An optimized artificial intelligence (AI) algorithm has accurately identified retinopathy of prematurity (ROP) in 100% of severe cases and 80% of more-than-mild cases, according to a new paper published in JAMA Ophthalmology.
Give me some background first.
ROP is caused by the abnormal development of retinal blood vessels as a complication of preterm (before 37 weeks) birth and/or low birth weight and can lead to partial or even complete vision loss.
Fact: Approximately 14,000 infants born in the U.S. each year will have ROP, and up to 10% of those will develop severe ROP.
Now, talk about the study.
This study's goal was to determine whether an AI algorithm could identify more-than-mild ROP (mtmROP—defined as “eyes with type 2 ROP or type 1 ROP or any eye with pre-plus disease”) and type 1 ROP.
Explain type 1 and 2 ROP.
Eyes diagnosed with type 1 ROP include those with:
- Zone 1 ROP with plus disease
- Zone 1, stage 3 ROP without plus disease
- Zone 2 , stage 2 or 3 ROP with plus disease
Type 2 ROP eyes are those with prethreshold ROP:
- Zone 1, stage 1 or 2 ROP without plus disease
- Zone 2, stage 3 ROP without plus disease
See here for details on each ROP classification.
Now tell me about this algorithm.
The iROP deep learning (DL) algorithm was originally developed as a plus disease classifier but has undergone consistent updates and optimization since then.
This particular study retrained the iROP DL using 2,530 examinations from 843 infants in the Imaging and Informatics in Retinopathy of Prematurity (i-ROP) study.
Who was included in the study?
The study itself evaluated the performance of the algorithm on two datasets:
- 6,245 examinations from 1.545 infants in the Stanford University Network for Diagnosis of ROP (SUNDROP)
- 5,635 examinations from 2,699 infants in the Aravind Eye Care Systems (AECS) telemedicine programs
These data were collected between January 2012 and July 2021, and the analysis was performed July to December 2023.
Findings?
For mtmROP, the sensitivity of the algorithm was 80%—and 100% for type 1 ROP.
Interestingly, one infant screened negative at the examination where type 1 ROP was diagnosed, but had screened positive previously.
“This case underscores four key aspects,” wrote the study authors: “The role of [field of view] in ROP diagnosis, the subjectivity in diagnosing plus disease, the vascular phases of ROP, and the need for clinical safeguards for implementing autonomous ROP screening.”
Expert opinion?
“While there are doctors who are skilled in ROP treatment in many parts of the world, there simply aren’t enough to screen all of the babies who are at risk,” stated the study’s corresponding author, J. Peter Campbell, MD, MPH.
“This paper demonstrates that AI can effectively replace the physician for bedside screening and refer the most urgent cases to a physician for treatment.”
Limitations?
In a response to the study, Heidi Bowie, OD, MPH, noted that: “AI alone does not replace an appropriate medical history from the parent to determine if there were other issues that might be related to retinal abnormalities, including the possibility of inherited retinal conditions that might affect infants.”
She added: “Co-morbidities, both ocular and systemic, likely should be incorporated into synchronous or asynchronous AI algorithms.”
Take home.
Currently, diabetic retinopathy is the only eye disease autonomously detectable by AI.
Should this technology bear fruit, it could mean expanded access to eyecare for vulnerable populations currently underserved.