# Pilot Evaluation of a Web Application for Amblyopia Risk Screening Integrating Parent-Reported Factors with AI-Assisted Strabismus Detection

## TL;DR

A web-based tool combining parent reports and AI detects children at high risk for amblyopia with high accuracy, offering a scalable solution for areas with limited eye care.

## Contribution

A novel hybrid screening tool integrating parent-reported factors and AI-assisted strabismus detection for amblyopia risk assessment in children.

## Key findings

- The tool achieved 100% positive predictive value in the high-risk category for amblyopia.
- Low-risk children were safely excluded with 100% negative predictive value.
- AI-assisted strabismus detection showed 96.9% confirmation in high-risk cases.

## Abstract

Amblyopia is the most common cause of visual impairment in children, and early detection is essential, yet screening remains limited in many settings, especially where access to eye-care specialists is scarce.

To evaluate the accuracy of a web-based screening tool that combines parent-reported risk factors with AI-assisted strabismus detection for identifying children at risk of amblyopia.

This pilot study included 105 children aged 3–10 years attending a public hospital in Morocco for their first ophthalmological evaluation. Parents completed an online screening tool consisting of eight validated amblyopia risk-factor questions and an automated strabismus analysis based on a frontal smartphone photograph. The AI module combined geometric measurements of pupil–nasal root symmetry with convolutional neural network (CNN) features such as corneal light reflex and gaze vector orientation. Each child received a total score (0–9), stratified into high-risk (6–9), moderate-risk (3–6), or low-risk (0–3) categories. A comprehensive ophthalmological examination, performed by a clinician blinded to the application results, served as the reference standard.

Of the 105 children screened, 32 were classified as high-risk, 62 as moderate-risk, and 11 as low-risk. The tool demonstrated perfect agreement in the high-risk category, with all 32 high-risk children clinically confirmed to have amblyopia (PPV = 100%). In the moderate-risk group, 30 of the 62 children were clinically confirmed (PPV = 48.4%). No child in the low-risk group had amblyopia (NPV = 100%). The AI-assisted strabismus module showed strong predictive accuracy in the high-risk category (96.9% confirmation). Statistical analyses showed no significant differences in diagnostic performance across age, gender, or urban/rural subgroups (p > 0.05).

The hybrid screening tool reliably identified children at high risk for amblyopia with complete concordance with a blinded clinical diagnosis, while safely excluding low-risk children. Although moderate-risk scores require cautious interpretation and clinical follow-up, this approach offers a low-cost, accessible, and scalable solution for paediatric vision screenings in resource-limited settings. Further large-scale community-based studies are warranted to validate generalisability.