Predicting Diabetic Retinopathy Using Machine Learning

Abstract

Over the last 3 decades, advances in laser surgery and intraocular drug delivery have decreased the risk of diabetic retinopathy. However, the use of artificial intelligence and other forms of data analysis in retinopathy screening has not had the same advances. This project aimed to analyze the National Health and Nutrition Examination Survey and to develop the best machine learning model to predict the disease. Violin plots were created to compare the distribution of diabetic retinopathy diagnosis with gender. The plot showed that females who had said they were not taking insulin were less likely to be diagnosed with diabetic retinopathy. Another violin plot showed that those with hypertension who were taking insulin were less likely to have diabetic retinopathy. Histograms were also created to show the distributions of the variables, color-coded by retinopathy diagnosis. Specifically, it was found that the bigger the time gaps since the diabetes diagnosis, the more likely a person suffers from diabetic retinopathy. Finally, multiple machine learning models were tested and these were the most accurate in predicting diabetic retinopathy with an 80% accuracy, LinearSVC, CalibratedClassifierCV, and Logistic Regression.