Predicting Vision-Threatening Retinopathy in Patients With Diabetes Using Patient-Reported Data: Impact on Reading Center Workflow Efficiency

Purpose: First-come, first-served (FCFS) image review may delay identification of patients with vision-threatening diabetic retinopathy (VTDR). We developed a machine learning triage model using patient-reported clinical data to predict VTDR and evaluated its potential effect on reading-center workflow. Methods: We developed an ensemble model using XGBoost, CatBoost, and LightGBM in 9403 patients with diabetes imaged at the Alexandria iCare Retina Reading Center. Analyses were performed at the patient level using a worse-eye definition, with VTDR defined as severe nonproliferative DR or worse or diabetic macular edema. Predictors included demographic variables, comorbidities, smoking status, diabetes therapy, and hemoglobin A1c (HbA1c) when available. Model performance was evaluated with and without HbA1c. After model development, workflow simulations compared standard FCFS review with model-based prioritization and model plus prior-treatment prioritization. Results: The ensemble model achieved an area under the receiver operating characteristic curve of 0.89 with HbA1c and 0.87 without. With HbA1c, sensitivity was 74.8%, specificity was 88.6%, positive predictive value was 80.3%, and negative predictive value was 85.0%. In the 6-month workflow simulation, the mean time to evaluation for confirmed VTDR decreased from 57.08 ± 18.11 minutes with FCFS review to 48.86 ± 16.02 minutes with model-only prioritization and 43.01 ± 15.89 minutes with model plus prior-treatment prioritization. Conclusions: A patient-reported machine learning triage model identified patients at higher risk of VTDR and improved prioritization in simulated reading-center workflows. Translational Relevance: Patient-reported triage may help reading centers to prioritize high-risk patients while preserving human grading.