Purchase this article with an account.
Praveer Singh, Aaron S Coyner, Brian K Jordan, Robison Vernon Paul Chan, Susan R Ostmo, Kemal Sonmez, Deniz Erdogmus, Cindy McEvoy, Michael F Chiang, J. Peter Campbell, Jayashree Kalpathy-Cramer; Can the eye be a window to the lungs? AI predicts Bronchopulmonary Dysplasia through Retinal Fundus Photographs. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2985 – F0255.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Bronchopulmonary Dysplasia (BPD) is the leading cause of serious pulmonary morbidity in premature infants. 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 ROP screening may predict a future diagnosis of BPD.
5255 RFPs were collected from 871 patients as part of a multi-institutional Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study. Apart from clinical information, the dataset comprised of Reference Standard diagnoses for Plus disease for all patients. All RFPs corresponding to either patients without a diagnosis of BPD or captured at PMA>=34 weeks were removed (since BPD is diagnosed at 36 weeks PMA), leaving 477 patients /1284 RFPs (Post Menstrual Age [PMA]- mean:32.43, std:0.88), which were then divided into Train, Val, and TestSets via 80:10:10 splits on patient level. A Deep Learning (DL) model was trained to predict BPD at 36 weeks using the TrainSet (1006 RFPs; BPD:429, Normal:577; Plus:13, Pre-Plus:23, Normal:970; [PMA]- mean:32.48, std:0.89). The best performing model with the highest AUC-ROC score on the ValSet (137 RFPs; BPD:61, Normal:76) was finally evaluated on the TestSet (141 RFPs; BPD:47, Normal:94; Plus:4, Pre-Plus:4, Normal:133; [PMA]- mean:32.27, std:0.74). To avoid the DL model learning any common biomarkers with ROP disease, a secondary model was trained with only Non- Plus/Pre-Plus images using a pruned TrainSet (970 RFPs; BPD:403, Normal:567; Plus:0, Pre-Plus:0, Normal:970; [PMA]- mean:32.47, std:0.89) and ValSet (127 RFPs; BPD:55, Normal:72), though the performance was reported on the original TestSet.
The model trained with original TrainSet, performs with an overall AUC-ROC of 0.82 (image-level) and 0.86 (patient-level) on the TestSet (Fig.1). Performance improves to 0.84 (image-level) and 0.87 (patient-level) when the model is trained on a pruned TrainSet with only Non- Plus/Pre-plus images (Fig.2).
We found that a DL model trained on RFPs could predict a future diagnosis of BPD, even in babies with no clinical signs of ROP. In other words, the model isn’t just learning that BPD and ROP often occur in the same babies. Early identification of babies at high risk for BPD may facilitate interventional trials to reduce morbidity from BPD in the future.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only