Skip to main content

  • Using AI to Screen for ROP in Low- and Middle-Income Countries

    By Jean Shaw
    Selected and reviewed by Neil M. Bressler, MD, and Deputy Editors
    Add to My Bookmarks

    Journal Highlights

    JAMA Ophthalmology, August 2022

    Download PDF

    Blindness from retinopathy of pre­maturity (ROP) can be prevented with timely screening; however, such screening presents a challenge in many parts of the world. Coyner et al. aimed to determine whether an artificial intelligence (AI)–based risk model, incorporated into existing telemedicine infrastructure, could reduce the screen­ing burden and identify infants who need treatment in low- and middle-in­come countries (LMICs). They found that their risk model could identify all at-risk infants up to 11 weeks before clinical diagnosis of treatment-re­quiring (TR)–ROP in three separate screening cohorts in Asia.

    In a previous study, the researchers used their model to identify at-risk infants in the United States. For this diagnostic study, they tailored it to the LMIC setting using a dataset acquired by an ROP telemedicine screening program in India. Using fivefold cross-validation, logistic regression models were trained on two variables (gestational age and vascular severity score) for prediction of TR-ROP.

    The researchers validated the model on datasets from India, Nepal, and Mongolia. Primary outcome measures included 1) sensitivity, specificity, and positive predictive value and negative predictive value for predictions of future occurrences of TR-ROP; 2) the number of weeks before clinical diagnosis when a prediction of TR-ROP was made; and 3) the potential reduction in number of examinations required.

    A total of 3,760 infants were included in the study. The infants’ median post­menstrual age was 37 weeks, and 1,950 (51.9%) were male. The diagnostic model had a sensitivity and specificity, respectively, as follows: India, 100% and 63%; Nepal, 100% and 77.8%; and Mongolia, 100% and 45.8%. Infants with TR-ROP were identified a median of 2 (0-11) weeks before diagnosis in India, .5 (0-2) weeks in Nepal, and 0 (0-5) weeks in Mongolia. Implement­ing this model with a single examina­tion for infants found to be at low risk of TR-ROP would reduce the overall number of examinations by 38.4% in Nepal, 45% in India, and 51.3% in Mongolia.

    The authors noted that “by focusing resources on those who screen positive [with the risk model], it is less likely TR-ROP will be missed or treated late.” They added that this is especially true for aggressive ROP (A-ROP), which is more common in LMICs. (Also see related commentary by Isdin Oke, MD, in the same issue.)

    The original article can be found here.