Diabetic retinopathy risk prediction for fundus examination using sparse learning: a cross-sectional study

Area of Science:

• Ophthalmology
• Data Science
• Public Health

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness in diabetic patients.
• Early detection and management of DR are crucial for preventing vision loss.
• Diabetic patients exhibit low adherence to routine ophthalmologic examinations.

Purpose of the Study:

• To develop and validate sparse learning models for identifying diabetic retinopathy (DR) risk.
• To assess the efficacy of machine learning in predicting DR using electronic health records.

Main Methods:

• Utilized health records from Korea National Health and Nutrition Examination Surveys (KNHANES) V-1 and V-2.
• Constructed and validated prediction models for DR using ridge, elastic net, and LASSO algorithms.
• Compared the performance of sparse learning models against traditional DR indicators.

Main Results:

• The LASSO model demonstrated the highest efficiency in predicting DR risk.
• LASSO achieved an Area Under the Curve (AUC) of 0.81 (internal validation) and 0.82 (external validation).
• LASSO model accuracy was 73.6% (internal) and 75.2% (external), significantly outperforming traditional indicators.

Conclusions:

• Sparse learning models, particularly LASSO, are effective for analyzing DR epidemiological patterns.
• This study introduces the first machine learning model for DR risk prediction using health records.
• LASSO is a valuable tool for analyzing high-dimensional electronic health records, balancing discriminative power and variable selection.