TRALI-like mechanisms upon transfusion [23–25].

Laboratory data showed that leukocytosis, neutrophilia, and elevated CRP were common in the BPD cohort, reflecting systemic inflammation. This inflammatory state often precedes chronic lung disease in preterm infants. Specifically, neutrophils release elastase, which degrades the lung’s extracellular matrix. This damage compromises alveolar architecture and hinders the septation process, directly impacting lung maturation [26,27].

Within our multivariate framework, mechanical ventilation emerged as a significant independent determinant of BPD development. This association underscores the precarious trade-off inherent to invasive support, where the maintenance of essential gas exchange is frequently achieved at the expense of ventilator-induced lung injury and the cumulative burden of hyperoxic stress. Collectively, these synergistic stressors drive the ‘new BPD’ phenotype, which is a pathology hallmarked by the premature arrest of alveolarization and the profound disorganization of microvascular maturation [13,28,29].

While ACS administration was recorded in 67.8% of the BPD cohort, this frequency likely functions as an index of clinical vulnerability rather than signifying a discrete causal link. Beyond these intrauterine insults, the maintenance of circulatory equilibrium stands as a pivotal driver of neonatal prognosis; specifically, the presence of a PDA constitutes a significant risk variable, corroborating earlier observations [22,30]. From a mechanistic standpoint, this association is mediated by the chronic left-to-right shunting that precipitates pulmonary overcirculation and attenuates lung compliance. This hemodynamic strain frequently necessitates protracted ventilatory assistance, thereby exposing the underdeveloped lung parenchyma to cumulative iatrogenic injury and, ultimately, the morphological remodeling characteristic of BPD.

HDP appear to exert a paradoxically protective influence against the development of BPD. This phenomenon, frequently termed the ‘HDP paradox’, suggests a distinct adaptive pathway within the intrauterine environment. Chronic maternal hypertension likely induces a state of sustained fetal stress that, paradoxically, serves as a catalyst for precocious pulmonary maturation. This intrauterine insult upregulates surfactant synthesis and enhances lung compliance prior to delivery; consequently, these neonates often manifest a physiological resilience that exceeds that of weight-matched infants born to normotensive mothers [31,32]. Such accelerated maturation confers a vital defensive buffer, mitigating the susceptibility to iatrogenic lung injury in the immediate postnatal period.

PDP analysis revealed that BPD risk increases sharply when birth weight falls below 1,430 g (Figure 6). This cutoff, similar to that reported by Yang and Huang, effectively divides neonates into different risk groups [33,34]. In the higher birth weight range, the estimated risk of BPD remains relatively low and stable; however, as birth weight decreases, the risk exhibits a progressively steeper upward trend. Although an inflection point in the risk trajectory was identified around 1,430–1,435 g, further data are required to verify whether this can directly guide clinical decision-making. Meanwhile, this finding suggests that the association between lower birth weight and BPD risk becomes more pronounced, further highlighting the nonlinear relationship between birth weight and BPD risk.

We anticipated that higher birth weights would correlate with lower BPD risks, as previously reported [35]. However, our data revealed an unusual trend where BPD incidence actually rose in neonates over 1,500 g. We suspect this inconsistency is due to random noise in our sub-sample or the inclusion of patients with atypical profiles, such as those with severe sepsis. When assessing the practical value of the GBM model through DCA, we found it consistently maintained an advantage over standard ‘treat-all’ and ‘treat-none’ strategies. Standard binary clinical protocols often lack the flexibility needed for nuanced patient care. Our algorithm addresses this limitation by delivering a steady net benefit across a wide range of thresholds, thereby supporting more tailored, risk-based interventions (Figure 4).

In terms of external validation, both the GBM and LR models exhibited varying degrees of overfitting in calibration performance (Table 5). Interestingly, the discriminative performance of all models was higher in the external cohort than in the internal validation cohort. This finding may partly reflect differences in case mix and predictor distributions between the two institutions. Greater separation of clinical characteristics between infants with and without BPD in the external cohort could increase the apparent discriminative performance of the models. Although the LR model demonstrated better calibration than the GBM model during external validation, the confidence interval from the DeLong test indicated a negligible difference (Table 6), suggesting that the practical improvement in clinical application might be limited. Given that the GBM algorithm possesses a stronger ability to capture non-linear features, it provides more aggressive and sensitive identification of extremely high- and low-risk pediatric patients compared to the overly conservative risk probability estimates of the LR model. This enhanced sensitivity facilitates earlier invasive protective interventions by clinicians. Therefore, we ultimately selected the GBM model as the optimal model (Figure S1).

Table 5.

Details of calibration in external validation.

	LR	GBM
Calibration intercept	1.036	0.824
Calibration slope	1.483	1.783
Calibration-in-the-large	0.937	0.425
Observed-to-expected ratio	1.182	1.115
Brier score with confidence intervals	0.0632 (95% CI: 0.0549–0.0718)	0.0756 (95% CI: 0.0673–0.0847)

Table 6.

De-Long test for LR and GBM.

De-Long test	Result
Z	−2.949
p-value	0.003188
95% CI:	(−0.0104, −0.0021)
GBM’s AUC	0.9706
LR’s AUC	0.9769

The GBM model achieves high accuracy but remains a static assessment tool. BPD is fundamentally a dynamic and evolving pathology. Future research must adopt time-series architectures to track longitudinal patient trajectories. Real-time clinical support requires the integration of high-frequency physiological data and continuous respiratory waveforms. This transition allows for medical interventions based on an infant’s immediate status rather than baseline markers alone.

Using a 15-year dataset provides us with strong statistical evidence, but we must consider the impact of time-related factors. Clinical practices in neonatology have shifted during this window, potentially influencing BPD outcomes in ways that static international guidelines cannot completely filter out. The current model lacks biological depth because it relies entirely on bedside parameters. Importantly, external validation alone does not establish clinical effectiveness. Future prospective implementation studies are needed to assess whether integrating this model into routine NICU practice improves risk recognition, appropriately guides clinical management, and ultimately translates into better neonatal outcomes. While we have preliminarily developed an online calculator to facilitate this, further refinement is still required before broader clinical application.

5. Conclusion

In summary, we developed and validated a machine learning model for early BPD identification in low birth weight infant. The GBM framework performed best by integrating hematological markers and clinical history into a reliable predictive tool. A key finding was the 1,430 g birth weight threshold, below which the risk of BPD increases sharply and non-linearly. By providing interpretable clinical markers, this tool supports a more personalized approach to neonatal care. Implementing these evidence-based strategies could help clinicians optimize respiratory management and improve long-term outcomes for high-risk infants.

Funding Statement

This work was supported by Medical Key Strategic Project of Wuxi Health Commission.

Ethics approval and consent to participate

This study was approved by the research ethics board of the Affiliated Women’s Hospital of Jiangnan University (YLSL2025-094). The study protocol was approved in compliance with the Declaration of Helsinki. Given the retrospective nature of this study, the requirement for informed consent was waived by the ethics board.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from Affiliated Women’s Hospital of Jiangnan University and the First Affiliated Hospital of Nanjing Medical University but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Affiliated Women’s Hospital of Jiangnan University and the First Affiliated Hospital of Nanjing Medical University.