Artificial intelligence for cataract diagnosis and referral using real-world database

The use of artificial intelligence (AI) for ophthalmology applications has shown g promising results worldwide; however, its performance is dependent on population groups and must be evaluated in real-world scenarios. We evaluated the use of AI for cataract diagnosis and referral to specialists using a real-world database consisting of 2642 eye images from a working telemedicine service in South Brazil. Our AI solution adopts an ensemble model to improve classifier performance. The best results showed an accuracy of 90.6% for cataract diagnosis with a corresponding area under the receiver operating characteristic curve (ROC AUC) of 96.7%. The accuracy for surgical referral was 86.5% with a corresponding ROC AUC of 94.3%. The results indicate that the use of an ensemble of models and training with a heterogeneous real-world clinical database enabled our solution to achieve superior performance compared to other works in the literature when evaluated on real-world data.