Accepted for/Published in: JMIR Medical Informatics
Date Submitted: Aug 9, 2026
Date Accepted: Oct 30, 2026
Image-based Artificial Intelligence for the Diagnosis of Retinopathy of Prematurity: Systematic Review and Meta-Analysis
ABSTRACT
Background:
Retinopathy of prematurity (ROP) is one of the major causes of preventable blindness in premature infants. Early screening, accurate diagnosis, and severity classification are crucial for reducing irreversible visual impairment. Recently, image-based deep learning (DL) models have been widely applied to ROP screening, diagnosis, and severity classification. However, the model architecture, data source, validation method, and diagnostic task vary across studies. Consequently, the overall diagnostic efficiency and clinical utility of DL remain to be systematically assessed.
Objective:
This study aims to evaluate the performance of DL in the diagnosis and grading of ROP, explore its clinical screening application value, and compare the diagnostic accuracy of deep learning, traditional machine learning algorithms, and ophthalmology clinicians.
Methods:
We performed a systematic search across Web of Science, Embase, IEEE Xplore, PubMed, and Cochrane Library as of June 12, 2026 to identify studies investigating image-based DL for ROP detection or clinical subtype classification. The risk of bias (RoB) of included studies was assessed using the QUADAS2 tool. Bivariate mixedeffects models were applied for data analysis, and publication bias was evaluated via Deeks' funnel plot.
Results:
Sixty-two studies were ultimately included, with 165 2×2 contingency tables. The pooled sensitivity was 0.95 (95% CI 0.93–0.96), the pooled specificity was 0.95 (95% CI 0.94–0.97), and the pooled area under the receiver operating characteristic curve was 0.98 (95% CI 0.97–0.99). Nevertheless, model performance varied across different validation conditions. Both internal validation and crossvalidation had a specificity of 0.96. In external validation, sensitivity was stable at 0.95, but specificity declined to 0.82. Models also exhibited reduced specificity for tasks with higher clinical requirements, including the identification of treatmentrequiring ROP (0.74) and classification of mild and severe ROP (0.78). Significant betweenstudy heterogeneity was observed for both sensitivity and specificity.
Conclusions:
Imagebased DL models achieve high pooled diagnostic accuracy in datasets dominated by retrospective studies and internal validation. However, the specificity of DL models declines in independent external validation and clinically complex classification tasks. Considering the substantial betweenstudy heterogeneity and the relatively limited evidence in external validation, DL models are more appropriate as auxiliary tools for ROP screening and clinical decisionmaking. More multicenter, prospective, independent external validation studies are required to further verify their realworld clinical utility. Clinical Trial: CRD420261418502
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.