Predicting diabetic retinopathy and identifying interpretable biomedical features using machine learning algorithms.

Area of Science:

• Ophthalmology
• Endocrinology
• Data Science

Background:

• Diabetic retinopathy (DR) risk factors require further clarification for targeted prevention.
• Accurate identification of DR risk factors can enable early intervention in high-risk diabetic populations.

Purpose of the Study:

• To develop a predictive model for DR in type 2 diabetes mellitus (T2DM).
• To compare the efficacy of various data mining techniques for DR prediction.

Main Methods:

• Utilized machine learning algorithms: support vector machines (SVM), decision trees, artificial neural networks, and logistic regressions.
• Employed data mining techniques to analyze clinical features associated with DR.
• Evaluated model performance using percentage split and three-way data split schemes.

Main Results:

• Support vector machines demonstrated superior prediction performance, achieving 79.5% accuracy and 0.839 AUC.
• A three-way data split (60% training, 20% validation, 20% testing) provided a more realistic performance evaluation.
• The developed model outperformed previous studies in most evaluation metrics.

Conclusions:

• Insulin use and diabetes duration were identified as significant, interpretable predictors of DR.
• Increased diabetes duration by one year elevated DR odds by 9.3%.
• Insulin use increased DR odds by 3.561 times compared to non-users, supporting clinical decision support system development.