Identifying Patients at High Risk of Lapses in Diabetic Retinopathy Care: A Machine Learning Study.

Area of Science:

• Ophthalmology
• Health Informatics
• Predictive Analytics

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Predicting and preventing lapses in DR care is essential for effective disease management.
• Existing prediction models may not fully capture the multifactorial nature of care gaps.

Purpose of the Study:

• To develop and validate machine learning algorithms for predicting lapses in diabetic retinopathy care.
• To identify key predictors of care lapses, including electronic health record (EHR) data and social determinants of health (SDoH).
• To assess the performance of different algorithmic approaches in forecasting DR care interruptions.

Main Methods:

• Retrospective cohort study involving adult patients with diabetes receiving DR care.
• Development of random forest (RF) and XGBoost (XGB) prediction models.
• Models incorporated EHR variables, location-based SDoH, and patient history of care lapses.

Main Results:

• The best performing model, RF-C, demonstrated strong predictive capability with an AUROC of 0.774 and AUPRC of 0.707.
• Models incorporating SDoH and history of lapses significantly outperformed those using EHR data alone.
• XGBoost models confirmed the superior performance of algorithms including comprehensive variable sets.

Conclusions:

• Machine learning models can effectively predict lapses in diabetic retinopathy care.
• Integrating EHR data, SDoH, and historical care patterns enhances prediction accuracy.
• These predictive tools can facilitate targeted interventions to reduce care gaps and prevent vision loss.