Deploying Artificial Intelligence For Assessing Burden Of Corneal Blindness And Population Need For Donor Cornea

To create an Artificial Intelligence (AI) based smartphone platform that can detect corneal opacities from smartphone images and identify patients requiring corneal transplantation, and to validate it in a community setting. This would help in filling the lacunae in countries with developing economies where Corneal Blindness is a major public health concern and there is limited community-based data on its burden and the need for keratoplasty at the population level.

This was a cross-sectional observational study to create an AI based application followed by a population based cross sectional survey to validate it. 2000 images of the cornea (1000 corneal opacities and 1000 normal eyes) were used for the training dataset. 4954 participants were enumerated through the population-based survey for validation of the AI platform.

2000 smartphone clicked images of cornea were pre-processed and used to train an AI model using 80% of the dataset, whereas 20% was used to internally validate the model. Once trained, the model was integrated into Android application along with a questionnaire and tested in a population-based survey to evaluate the applicability of the platform. The sensitivity, specificity, positive predictive value (PPV) negative predictive value (NPV) and accuracy were calculated in the internal validation of the model. On field testing in the real world, the accuracy of the AI model in identifying corneal opacities and predicting the need for keratoplasty was calculated and Cohen’s Kappa was determined for agreement with the ophthalmologist diagnosis.

The trained model achieved an accuracy of 92.25% with sensitivity, specificity, PPV and NPV above 90%. In the field, the model demonstrated an accuracy of 94.59% in identifying corneal opacities, with a sensitivity and PPV of 72.4% (95% CI 52.8% to 87.3%) and specificity and NPV of 97% (95% CI 94.2% to 98.7%). Cohen’s Kappa was determined to be 0.6942 which indicated a significant agreement of the AI model diagnosis with that of the ophthalmologist. The need for keratoplasty was determined to be 0.07%, 7 donor corneas per 10000 population, with a 100% correlation between the model and the ophthalmologist. 

Smartphone image-based AI systems demonstrated have a high diagnostic accuracy to estimate the burden of corneal opacities and determining the unmet need for donor corneas. It can be used by healthcare workers at the community level with ease. Such widely accessible and scalable Smartphone image-based AI systems can be used to improve screening accessibility and connect marginalized groups to sight-restoring interventions.