Telemedicine for ROP Diagnosis in a Real-World System: Feasibility of Implementing Artificial Intelligence for Disease Screening

Abstract

Purpose : We performed a retrospective observational clinical study to evaluate the feasibility of Deep Learning to remotely evaluate infants at risk of ROP and detect treatment-requiring (TR) disease in an operational telemedicine system.

Methods : In this study, subjects were enrolled who participated in the Oregon Health & Science University ROP Telemedicine program with Salem Hospital from 2015-2018. Medical charts were reviewed for clinical demographics for all patients screened by the program using digital fundus imaging with remote telemedical interpretation. Patients were transferred to our tertiary care hospital for any TR-ROP. All images were analyzed using the Imaging and Informatics in ROP (“i-ROP”) DL system and a vascular severity score (“i-ROP score”) from 1-9 was assigned to each image using methods previously published. We compared the i-ROP score distribution for each International Classification of ROP (ICROP) disease category (no ROP, mild ROP, type 2 ROP or pre-plus, and type 1/ TR- ROP).

Results : 81 infants were included in this analysis. The mean (± standard deviation) gestational age at birth was 29.2 ± 2.1 weeks, with mean birth weight 1240 ± 235 grams. 613 individual eye telemedicine examinations were included (median 3 examinations per infant, range 1-10). Two patients (four eyes) developed TR-ROP during the remote screening and were transferred to our institution for treatment. The mean gestational age for the two infants that required treatment was 25 weeks with mean birth weight of 795 g. Figure 1 demonstrates the mean i-ROP score distribution by disease category (p <0.001). The area under the receiver operating characteristic curve for detection of TR-ROP by deep learning was 0.99.

Conclusions : Remote evaluation of retinal photos using a deep learning image assessment system may be an effective method for automated detection of TR-ROP. These findings suggest the feasibility of incorporating artificial intelligence systems into clinical practice for remote evaluation of ROP.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Figure 1. Clinical exam findings for 613 telemedicine ROP eye examinations with the corresponding i-ROP deep learning image assessment system score for these encounters.