Multisource Machine Learning Model for Detecting Referral-Warranted Retinopathy of Prematurity.

Purpose: To develop a multisource machine learning model for detecting referral-warranted retinopathy of prematurity (RW-ROP) using retinal images and demographics. Design: Secondary analysis of data from the Telemedicine Approaches to Evaluating Acute-Phase Retinopathy of Prematurity Study. Subjects: One thousand two hundred fifty-seven premature infants (mean birth weight 864 g; mean gestational age 27 weeks; 19.4% with RW-ROP) enrolled from 12 clinical centers in North America. Methods: A multisource ROPNet (MS-ROPNet) model that combines a VGG-Swin Transformer model to extract features from retinal images and a random forest model that captures patterns of demographics was developed using central-view retinal images from 7741 eye visits with concurrent clinical eye examinations and demographic characteristics (birth weight, gestational age, sex, ethnicity, and age at retinal image). The MS-ROPnet model was compared to several existing machine learning models for detecting RW-ROP. Main Outcome Measures: Model performance metrics including the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), sensitivity, specificity, and accuracy from fivefold cross-validation, with RW-ROP diagnosed by certified ophthalmologists as the reference standard. Results: The MS-ROPNet achieved an AUROC of 95.0 ± 0.7% (mean ± standard deviation), AUPRC of 78.0 ± 2.8%, sensitivity of 81.5 ± 4.0%, specificity of 94.3 ± 1.4%, and accuracy of 93.0 ± 0.9% under the default cutoff of 0.50 in predicted probability. After adjusting the cutoff to achieve higher sensitivity, MS-ROPNet had 90% sensitivity and specificity of 84.8 ± 2.5% (using cutoff 0.3504), and 95% sensitivity with specificity of 72.8 ± 6.9% (using cutoff 0.2074), which outperformed both existing multisource models and the best single-source model by ≥3.7% in specificity at the same sensitivity. Conclusions: The MS-ROPNet achieved high performance in RW-ROP classification by effectively integrating retinal images with demographics, demonstrating its potential for accurate risk stratification of RW-ROP. Financial Disclosures: The authors have no proprietary or commercial interest in any materials discussed in this article.