Assessment of artificial intelligence to determine referral pathways in a real-world population of the English Diabetic Eye Screening Programme (DESP)

Abstract

Purpose : The Artificial Intelligence (AI) in Health and Care Award aims to benefit patients by combining the power of AI with the expertise of the National Health Service (NHS). National screening programmes for diabetic retinopathy (DR) are effective, but also labour and capital intensive. This study was supported by the NHS AI Award to investigate the clinical and cost effectiveness of an alternative approach to grading of fundus retinal images for DR and diabetic macular oedema (DME) within a retrospective and historical prospective population of the English DESP using automated software (Optos AI for DR developed with Google and Verily).

Methods : 100,296 retrospective and 7,953 historical prospective imaging sessions of patients who have attended DESP appointments underwent automated grading by Optos AI for DR. The referral outcomes between Optos AI for DR and the DESP final grading were compared and disagreements underwent adjudication by DESP senior graders to establish the ground truth for analyses.

Results : The overall sensitivity and specificity of Optos AI for DR for referral recommendation in the DESP pathway in the retrospective dataset was 82% and 85%, respectively. The sensitivity and specificity of Optos AI for DR on the DR/No DR outcome was 86% and 83%, respectively. Optos AI for DR’s agreement with the DESP referral recommendation for urgent referrals was 99%. Post adjudication, the overall sensitivity for referral recommendation increased to 88% and Optos AI for DR did not undercall or miss any sight-threatening DR/DME. An economic evaluation highlighted that DESP human grading would make an additional correct referral decision as an extra cost of £766 compared with Optos AI for DR. The historical prospective study successfully proved the potential for live implementation. Patient and public involvement data highlighted a general feeling of positive expectation on the implementation of automated grading within DESP.

Conclusions : The results of this retrospective study within a real-world database are encouraging, with consistent performance of the algorithm across different racial groups, showing no bias within the screened population. Based on the health economics analysis, the implementation of Optos AI for DR within DESP was found to be a cost saving, alleviating the burden on human graders.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.