Usefulness of Machine Learning for Identification of Referable Diabetic Retinopathy in a Large-Scale Population-Based...

Area of Science:

• Ophthalmology
• Medical Informatics
• Public Health

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection and management are crucial for preventing blindness.
• Non-ocular metrics offer a potential screening avenue where ophthalmic infrastructure is limited.

Purpose of the Study:

• To develop and validate machine learning (ML) classifiers for detecting referable diabetic retinopathy (RDR).
• To utilize simple, non-ocular metrics for RDR detection in a large Chinese population.
• To identify key non-ocular predictors for RDR occurrence.

Main Methods:

• Eight ML algorithms (XGBoost, Random Forest, etc.) were trained and validated on data from 1,418 diabetic patients.
• Models were developed using 10 key non-ocular variables: diabetes duration, HbA1c, blood pressure, etc.
• Performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC) with cross-validation.

Main Results:

• The XGBoost model demonstrated the highest discriminative performance with an AUC of 0.816.
• Key predictors included diabetes duration, HbA1c, systolic blood pressure, and triglyceride levels.
• Other significant variables were BMI, serum creatinine, age, education, hypertension duration, and income.

Conclusions:

• ML classifiers using easily obtainable non-ocular variables can effectively detect RDR.
• Variable importance scores offer insights into RDR prevention strategies.
• ML-based RDR screening is a valuable complementary tool for resource-limited settings.