In Neonatal Intensive Care Units (NICUs), preterm infants constitute a population particularly well-suited for AI-based monitoring of BPD. Currently, predictive models frequently incorporate physiological parameters such as heart rate, respiratory rate, oxygen saturation, and ventilator settings (7, 8). While NICU equipment facilitates continuous acquisition of large-scale physiological data, conventional monitoring techniques often lack the capacity to deliver accurate and real-time analyses due to the high variability of these indicators in preterm infants.

AI offers significant advantages in the prediction and management of BPD by enabling the integration and interpretation of complex, multidimensional clinical datasets. For instance, AI models embedded within electronic health record systems have been utilized to identify early risk signals for BPD from vital sign trends, ventilator metrics, and additional physiological variables. These models support clinical decision-making by contributing to early risk stratification, prognostic assessment, and the development of individualized treatment strategies (9, 10). Moreover, AI systems are capable of integrating diverse data sources, including imaging, genomic data, and biochemical profiles, to examine inter-organ interactions, thereby enhancing both the accuracy and comprehensiveness of predictive models (11, 12). The development of clinically applicable BPD risk assessment tools requires minimization of false-positive rates and maximization of clinical interpretability. It remains essential that healthcare professionals synthesize AI-generated outputs with comprehensive clinical evaluations to ensure that decision-making remains clinically grounded and contextually appropriate.

In the context of BPD prediction, supervised learning is the predominant methodology. That is, based on the data with known labels, the model is trained to achieve the mapping between features and outcomes. In terms of application division of labor: structured clinical variables and laboratory data often adopt machine learning approaches (12–15). Medical imaging and high-dimensional continuous monitoring signals are more suitable for deep learning (13). During the model development phase, hyperparameter optimization is typically used to improve generalization performance. For model performance evaluation, in addition to AUROC (Area Under the Receiver Operating Characteristic Curve), it is recommended to report AUPRC (Area Under the Precision-Recall Curve) simultaneously to better reflect the ability to identify positive cases. Meanwhile, the Brier score is used to assess the consistency between probabilistic predictions and actual outcomes (16). The model validation phase mainly includes: internal validation (e.g., cross-validation, bootstrapping), which is used to evaluate consistency and robustness; external validation (using data from different institutions or time periods), which is used to test transferability across regions and systems (17). At present, most BPD-related AI models are still in the development or internal validation stage, lacking cross-center external validation and real-world data testing, which limits their feasibility for clinical application. Ultimately, evidence still relies on randomized controlled trials (RCTs) to assess their real-world benefits and safety for clinical outcomes, thereby supporting routine application.

Currently, RTI BPD Outcome Estimator (https://neonatal.rti.org) is a widely used clinical tool for BPD prediction, featuring clear parameter settings, transparent logic, and a user-friendly interface, making it convenient for bedside rapid application (18). However, its limitation lies in that the model is constructed based on a small number of static variables (such as gestational age, birth weight, and oxygen concentration), and it is difficult to incorporate continuous physiological signals, medical images, and multi-omics data in the NICU. It also lacks a dynamic update mechanism and adaptive capabilities, which limits its adaptability in complex scenarios.

In recent years, the application of AI in BPD research has become increasingly widespread. Lei et al. used a random forest model with 6 variables to predict BPD, achieving an AUC of 0.929 (19). Dai et al. combined clinical and genetic information, and the AUC values for predicting BPD and severe BPD were 0.915 and 0.907, respectively—outperforming the model that used only clinical data (20). A study based on chest x-rays demonstrated that the AI model can accurately predict BPD within 24 h of birth, outperforming expert diagnosis (21); another transcriptomics study also showed that a small number of genes can effectively predict BPD, expanding the data dimension (22). However, a high AUC alone is insufficient to support clinical translation. First, in the context of class imbalance, AUC tends to be overestimated. It is recommended to report AUPRC simultaneously to reflect the value of identifying minority classes (patients with diseases). Second, features must be accessible at the time of prediction to avoid data/label leakage. Recent reviews and empirical studies have shown that leakage significantly exaggerates effect estimates, especially in small sample sizes (23); third, external validation must be completed across three dimensions—geography, time, and equipment—and the sample size, distribution of gestational age/body weight, definition of BPD, and its sensitivity analysis should be transparently disclosed; fourth, in terms of evaluation dimensions, the focus should expand from discrimination to calibration and clinical net benefit: beyond AUC/AUPRC, regularly report the Brier score, expected calibration error (ECE), and calibration slope/intercept, use Decision Curve Analysis (DCA) to quantify net benefits at preset adoption thresholds (24), and simultaneously provide the PPV (Positive Predictive Value), NPV (Negative Predictive Value), and alert rate at these thresholds to align with the action thresholds of NICUs; the aforementioned practices collectively enhance the model's robustness and adoptability, serving as a prerequisite for its entry into real-world evaluation and clinical trials.

BPD is recognized as a multifactorial disorder, and reliance on individual predictors is insufficient for accurate risk estimation, multivariate integration offers greater advantages (26). Observable clinical information corresponding to alveolar simplification is mainly reflected in the distribution and fluctuation of FiO₂ and SpO₂, the frequency and duration of intermittent hypoxemia (IH), and the oxygen saturation index (OSI). Phenotypes associated with increased pulmonary vascular load often manifest as persistently high oxygen demand under similar ventilation settings and difficulty in weaning from ventilation. In NICUs, high-resolution data (e.g., ventilator parameters, continuous SpO₂, blood gas results, OSI) and low-resolution data (e.g., demographics, medical history, medication records) together form longitudinal evidence, enabling early identification of ventilation-oxygenation mismatch and trends in respiratory work (27). Based on this, it is recommended to extract robust temporal features (median, variability, trend slope, etc.) from EHRs across three clinical time windows: ≤24 h after birth, 3–7 days after birth, and approximately 28 days after birth, with the features limited to those accessible at the time of prediction. The model output should provide calibrated individualized risk and report clinical net benefit under preset thresholds, facilitating direct alignment with ventilator parameter fine-tuning, oxygen therapy target setting, and weaning assessment.

It is important to emphasize that despite the richness of available clinical data, multiple challenges persist, including the labor-intensive nature of manual data entry, missing values, changes in disease definitions over time, difficulties in capturing complete longitudinal records, potential biases in data collection, and concerns related to data governance and privacy. These limitations underscore the necessity of advancing data acquisition and storage methodologies to enhance the robustness of BPD-focused research and model development.

It complements the sequential variables such as EHR and enables repeatable quantification. In the NICU, chest X-ray (CXR) and bedside lung ultrasound (LUS) are the most accessible: CXR is convenient for longitudinal follow-up of lung gas volume and interstitial changes, while LUS can be frequently performed bedside and sensitively reflect lung gas volume and edema status (28). CT and MRI are more suitable for evaluating the pulmonary microstructure and ventilation/perfusion conditions in critical decision-making or research scenarios, but they are often limited by factors such as radiation exposure, time requirements, and the need for sedation (29).

The imaging AI technology is mainly divided into radiomics based on manual features and end-to-end deep learning. Usually, it needs to be combined with lung field segmentation, quality control and clinical variables for multimodal fusion to enhance the robustness and interpretability of the model (30). Research shows that the deep learning model based on neonatal CXR can output risk signals comparable to or even better than those of expert interpretations at an extremely early stage of birth. However, cross-center external validation is still insufficient, and it is urgently necessary to evaluate its generalizability under a unified annotation and standardized collection framework (21). Based on this, this article further proposes: The imaging AI evidence should be integrated into specific clinical time windows and intervention nodes, rather than just being a comparison of methodological performance. Thus, the experimental “imaging-AI” is transformed into a clinical pathway of “imaging-AI-guided intervention pathway”.

Blood biomarkers are easy to standardize and structurally input, making them important potential input features for AI models. Existing evidence shows the following: Angiogenic factors (VEGF, PLGF) are crucial for pulmonary vascular and lung development (31). Serum VEGF/PLGF is expected to be used for early risk identification and intervention stratification (32). Inflammatory factors (IL-4/6/8, TNF-α, IL-33) reflect the persistence of inflammation and are associated with a high risk of BPD (33, 34). Systemic immune-inflammatory indices (e.g., NLR, SIRI) are significantly elevated at birth and at 36 weeks of gestational age in infants with moderate-to-severe BPD, indicating an increased systemic inflammatory load (35, 36). Among lung injury biomarkers, KL-6 indicates alveolar epithelial and pulmonary interstitial damage (37) and its level at 14 days after birth shows good discriminative power for BPD (AUC = 0.88) (38). NT-proBNP reflects volume load/pulmonary vascular stress (39); in very low birth weight infants, NT-proBNP > 2.264 ng/L at 14 days after birth can early indicate the risk of BPD (sensitivity: 100%, specificity: 86%, AUC = 0.93) (40).

For clinical application, biomarker modeling should align with disease course time windows and pathological pathways: Within ≤24 h after birth, prioritize indicators reflecting alveolar/interstitial damage and early inflammation (e.g., KL-6, IL-6/IL-8) for extremely early risk stratification; from 3 to 7 days after birth, combine NT-proBNP with inflammatory profiles to identify the possible “re-elevation of inflammation-pulmonary vascular load,” so as to indicate the timing of respiratory support and fluid/pharmacological intervention; at approximately 28 days after birth, integrate persistent oxygen dependence, nutritional/growth indicators, and VEGF/PLGF to assess phenotypes of persistent damage and microvascular development disorders and plan follow-up. Integration with other modalities should follow the principle of “starting with the easy and moving to the difficult”: When sample size is limited or modalities are frequently missing, adopt late fusion (train each modality independently, then perform confidence weighting or stacking integration after outputting calibrated probabilities)—this facilitates external validation and allows the use of Decision Curve Analysis (DCA) to demonstrate clinical adoptability; after multi-center data and annotations are improved, transition to mid-level fusion (jointly model “pathway-level” biomarkers with imaging and temporal physiological data within the same time window) to obtain cross-modal information gain and mechanistic interpretability (41).

Although the above studies mostly employed traditional statistical models, their data structure is highly suitable for being incorporated into AI systems for multi-feature modeling. Although the blood biomarkers have shown certain potential initially, due to the small sample size of the research and the high clinical heterogeneity, they are currently still in the research stage and have not yet entered the routine clinical application process. Larger-scale multi-center studies are needed to verify their efficacy (36).

Multi-omics provides a key supplement for revealing the molecular and pathway-level mechanisms of BPD. At the metabolomics level: Premature infants with BPD show elevated levels of alanine, serine, taurine, and citrulline in bronchoalveolar lavage fluid on the first day after birth (42). Meanwhile, studies have shown that the levels of hydroxy fatty acids and oxidized fatty acids in the amniotic fluid of premature infants with BPD are higher than those in the non-BPD group (43, 44). At the microbiomics level: Research indicates that compared with the non-BPD group, the pulmonary microbial community of the BPD group shows a trend of increasing proportion of Firmicutes and decreasing proportion of Proteobacteria over time (45). Additionally, gut microbiota may affect BPD development through modulation of systemic immunity and inflammatory responses.

At the genomics/transcriptomics level: Variations in angiogenesis pathways (e.g., VEGF) and abnormalities in airway exosomal miRNAs (such as miR-876–3p, miR-20a-5p) are associated with BPD susceptibility (43, 46).

To advance omics evidence from “description” to “computational evidence usable for prediction and stratification,” two types of alignment need to be achieved. First is methodological alignment: perform batch effect correction and scale unification for multi-center/multi-platform data, and reduce high-dimensional omics features to pathway/module-level representations (e.g., aggregating genes/metabolites into activity scores for pathways such as “inflammation, oxidative stress, and angiogenesis”) to lower dimensionality and facilitate mechanism-based interpretation. Second is temporal alignment: align omics data with concurrent imaging data and continuous physiological signals according to clinical prediction windows (0–24 h, 3–7 d, ≈28 d) for joint modeling, ensuring features are accessible at the time of prediction and avoiding data/label leakage.

For integration strategies, progression should follow the principle of “late fusion first, then mid-level fusion.” When sample size and modality accessibility are limited, late fusion should be adopted first. As multi-center samples and annotation standards are gradually refined, transition to mid-level fusion: jointly characterize pathway-level omics, imaging, and other data within the same time window, and use multi-modal Transformers (with cross-modal attention and mask training to enhance robustness to missing modalities) or heterogeneous graph neural networks (constructing graphs of “gene/metabolite/microbiota—pathway—phenotype” and performing message passing for importance attribution) to improve mechanistic interpretability (47, 48). Existing small-sample prototypes have shown that combined metabolomic-microbiomic multi-omics exhibit cross-modal consistent signals in inflammation/oxidative stress pathways (49). However, regardless of the approach, multi-omics integration can only be transformed into interpretable, verifiable, and implementable decision support in NICU settings if discrimination, calibration, net benefit, and robustness all meet the required standards.

The implementation of DL models within clinical settings necessitates seamless integration with healthcare information systems, including EHRs and imaging databases. However, this process often encounters challenges such as high technical barriers, long implementation periods, large capital investment, and complex compliance requirements. Especially in developing regions with lower levels of informatization, the data foundation is weak, and obtaining high-quality samples is limited. Moreover, the training data has insufficient representativeness for some populations, all of which affect the generalization performance of the model (50). Furthermore, medical records often suffer from missing information and incorrect input. The data structures from different sources such as EHR, wearable devices, and medical images vary greatly, making it difficult to integrate across institutions, which undermines the stability and accuracy of the model (51–54). Currently, the related research on BPD lacks open and shared datasets and code libraries, which limits the verification of results and the re-development of models, and hinders its generalizability across different institutions.

Currently, the research on AI in BPD management mainly focuses on risk prediction, but there is less discussion on how to convert the prediction results into specific intervention measures (60). Based on existing systematic evidence, this review proposes a Time Window–PALM (Predict–Act–Learn–Monitor) framework, which takes ≤24 h, 3–7 days, and ≈28 days after birth as key nodes in the disease course, requires the model to output calibrated individualized risk at each node and use Decision Curve Analysis (DCA) to provide net benefits under acceptable thresholds, and maps risk thresholds to specific intervention measures (e.g., ventilator fine-tuning, oxygen therapy targets, medication/fluid management, imaging re-evaluation); within fixed time windows, the following should be routinely reported: calibration and net benefit, alert rate, and robustness across subgroups (e.g., by gestational age or center), which are used to trigger or delay interventions, and subsequently, real-world outcomes are fed back to recalibrate the model and conduct incremental update training while continuous monitoring of model performance and drift is required to ensure traceability and auditability.

The clinical value of AI lies in achieving refined, phased, and individualized treatment guidance (6). Within the first 24 h after birth, the model provides calibrated risk, net benefits at selected thresholds, and simultaneously presents PPV, NPV, and alert rate; if the risk exceeds the center's threshold, parameter optimization for lung-protective ventilation is first implemented (limiting tidal volume within the allowable oxygenation/carbon dioxide targets, setting appropriate PEEP, and reducing high-pressure impact), and lung recruitment strategies and oxygen therapy targets are reviewed based on CXR/LUS, followed by re-evaluation of risk trajectory and adverse events at 24–48 h to verify the immediate effect.

From 3 to 7 days after birth, calibrated individualized risk assessment is updated under the same methodological framework and interpreted in conjunction with inflammatory profiles/cardiopulmonary load signals (e.g., CRP/IL-6/IL-8, NT-proBNP, LUS scores); when both risk and net benefits meet the standards, stratified medication and fluid management can be initiated: for infants with severe inflammatory phenotypes and weaning difficulties, glucocorticoid regimens are selected based on evidence, while infants with suspected increased fluid/pulmonary vascular load receive fluid restriction/diuresis and supportive therapy, with re-testing at 48–72 h to evaluate efficacy and safety.

At approximately 28 days after birth, if infants still have persistent oxygen dependence, longitudinal oxygen demand curves, weight gain/nutrient intake, KL-6/NT-proBNP, and imaging features are comprehensively interpreted to distinguish phenotypes dominated by interstitial damage/alveolar simplification or accompanied by significant pulmonary vascular load: the former prioritizes nutritional and antioxidant support, rehabilitation, and weaning assessment, while the latter undergoes multi-disciplinary cardiopulmonary evaluation and necessary advanced imaging/functional tests, with the establishment of a long-term follow-up plan.

To facilitate understanding of practical implementation, two contextual examples are provided below: A very preterm infant was identified as high-risk by the model 6 h after admission, with positive threshold net benefits shown by DCA; the team accordingly adjusted down ventilator parameters and calibrated oxygen therapy targets, and re-evaluation at 48 h showed decreased OI/OSI and controllable alert rate, so the intervention was retained and documented. On day 5, NT-proBNP increased and LUS score worsened, with risk exceeding the threshold again; short-course low-dose glucocorticoids and fluid restriction were administered in accordance with prescription guidelines, and re-evaluation at 72 h determined whether to upgrade imaging or discontinue pharmacological intervention.

Predict: Use features accessible at the time of prediction, and routinely report calibration and threshold ranges indicated by DCA; it is recommended to centrally establish time-windowed risk control tables (including: risk, threshold, net benefit, preset actions, and re-evaluation time points) (61).

Act: Solidify “threshold-action” into executable medical orders/pathways (ventilator parameter targets, oxygen therapy ranges, recruitment strategies, inclusion/exclusion criteria for glucocorticoid and fluid management); run in parallel (shadow mode) in the initial launch phase to prevent premature automation (62).

Learn: Conduct incremental/periodic retraining and threshold recalibration based on real-world outcomes; supplement with feature ablation and key subgroup reviews to maintain stability (62).

Monitor: Continuously monitor performance, fairness, and drift; retain audit logs and version control; align with incident reporting-rollback processes, as well as information notification and appeal channels for family members and medical staff to ensure traceability and accountability (63).