While eye examination is routinely performed in preterm neonates below 34 weeks gestational age, it is not commonplace in full-term neonates. For full-term and stable late preterm neonates, various guidelines recommend universal eye screening using a red reflex examination. However, the red reflex examination is not sensitive enough to detect all potentially sight-threatening disorders in newborns. Several pilot studies of universal newborn screening with wide-field retinal imaging (WFRI) were conducted in the last decade to fill this gap. Retinal haemorrhages were the most common abnormalities detected, most of which were benign. In addition, several abnormalities requiring medical and/or surgical attention were detected, including congenital cataracts, congenital glaucoma, vitreous haemorrhages, cytomegalovirus retinitis, retinoblastoma, posterior uveitis, etc. Data from Indian studies found that about 1% of neonates had ocular conditions requiring medical/surgical attention. If extrapolated to annual births in India (about 25 million), about 2.5 lakh neonates born annually could have some abnormality requiring evaluation or therapy. These findings highlight that a universal newborn eye screening program can identify several neonates with sight-threatening conditions. However, screening such many babies requires many ophthalmologists and may not be feasible. Hence, alternative approaches must be considered, driven by the paediatrician/neonatologist, nurse, optometrist, or paramedical personnel. In this pilot study, we plan to assess the feasibility and outcomes of an optometrist-driven model for universal newborn eye screening. The gold standard for retinal imaging, globally, is the RetCam. However, a low-cost alternative was evaluated by Vinekar et al. It was found to have good accuracy in diagnosing ROP in preterm neonates. The cost is almost one-fifth of the RetCam III, so it is more likely to be financially viable in India. We plan to use the indigenous 3NetraNeo for eye screening in this pilot study.

Objectives:

1)    To assess the feasibility of optometrist-driven universal newborn eye screening in a tertiary care setting.

2)    To develop and validate an artificial intelligence-based model for newborn eye screening

Outcome measures:

a. Prevalence of newborn eye conditions requiring medical/ surgical intervention in healthy late preterm and term neonates

b. Diagnostic accuracy of AI model in detecting retinal image abnormalities in healthy late preterm and term neonates

Statistics:

Objective 1: Descriptive statistics will be used to describe the population characteristics (maternal and neonatal). We will calculate the prevalence of ocular disorders requiring intervention (evaluation or therapy). SPSS software will be used for analysis. The normally distributed data will be expressed as mean ± standard deviation, and the t-test will be used for pairwise comparison between groups. Skewed distribution data will be expressed as median (interquartile range), and pairwise comparison between groups will be conducted using the Mann-Whitney U test. A P-value of 0.05 will be considered significant in all analyses.

Objective 2: The proposed system is a computational system leveraging AI models to automatically learn disease-causing features from the diagnostic data. Hence, statistical studies do not apply to the initial phase of the study (model development phase). Standard evaluation metrics such as F1 score, sensitivity, specificity, area under the receiver operating characteristic (AUROC), etc., will be used to evaluate the automated system’s performance for validating system performance regarding expert-annotated and verified cases. The retinal images, as classified by the ophthalmologist, will be considered the gold standard for classification. True positive (TP)- abnormal classification by both ophthalmologist and AI-model; True negative (TN)- normal classification by both ophthalmologist and AI-model; False positive (FP)- normal classification by ophthalmologist, abnormal classification by AI-model; False negative (FN)- abnormal by ophthalmologist and normal by AI-model. Sensitivity will be calculated as TP/(TP + FN), specificity as TN/(TN + FP), Positive predictive value (PPV) as TP/(TP + FP), and Negative predictive value (NPV) as TN/(TN + FN). The ROC algorithm of SPSS software will be used to calculate AUROC. 

Detailed description of procedure/processes:

Objective 1: In the project’s initial phase, an optometrist will be trained to use 3NetraNeo (1300 lens) and acquire images from newborn babies. All full-term neonates and healthy late preterm neonates admitted to the neonatal unit during the study period will be considered eligible for inclusion. We will include both intramural and extramural births. The examinations will be conducted within 72 hours of birth or before discharge from the hospital. For sick neonates admitted to neonatal intensive care, the examination will be performed after the neonate is stabilized and shifted to step down unit. After parental consent, pupillary dilatation will be achieved using a combination of cyclopentolate 0.5% and 0.5% phenylephrine. Topical anaesthetic proparacaine 0.5% will be instilled into both eyes. A sterile infant wire speculum will be used to separate the conjunctiva. An inert jelly will be applied to the cornea, followed by the placement of the camera. An anterior segment image will be obtained initially. Five-directional fundus images will be obtained from the posterior pole, optic nerve centred, optic nerve superior, optic nerve inferior, and optic nerve nasal views. The photos will be reviewed twice weekly by a paediatric ophthalmologist. The various conditions that will be looked for include anterior segment abnormalities like cataracts, vitreal abnormalities like vitreal haemorrhages, and retinal abnormalities like haemorrhages, uveitis, vascular ridges, colobomas, scars, vasculitis, retinoblastoma, retinal dysplasia, etc. The neonates with abnormal findings will be referred to the Ophthalmologist for further evaluation and management. The neonatal unit will conduct the review if medical evaluation or management is required for disorders like suspected intrauterine infection. After discharge, all neonates will be telephonically followed by optometrists for seven days and assessed for the development of conjunctivitis.

Objective 2: The images obtained during the screening program will be categorised as normal, abnormal, and non-gradable. The annotated dataset containing all three classes (normal, abnormal and non-gradable) is aggregated. The initial images will be used to develop the artificial intelligence model among the aggregated datasets. As per the experience of the technology team, at least 400 images of each class are required for model development. The remaining dataset will be used for testing/ validation. About 20% of neonates have some abnormality on fundus examination (based on data from previous studies). About 10% of images may be non-gradable, and the remaining 70% will likely be normal. Assuming that AI software requires at least 400 images of each class for saturation, we need 400 neonates (800 eyes) with non-gradable photos (as this class has the lowest percentage). This will be attained after screening 4,000 neonates (8,000 eyes). This initial period will help develop the AI model. This phase will be performed by three experts in artificial intelligence with experience in retinal image analysis. In the subsequent phase (second half of the study- the validation phase), the images obtained by the optometrist will be seen independently by the neonatologist (PI) and the pediatric ophthalmologist. The neonatologist will use the AI model to categorise the images into three classes. Simultaneously, all the photos will be seen by the paediatric ophthalmologist, who remains blind to the AI-model findings. The clinical management of the neonates will continue as per the ophthalmologist’s review. The sensitivity, specificity, positive, and negative predictive values of the AI-model findings will be estimated.