Prediction of Incident Diabetic Retinopathy in Adults With Type 1 Diabetes Using Machine Learning Approach: An...

Area of Science:

• Endocrinology
• Ophthalmology
• Data Science

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in adults with type 1 diabetes (T1D).
• Early detection and intervention are critical for preventing vision-threatening complications.
• Continuous glucose monitoring (CGM) data offers a rich source of information for risk prediction.

Purpose of the Study:

• To explore machine learning (ML) models for identifying individuals with T1D at risk of DR.
• To identify key glycemic factors from CGM data that influence DR development.
• To predict high-risk individuals for timely intervention and vision preservation.

Main Methods:

• Retrospective analysis of CGM data from adults with T1D (n=60) with and without incident DR.
• Training and evaluation of three ML algorithms using glycemic features and principal components (PCs) derived from CGM data.
• Model performance assessed using 10-fold cross-validation and AUC-ROC, comparing two scenarios: raw glycemic features vs. two PCs representing hyperglycemia and hypoglycemia risk.

Main Results:

• Classifiers utilizing two PCs (hyperglycemia and hypoglycemia risk) significantly outperformed models using raw glycemic features (Scenario 2 vs. Scenario 1).
• An average AUC-ROC of 0.92 was achieved by two ML models in Scenario 2, indicating strong discriminative ability.
• The two PCs effectively captured vital classification data, enhancing predictive performance for DR risk.

Conclusions:

• Machine learning applied to CGM data shows promise for identifying adults with T1D at elevated risk of DR.
• This predictive capability can facilitate targeted interventions to prevent vision-threatening DR.
• Further research can refine ML models for more precise DR risk stratification in T1D.