A Deep-Learning Model May Provide Effective Screening for More-Than-Mild ROP

An autonomous artificial intelligence (AI)–based screening model for retinopathy of prematurity (ROP) effectively detected moderate or worse ROP in 2 large datasets in the United States and India. With additional validation, such a program could have screening value especially in telemedicine applications, potentially significantly reducing burden on clinicians.

Study Design

This diagnostic, external validation study evaluated the performance of an automated, AI-based algorithm for the purpose of identifying more-than-mild retinopathy of prematurity (mtmROP) and type 1 ROP. The authors used “more-than-mild ROP" to encompass all eyes with type 1 ROP (worst disease, treatment warranted), type 2 ROP (moderate disease, treatment often recommended), and any eyes with pre-plus disease (abnormal vascular changes present but below the threshold of severe disease). The deep learning algorithm was trained using retinal fundus imaging data from the Imaging and Informatics in the Retinopathy of Prematurity study (843 infants) and was tested on the Stanford University Network for Diagnosis of ROP and Aravind Eye Care Systems telemedicine datasets.

Outcomes

The AI-based algorithm had more than 80% sensitivity and specificity for detecting mtmROP and type 1 ROP at the examination level. At the patient level, the algorithm identified all infants who developed type 1 ROP (100% sensitivity).

Limitations

The deep learning–based algorithm may not be trained to detect ROP in all clinical scenarios. For example, no cases of retinal detachment were present on initial screening in the datasets used to train or validate the algorithm. In addition, if camera technology or image acquisition protocols change over time, the performance of the algorithm would need additional validation. Lastly, precisely when to initiate and terminate screening using the algorithm will be important when determining if such AI-based algorithms can be effectively employed in the clinical setting, which may vary among differing health systems.

Clinical Significance

Retinopathy of prematurity is a significant cause of vision loss in infants worldwide. This study demonstrated that an autonomous, AI-based algorithm may be an effective screening method for detecting mtmROP and type 1 ROP in areas where ROP telemedicine programs are implemented. This may further reduce the clinician workload needed to identify affected individuals and help to minimize vision loss from ROP.

Financial Disclosures: Dr. M. Ali Khan discloses financial relationships with Allergan, Apellis Pharmaceuticals, Genentech (Consultant/Advisor); Regeneron Pharmaceuticals (Grant Support).