(061) Predictive Model for Sight-Threatening Diabetic Retinopathy in People with Type 2 Diabetes Mellitus

Background: Diabetic retinopathy (DR) is a common microvascular complication of diabetes, which may lead to significant visual impairment and potential blindness. The timely identification of people at risk of rapid progression to sight-threatening diabetic retinopathy (STDR) is crucial for prevention, early detection, and treatment to maintain vision.

Objectives: To develop a machine learning model that predicts the 3-year risk of STDR in people with type 2 diabetes (T2DM).

Methods: The study included individuals with T2DM who had undergone at least one fundoscopic examination, as recorded in the electronic health records of Seoul National University Hospital, from January 2005 to December 2017. The development of STDR was monitored up to three years following their first fundoscopic examination. STDR was defined as the presence of proliferative diabetic retinopathy (PDR), macular edema (ME), or the initiation of treatment for diabetic retinopathy, including laser photocoagulation, anti-vascular endothelial growth factor injection, and vitrectomy. Variables considered in the predictive model were collected within two years prior to the first fundoscopic examination and included age, sex, biochemical test, history of early stage DR, comorbidities, medications, and healthcare utilization. The dataset was randomly divided into training and test sets in an 80:20 ratio. The XGBoost classifier was employed for the model, with hyperparameters optimized through a grid search by cross-validation. The performance of the model was assessed using the Area Under the Receiver Operating Characteristic curve (AUROC) and the Area Under the Precision-Recall Curve (AUPRC). In addition, Shapley Additive exPlanations (SHAP) values were calculated to identify the most important predictors.

Results: The study cohort consisted of 6,906 people with T2DM, of whom 269 (3.90%) developed STDR over a three-year period. After data preprocessing, 164 variables were included in the predictive model. The performance of the predictive model on the test set yielded an AUROC of 0.851 and an AUPRC of 0.214. SHAP values revealed that the history of early stage DR , HbA1c levels, and insulin therapy were the three most significant predictive factors. The model predicted a higher probability of developing STDR onset within three years for patients with existing DR, elevated HbA1c levels, and those receiving insulin treatment.

Conclusions: We formulated a predictive model with strong performance to assess the risk of STDR over a three-year period in people with T2DM. Additionally, we identified key predictors to enhance interpretability. Further investigations are required, particularly regarding causal inference and generalizability.