Comparison of strategies for grading retinal images of premature infants for referral warranted retinopathy of prematurity

Purpose To determine the accuracy of identifying referral-warranted retinopathy of prematurity (RW-ROP, defined as any zone I ROP, stage 3 or worse, or plus disease) from retinal image sets using three grading protocols: a single optic disk–centered image, a set of 3 horizontal images, and a 5-image set. Methods In this secondary analysis of images from the e-ROP study, a weighted sample of 250 image sets from 250 infants (125 with RW-ROP and 125 without RW-ROP) was randomly selected. The sensitivities and specificities for detecting RW-ROP and its components from a single disk center image, along with nasal and temporal retinal images, were calculated and compared with the e-ROP grading of RW-ROP of all 5 retinal images (disk center and nasal, temporal, superior, and inferior retinal images). Results RW-ROP was identified with a sensitivity of 11.2% (95% CI, 6.79%-17.9%) using a single disk center image, with a sensitivity of 70.4% (95% CI, 61.9%-77.9%) using 3 horizontal images, and a statistically higher sensitivity of 82.4% (95% CI, 75.0%-89.0%) using all 5 images ( P = 0.002). The specificities were 100%, 86.4%, and 90.4%, respectively. For grading using 3 horizontal images, sensitivity was 14.3% for plus disease, 25% for zone I ROP, and 71.2% for stage 3 or worse compared to 40.8%, 50%, and 79.8% for grading using 5-image sets, respectively. Conclusions Both a single, disk-centered, posterior pole image and 3 horizontal images were less effective than a 5-image set in determining the presence of RW-ROP on qualitative grading by trained readers.