Retinopathy of prematurity (ROP) is a proliferative vascular disease that occurs in areas of the retina where vascular development is incomplete in preterm infants. It is characterized by abnormal blood vessel proliferation, which can lead to retinal detachment and blindness (1). A meta-analysis involving 121,618 preterm infants reported that the global incidence of ROP in preterm infants is 31.9%, with the incidence of severe ROP at 7.5% (2). A cohort study (3) conducted in the United States from 2003 to 2019 involving 125,212 cases of ROP showed that the overall incidence of ROP in premature infants increased from 4.4 to 8.1%; the incidence in African-American infants rose from 5.8 to 11.6% and that in infants from low-income families rose from 4.9 to 9.0%; and the increase was the greatest in the southern region (3.7 to 8.3%). Therefore, there are certain differences in the incidence of ROP in various regions and populations, and it is necessary to develop risk prediction models for different groups. Clinically, the diagnosis and severity assessment of ROP are based on the International Classification of Retinopathy of Prematurity, 3rd Edition (ICROP3), released in 2021. This classification systematically defines the lesion location, progression extent and risk level of ROP through indicators including lesion zones I-III, disease stages 1-5 and plus disease (severe retinal vascular dilation and tortuosity in the posterior pole), thereby providing a foundation for clinical diagnosis (4).

Current research suggests that the pathogenesis of ROP is characterized by a biphasic pattern of retinal vascular development: An initial phase of reduced retinal vascular growth, followed by excessive vessel proliferation into the vitreous body (5). Vascular endothelial growth factor (VEGF) serves a critical role in retinal neovascularization. In recent years, therapeutic approaches have shifted from laser therapy to anti-VEGF therapy (6). The RAINBOW Extension study confirmed that ranibizumab, when administered to preterm infants with ROP, significantly decreases the risk of high myopia (7). However, follow-up studies (8,9) have indicated that infants receiving anti-VEGF therapy may have a higher risk of adverse neurodevelopmental outcomes. Furthermore, in remote areas, the lack of medical resources and equipment can lead to worsening conditions in preterm infants, resulting in blindness (10). Therefore, compared with single biomarkers or imaging examinations, ROP prediction models have value in integrating multi-dimensional clinical data to achieve dynamic risk stratification. Particularly in resource-poor regions, these models may compensate for the insufficiency of fundus screening equipment, enabling the rapid identification of high-risk preterm infants through basic indicators. To obtain more practical evidence for ROP risk prediction, current research (11) focuses on early disease risk prediction, conducting large-scale clinical cohort studies, and developing new artificial intelligence/machine learning algorithms to enhance the accuracy and practicality of ROP risk prediction. However, with the development of computer vision and deep learning algorithms, despite the continuous emergence of various prediction models and algorithms, challenges such as overfitting and insufficient data diversity remain. There is significant heterogeneity among studies (12,13) in terms of sample size, predictors, model construction methods and reported performance metrics, and there are relatively few systematic reviews on the diagnostic performance of these models (14,15).

Almutairi et al (10) performed a meta-analysis that focused on the association between platelet count, thrombocytopenia and severe ROP, and explored the potential mechanisms linking a single biomarker to the disease through methods such as Bayesian model averaging. However, this previous study was limited to the causal exploration of a single risk factor, failing to involve the systematic integration and methodological evaluation of existing multifactorial ROP risk prediction models. It also did not assess differences in applicability of various predictive tools in clinical settings, and did not provide a comprehensive understanding and basis for selecting ROP risk screening tools. The present study aimed to overcome the limitations of single-factor association studies by conducting a systematic review and meta-analysis on risk prediction models for ROP in preterm infants. It systematically organizes, synthesizes and evaluates the methodological characteristics and application potential of existing models, thereby providing a design framework for the rational selection and optimization of ROP risk prediction models in clinical practice.

Effective screening combined with timely intervention can notably decrease the blindness rate of ROP (12). Novel imaging technologies and internet connectivity have transformed the ROP screening model, with artificial intelligence-supported ROP screening becoming a research hotspot (53,54). However, existing predictive models have limitations and are not well-suited for areas with quality poor neonatal care (55,56). The present study conducted a meta-analysis to evaluate the performance of current ROP prediction models and improve ROP risk prediction models.

The present study included 28 ROP risk prediction models, among which six studies did not provide the 95% CI values for the AUC. Pooling results from the remaining 22 studies, the AUC of ROP prediction models was 0.87 (0.34; 0.99), indicating that the overall performance of ROP risk prediction models is favorable, with good discriminative ability. However, there was high heterogeneity between the models. When the models were sub-grouped into traditional statistical and machine learning models, high heterogeneity remained within each subgroup. The causes of heterogeneity were considered to be differences in covariate selection, sample characteristics and outcome definitions and a lack of unified standards for data preprocessing and validation protocols. Regional subgroup analysis showed that studies from South America had no significant heterogeneity, while those from Asia, North America and Europe exhibited high heterogeneity. It was hypothesized that studies in South America showed consistent results due to the uniformity of population characteristics and medical standards within the region; by contrast, the high heterogeneity in Asia, North America and Europe may be attributed to notable differences in cross-regional population characteristics and medical systems. Nevertheless, funnel plot symmetry and Peters' bias test indicated no significant publication bias, and sensitivity analysis showed stable results, supporting the reliability of the conclusions. Future studies should reduce heterogeneity by unifying methodological standards and conducting multicenter research.

The risk of bias for all the models was unclear. Badriah et al (27) and Park et al (32) adopted retrospective designs, relying on medical record reviews for data collection. This may introduce information bias due to incomplete record-keeping or subjective data assignment. The study by Park et al (32) was a single-center study, with sample selection limited to a specific medical setting, which may decrease the generalizability of the results.

In terms of variable measurement, there were differences in the definitions and measurement methods of key predictors across studies. Filho et al (25) used ‘weight gain proportion at 6 weeks after birth’ as an indicator, while Cerda et al (31) employed changes in z-scores based on the Fenton growth curve. Such inconsistent standards limited data comparability. While most studies (25-52) referenced international classification criteria, Blue et al (38) and Chen et al (44) did not explicitly mention international ROP severity assessment criteria in their relevant evaluations. In addition, Filho et al (26) and Gerull et al (30) did not explicitly describe the implementation of assessor blinding, which may introduce subjective bias, however, the core staging criteria remained consistent.

The multicenter study by Ying et al (33) covered 29 hospitals in North America. Differences in medical care practices between institutions may have interfered with the identification of ROP risk factors, but the study adjusted for key variables such as gestational age and birth weight through multivariate regression, decreasing confounding effects to a certain extent. Park et al (32), Hari et al (50), and Shi et al (51) had relatively small sample sizes, which may pose a risk of model instability. However, owing to the explicit and rigorous definition of outcome events, the risk of overfitting (where a model or analytical approach exhibits excessive adaptation to the original dataset while lacking generalizability to new data-was kept at a controllable level. While specific biases were present across multiple dimensions of the study, they did not substantially compromise the overall validity of the findings. Consequently, the overall risk of bias was evaluated as unclear, indicating that although biases exist, they are not severe enough to invalidate the core conclusions of the study.

In predictor selection, the included studies did not report the effectiveness of predictors in practical applications or clearly specify whether the predictive ability of predictors was independently evaluated under blinded conditions. In addition, there were notable differences in terms of predictor selection and their determination methods. Some studies (45,46) failed to fully consider potential interactions between predictors. Multiple included studies consistently identified oxygen therapy as a key risk factor for ROP; however, the studies did not conduct in-depth exploration of interaction effects between factors such as different oxygen therapy modalities and fluctuations in oxygen concentration. Kubota et al (45) focused on analyzing the association between oxygen saturation (SpO₂) fluctuations and ROP risk by calculating the total difference in SpO₂ values over the total effective time. On the other hand, Lin et al (46) mainly focused on relevant indicators of fraction of inspired oxygen (FiO₂), including the average FiO₂ and the coefficient of variation of FiO₂, aggregating daily data to smooth short-term fluctuations and decrease noise. Unlike other studies (26,27,30,37,40,50) that may focus only on oxygen concentration or oxygen duration at a fixed point in time, this previous study emphasized the impact of the trend in FiO₂ changes over time on ROP. In practical applications, there may be differences in the frequency and method of FiO₂ monitoring and recording between different medical institutions.

Future studies should adopt prospective designs and increase sample sizes. This would allow for more precise control of influencing factors during data collection, enable tracking and observation of changes in the target population, provide more representative study samples, and avoid information and selection biases inherent in retrospective studies. Regularization methods should be used to limit model degrees of freedom and prevent overfitting. When constructing prediction models, potential confounding factors should be controlled.

Of 28 studies, 16 did not mention model validation, while five conducted internal validation. The internal validation sample sizes in the studies by Lin et al (46) and Takeda et al (52) were relatively small, which may not fully capture the various characteristics and associations in the data. While the models performed well on the training sets, they may not perform as well in real-world applications. In the study by Chen et al (44), five-fold cross-validation was performed on data from 22,569 patients, improving the training effect of the model and enhancing its stability and reliability. However, the occurrence and development of ROP may be influenced by factors such as medical conditions and environmental factors in different regions. Therefore, it is essential to include data from different regions and medical centers for external validation to ensure the general applicability of the model. The present study combined the external validation results from four studies (40,41,47,48) with an external validation model AUC of 0.90 (0.76; 0.96), showing good external validation performance. However, in clinical applications, the practical value of the model should be verified by incorporating other evaluation indicators.

Additionally, in the statistical analysis, certain studies (25,26) had issues with improper handling of missing data and insufficient assessment of collinearity. In the study by Takeda et al (52), five machine learning methods (decision trees, random forests, gradient boosting trees, neural networks and naive Bayes) were used to construct the models. Random forests and naive Bayes models performed well. Non-imaging machine learning models demonstrated high performance in predicting ROP occurrence, providing a feasible predictive approach for hospitals that lack access to retinal images or pediatric retinal cameras. However, this previous study acknowledged the small sample size and issues with variable collinearity. In the future, combining LASSO regression with embedded feature selection techniques may help control for collinearity between variables, select a stable and effective set of features and improve the robustness of the models.

Of 28 studies, 24 reported (25-27,29-35,37,39,40-48,50-52) that low birth weight in preterm infants is a risk factor for the occurrence of ROP. Yildirim et al (57) found that, compared with preterm infants without ROP, the average weight gain in the third week after birth was significantly lower in infants with ROP; Preterm infants with low birth weight have underdeveloped retinal vascular systems, and their blood vessels are structurally and functionally fragile, making it difficult to cope with abnormal vascular proliferation, fibrosis and other pathological changes, thereby increasing the risk of ROP (58).

Additionally, eight studies (33,34,37,42,47,48,49,52) used the Apgar score to establish prediction models. The Apgar score primarily evaluates neonatal health status based on heart rate, respiration, muscle tone, reflexes and skin color. A low Apgar score often indicates that the newborn experienced asphyxia or hypoxia at birth, which affects the normal development of the retinal blood vessels and increases the risk of ROP (59).

Multiple pregnancies are also a risk factor for ROP. In multiple pregnancies, the blood supply from the placenta is distributed unevenly, leading to insufficient fetal nutrition, and restricted growth and development, which often results in low birth weight (60)

Furthermore, studies (23,27) have shown that Caucasian infants have a higher risk of developing ROP compared with other ethnicities. This is because Caucasian individuals have less pigmentation, making the retinas more sensitive to oxygen and light damage. Under oxygen therapy or high-oxygen exposure conditions, the retina is more prone to vascular development disorders (61). However, a 2006-2017 study (62) of 41 North American hospitals found African-American infants had lower birth weight and gestational ages than white and Asian infants; African-American and Asian infants had significantly lower daily weight gain than Caucasian infants 31-40 days after birth and after adjusting for birth weight and gestational age, African-American infants had a lower incidence of severe ROP than Caucasian and Asian infants. There were no differences in the incidence or timing of severe ROP between different ethnicities; this mechanism requires further exploration.

Although the aforementioned factors have been confirmed to be associated with ROP in most studies, certain research (27,32,34,42,43,46,49,50-52) still has the limitation of having a relatively small sample size. Additionally, there are notable differences in medical standards, preterm infant care and environmental factors across regions, and the level of care provided to infants, including nutritional support and oxygen therapy management, varies (27,33,35).

It is recommended that pregnant patients at risk of preterm birth receive prenatal corticosteroids: these medications accelerate fetal lung maturation, reducing the incidence and severity of respiratory distress syndrome in preterm infants (63). Additionally, unnecessary postnatal ventilation and oxygen therapy for neonates should be avoided, as excessive oxygen exposure directly disrupts retinal vascular development. Neonates should also receive high-quality care and providing targeted management of comorbidities like bronchopulmonary dysplasia.

The present study had certain limitations: the systematic review was not pre-registered on the PROSPERO platform, which decreases the traceability of the research design and implementation process, and may affect methodological transparency and the validation of the standardization of the research protocol.

Future studies should consider combining fundus images with clinical data and use deep learning methods such as convolutional and recurrent neural networks, or ensemble learning algorithms such as random forest, gradient boosting machine and XGBoost to construct ROP prediction models (58,64). Such an approach could potentially handle large volumes of complex non-linear relationships and enable automated feature extraction, which might provide a more accurate basis for early screening-though this remains to be validated in future research.

In summary, existing ROP prediction models have potential in discriminative ability and may provide references for preliminary clinical screening. However, their application is limited by insufficient external validation and inadequate sample representativeness, and their generalization ability needs to be improved. Future model development should prioritize multicenter, large-sample prospective designs, control confounding factors, and systematically incorporate data from populations with different medical resource backgrounds and ethnic groups to enhance the generalizability of results. Meanwhile, it is necessary to strengthen external validation processes, explore multimodal fusion of fundus images and clinical indicators, introduce deep learning or ensemble learning methods to handle complex data associations and explore potential biomarkers associated with ROP, thereby providing more reliable tools for early accurate screening and intervention for neonatal ROP.