Vision threatening disease triage using teleophthalmology during COVID-19 in the emergency department: A pilot study

ABSTRACT

Purpose: The Centers for Disease Control reports 28.2% of surveyed US adults had reduced access to medical care (June/August 2020) due to the COVID-19 pandemic, with 8.9% reporting reduced access to vision care. A non-mydriatic digital retinal camera was piloted for deployment to the Emergency Department (ED) to help address this gap in vision care. Referrals for clinical follow-up in vision threatening diseases (VTDs) such as age-related macular degeneration, cataracts, diabetic retinopathy (DR), and glaucoma were assessed with human readers. Artificial Intelligence (AI) deep learning software was evaluated in known DR cases. Methods: 33 patients with known VTDs (48.48% male, avg 59.33 years) and 36 control subjects (41.67% male, avg 31.33 years) were included in tele-ophthalmology screening. A Canon CR-2 Plus AF non-mydriatic retinal camera captured 45-degree angle color and auto-fluorescence images of the eyes. Images (136 eyes) were graded by a certified telemedicine reader on site and an off-site clinical ophthalmologist following International Clinical Diabetic Retinopathy Disease Severity Scale (ICDRSS). Intergrader agreement between readers was evaluated with Cohen's kappa. An automated deep learning screening software optimized for DR (SELENA+, EyRIS Pte Ltd, Singapore) performed independent validation of readable color fundus images (17 eyes). Results: 5.07% of images were deemed unreadable by graders due to poor quality. Intergrader agreement for subject referral was κ = 0.710 (95% CI 0.545-0.875, p<.0005), with the clinical ophthalmologist generating more referrals than the telemedicine reader. Readers had 96.97% sensitivity (95% CI 91.12-1.028) and 72.22% specificity (95% CI 57.59- 86.85) in detecting referable disease. Positive predictive value was 76.19% (CI 63.31%- 89.07%) and negative predictive value was 96.30% (CI 89.17%- 1.034%). Of the 10 false positives, 6 were referred for rule out of glaucoma. Four had early stage cataracts that were deemed nonurgent. SELENA+ referred 100% of the known 9 DR patients. Conclusions: Tele-ophthalmology deployment in the ED helps limit patient and staff exposure to SARS-CoV-2 without sacrificing evaluation for VTDs. Tele-ophthalmology readers err on the side of caution to avoid missing VTD in a given patient. Use of AI can help keep strict adherence to referral guidelines.