Pilot: A validation study of a novel, fully-automated, disease grading and clinical decision support system for screening and telemedicine-based diagnosis of the sight threatening condition of diabetic retinopathy (DR)

Abstract

Purpose : Regular screening and early detection of diabetic retinopathy (DR) in patients with diabetes are paramount in reducing preventable blindness and impaired vision associated with DR. Due to limited ophthalmology resources at public hospitals in Saudi Arabia, screening all patients with diabetes becomes costly and inefficient, with increasing inappropriate referrals and a prolonged waiting list for ophthalmology consultations. The recent emergence of artificial intelligence (AI) – based, deep learning algorithms enhance the opportunity for early detection of DR from retinal images without the need for a specialist opinion and refer only those requiring treatment

Methods : A total of 101 patients with diabetic retinopathy at King khaled Eye Specialized Hospital (KKESH) were enrolled in this pilot study. Adult patients with diabetes were included. Pregnant patients and Eyes with media opacities were excluded. Retinal images collected from patients stored in the medical images server located at KKESH were assessed and graded for DR by two sets of ophthalmologists to create a 'Gold Standard'. Same images were run through AI-based DR detection system (TeleMedC DR grader) and results compared with the Gold standard. The degree of agreement between automated analysis and manual grading was quantified and assessed using multiple statistical analysis tools for sensitivity, specificity, and DR grading accuracy. For all statistical tests, p-value <0.05 were considered significant.

Results : In this pilot, 101 images were obtained, 8 of which were excluded for poor quality for AI based grading and removed from calculations by the AI based quality control system
Compared with the Gold Standard, Sensitivity for detecting DR was 97.7% . All 3 images with no disease have been graded correctly by AI (TeleMedC DR grader)
AI grading Accuracy was found to be 97.85% compared to the Gold standard.

Conclusions : AI-based DR detection systems are becoming a powerful tool for grading and clinical decision support system for screening and telemedicine-based diagnosis and referral for DR. The Validation process is being continued with a larger sample size, and when completed, the accuracy, and time and cost saving benefits will lead to adoption by national DR screening programs in large populations

This is a 2020 ARVO Annual Meeting abstract.