Surrogate Augmented Deep Predictive Learning for Retinopathy of Prematurity

点击翻译按钮获取中文摘要

Surrogate Augmented Deep Predictive Learning for Retinopathy of Prematurity ABSTRACT This proposal aims to develop novel surrogate augmented deep predictive learning algorithms for predicting retinopathy of prematurity (ROP). The proposal directly addresses a critical clinical burden in ophthalmology that limited ROP experts are available in the United States and worldwide, yet the early detection of ROP for timely treatment has tremendous clinical benefit for infants in preventing childhood blindness. Using a unique and massive dataset with 7905 image sets collected from a longitudinal observational study of 1257 premature infants from 13 centers in North America, we plan to develop, validate, and evaluate novel analytic algorithms that hold a promise of directly improving clinical practice in the ROP care of premature infants. The overarching goals of this proposal are: (1) to develop novel methods for performing risk stratification through the surrogate augmented deep predictive learning of earliest retinal images (prior to 34 weeks postmenstrual age [PMA]) and the most important ROP risk factors (birth weight, gestational age) for early prediction of referral-warranted ROP (RW-ROP), defined as plus disease, ROP in zone I, or stage 3 ROP or greater; and (2) to optimize the ROP schedule through the surrogate augmented deep predictive learning of accumulated longitudinal retinal images for the dynamic prediction of RW-ROP. Our methods, after proper validation in future prospective studies, may serve as a useful tool for ROP risk stratification and optimization of scheduling of ROP examinations, which can reduce the burden of ROP examination for both infants and ophthalmologists while improving the eye care of premature infants for the prevention of childhood blindness. The Specific Aims to achieve these goals are: Aim #1: Develop and evaluate the surrogate augmented deep predictive learning of the earliest retinal image sets taken prior to 34 weeks PMA and demographic factors to predict RW-ROP. Accurate risk stratification through earlier prediction of RW-ROP will help identify high-risk infants for close follow-up by ophthalmologists for early detection and timely treatment of ROP, and low risk infants who are currently receiving unnecessary physically stressful retinal examinations for less frequent ROP examinations. Aim #2: Implement the surrogate augmented deep predictive learning of accumulated retinal images over time to dynamically predict RW-ROP. The dynamic prediction of the future course of ROP by deep learning of longitudinally accumulated retinal images will help optimize the schedule of ROP examinations by ophthalmologists, thus reduce the burden of ROP examinations for both infants and ophthalmologists. The successful completion of this project will lead to novel analytic algorithms of retinal images for early identification of high-risk infants for close follow-up and for optimization of ROP exam schedule, which will lead to earlier detection and timely treatment of ROP while minimizing the number of ROP exams. This research is highly feasible and potentially transformative in its global impact on the ROP care of premature infants.