Development and Validation of a Machine Learning Algorithm for Predicting Diabetes Retinopathy in Patients With Type...

Area of Science:

• Ophthalmology
• Medical Informatics
• Data Science

Background:

• Diabetic retinopathy (DR) is a leading cause of preventable blindness globally.
• Machine learning (ML) offers potential for enhancing DR screening in community settings.
• Predictive model performance for ML in DR screening requires further determination.

Purpose of the Study:

• To assess ML-based risk prediction for DR development in adults with type 2 diabetes mellitus (T2DM).
• To develop a universally applicable and accurate DR risk prediction model using South Korean healthcare data.

Main Methods:

• Utilized electronic medical records from 3 South Korean university hospitals (discovery cohort: n=14,694; validation cohort: n=1856).
• Developed and tuned various ML models, selecting the best performing one based on the area under the receiver operating characteristic (ROC) curve.
• Primary outcome: presence of DR at 3 years.

Main Results:

• The extreme gradient boosting (XGBoost) model achieved 75.13% accuracy in the discovery cohort and 65.14% in the validation cohort.
• Key predictors identified by XGBoost include dyslipidemia, cancer, hypertension, chronic kidney disease, neuropathy, and cardiovascular disease.
• DR was diagnosed in 2.37% of the screened patients.

Conclusions:

• ML-based risk prediction, particularly using XGBoost, shows potential for timely DR intervention in T2DM patients.
• The model can improve understanding of DR contributing factors and reduce complications.
• The proposed model is anticipated to be cost-effective for primary care settings in South Korea.