Abstract
Purpose :
To identify children at risk of developing high myopia for timely assessment and intervention, in order to prevent myopia progression and development of complications in adulthood.
Methods :
Using school-based cohort in Singapore, comprising of 998 children, aged 6-12 years old, we trained and performed primary validation of a deep learning system using 7456 baseline fundus images of 1878 eyes and performed external validation using an independent test dataset comprising of 821 baseline fundus images of 189 eyes; with clinical data (age, gender, race, parental myopia and baseline spherical equivalent (SE)). Three distinct algorithms: image, clinical and mixed (image + clinical) models, comprising of fundus photo, age, race, gender and SE, were derived to predict the development of high myopia (SE ≤ -6.00 dioptre) during teenage years (5 years later, aged 11-17). Model performance was evaluated based on area under the receiver operating curve (AUC).
Results :
All baseline models achieved clinically acceptable performance, with image models (AUC 0.93-0.95 in primary dataset; 0.91-0.93 in test dataset), clinical models (AUC 0.90-0.97 in primary dataset; 0.93-0.94 in test dataset) and mixed (image + clinical) models (AUC 0.97 in primary dataset; 0.97-0.98 in test dataset).The addition of 1 year SE progression achieved a marginal improvement or decline in performance for the clinical model (AUC 0.98 versus 0.97 in primary dataset; 0.97 versus 0.94 in test dataset) and mixed model (AUC 0.99 versus 0.97 in primary dataset; 0.95 versus 0.98 in test dataset).
Conclusions :
Our deep learning system was able to predict the development of high myopia in school-going children by teenage years, which has the potential to be utilized as a clinical-decision support tool to identify ‘at risk’ children for early intervention.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.