A risk model for early detection of treatment-requiring retinopathy of prematurity using a deep learning-derived vascular severity score

Abstract

Purpose : Retinopathy of prematurity (ROP) is a leading cause of blindness in children, although it is often preventable with accurate and timely diagnosis and treatment. ROP screening guidelines are designed to be highly sensitive to avoid missing cases of treatment-requiring (TR-) ROP; consequently, approximately 80% of exams in a screening population have no or mild disease. Current ROP risk models require multiple predictors and/or exams, and performance often decreases significantly when applied to more diverse populations. We aimed to develop a risk model that could reduce the screening burden without missing cases of TR-ROP by using demographic risk factors and a deep learning-derived vascular severity score (VSS, all of which can be evaluated during a single exam) using a large cohort of North American infants.

Methods : A multi-institutional ROP dataset consisting of retinal fundus images and clinical factors for 852 subjects was collected as part of the Imaging and Informatics in ROP (i-ROP) study. A reference standard ROP diagnosis was provided for each exam. Posterior pole images were assigned a vascular severity score ranging from 1.0 to 9.0. Considering that infants who develop TR-ROP often have increasing VSS prior to the diagnosis of TR-ROP, we developed a risk model based on demographic risk factors and the VSS at 32-33 weeks post-menstrual age. Using all combinations of birth weight, gestational age (GA), and VSS, 7 ElasticNet logistic regression models were tuned via 5-fold cross-validation. The best-performing model was evaluated using the held-out i-ROP test dataset consisting of 121 infants, and an independent dataset of 30 infants screened as part of a telemedicine program in Salem, OR.

Results : The best performing model used GA and VSS, based on the area under the precision-recall curve (Table 1). On each independent test set, the model achieved sensitivity of 100% with a positive predictive value ranging from 12% to 18%, and specificity ranging from 55% to 68% with a negative predictive value of 100% (NPV, Table 2).

Conclusions : This model, with just two predictors which can be collected during a single exam, can identify all subjects who will eventually develop TR-ROP, while correctly ruling out, with 100% NPV, more than half of those who will not.

This is a 2021 ARVO Annual Meeting abstract.

 

Table 1: 5-fold cross-validation results for tuned ElasticNets.

 

Table 2: Test set results.