Diabetes is a long-term illness that poses a significant challenge to global healthcare. It is associated with various complications that impact different organs and systems in the body. The American Diabetes Association (ADA) has identified Diabetic Retinopathy (DR) as a significant complication of diabetes, which can result in damage to the retina. In addition to the rapid growth of eye healthcare demands, the existing screening tools and procedures are associated with high costs, prolonged waiting times, and a need for professional healthcare resources. Therefore, this study aims to develop a machine learning predictive model that can utilize routine lab results, enabling more accessible access to DR screening and reducing the burden on specialized ophthalmic services. It provides medical doctors with simple rules to help them accelerate their decision-making. This study demonstrated the RuleFit model to predict diabetic retinopathy using an Electronic Health Records (EHR) dataset. The methodology of this study focused on training a RuleFit model using a dataset obtained by Cerner and then compared it with other models: DT, K-NN, RF, and LR, based on the performance measures of each model, including accuracy, Precision, F1-Score, Recall, and the area under the Receiver operating characteristic curve (AUC-ROC). Then the trained models were validated on an unseen dataset obtained from KUMC. We pruned and combined the rules out of the RuleFit model and got four simple combined rules. The combined rules were then validated on Internal and external datasets.

The RuleFit model exhibited outstanding performance with an AUC of 0.99, a recall of 0.579, and an impressive F1-score of 0.719. This highlights the ability of the RuleFit model to accurately classify and predict diabetic retinopathy using routine lab results. The combined rules results showed an accuracy range between 69% - 96%. This underscores the effectiveness and reliability of the combined rules in accurately classifying DR samples on both datasets and emphasizes their potential for improving diabetic retinopathy screening and diagnosis. In conclusion, the RuleFit model, along with its integrated combined rules, shows great potential in improving healthcare outcomes and tackling the challenges faced by specialized ophthalmic services.