Predictive Models for Diabetic Retinopathy from Non-Image Teleretinal Screening Data.

Area of Science:

• Ophthalmology
• Medical Informatics
• Public Health

Background:

• Diabetic retinopathy (DR) detection is challenging in underserved US areas due to limited access to eye specialists.
• Teleretinal screening programs aim to bridge this gap in care.
• Early detection of DR is crucial for preventing vision loss.

Purpose of the Study:

• To evaluate machine learning (ML) methods for predicting diabetic retinopathy.
• To assess the feasibility of using routinely collected clinical data for DR prediction.
• To identify high-risk diabetic patients in medically underserved settings.

Main Methods:

• Utilized data from 27,116 diabetic patients in a Los Angeles County teleretinal screening program.
• Included variables such as ethnicity, gender, age, HbA1c, insulin dependence, and diabetes duration.
• Compared six ML classifiers on a dataset with class imbalance.

Main Results:

• Six ML classifiers demonstrated predictive capability for diabetic retinopathy.
• The best-performing model achieved an Area Under the Curve (AUC) of 0.754.
• This model had a sensitivity of 58% and a specificity of 80%.

Conclusions:

• Predicting diabetic retinopathy from clinical data is feasible and can support underserved populations.
• This ML approach can help identify patients needing further eye examinations.
• The study represents a step towards improving DR screening accessibility.