Although progress has been made in the research of BPD, the pathogenesis and etiology of the disease remain elusive. Stratified risk assessment for BPD in preterm infants is crucial. Through such assessments, we can identify risk factors and stratify BPD risk, giving clinicians a clearer management target for high-risk infants. This enables the formation of differential prevention and intervention strategies, thereby improving the treatment and quality of life for these infants.
Currently, logistic regression dominates the field of BPD predictive modeling [8]. However, the pathogenesis of BPD is complex, and traditional statistical analysis methods may overlook important information variables. Machine learning algorithms, with their superior handling of complex data, are increasingly favored for BPD prediction studies. Yet, research constructing stratified prediction models for BPD using these algorithms is scarce [11, 12]. This study aims to fill that gap by utilizing machine learning algorithms at different postnatal time points within the first week to construct stratified prediction models for BPD in extremely premature/very low birth weight infants. The optimal models for these specific time points were chosen to provide clinical guidance for early stratified prediction of BPD.
In this study, the independent risk factors for the severity of BPD at three postnatal time points included gender, GA, BW, use of antenatal steroids, interruption of umbilical blood flow and FIO2. GA and BW are widely agreed upon as key factors for the occurrence of BPD [13]. Previous prediction models [14] have also identified GA, BW, and gender as critical early predictors for BPD, which is consistent with our findings
The regulated use of antenatal corticosteroid (ACS) can reduce intrauterine inflammation, promote surfactant synthesis, stabilize alveolar architecture, minimize lung injury, foster pulmonary development, promote the maturation of fetal lungs and significantly reduce the incidence rate of BPD [15–18]. Studies by Greenberg et al [19], have highlighted the predictive role of ACS on the first postnatal day for BPD, aligning with the NEOCOSUR model [20], where ACS use is an early predictor for severe BPD or mortality postnatally. Our findings corroborate this. Umbilical blood flow abnormalities can cause various in utero anomalies, leading to fetal hypoxia and inflammatory responses, which through multiple pathways and cytokines, hinder alveolar and pulmonary vascular development, culminating in the development and progression of BPD [20, 21]. FiO2 was independently associated with BPD severity, consistent with BPD predictive models by Laughon [14] and Greenberg [19]. The heavier the infant’s condition, the higher the required respiratory support mode, oxygen concentration, and consequent lung damage, thus promoting the development and progression of BPD [22].
As severe preeclampsia was a significant risk factor for BPD severity on the1st and 3rd postnatal days in the present study, consistent with Shim research model [23]. Additionally, prenatal infections are identified as independent predictors of BPD severity on the 7th postnatal day. Prenatal infections not only result in a reduction of GA and BW, but also amplify the likelihood of sepsis in infants [24]. Moreover, prenatal infections activate the fetal immune response, trigger an inflammatory reaction, induce the excessive release of cytokines, provoke inflammatory cascades, worsen lung inflammation, impair lung development [25], and elevate the risk of BPD onset and progression [26], consistent with the predictive model for BPD developed by Zhang [27].
Underdevelopment of the lungs, lung injury, and abnormal repair post-injury represent three pivotal stages in BPD pathogenesis [28], involving multiple mechanisms such as oxidative stress, inflammatory damage, immune dysregulation, and nutritional imbalance [29, 30], and involving various biomarkers. Inflammation affects the generation and maturation of red blood cells (RBC), RBC impacting immune function by inducing vascular dysfunction [31]. C-reactive protein (CRP) is an established marker of inflammation and has been validated for its value in early prediction of BPD and its severity [32]. SIRI derived from relevant inflammatory cells in peripheral blood, offers a more stable measure and compensatory function among neutrophils, lymphocytes, and monocytes, thereby providing a comprehensive reflection of the body’s inflammatory and immune status. Previous research has identified SIRI as a potential biomarker for predicting BPD [34]. PMI reflects the body’s inflammatory and immune statuses [33], consistent with Sati research [35].
The PNI amalgamates albumin and peripheral blood lymphocyte levels, providing a cost-effective and utilitarian biomarker reflecting the body’s nutritional and immune status. A higher PNI suggests a favorable prognosis, whereas a lower index indicates a poorer outcome [36]. PNI was inversely proportional to the severity of BPD.The findings may epitomize the collective impact of nutritional and immune imbalances on the onset and progression of BPD [37].
Blood gas analysis serves as a critical component of neonatal respiratory management by directly measuring pulmonary ventilation and oxygenation functions. OI is indicative of an infant’s alveolar oxygenation efficiency. OI independently affects the severity of BPD, with a negative correlation to the disease’s severity, which corroborates previous studies [38]. A-aDO2 combines the inhaled oxygen concentration with carbon dioxide and oxygen partial pressures, serving as an index of pulmonary ventilation and alveolar oxygenation ability [39]. Elevated A-aDO2 levels, suggesting hypercapnia and hypoxic states, reflect impaired pulmonary function and extensive lung injury, suggesting that infants consequently require higher oxygen concentration and respiratory support to maintain normoxic conditions. Such injury may lead to lung damage via oxidative stress pathways, ultimately causing the development and progression of BPD [5].
With the advent of the big data era and the progression of artificial intelligence technologies, predictive modeling has advanced significantly, utilizing machine learning and other sophisticated data analytics techniques to establish more precise models. Studies employing machine learning algorithms for the prediction and management of BPD have shown substantial advancements [40]. Nevertheless, there persists contention regarding the performance of machine learning models in predicting BPD, where they haven’t conclusively outperformed traditional LR models [11, 41, 42]. Our research, by constructing LR, RF, GBDT, and XGB models at various timepoints within the first postnatal week, identified that the LR and XGBoost models in particular perform well on days 1, 3, and 7 for early stratified prediction of infants at high risk for BPD. In contrast with former machine learning models, our models have individually assessed the contribution of risk factors to predictive efficacy at each time point. Hence, the inclusion of different variables at distinct time points tailored to BPD’s dynamic nature has achieved effective dynamic stratified prediction within the first week postnatal with considerable predictive performance. This translation into professional academic language suitable for an SCI journal publication should undergo peer review for terminological accuracy and adherence to publication standards.
The dataset used to construct the models in this study was derived from a single-center, retrospective investigation with an extensive time span, which does not guarantee complete homogeneity of all indicators and presents a multitude of confounding factors that might result in bias. Future multicenter, prospective studies to build models may enhance the performance and predictive power of these models.
In conclusion, in this study developed various types of machine learning predictive models at different time points within the first postnatal week, which serve as useful tools for early stratified prediction of BPD in clinical practice. These models aid clinicians in the early personalized intervention of high-risk groups, potentially reducing the incidence of BPD.