Can the eye be a window to your body? AI predicts Neonatal Pulmonary Hypertension through Retinal Fundus Photographs.

Abstract

Purpose : Pulmonary Hypertension (PHT) is a serious morbidity in premature infants that adds significant complexity to the clinical care of affected infants, with outcomes as grim as 50% mortality within the first 2 years. Recent work has found that retinal fundus photos (RFPs) can contain information relevant to systemic health in adults. In this study, we evaluated the hypothesis that RFPs obtained as part of Retinopathy of Prematurity (ROP) screening may predict PHT diagnosed at 34 weeks Postmenstrual Age (PMA).

Methods : 5255 RFPs were collected from 871 patients as part of a multi-institutional Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study. Only one institution (Oregon Health and Science University) had PHT labels, leaving us with 2331 RFPs from 256 patients.
After removing duplicates as well as RFPs captured at PMA>34 weeks (since PHT was diagnosed at 34 weeks), we were left with 184 patients / 616 RPFs, which was then divided into Train+Val and Test splits via 80:20 splits on patient level. We performed a 5-fold cross validation on the Train+Val set. We extracted Deep Learning (DL) features from 616 RFPs using a pretrained Resnet18 model architecture, and trained 5 different SVMs (corresponding to 5-fold cross validation) to predict PHT at 34 weeks. We finally used this ensemble of 5 models to evaluate on the held out TestSet. Again to avoid any confounding with ROP, a secondary model was trained with only Non-Plus/Pre-Plus images using a pruned Trainset, though the performance was reported on the original held out TestSet.

Results : The ensemble PH model, trained via 5-fold cross validation on the original Train+Valset resulted in an overall AUC-ROC of 0.79 (image-level) and 0.96 (patient-level) on the held out Test Set (Fig.1). Performance was 0.74 (image-level) and 0.96 (patient-level) when the model is trained on pruned Trainset with only Non-Plus/Pre-plus images (Fig.2).

Conclusions : We found that a DL model trained on RFPs could predict PHT, even in babies with no clinical signs of ROP. In other words, the model isn’t just learning that PHT and ROP often occur in the same babies. Early identification of babies at high risk for PHT may facilitate interventional trials to reduce morbidity from PHT in the future.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.