EyeWorld Asia-Pacific September 2020 Issue

EWAP SEPTEMBER 2020 21 SECONDARY FEATURE real-world deployment of AI algorithms exceeding clinically acceptable performance. 4 Some recent studies have further demonstrated cost- effectiveness of using AI as a complementary approach of semi-automated DR screening wherein both human graders and AI algorithms work hand- in-hand. 5 Other AI applications include predicting individual risk of DR or glaucoma, as well as predicting treatment response to intravitreal anti-vascular endothelial growth factor (anti- VEGF) therapy in age-related macular degeneration and diabetic macular edema. 6 New digital ophthalmology models of care During COVID-19, the priorities are to reducing non-urgent and non-essential ophthalmology consultations, reduce patient time in clinic and consultation, and to improve community and home monitoring of eye diseases. Safety for ophthalmologists are important as they are at increased risk of contracting COVID-19 from asymptomatic carriers due to high outpatient clinical loads and the close proximity with patients required for eye examinations. 7 So what are possible new digital models of care for COVID-19? Digital ophthalmology prior to eye consultations First, digital technology can be applied before the patient seeks ophthalmology care at the eye clinic or eye hospital (“out of hospital” models). One such model involves the use of digital technology to triage patients or provide tele-support recommendations for self- management or to seek detailed reviews with ophthalmologists where required. Various AI- based conversational chatbots have been applied in pilots, although none has completed clinical validation. 8,9 This model has also been applied for AI systems providing automated classification of DR and other eye diseases in community screening. 10 This “internet hospital” model by Zhongshan Ophthalmic Center in China was used to scale-up responses during COVID-19. 11 An alternative is the “hub and spoke” model in which a tertiary eye center (“hub”) provides clinical service provided in geographically remote settings (“spokes”); this can be through patient-general ophthalmologist or general ophthalmologist- sub-specialists (e.g., retinal specialists) based in the “hub.” 1 This model can be synchronous (“real time”), or can be asynchronous (“delayed”) with a later follow-up consultations after patients complete tests such as fundus imaging in the “spokes.” 12 Asynchronous consultations lower operating costs as ophthalmologists do not need to “stand by” waiting for virtual consultations. These models have been trialled in COVID-19. For example, in Hong Kong, a single dedicated staff performed all assessments and relayed information to an ophthalmologist remotely for triage and advice for all patients deemed to be of high risk. 13 The benefits of these digital models during the COVID-19 pandemic has enhanced right-siting of patients and optimized patient flow to eye clinics and hospitals. Digital ophthalmology in eye clinics Secondly, digital models can be used in eye clinics during a patient’s visit. Various front-line settings present opportunities for the adoption of digital technology. The aim is to reduce the physical time in the clinic. One model is to have assessment of vision and other tests (e.g., imaging, visual fields) in the community before a visit to the eye clinic. This can be augmented with end-to-end digital service with scheduled asynchronous consultations plus synchronous and/or on-demand consultations as needed. One such example in Singapore is a follow-up video consultation service for patients with stable glaucoma to reduce hospital encounters. Patients receive investigations in the community with results sent for asynchronous review by specialists in video consultation clinics. 14 Patients with no change in management plans are notified by email, whereas those that require a change receive a synchronous video consultation. Digital ophthalmology at home Finally, digital health solutions for continuous monitoring can be used for patients at home. Patient-reported solutions can be used to directly provide critical insights to patients, such as the “AllEye” self-monitoring application that characterizes metamorphopsia for early detection of wet age-related macular degeneration. 15 Home vision function tests like these can play a critical role for patients with missed follow- up appointments due to the pandemic, 7 to facilitate early detection of progression or complications that may require urgent treatment. Challenges limiting the effectiveness of digital health Despite the promise, there are significant challenges to widespread adoption and effective deployment of digital technology in ophthalmology, even during a crisis as significant as COVID-19. The adoption of digital health has been hampered by systemic barriers to sustained adoption, a short term view of the technology, the potential impact on supporting clinical and healthcare systems, and a lack of pragmatic validation for operational models. While COVID-19 aligns the priorities of many stakeholders to facilitate adoption of digital health, the same factors that have hampered adoption will continue to dampen the effectiveness of digital health solutions in the ongoing pandemic, and potentially undo any progress in adoption once operational demands revert to “normal” after the crisis. How do we overcome such barriers to adoption? A system level approach is needed to identify barriers, and then design potential ways to address them, including monitoring data and measures for adoption. For example, in testing AI algorithms for DR screening, decision trees that impact performance (e.g., sensitivity, specificity) may need to be configured for different countries based on cultural factors and availability of resources. This requires consideration of cost structure, acceptance of false negative results with screening, as well as capacity to follow-up on referred cases (e.g., availability of ophthalmologists) considering the false positive rate. 5,16 Finally, algorithmic accuracy in real-world application could potentially be enhanced through calibration using datasets from target deployment settings, using techniques such as transfer learning. 16 Even then, programs to drive digital adoption may meet unanticipated challenges if the needs of different relevant stakeholders are not considered holistically. 17 For example, patients should have sufficient digital literacy and should understand information and implications of such information and be able to use it safely. Video consultations is an example whereby global adoption has been relatively slow as its requirements for operationalisation have been grossly under-estimated. Many ophthalmologists find it difficult coping with interruptions to existing packed clinics with ad hoc on-demand teleconsultation requests from remote patients due to IT and other technical issues. 18 Without established work flows and priorities, the knock-on effects of these interruptions can lead to delays that overwhelm