Skip to main content

  • Artificial intelligence models show promise in predicting ROP requiring treatment

    Pediatric Ophth/Strabismus, Retina/Vitreous

    Add to My Bookmarks

    Review of: External validation of a retinopathy of prematurity screening model using artificial intelligence in 3 low- and middle-income populations

    Coyner A, Oh M, Shah P, et al. JAMA Ophthalmology, in press

    An artificial intelligence (AI) risk model previously used to help predict treatment-requiring retinopathy of prematurity (TR-ROP) in the United States was adapted to predict TR-ROP in low- and middle-income countries.

    Study design

    Fundus photos from 3760 premature infants from the first exam after 30 weeks’ gestation were collected from India in order to establish an AI-derived vascular severity score (VSS). A model based on gestational age and VSS was created to predict which infants would develop TR-ROP; this model was then validated on another dataset. Finally, the model was applied and tested on datasets from India, Mongolia, and Nepal.

    Outcomes

    The AI model for the validation and test datasets showed 100% sensitivity in all 3 countries, while specificity ranged from 46% to 78%. The model identified those infants with TR-ROP at a median of 0 weeks before diagnosis in Mongolia, 0.5 weeks in Nepal, and 2.0 weeks in India.

    Limitations

    This AI model was based on digital photographs from a specific fundus camera (RetCam) and thus may not be applicable to other imaging modalities. Access to this technology may be limited in lower-income countries. Lastly, the model would need to be adapted and validated for different populations, though it is not clear what the criteria are for demarcating different populations. Although the current study is country-based, there could be wide variations in populations within a country based on access to medical care and differences in premature infant care standards.

    Clinical significance

    Artificial intelligence is being used to create models for stratifying risk for TR-ROP. Traditional ROP screening is labor- and resource-intensive, so improved models with high sensitivity and greater ability for predicting which infants will be at high risk will help increase efficiency and decrease cost without sacrificing patient care.