A recent study examined whether an artificial intelligence (AI) trained risk model could screen for retinopathy of prematurity, a leading cause of preventable blindness in infants in low- and middle-income countries, via telehealth.
A study published in JAMA Ophthalmology found that a retinopathy of prematurity (ROP) artificial intelligence (AI) risk model could be used to ensure early diagnosis and timely treatment for infants with treatment-requiring ROP (TR-ROP) while also reducing the burden of exams on ophthalmologists.
ROP is a leading cause of preventable blindness in infants in low- and middle-income countries. The condition is rare in the United States but on the rise in India, which has one of the highest rates of premature births in the world. The model described in the paper was originally trained in the United States but then retrained using data sets from India.
Infant retinal fundus images were collected as part of an Indian ROP telemedicine screening program and an AI-derived vascular severity score (VSS) was obtained from images from the first examination after 30 weeks postmenstrual age.
The study used data from February 1, 2019, to December 31, 2020, when infants in India were longitudinally screened for ROP. A follow-up data set for these infants was collected from January 1, 2021, to June 30, 2021, and was used as a test data set. The model was validated on infant screening data from Mongolia and from Nepal.
Infants in India and Mongolia were screened if they had a birth weight of 2000 g or less or gestational age of less than 37 weeks. Infants were screened in Nepal if they had a birth weight of 1700 g or less or gestational age less than 36 weeks.
There were a total of 3760 infants included in this study; 1950 (51.9%) were male and they had a median (IQR) postmenstrual age of 37 (5) weeks.
The datasets revealed that infants who had developed TR-ROP had lower birth weight and gestational age at birth. Infants with TR-ROP had higher VSS in all 3 data sets even before an official diagnosis. Infants with TR-ROP were identified a median (IQR) of 2.0 (0-11) weeks before TR-ROP diagnosis in India, 0.5 (0-2.0) weeks in Nepal, and 0 (0-5.0) weeks in Mongolia with the AI model.
Notches between box plots of the VSSs of infants who developed TR-ROP vs those who did not had stopped overlapping at 30 weeks postmenstrual age, which suggested that the median scores were different. The mean (SD) of the area under the receiver operating characteristic curve was 0.919 (0.034). The eye-level receiver operating characteristic curve was 0.944.
The diagnostic model proved to have a sensitivity and specificity, respectively, of 100.0% (95% CI, 87.2%-100.0%) and 63.3 (95% CI, 59.7%-66.8%) in India; 100.0% (95% CI, 54.1%-100.0%) and 77.8% (95% CI, 72.9%-82.2%) in Nepal; and 100.0% (95% CI, 93.3%-100.0%) and 45.8% (95% CI, 39.7%-52.1%) in Mongolia.
The researchers predicted that implementing this AI model would reduce the number of examinations by 45.0% in India, 38.4% in Nepal, and 51.3% in Mongolia.
There were some limitations to this study. The system was validated with the RetCam camera and future research must validate the effectiveness of the algorithm using other devices. Access to digital fundus photography and accurate measure of gestational age are required for this model to work, which may make it difficult for screening outside of telemedicine. The AI algorithms may be limited by photo quality. The model will also need to be adapted to other populations.
The researchers concluded that their study demonstrated that the AI-based risk model use telemedicine could reduce screening burden and boost early diagnosis in infants with TR-ROP, which could lead to improved visual outcomes for premature infants.
Coyner AS, Oh MA, Shah PK, et al. External validation of a retinopathy of prematurity screening model using artificial intelligence in 3 low- and middle-income populations. JAMA Ophthalmol. Published online July 7, 2022. doi:10.1001/jamaophthalmol.2022.2135