Abstract
Purpose :
Recent genome-wide association studies (GWAS) in multiethnic cohorts have identified two loci significantly associated with severe retinopathy of prematurity (ROP) (≥ stage 3), highlighting the potential value of genetics in screening. However, routine DNA sample extraction may be financially and logistically prohibitive in infants. Here, we aim to predict the probability an infant carries the most significant ROP-related locus, rs2058019, using deep learning models analyzing retinal fundus images. We also investigate the potential of combining this predicted genetic risk score (GRS) with the vascular severity score (VSS) and gestational age (GA) to improve the risk model for treatment-requiring ROP (TR-ROP).
Methods :
Our study spanned eight North American centers, encompassing 1,800 infants, 920 of whom had genetic risk information. Given previous GWAS findings linking the rs2058019 risk locus to Hispanic and Caucasian infants, and the predictive value of early retinal fundus images for TR-ROP, we applied the following inclusion criteria: 1) Caucasian race, 2) first eye examination, and 3) infants not diagnosed with TR-ROP within the examination window. To infer the genetic risk label, we utilized a pre-trained CNN model, EfficientNetV2. The dataset was split into training, validation, and testing sets (70%/15%/15%), at the patient level. We developed 4 logistic regression models with L2 regularization to explore the combined effects of GA, VSS, and GRS, and employed five-fold cross-validation with a randomly repeated 100-time approach.
Results :
Of the 920 infants with genetic risk information, 452 (4,155 images) met inclusion criteria, with 39 positive for the rs2058019 risk locus, and 413 negative. The fine-tuned EfficientNetV2 model demonstrated accuracy at the patient level, achieving an AUROC of 0.877, AUPRC of 0.405, and F1 score of 0.444. For the TR-ROP risk model, we included 273 infants who met criteria. The model that combined GA, VSS, and GRS showed better performance compared to models that used GA and VSS or GA alone. (Table 1, mean AUPRC = 0.501±0.020 vs. 0.493±0.020 vs. 0.387±0.017).
Conclusions :
A deep learning model can infer a genetic risk locus from retinal fundus images, narrowing down the population who might need genetic screening. Furthermore, integrating the predicted GRS with VSS and GA improves the prediction accuracy for TR-ROP.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.