Prediction of Coronary Artery Calcium using Retinal Photographs via Deep Learning: Korean, Spanish and Indian populations
Abstract
Introduction:
Cardiovascular diseases (CVD) are the leading cause of death in developed countries. Coronary artery calcium (CAC) is a clinically validated strong marker of CVD, and previous studies suggest that retinal blood vessels provide relevant information. This study aimed to validate the Reti-CVD model, developed for predicting CAC score through retinal photographs, using datasets from Spain, Korea and India.
Hypothesis:
We proposed the hypothesis that Reti-CVD model can accurately predict CAC scores from retinal photographs in the Spanish, Korean and Indian populations.
Methods:
The Reti-CVD model was applied to the Spanish Vall d’Hebron Institut de Recerca (VHIR) dataset (n=76), the Korean GreenCross Center dataset (n=3999), Korean Philip Screening Center dataset (n=5010) and the Indian population dataset (n=90). Key performance metrics were calculated to assess the model’s effectiveness, including specificity, sensitivity, accuracy, and the area under the curve (AUC). Bootstrap replicates of 2000 were used to determine confidence intervals (CI) for the AUC, and the optimal thresholds were searched using Youden index method.
Results:
In the Spanish VHIR dataset, the Reti-CVD model showed strong predictive performance. The model achieved a specificity of 78.57%, sensitivity of 85.29%, and overall accuracy of 81.58%. The AUC was 0.8508, with a 95% CI of 0.7556-0.9307. In the Korean GreenCross dataset, the model achieved a specificity of 65.52%, sensitivity of 82.11%, and overall accuracy of 72.09%. The AUC was 0.8084, with a 95% CI of 0.7948-0.822. In the Korean Philip Screening Center dataset, the model achieved a specificity of 69.33%, sensitivity of 85.53%, and overall accuracy of 73.01%. The AUC was 0.845, with a 95% CI of 0.8329-0.8571. In the Indian population dataset, the model achieved a specificity of 55.77%, sensitivity of 75.00%. The AUC was 0.72.
Discussion:
The validation results indicate that the Reti-CVD model effectively predicts CAC scores using retinal photographs in the Spanish, Korean and Indian populations. These findings validate the model’s robustness and generalizability and support the model’s potential for non-invasive CVD risk screening within different ethnicities.