Deep learning models trained on retinal images obtained during routine screening for retinopathy of prematurity were able to predict bronchopulmonary dysplasia and pulmonary hypertension in premature infants, according to a diagnostic study published in JAMA Ophthalmology.

Researchers evaluated whether posterior pole fundus photographs — captured as part of standard neonatal intensive care unit screening — contained nonocular signals associated with cardiopulmonary disease. The work builds on the emerging field of “oculomics,” in which retinal imaging is used to infer systemic disease risk. In this setting, the potential advantage is practical: extremely premature infants already undergo repeated retinal imaging for retinopathy of prematurity screening, offering a built-in pathway for deployment if predictive models prove clinically useful.

The analysis used images from infants enrolled in the multicenter Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study, which recruited infants from seven neonatal intensive care units between 2012 and 2020. For the current analysis, images were limited to those taken at 34 weeks’ or less postmenstrual age to ensure that retinal data preceded the clinical diagnosis of bronchopulmonary dysplasia or pulmonary hypertension.

The study included 493 infants in the bronchopulmonary dysplasia cohort. A subset of 184 infants from a single site, Oregon Health & Science University, comprised the pulmonary hypertension cohort, where additional cardiology labels were available. Bronchopulmonary dysplasia was defined by oxygen requirement at 36 weeks’ postmenstrual age. Pulmonary hypertension was defined based on echocardiography performed at 34 weeks’ postmenstrual age.

Researchers compared three approaches: a model trained on image features alone, a model trained on demographic risk factors alone (gestational age, birth weight and postmenstrual age), and a multimodal model that combined both image features and demographics. Retinal image features were extracted using a ResNet18 deep learning architecture and classified using a support vector machine. For the pulmonary hypertension analysis, investigators used ImageNet-pretrained ResNet18 features without fine-tuning because of the smaller data set size.

For bronchopulmonary dysplasia, the multimodal model performed better than either demographics alone or imaging alone. In the held-out test set, the multimodal model achieved an area under the receiver operating characteristic curve of 0.82 compared with 0.72 for both the demographics-only and imaging-only models.

For pulmonary hypertension, imaging appeared to carry particularly strong diagnostic signal. The imaging-only model achieved an area under the receiver operating characteristic curve of 0.91 compared with 0.68 for the demographics-only model. The multimodal model also achieved an area under the receiver operating characteristic curve of 0.91, suggesting that adding demographic variables did not meaningfully improve performance beyond the retinal features.

Because prematurity-related conditions tend to cluster, researchers examined whether the findings were confounded by the presence of retinopathy of prematurity itself. Secondary models restricted to images without visible retinopathy signs yielded consistent results.

In the discussion, researchers suggested several possible biological explanations. For bronchopulmonary dysplasia, they proposed that oxygen exposure, mechanical ventilation or continuous positive airway pressure might influence retinal or choroidal vasculature in ways detectable in fundus images. For pulmonary hypertension, they hypothesized that elevated right-sided pressures could contribute to venous congestion or altered retinal vascular drainage, similar to changes described in adults with pulmonary hypertension.

The researchers noted that the findings are proof of concept and hypothesis generating. They noted several limitations, including the relatively small pulmonary hypertension cohort, lack of external validation across different imaging devices and absence of model explainability analyses. They also cautioned that deep learning models may perform poorly when applied to out-of-distribution images, such as those acquired in low-resource settings or using different camera systems.

Nevertheless, the study raises the question of whether retinal imaging already embedded in care pathways could eventually support earlier identification of infants at high risk for severe cardiopulmonary complications.

Disclosures can be found in the study.

Source: JAMA Ophthalmology