Multicenter study to validate an artificial intelligence algorithm for the screening of diabetic retinopathy: the CARDS study

Abstract

Purpose : Diabetic retinopathy (DR) is the main cause of visual impairment in middle-aged and elderly people. Early diagnosis is the best strategy to prevent or delay vision loss. The purpose of this study was to validate a deep learning-based artificial intelligence (AI) algorithm (LuxIA) for the screening of more-than-mild DR (mtmDR).

Methods : This is a multicenter, cross-sectional, observational study. We collected images of adult diabetic patients (≥18 years old) with or without DR in Spanish routine clinical practice from five University Hospitals. One eye was selected from each patient. One fundus photograph per study eye was captured using a commercially available, CE-certified, non-mydriatic ocular fundus camera (NW200, NW400 or equivalent Topcon cameras). Input data collection and filtering were done through the Discovery® platform (RetinAI) and electronic Consent Report Forms (eCRF) were specifically designed for the study. A retina specialist verified that the image quality satisfied the requirements for the LuxIA use. A committee of experts (3 retina specialists experienced in DR diagnosis and management) independently evaluated the images through the Discovery® platform, according to the simplified international classification guidelines. Experts were blinded to the algorithm output for a given image. Findings not consistent with DR were reported. Primary endpoint: LuxIA sensitivity and specificity for detection of mtmDR in the primary care setting (referrable patients).

Results : Eight hundred and twenty-nine evaluable images were included. LuxIA algorithm detected mtmDR with a sensitivity of 0.9714 and a specificity of 0.9478. The area under the receiver-operating characteristic (ROC) curve (accuracy measure of the test) value was 0.9596 (Figure).

Conclusions : AI is increasing the testing feasibility for medical professionals in DR screening. This study validates the use of LuxIA algorithm for mtmDR screening in a real-world primary care setting.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

ROC curve