SCH: Harnessing Tensor Information to Improve EHR Data Quality for Accurate Data-driven Screening of Diabetic Retinopathy with Routine Lab Results - Project Summary This application is being submitted in response to the Notice of Special Interest (NOSI): NOT-RM-24-013. Despite the high prevalence of diabetic retinopathy (DR), the recommended annual ophthalmic exam for diabetic patients has an alarming low compliance rate, at only 43%. Many patients do not seek proper medical attention because DR is asymptomatic in the early stages, thus missing the most effective period to halt DR progression and prevent vision loss. Moreover, ophthalmic equipment is predominantly limited to urban areas, restricting access by rural patients. About 25% of DR patients and 19% with sight-threatening DR remain undiagnosed. Efforts to improve the adherence rate by bringing fundus-image-based DR screening tools to primary health settings have been hindered by the relatively high cost of retinal cameras. Moreover, retinal imaging is technically challenging, potentially hindering its adoption in resource-limited settings, like rural clinics. All of these issues create an urgent need for cost-effective, widely-available tools that enable early detection of DR. Our team has developed machine-learning (ML) models for non-image-based DR screening using existing comorbidity data and routine lab results collected from diabetic patients’ semiannual primary care visits. It helps physicians identify at-risk patients and recommend the ophthalmic exam confidently. Patients will pay more attention when they know that the recommendations are personalized and data-driven. It can also improve the compliance rate by providing DR screening for rural and underserved populations without extra cost or tests. Our approach differs from the existing AI-based DR detection tools (e.g., IDx-DR and Google’s Verily) that require fundus images taken by expensive specialized high-resolution retinal cameras. Our model achieved 92.48% of AUC (area under curve), comparable to the fundus-image-based AI tools for DR screening approved by FDA, e.g., IDx-DR (AUC=95%). Our approach is promising to break the barrier to ubiquitous diabetic eye care in rural communities and increase the compliance rate of the ophthalmic exams among asymptotic patients. Early identification of the asymptomatic patients can help them to seek timely treatment prior to vision loss, and thus save thousands of people from blindness. To improve the explainability of the model, we developed an easy-to-use risk index for calculating the risk of developing DR for diabetic patients. We would like to collaborate with an independent contract research organization (CRO) to replicate our model and conduct training and validation with an independent EHR dataset. The CRO must have access to high- quality, diverse EHR data, along with the necessary technical infrastructure, software tools, and expertise in data science, clinical medicine, and regulatory compliance. The PI will provide the original model, source codes, along with associated documents that describe the experimental setup, hyperparameters, and data processing steps.
