Diabetes Prediction without Blood Tests—The Role of Retinal Imaging and Artificial Intelligence

Introduction and Objective: 

This study evaluates a non-invasive method for diabetes prediction using retinal imaging features—Reti-DR (probability of diabetic retinopathy) and Reti-HbA1c (probability of HbA1c > 6.5)—alongside demographic and clinical data. The objective was to determine the predictive accuracy of this approach without blood-based tests.

Methods: 

Cross-sectional data from 1,191 CMERC-HI participants (732 non-diabetic, 459 diabetic), a cohort at high risk of cardiovascular events, were analyzed. Independent variables included Reti-DR, Reti-HbA1c, age, gender, and BMI. Scores were averaged across fundus images. Reti-DR and Reti-HbA1c models were pre-trained on ~0.2M and 325,177 fundus images, respectively, from a health screening dataset. Data were split into training (80%, n=952) and testing (20%, n=239) sets. Features were standardized, and logistic regression was evaluated using AUC and positive predictive value (PPV)/sensitivity at the optimal threshold.

Results: 

The model achieved an AUC of 0.81. At an optimal threshold of 0.33, PPV was 0.58, and sensitivity was 0.80. Odds ratios showed Reti-HbA1c (2.24) and Reti-DR (1.75) as the strongest predictors.

Conclusion: 

This study demonstrates a feasible, non-invasive method for detecting diabetes, including in individuals on medications. Integrating Reti-DR and Reti-HbA1c into models offers an alternative to blood tests.