Artificial Intelligence in Bronchopulmonary Dysplasia: A Review of the Literature

Abstract

Bronchopulmonary dysplasia (BPD) is a neonatal lung condition predominantly affecting preterm infants. Researchers have turned to computational tools, such as artificial intelligence (AI) and machine learning (ML), to better understand, diagnose, and manage BPD in patients. This study aims to provide a comprehensive summary of current AI applications in BPD risk stratification, treatment, and management and seeks to guide future research towards developing practical and effective computational tools in neonatal care. This review highlights breakthroughs in predictive modeling using clinical-, genetic-, biomarker-, and imaging-based markers. AI has helped advance BPD management strategies by optimizing treatment pathways and prognostic predictions through computational modeling. While these developments become increasingly clinically applicable, numerous challenges remain in data standardization, external validation, and the equitable integration of AI solutions into clinical practice. Addressing ethical considerations, such as data privacy and demographic representation, as well as other practical considerations will be essential to ensure the proper implementation of AI clinical tools. Future research should focus on prospective, multicenter studies, leveraging multimodal data integration to enhance early diagnosis, personalized interventions, and long-term outcomes for neonates at risk of BPD.

1. Background

1.1. Bronchopulmonary Dysplasia

Bronchopulmonary dysplasia (BPD) was first described by Northway et al., 1967 as a lung disease in preterm infants who initially had hyaline membrane disease, now commonly known as respiratory distress syndrome (RDS) [1]. Over the subsequent decades, the clinical manifestation and understanding of BPD have undergone substantial transformation, paralleling advances in neonatal intensive care and respiratory therapeutic interventions.

BPD constitutes a complex clinical syndrome of postnatal lung injury characterized by disrupted alveolarization and microvascular development [2]. Pathological examinations reveal distinct anatomical alterations, including regions of cystic emphysema, fibrotic changes, diminished alveolar formation, and structural airway modifications [2,3,4]. These morphological changes result in profound physiological consequences, such as impaired gas exchange, compromised pulmonary mechanics, and potential long-term sequelae, such as persistent hypoxia, reduced exercise tolerance, and increased susceptibility to chronic respiratory conditions [2].

Epidemiological investigations consistently demonstrate prematurity and low birth weight as the most significant predictive risk factors. According to Younge et al., 2021, approximately 80% of infants born at 22–24 weeks of gestation develop BPD, contrasting with a 20% incidence among infants born at 28 weeks [5]. Additional risk factors are outlined in

Table 1. Risk factors for the development of bronchopulmonary dysplasia in neonates.

Category	Specific Factors
Perinatal Factors	Extreme prematurity (<28 weeks gestational age) [6] Very low birth weight (<1500 g) [7] Intrauterine growth restriction [8] Male sex [9] Multiple gestation [6] Chorioamnionitis [10] Preeclampsia [11]
Genetic Factors	Family history of asthma or atopy [12] Genetic polymorphisms (TGF-β, IL-4, IL-10) [13] Ancestry-specific genetic variants [6]
Respiratory Factors	Respiratory distress syndrome [6] Mechanical ventilation > 7 days [7] High oxygen requirements (FiO2 > 0.4) [14] Ventilator-induced lung injury [6] Patent ductus arteriosus [19]
Inflammatory Factors	Systemic inflammation [6] Early-onset sepsis [6] Late-onset sepsis [15] Elevated inflammatory markers (IL-6, IL-8, TNF-α) [22]
Nutritional Factors	Vitamin A deficiency [16] Fluid overload in first week of life [17] Inadequate protein intake [6] Poor postnatal growth [18]

[2,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22]. Emerging research suggests a significant genetic component, with ongoing investigations exploring potential genetic markers and molecular mechanisms underlying BPD pathogenesis [20,21,22].

The pathophysiological mechanism of BPD is multifactorial. Postnatal respiratory function critically depends on the interface between pulmonary alveolar and endothelial cells [2]. A deficiency of pulmonary surfactant after birth is a primary cause of RDS in preterm infants [23]. However, therapeutic modalities, such as mechanical ventilation, continuous positive airway pressure, and supplemental oxygen, while essential for initial stabilization, simultaneously introduce potential mechanisms of pulmonary inflammation and pulmonary developmental arrest [24].

The diagnostic criteria of BPD have changed over the years. The previous definition was based on the persistent need for supplemental oxygen at a postnatal age of 28 days or a postmenstrual age (PMA) of 36 weeks; newer definitions are more nuanced [1]. In 2018, the National Institute of Child Health and Human Development (NICHD) revised the definition of BPD by eliminating the requirement for 28 days of oxygen therapy before 36 weeks PMA, adding a requirement for the radiographic confirmation of parenchymal lung disease and using a severity grading based on supplemental oxygen to incorporate newer modes of non-invasive ventilation [14]. In 2019, Jensen et al. proposed modifications to the NICHD definition. Their scale defined the stages of BPD, at 36 weeks PMA, as follows: (1) no BPD (no support required), (2) grade 1 (nasal cannula ≤ 2 L/min), (3) grade 2 (nasal cannula > 2 L/min or non-invasive positive airway pressure), and (4) grade 3 (invasive mechanical ventilation) [7].

Contemporary clinical interventions focus on minimizing pulmonary trauma and optimizing respiratory outcomes. The Caffeine for Apnea of Prematurity (CAP) trial by Schmidt et al. demonstrated significant reductions in ventilator dependency [25]. At the same time, targeted oxygen saturation management (91–95%) has shown promise in mitigating oxidative lung damage. Emerging therapeutic strategies, including less invasive surfactant administration and standardized glucocorticoid protocols, represent additional advances in BPD management [26,27,28].

Despite advances in early respiratory care, some preterm infants develop chronic respiratory disease and related comorbidities that persist after discharge [2,29]. These patients may require additional interventions, such as continued supplemental oxygen therapy, respiratory medications, or in severe cases, tracheostomy placement along with home mechanical ventilation. Approximately 50% of patients experience rehospitalization within their initial two years, with heightened vulnerability to respiratory viral infections [30,31]. Long-term sequelae include increased risks of reactive airway disease, potential obstructive or restrictive lung pathologies, and associated developmental complications as a result of chronic hypoxia [32,33,34,35]. The trajectory of BPD research underscores the complexity of premature infant pulmonary development. Continued interdisciplinary investigation remains critical to elucidating comprehensive prevention and management strategies for this challenging clinical syndrome. New approaches are needed to better optimize early BPD diagnosis and the management of patient care, and advancements in artificial intelligence offer promising tools to address these challenges.

1.2. Artificial Intelligence and Applications in Healthcare

Artificial intelligence (AI) is the simulation of human intelligence by machines to perform tasks typically requiring human cognitive functions, such as problem solving, learning, and decision making [36]. AI systems utilize algorithms and computational models to interpret data and execute actions autonomously or semi-autonomously. AI technology holds promise across various sectors, including healthcare, due to its ability to manage large datasets, provide real-time analysis, and improve efficiency and accuracy in decision making [36]. The field of AI consists of a vast number of subfields, each with specific approaches, methodologies, and use cases, as outlined in

Table 2. Subfields of artificial intelligence and healthcare applications.

AI Type	Core Methodology	Examples of Healthcare Applications
Machine Learning [37]	Pattern recognition from provided features about the data	Disease prediction Risk stratification Treatment response forecasting
Deep Learning [37]	Leveraging complex neural network-based architectures to identify features and make predictions	Medical image analysis Radiology diagnostics Genomic interpretation
Expert Systems [38]	Using rule-based reasoning to provide solutions in specific domains	Clinical decision support Diagnostic assistance Treatment recommendation
Natural Language Processing [39]	Linguistic processing and feature extraction of human language to analyze, classify, or generate text	Electronic health record analysis Clinical documentation Patient communication

[37,38,39].

AI involvement in healthcare dates back to the 1970s, with early attempts focusing on rule-based systems. Programs such as MYCIN, developed at Stanford University, aimed to assist physicians in diagnosing infectious diseases by offering antibiotic recommendations [40]. The field gained momentum in the 2000s with the advent of more advanced computational tools, larger datasets, such as electronic health records (EHRs), and breakthroughs in machine learning algorithms [41]. Advances in deep learning (DL) further propelled AI’s role in healthcare, particularly in medical imaging. For instance, DL algorithms have been shown to perform at or above the level of human experts in identifying conditions such as pneumonia in radiographic images [42]. Today, AI is used in various applications, such as medical imaging, predictive analytics, clinical decision support systems, and personalized medicine [41,42,43]. Figure 1 highlights the conventional research workflow to developing, optimizing, and deploying a predictive ML model.

The rapid expansion of the AI field has also made its way into neonatology and has impacted patient care. Numerous neonatology researchers in recent years have employed ML technology to assist in the risk stratification, early diagnosis, and treatment management of BPD patients. These studies have helped researchers construct a deeper understanding of the disease mechanism of BPD and have assisted in the proper management of patient care. This review aims to synthesize existing models to evaluate the role of AI in addressing the challenges of BPD patient care and management. Specifically, this study summarizes the existing applications of AI in stratifying BPD risk, predicting BPD diagnosis, and managing BPD treatment and identifies areas for further research in order to develop more practical and effective computational tools for neonatal care.

1.3. Methods

To systematically assess the role of AI in predicting BPD, we conducted a review of the literature. A search was performed across the following databases—PubMed, CINAHL, Research Rabbit, Web of Science, and Scopus—to identify relevant studies published between 1 January 2021 and 15 March 2024. To ensure the inclusion of the most recent advancements in AI-driven BPD prediction, we focused on studies published from 2021 onward. This time frame captures the rapid evolution of deep learning techniques and external validation strategies that were not well-established in earlier studies. The following Boolean search terms were used: (“Bronchopulmonary Dysplasia” OR “BPD”) AND (“Artificial Intelligence” OR “Machine Learning” OR “Neural Networks”) AND (“Premature” OR “Neonate” OR “Newborn”).

Studies were included if they (1) developed or validated an AI/ML model for predicting BPD in neonates, (2) utilized clinical, laboratory, or imaging-based features, and (3) reported performance metrics, such as AUC, sensitivity, specificity, or predictive accuracy. Studies were excluded if they were animal studies, in vitro experiments, case reports, or conference abstracts. Data extraction focused on study characteristics, AI model types, dataset sources, validation strategies, and performance metrics. Heterogeneity in study design, BPD definitions (2001 NICHD, 2018 NICHD, Jensen criteria), and dataset size were also analyzed.

2. Artificial Intelligence in Analyzing BPD Risk

New approaches using ML have emerged as powerful tools for identifying predisposing risk factors and developing predictive models for BPD. While traditional statistical methods have long been employed in clinical research, recent studies demonstrate the potential of ML techniques to uncover nuanced predictive markers and stratify risk more precisely and at an earlier age. For additional reference, the applications of various AI algorithms in neonatology are outlined in

Table 3. Summarization of artificial intelligence algorithms, clinical use cases, strengths, and weaknesses.

AI Model	Use Cases	Strengths	Weaknesses
Logistic Regression [44]	Binary outcomes (e.g., mortality, BPD diagnosis)	Simple to interpret; works well with small datasets.	Limited to linear relationships; struggles with complex interactions.
Random Forest (RF) [44]	Classifying neonates by risk level or predicting categorical outcomes (e.g., sepsis risk)	Handles non-linear data well; robust to overfitting; interpretable via feature importance.	Computationally intensive with large datasets; less transparent than simpler models.
XGBoost [44]	Predicting rare outcomes (e.g., long-term complications) or improving accuracy on structured data	High accuracy; robust with imbalanced data; interpretable using SHAP values.	Requires parameter tuning; can be computationally expensive.
CatBoost [44]	When dataset has categorical features (e.g., feeding methods, medications)	Optimized for categorical variables; fast and handles missing data.	Limited documentation compared to other models; requires careful preprocessing.
Support Vector Machine (SVM) [44]	Identifying patterns in small, high-dimensional datasets (e.g., gene expression for BPD)	Effective in high-dimensional spaces; works well with limited data.	Difficult to interpret; computationally expensive with large datasets.
Neural Networks (NNs) [44]	Complex problems with large datasets (e.g., image analysis, time-series prediction)	Capture complex non-linear relationships; adaptable to various data types.	Require large datasets; lack interpretability; prone to overfitting without proper tuning.
k-Nearest Neighbors (kNNs) [44]	Simple classification tasks (e.g., identifying patient subgroups)	Easy to implement; no training phase; work well with small datasets.	Computationally expensive during prediction; sensitive to irrelevant features.
Clustering [44]	Identifying subgroups in neonates (e.g., phenotyping BPD or sepsis)	Unsupervised learning; helps explore hidden patterns; simple to implement.	Requires pre-specification of cluster numbers; sensitive to outliers.
Linear Discriminant Analysis (LDA) [44]	Classifying outcomes with clear group separations (e.g., birthweight categories)	Simple and interpretable; works well with small datasets.	Assumes linear separability; limited use with complex data.
Recurrent Neural Networks (RNNs) [44]	Time series predictions (e.g., vital signs monitoring for sepsis or NEC)	Capture temporal patterns; useful for sequential data.	Require large datasets; risk of vanishing gradients; computationally expensive.
Convolutional Neural Networks (CNNs) [44]	Image-based tasks (e.g., detecting pneumothorax or brain abnormalities in imaging)	Excellent for image analysis; captures spatial relationships well.	Require large datasets; limited interpretability; resource-intensive training.
Ensemble Models [44]	Combining predictions from multiple models for tasks like risk scoring	Improve accuracy and robustness; reduce overfitting.	Can be complex to implement; less interpretable than single models.

[44].

Studies included datasets from seven countries, with Korea contributing the largest proportion (49.3%). While this diversity enhances generalizability, the predominance of data from high-resource settings may limit applicability to low-resource neonatal units. Logistic regression was frequently used for its interpretability, while random forest and XGBoost models were commonly used due to their robustness in handling nonlinear relationships and missing data. SVM was favored for small datasets, while deep learning models, like CNNs, were less prevalent due to their high data requirements. Deep learning models, such as CNNs, were used in imaging-based approaches due to their high accuracy in pattern recognition.

Lei et al., 2021 initiated this approach by constructing random forest (RF) models to evaluate feature selection’s impact on predictive accuracy [45]. Their study analyzed clinical data from 648 preterm infants and compared two ML models with differing feature sets to optimize performance, while limiting data requirements. They began by using the Boruta algorithm to select for clinical features. This algorithm duplicates and randomizes feature data, generates an RF decision tree-based classifier to assess feature importance, and marks important clinical features. This process is iterated multiple times to ensure stable results and returns the 12 most important features. The researchers fed these features into an RF model for evaluation and further isolated the six most important features to create a separate RF model for comparative analysis. Despite using fewer clinical variables (6 vs. 12), the 6-feature model maintained comparable accuracy (0.8958 vs. 0.9167) and showed a marginal improvement in area under the curve (AUC) from 0.922 to 0.929. However, the study was limited by its single-center design and geographical specificity to the Liuzhou Hospital in Western China, potentially restricting broader generalizability. The single-center design potentially limits generalizability as the predictive model and features identified may be due to bias within the patient population.

Building upon clinical variable approaches, Moreira et al., 2023 explored transcriptomic signatures using multiple machine learning algorithms [46]. They began by collecting microarray data for 97 very low birth weight (VLBW) infants on day 5 of life and filtered relevant genes through expression levels and statistical methods to isolate 4523 genes. By employing k-nearest neighbors (kNNs), support vector machines (SVMs), neural networks (NN)s, and extreme gradient boosting (XGBoost), they identified 14 potential gene signatures, ultimately focusing on five most predictive genes: PNPO, MSANTD2, CD4, SNX1, and P2RX7. These genes suggest a dysregulation of T cell development and function in BPD. Their XGBoost model demonstrated exceptional predictive capacity, achieving an AUC of 0.961. Remarkably, the model’s performance in diagnosing severe BPD reached an impressive AUC of 0.998. This study, however, was limited by its use of single-center data points and microarray data that are not always readily available in the clinic at birth. Additionally, the study’s performance could not be compared to existing BPD outcome estimators due to the lack of clinically relevant data points, such as respiratory support.

Montagna et al., 2024 provided a critical analysis between conventional statistical methods and ML approaches [47]. Traditional statistical analysis highlighted prenatal variables, like absent or reversed end-diastolic flow velocity (AREDV), ductus venosus flow alterations, and chorioamnionitis, as key predictors. In contrast, their ML models, particularly the XGBoost classifier, identified a slightly different set of most relevant variables, as follows: extremely low birth weight, gestational age, AREDF status, magnesium sulfate prophylaxis, and mechanical ventilation status. The convergence of both methodologies on certain predictors—extremely low birth weight, gestational age, mechanical ventilation, and AREDF status—strengthens the reliability of these risk factors. The additional risk factors identified through ML methods were considered less reliable due to the poor precision of the algorithms.

Using a different type of ML, Moreira et al., 2023 expanded their investigation on BPD by using unsupervised clustering techniques [48]. They collected microarray data and identified differentially expressed genes based on “BPD subgroups”. They separated BPD cases into four distinct endotypes (A, B, C, and D) based on 7319 genes. By creating clusters or groupings of unlabeled datapoints, they identified a set of 1207 genes that overlapped across all clusters. Afterwards, supervised learning identified 20 gene markers that could predict the different endotypes, BPD clusters based on genetic data. Despite these promising findings, the clinical translation of such complex genetic stratification remains challenging and requires further investigation. Their model also focuses on retrospective data analysis, and no current studies evaluate prospective BPD risk through endotyping.

These studies collectively demonstrate the potential of ML in BPD research. Advanced computational techniques allow researchers to identify more precise and personalized risk prediction models. However, significant challenges remain, including the need for larger, more diverse datasets, comprehensive validation across different clinical settings, and the development of clinically implementable predictive tools.

3. Artificial Intelligence in BPD Diagnosis

Emerging AI technology has also aided in the early diagnosis of BPD. Using patient data available at birth or during pre-birth screenings, researchers can now leverage computational approaches to develop models that explore clinical, genetic, and imaging-based markers to accurately predict BPD diagnosis.

In one of the earlier BPD predictive models, Jassem-Bobowicz et al., 2021 developed a multivariate risk assessment approach using 278 preterm infants [49]. Their model focused on clinical variables, including gestational age, red blood cell transfusions, surfactant administration, and patent ductus arteriosus. Although the researchers did not leverage AI technology, they achieved a high AUC of 0.932, demonstrating the potential of carefully selected predictive variables in BPD risk assessment. Their study created a risk-scoring system for BPD and maintained a high AUC, while being simple and feasible in clinical practice. Their predictive model is, however, limited by their low sample size of patient data in the different subcategories split by clinical variables.

Expanding on clinical data integration, Verder et al., 2021 introduced an innovative approach by combining clinical information with gastric aspirate sample analysis through Fourier transform infrared (FTIR) spectral analysis [50]. FTIR spectral analysis is an analytic laboratory technique that characterizes substances based on their interactions with infrared radiation. The researchers fed FTIR data alongside clinical data available at birth into an SVM model to classify each patient based on BPD diagnosis. The SVM model integrates clinical and FTIR data to position each patient within a multidimensional feature space and generates a hyperplane to differentiate between patients with and without BPD. This hyperplane facilitates the classification of patients and enables the prediction of BPD diagnoses. The resulting model achieved remarkable diagnostic performance, with 88% sensitivity and 91% specificity, offering a near-immediate predictive capability immediately after birth.

Genetic factor integration represented another significant advancement in BPD prediction. Dai et al., 2021 demonstrated the substantial predictive power of incorporating genetic risk factors into ML models [51]. Analyzing a dataset of 245 premature infants, they observed substantial improvements in model performance. They first created a risk-scoring system to identify genetic markers that are more strongly associated with both mild and severe BPD diagnoses. They identified 30 risk genes for mild BPD and 21 genes for severe BPD, including OBSL1, GNAS, TCIRG1, and C5 among others. Most of these genes are primarily involved in susceptibility to infection, immune regulation, and cellular biological function. They also collected clinical information, such as birth history, morbidities in the first week of life, early treatment after birth, morbidities during hospitalization, and treatment. The AUC increased from 0.826 to 0.907 for severe BPD prediction and from 0.814 to 0.915 for mild-to-moderate BPD classification, when genetic factors were integrated, highlighting the potential of comprehensive data approaches. While promising, their study has yet to be validated on an external dataset of patients, so it still has limited predictive value.

Biomarker-based prediction emerged as another promising avenue. Gao et al., 2023 investigated umbilical cord blood interleukin-6 (UCB IL-6) as a predictive marker across 414 preterm infants categorized by BPD severity [52]. Comparing four ML algorithms—XGBoost, CatBoost, LightGBM, and RF—they consistently identified UCB IL-6 as the most prominent feature. The XGBoost, CatBoost, and LightGBM models are all decision tree-based algorithms that use boosting. These models are trained iteratively in which successive decision trees are constructed to progressively enhance the model’s performance, with the final classification determined by an ensemble of trained decision trees. In contrast, the RF algorithm employs a bagging technique, combining the predictions of multiple decision trees to generate the classification outcome. All of the models achieved impressive AUCs ranging from 0.841 to 0.870, with CatBoost performing optimally. The study established a strong association between elevated UCB IL-6 levels and advanced BPD grades, highlighting the potential of specific biomarkers in risk stratification. Similar to the previous studies, their study was limited by a small sample size, which disrupted the model’s ability to predict the grade of BPD severity.

Deep learning models, particularly U-Net and ResNet, have demonstrated promising results in neonatal chest radiograph analysis, offering improved accuracy and reduced observer variability compared to traditional manual interpretation. U-Net excels in segmenting lung regions by extracting and highlighting key image features, while ResNet, a pre-trained convolutional neural network (CNN), is commonly employed for classification tasks due to its ability to efficiently process large images and optimize training by skipping redundant layers. Chou et al. (2024) introduced a novel AI-driven BPD prediction model that leverages U-Net for lung segmentation and ResNet for classification, using radiographs taken within 24 h of birth from preterm and term infants at the National Cheng Kung University Hospital [53]. Their model achieved an average F1 score of 0.80, consistently outperforming human prediction in BPD diagnosis. Despite these promising results, the study was constrained by its retrospective, single-center design, limited dataset, and a lack of external validation. Furthermore, neonatal lung image analysis presents inherent challenges, including variability in radiograph quality and anatomical differences in preterm infants. These findings emphasize the need for multicenter collaborations and adaptive machine learning approaches to enhance model generalizability and clinical applicability. While AI models often outperform clinical scoring systems and logistic regression in predictive accuracy, the black-box nature of deep learning remains a significant hurdle. Balancing AI’s predictive power with clinician trust through explainability techniques will be crucial for successful real-world implementation.

These studies collectively illustrate the capability of AI in BPD diagnostics. The models are summarized in

Table 4. Predictive ML models for BPD diagnosis.

Study	Geographic Area	Sample Size	Algorithms Employed	Notable Features	Validation	Performance Metrics
Verder et al. [50]	Denmark	61	SVM	FTIR spectral analysis	Training, test split	88% specificity 91% sensitivity
Dai et al. [51]	China	245	LASSO	High risk genes OBSL1, GNAS, TCIRG1, C5, and others	10-fold CV	0.907 AUC, severe BPD 0.915 AUC, mild BPD
Gao et al. [52]	China	414	XGBoost, CatBoost, Light GBM, RF	UCB-IL6	10-fold CV	0.870 AUC
Chou et al. [53]	Taiwan	380	U-Net, ResNet	Chest radiograph images	5-fold CV	0.80 F1 score

[50,51,52,53]. Through the integration of diverse data sources—clinical variables, genetic markers, biomarkers, and imaging data—researchers are developing increasingly precise predictive models. Areas of future research include large-scale validation, the standardization of methodologies, and the translation of these computational approaches into clinical practice.

4. Artificial Intelligence in BPD Management and Treatment

Modern computational ML approaches are also being leveraged to address critical challenges in BPD management, offering unprecedented insights into patient outcomes, disease severity prediction, and treatment optimization strategies.

Leigh et al., 2022 conducted a comprehensive computational analysis of 689 preterm infants, employing ensemble modeling techniques to elucidate BPD-free survival predictors [54]. Their statistical modeling achieved a robust AUC of 0.899, systematically demonstrating that reduced gestational age and prolonged intubation durations significantly correlated with diminished survival probabilities. The research identified continuous positive airway pressure post-extubation as a superior intervention compared to non-invasive mechanical ventilation, offering evidence-based guidance through AI tools for BPD respiratory support management in clinical practice.

Extending computational methodologies to radiographic analysis, Xing et al., 2022 implemented advanced DL architectures for early BPD severity [55]. Utilizing transfer learning and sophisticated NN segmentation, they developed a predictive model capable of determining BPD severity at 28 days—substantially earlier than traditional 56-day clinical assessments. Through a comprehensive evaluation of six NN architectures, the VGG-16 network demonstrated an exceptional diagnostic accuracy of 95.58%, substantiating the potential of advanced computational imaging techniques in neonatal respiratory diagnostics. Similar to the ResNet model, VGG-16 is a pre-trained CNN that is known for its simplicity and depth and its ability to process and classify images. This study highlights the application of DL approaches in the predictive abilities of BPD severity.

Wu et al., 2023 developed a comprehensive prognostic risk stratification model utilizing the Taiwan Neonatal Network’s (TNN’s) extensive dataset of 3200 infants [56]. They collected 24 early-life clinical characteristics for features for the model and graded the severity of the initial neonatal resuscitation. The researchers then systematically evaluated seven ML algorithms and constructed a predictive framework with logistic regression, exhibiting the most statistically robust performance. The model’s AUC values ranged from 0.763 to 0.881 across various outcome categories, with important predictive features including birth weight, gestational age, initial resuscitation intubation characteristics, early sepsis manifestations, and surfactant administration. This study is, however, limited by the geographical specificity of the TNN database and potential collinearity of clinical variables that might serve as a confounding variable when identifying the critical predictive features of late respiratory support.

Addressing practical clinical challenges, Tao et al., 2024 developed a targeted ML algorithm for predicting extubation failure in BPD-diagnosed neonates [57]. They first separated their dataset based on the outcome of extubation (success or failure) and leveraged conventional statistical methods to compare clinical data points. Analyzing 284 mechanically ventilated infants, the researchers employed six computational approaches, with XGBoost demonstrating superior predictive capabilities, achieving an AUC of 0.873. The study identified partial oxygen pressure, hemoglobin levels, and mechanical ventilation status as the most significant parameters. This study is limited due to conflicting definitions of extubation failure and also must be validated through a multicenter prospective study to assess generalizability.

These computational investigations collectively demonstrate the power of ML methodologies in BPD management. Researchers are integrating complex multidimensional clinical, physiological, and imaging datasets and are developing increasingly sophisticated predictive models that promise to revolutionize neonatal respiratory care strategies. However, there are still numerous obstacles to fully implementing AI in clinical treatment.

Multimodal Data Integration in BPD Prediction

BPD is a multifactorial disease influenced by clinical, imaging, genetic, and physiological factors, making multimodal data integration a critical approach for improving predictive accuracy. Traditional ML models often rely on single-source data, such as EHRs or imaging, limiting their predictive capability. Recent advancements in AI and deep learning have enabled models to integrate multiple data types, improving their robustness and applicability in neonatal care. Clinical AI models incorporating gestational age, birth weight, ventilatory support settings, and laboratory markers have demonstrated superior predictive power compared to conventional logistic regression models [7,14]. For instance, deep learning-based radiomics has enabled the extraction of quantitative lung imaging biomarkers from neonatal chest radiographs, facilitating early disease detection [4,50]. A CNN model combining chest X-ray features with clinical data improved BPD risk stratification, outperforming models trained on a single modality [53,55]. Beyond imaging, genetic and transcriptomic data are emerging as important components of BPD prediction. Genome-wide association studies (GWAS) have identified genetic loci related to lung development and inflammatory responses, such as VEGF, IL-6, and ANGPT2, which are implicated in BPD pathogenesis [12,13]. Multiomics approaches integrating RNA sequencing and proteomics data with clinical risk factors have demonstrated potential in identifying high-risk neonates [46,51]. Additionally, continuous physiological monitoring, including oxygen saturation trends, heart rate variability, and ventilatory patterns, has been proposed as a valuable predictive tool, allowing AI-driven models to assess BPD risk dynamically [11,56].

Despite its promise, multimodal AI integration presents several challenges, including data heterogeneity, computational costs, and the need for standardized validation frameworks. Differences in EHR formats, imaging acquisition protocols, and genomic sequencing methods can hinder model generalizability [23,49]. To address these issues, federated learning techniques have been introduced, allowing AI models to be trained across multiple centers, while preserving patient privacy and compliance with HIPAA and GDPR regulations [6,58]. Furthermore, explainability in AI models remains a significant barrier to clinical adoption. Advanced interpretability techniques, such as SHAP (Shapley additive explanations) and attention-based neural networks, are being explored to improve transparency and clinician trust in AI-generated predictions [47,59]. Future research should focus on the prospective validation of multimodal AI models through large-scale, multicenter clinical trials to ensure regulatory approval and real-world applicability. By integrating clinical, imaging, genetic, and physiological data, AI-driven models can provide more accurate and individualized risk assessments, ultimately advancing early BPD prediction and neonatal care.

5. Discussion

5.1. AI vs. Traditional Statistical Methods and Research Recommendations

AI-based models have demonstrated higher predictive performance compared to traditional statistical approaches for BPD risk assessment. In multiple studies, machine learning models, such as random forest, XGBoost, and SVM achieved higher AUC values (ranging from 0.80 to 0.97) than conventional logistic regression models, which often plateau around 0.70 to 0.85. Additionally, deep learning architectures, like CNNs, have shown success in processing imaging data to predict BPD severity, an area where traditional regression-based methods fall short due to their inability to capture spatial features.

Despite these advantages, logistic regression remains widely used in clinical settings due to its interpretability and transparency. Unlike black-box AI models, logistic regression provides clear coefficients and direct associations between input variables and outcomes, making it easier for clinicians to understand and validate predictions. Moreover, logistic regression models require fewer computational resources and are less susceptible to overfitting, especially when dealing with smaller datasets or limited features.

To facilitate real-world AI adoption, several key challenges must be addressed:

• External validation and generalizability: Many AI models for BPD prediction are developed using single-center datasets, limiting their applicability across diverse patient populations. Future research should prioritize multicenter collaborations, where models are trained and tested on heterogeneous datasets from different NICUs to enhance robustness.
• Federated learning for data privacy: Due to strict data-sharing regulations (HIPAA, GDPR), hospitals are often unable to share patient data for AI model training. Federated learning (FL) provides a potential solution by enabling decentralized model training across multiple institutions without transferring sensitive patient data. This approach can improve model accuracy, while maintaining data security and compliance with privacy laws.
• Standardization of AI model reporting: A major barrier to AI integration in neonatal care is the lack of uniform reporting standards for model performance metrics. While AUC, sensitivity, and specificity are commonly reported, calibration plots, confidence intervals, and decision thresholds are often omitted, making it difficult to compare models. Future research should follow established guidelines, such as TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) to ensure reproducibility and facilitate regulatory approval.
• Improving model explainability: One of the major limitations of deep learning models in BPD prediction is their lack of transparency. Implementing post hoc interpretability methods, such as SHAP (Shapley additive explanations) and LIME (local interpretable model-agnostic explanations) can help clinicians understand which features contribute most to AI predictions. For example, if an AI model predicts a high risk of BPD, SHAP can highlight whether gestational age, oxygen therapy duration, or mechanical ventilation settings were the primary contributing factors.
• AI Integration into clinical workflows: Even the most accurate AI models hold little clinical value if they are not integrated into real-world NICU workflows. Future research should focus on developing user-friendly AI-based decision support tools that seamlessly integrate with EHRs, providing real-time risk predictions at the bedside.

By addressing these challenges, AI-driven BPD prediction models can transition from research settings to clinical practice, ultimately improving early detection and long-term outcomes for preterm infants.

5.2. Ethical and Practical Considerations

AI holds the potential to transform healthcare, particularly in the management of complex conditions, like BPD. However, the adoption of AI in NICUs raises several ethical and practical considerations that must be addressed to ensure safe, effective, and equitable care. Although AI models have demonstrated remarkable predictive power, their clinical translation remains hindered by concerns regarding data privacy, fairness, and model bias. Many AI algorithms are trained on datasets that may not be representative of diverse patient populations, leading to disparities in model performance. Training datasets must be diverse and representative of the populations affected by BPD to avoid bias. This includes data from infants of different racial and ethnic backgrounds, as well as those from varying socioeconomic and geographic contexts. Ensuring demographic diversity in the data can improve the generalizability of AI models and reduce the risk of biased or unequal treatment.

Data quality, privacy, and interoperability are crucial in ensuring the quality of AI models, which are limited by the datasets they are trained on. Data quality can vary significantly between institutions, resulting in incomplete data or inconsistent datasets. Missing data, signal noise, or variations in clinical practice can lead to poor model performance or inaccurate predictions. Healthcare institutions often use different electronic health record (EHR) systems, medical devices, and data formats, which creates challenges in integrating AI solutions. Data must be interoperable, so data can flow between departments and devices. Patient data are sensitive, and patient privacy must be conserved, as regulated by the Health Insurance Portability and Accountability Act (HIPAA) in the U.S. and similar regulations worldwide [59]. AI algorithms must meet stringent privacy standards to prevent unauthorized use or misuse of health information. This includes anonymizing datasets to protect patient identities, ensuring secure data storage, and regulating access to AI-generated insights. Sharing data between institutions or integrating third-party AI tools requires secure data-sharing protocols to prevent security breaches and navigate complex legal and regulatory frameworks [58].

Equally important are the ethical issues surrounding bias, demographic representation, and the equitable delivery of care, which impact how AI can be used to serve diverse patient populations. AI systems can inadvertently perpetuate or even exacerbate biases present in the training data [57]. For instance, if an AI model used to predict BPD risk is trained primarily on data from NICUs in high-resource settings, it may not generalize well to infants born in lower-resource settings or those from underrepresented racial or ethnic groups. This bias can lead to disparities in care, with AI potentially misclassifying or underestimating risks for certain populations. Training datasets must be diverse and representative of the populations affected by BPD to avoid bias. This includes data from infants of different racial and ethnic backgrounds, as well as those from varying socioeconomic and geographic contexts. Ensuring demographic diversity in the data can improve the generalizability of AI models and reduce the risk of biased or unequal treatment. Ensuring equitable access to AI-driven technologies is a key ethical consideration to prevent widening the gap between healthcare providers in different regions or socioeconomic environments.

Finally, practical barriers, such as the readiness of healthcare institutions, acceptance by clinical teams, and cost, must be carefully navigated to realize the full potential of AI in BPD management. Many hospitals lack the infrastructure, such as advanced computing resources or integrated data systems, to support AI applications. Even when the technology is available, there may be a lack of trained personnel to manage and interpret AI outputs.

Hospitals must invest in staff training and robust infrastructure to further AI research and applications. This may be costly, particularly for smaller hospitals or healthcare systems with limited budgets. Initiatives or policies that support equitable access to AI technologies, such as government funding or public–private partnerships, can aid in the broader adoption of AI technology. AI models should also fit seamlessly into existing clinical workflows, avoiding disruptions that could lead to inefficiencies or errors in care. Involving clinicians in developing and testing AI models can increase the trust and acceptance of the models by multidisciplinary clinical teams [37].

5.3. Future Directions and Research Opportunities

The continuous development of advanced DL models presents unprecedented opportunities for enhanced BPD risk prediction and personalized management. Building upon Mohsen et al.’s, 2022 demonstration of multimodal data integration, future research should pursue comprehensive, multidimensional approaches to predictive modeling [60]. A promising avenue involves designing prospective, multicenter randomized studies that compare traditional single-modal predictive models with integrated models combining chest radiograph transformer models, comprehensive intensive care unit clinical datasets, and genetic risk factor analyses.

Such research should aim to develop a predictive framework that integrates advanced DL processing of chest radiographs with detailed clinical data, with the specific objective of validating early-onset BPD risk assessment before 28 days of life. Inspired by personalized medicine approaches, like Liu et al.’s, 2019 radiotherapy optimization, researchers could develop longitudinal cohort studies incorporating genetic testing, comprehensive EHRs, and ML-driven treatment stratification [61]. This approach would involve creating a risk stratifier that analyzes genetic risk markers, physiological parameters, and historical treatment response data to generate individualized treatment risk scores and recommend tailored intervention strategies.

Genomic research represents another frontier for BPD predictive modeling. Large-scale multicenter genetic sequencing of preterm infants could enable the development of models capable of identifying high-risk genetic markers, predicting BPD progression, and recommending preventative interventions. Additionally, advanced imaging predictive techniques offer additional research opportunities to investigate transfer learning approaches in radiographic analysis, deep CNN customization, and automated lung development progression tracking. Researchers should develop standardized imaging protocols for preterm infant lung assessment and create ML models capable of predicting lung development trajectories with unprecedented precision.

These proposed research designs leverage the transformative potential of AI in personalized medicine. By integrating advanced computational techniques with comprehensive clinical datasets, researchers can develop more precise, individualized approaches to BPD management. However, the potential to revolutionize neonatal respiratory care through sophisticated ML approaches represents a critical advancement in pediatric medical research, promising more targeted, effective interventions for preterm infants at risk of BPD.

5.4. Limitations

This study has several limitations. Although studies included data from multiple countries, there was an overrepresentation of data from high-resource settings, particularly Korea and China. Future research should prioritize data collection from diverse healthcare environments to enhance model generalizability. Additionally, standardization remains a major challenge, as included studies utilized different datasets with variable feature selection, preprocessing steps, and outcome definitions. Harmonizing data across centers is necessary to improve AI model reproducibility and reliability. Selection bias and small sample sizes also impact model performance. Techniques such as oversampling, synthetic data generation, and transfer learning could help address these challenges.

5.5. Conclusions

Modern computational technology, such as AI and DL, have revolutionized how we understand BPD, in many cases, surpassing traditional diagnostic methods. Notable advancements include early diagnostic capabilities, with DL models predicting BPD risk up to four weeks earlier than clinical assessments, and integrated genetic and clinical models achieving AUC values above 0.85. Studies have identified the associations of BPD with genes involved in T cell development and have developed predictive models that use these gene markers with high AUCs. Other studies have identified unique clinical features, such as absent or reduced end-diastolic flow velocity in the umbilical artery as highly predictive features of BPD. Beyond risk stratification and early BPD prediction, researchers have leveraged AI to predict the course of BPD severity and post-extubation failure, and others have identified superior ventilation treatments for BPD patients. These tools offer personalized insights into disease mechanisms and interventions, leveraging data from EHRs, genetic testing, and radiological imaging for a more comprehensive approach to neonatal care. Despite these breakthroughs, challenges remain in standardization, external validation, and clinical translation. AI and ML models show promise for BPD prediction, but real-world application requires improvements in explainability and cross-institutional collaborations. Future efforts must prioritize generalizable models, diverse demographic representation, and interdisciplinary collaboration to fully realize the potential of computational approaches in improving neonatal respiratory care.