Effectiveness of an artificial-intelligence based single-exam retinopathy of prematurity prediction model in three Asian populations

Abstract

Purpose : Previous work has demonstrated that an artificial-intelligence (AI) based retinopathy of prematurity (ROP) risk model could provide both early prediction of treatment-requiring (TR) ROP and reduce the overall number of examinations in a US population. However, significant demographic differences exist between US and low and middle income countries (LMIC), which may limit translation of risk models based on demographics to other populations. In this work, we evaluate whether a modified AI-based risk model could perform with high sensitivity despite demographic differences in 3 LMICs.

Methods : Retinal fundus images were collected from 2351 subjects as part of an Indian ROP telemedicine screening program. A VSS was derived from the first exam after 30 weeks postmenstrual age. Using five-fold cross-validation, logistic regression models were trained and validated on 2 variables (gestational age at birth and VSS at first exam) for the eventual outcome of TR-ROP. Sensitivity, specificity, and respective 95% confidence intervals (CIs) were evaluated on test datasets acquired from Indian (n = 762 subjects, 27 TR-ROP), Nepalese (n = 330 subjects, 3 TR-ROP), and Mongolian babies (n = 319 subjects, 53 TR-ROP). All 3 test sets were consecutive population-based cohorts imaged as part of ROP telescreening programs.

Results : For the Indian, Nepalese, and Mongolian test datasets, the model had sensitivity [95% CI] equal to 100.0% [87.2%, 100.0%], 100.0% [29.2%, 100.0%], and 100.0% [93.3%, 100.0%], respectively. Specificity was 63.3% [59.7%, 66.8%], 77.1% [72.27%, 81.6%], and 45.8% [39.7%, 52.1%]. The mean ± standard deviation weeks subjects were identified prior to TR-ROP diagnosis was 2.1 ± 2.5, 0.7 ± 1.2, and 4.7 ± 5.1 weeks. Post-hoc analysis of the Indian dataset determined that, if those who screened negative were to never be screened again, the total number of examinations could be reduced by 45.0%.

Conclusions : We found 100% sensitivity and moderate specificity of a modified AI-based ROP risk model in 3 LMIC populations. There are two potential advantages to implementation of this risk model in ROP telescreening programs: (1) high-risk infants could be identified well before TR-ROP diagnosis, potentially reducing the risk of late treatment, and (2) the number of exams required to effectively screen a population could be significantly reduced.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.