Biomed Rep. 2025 Oct 22;23(6):195. doi: 10.3892/br.2025.2073. eCollection 2025 Dec.
ABSTRACT
Retinopathy of prematurity (ROP) is a proliferative vascular disease affecting preterm infants with incompletely developed retinal vasculature, characterized by abnormal vascular proliferation that can lead to retinal detachment and blindness. Given its impact on neonatal visual health, developing reliable risk prediction models for ROP has become crucial for optimizing clinical screening and intervention strategies. However, existing models exhibit substantial heterogeneity in methodology, validation, and performance, limiting their generalizability across diverse clinical settings. The present study aimed to evaluate and summarize the effectiveness of existing ROP risk prediction models in preterm infants through a systematic review and meta-analysis, with the goal of providing reliable clinical screening tools based on effectiveness metrics. A systematic search was conducted across PubMed, Cochrane Library, Web of Science and Embase databases using a strategy that combined MeSH terms and free-text words to identify literature associated with risk prediction models for ROP in preterm infants. The risk of bias was assessed using the PROBAST tool. Statistical analysis involved data synthesis, heterogeneity testing, subgroup and sensitivity analyses, and publication bias assessment. A total of 492 relevant articles were retrieved; following deduplication and screening, 28 articles involving ROP risk prediction models were included. The included studies were published between 2009 and 2025, with sample sizes ranging from 90 to 22,569 participants, and a total sample size of 72,991. A total of 16 studies did not specify the validation method, five conducted external validation, two performed both internal and external validation, and five performed only internal validation. PROBAST assessment revealed that all included models had a moderate risk of bias, primarily attributed to the retrospective nature of the study design, inconsistent variable measurement and inadequate control of confounding factors. Meta-analysis showed that the pooled area under the receiver operating characteristic curve (AUC) was 0.87 (95% CI: 0.34; 0.99), indicating good discriminative ability of the models. However, significant heterogeneity was observed (I²=99.2%, P<0.05). Subgroup analysis by model type demonstrated significant heterogeneity in both traditional statistical (I²=92.2%) and machine learning models (I²=97.3%). Subgroup analysis by study region showed no significant heterogeneity in studies from South America (I²=0%), while high heterogeneity was found in studies from Asia and North America + Europe (I²=96.6 and 93.6%, respectively). This may be associated with cross-regional differences in population characteristics (such as ethnicity and disease spectra) and variations in medical standards. Funnel plot and Peters’ bias test indicated high reliability of the overall study conclusions, and the results of the sensitivity analysis were stable. However, some studies had small sample sizes and single-center designs, leading to selection bias. Additionally, multiple studies lacked model validation, and samples were limited to specific regions, failing to cover diverse healthcare settings and ethnic groups. In conclusion, current ROP risk prediction models for preterm infants exhibit good clinical application potential, with certain discriminative and predictive abilities, which can provide references for clinical screening. However, the risk of bias and insufficient validation limit their generalization ability. Future studies should expand sample sizes through prospective designs, strengthen external validation and optimize model development to improve prediction accuracy and universality, addressing the identified risks of bias and limited generalizability.
PMID:41221538 | PMC:PMC12598925 | DOI:10.3892/br.2025.2073
