Deep Learning-Based Prediction of Cardiopulmonary Disease in Retinal Images of Premature Infants.

Importance: Bronchopulmonary dysplasia (BPD) and pulmonary hypertension (PH) are leading causes of morbidity and mortality in premature infants. Objective: To determine whether images obtained as part of retinopathy of prematurity (ROP) screening might contain features associated with BPD and PH in infants and whether a multimodal model integrating imaging features with demographic risk factors might outperform a model based on demographic risk alone. Design, Setting, and Participants: A deep learning model was used to study retinal images collected from patients enrolled in the multi-institutional Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study. The analysis included infants at risk for ROP undergoing routine ROP screening examinations from 2012 to 2020. Infants were recruited from 7 neonatal intensive care units. Images were limited to 34 weeks' or less postmenstrual age (PMA) so as to precede the clinical diagnosis of BPD or PH. The dataset included the period from June 2015 to April 2020. Data were analyzed from April to June 2025. Exposures: BPD was diagnosed by the presence of an oxygen requirement at 36 weeks' PMA, and PH was diagnosed by echocardiogram at 34 weeks. A support vector machine model was trained to predict BPD or PH diagnosis using (1) image features alone (extracted using ResNet18), (2) demographics alone, or (3) image features concatenated with demographics. To reduce the possibility of confounding with ROP, secondary models were trained using only images without clinical signs of ROP. Main Outcomes and Measures: For both BPD and PH, performance was reported on a held-out test set and assessed by the area under receiver operating characteristic curve (AUROC). Results: A total of 493 infants (mean [SD] gestational age, BPD, 25.7 [1.8] weeks; normal, 27.3 [1.8] weeks; 267 male [54.2%]) were included in this analysis. Performance was reported on a held-out test set (99 patients from the BPD cohort and 37 patients from the PH cohort). For BPD, the multimodal model showed higher accuracy (AUC, 0.82; 95% CI, 0.72-0.90) than demographics-only (0.72; ∆AUC, 0.1; 95% CI, -0.008 to 0.21; P = .07) or imaging-only (0.72; ∆AUC, 0.1; 95% CI, 0.04-0.16; P = .002) models. For PH, multimodal AUC was 0.91 vs the demographics-only 0.68 (∆AUC, 0.14; 95% CI, 0.006-0.27; P = .04) and imaging-only 0.91 (∆AUC, -0.09; 95% CI, -0.3 to 0.12; P = .40) models. Results persisted when trained on images lacking clinical ROP signs. Conclusions and Relevance: Results suggest that retinal images obtained during ROP screening may be used to predict the diagnosis of BPD and PH in preterm infants, which may lead to earlier diagnosis and avoid the need for invasive diagnostic testing in the future.