A Deep Learning Framework for Diagnosing Diabetic Retinopathy and Ranking its Risk Factors

The biggest cause of preventable blindness worldwide is diabetic retinopathy (DR). Individual risk of DR development vary, making risk prediction and screening intervals problematic. We used various machine learning models to rate risk factor relevance on the dataset. A t-test and Shapely-additive-explanations(SHAP) rank risk factors. It then employs various ML-strategies to identify insignificant risk factors based on the AUC criterion to forecast DR. Risk factors’ significance is determined by a weighted average of classifier significance. They rank risk variables for machine learning models. The highest AUC risk factors are used to build a DR prediction model. After ranking by t-test, all five ML-techniques included glycosylated-hemoglobin and systolic-blood-pressure in the top three DR risk variables. The ML-strategies using SHAP ranked-systolic BP and history of hypertension in the five critical risk attributes for DR. Finally, SBP and diabetes mellitus duration were in the top four risk attributes for DR when an ensemble of five ML-strategies was used separately with the t-test and SHAP. DT and KNN have the highest AUCs (t-test 0.79, SHAP 0.77). KNN anticipated DR with 82.6% (t-test) and 78.3% (SHAP) accuracy.