Abstract
Purpose :
Recent advancements have led the American Heart Association to develop the cardiovascular-kidney-metabolic (CKM) syndrome as a more holistic approach to staging cardiovascular health than traditional models. A key challenge in staging CKM Syndrome is the extensive set of biomarkers and more complicated decision thresholds than traditional CVD risk equations. We hypothesize that retinal features extracted through deep learning (DL) may also be of value for CKM staging.
Methods :
From the UK Biobank dataset of 49,316 participants (age 40-76, 55.3% female) and 85,526 retinal images with sufficient data to determine CKM stage, the efficacy of DL-extracted retinal features on two tasks was examined: 1) CKM stage prediction up to stage 3 (0: Healthy, 1: Adiposity, 2: Risk Factors, 3: Subclinical CVD) and 2) binary prediction of CKM stage 0-2 vs CKM stage 3. CKM stage 4 was not included as it represents already known clinical CVD. For binary CKM stage prediction (0-2 vs. 3), a model was developed combining previously validated DL models for hypertension, ASCVD risk and CKD.
All models were developed separately for males and females. 5-fold cross-validation was used to develop and validate the models. Multi-class accuracy scores (the ratio between number of instances assigned the correct class divided by the total number of instances) were used to measure the performance of the models. For the binary CKM stage 0-2 vs stage 3, sensitivity and specificity were used to measure model performance.
Results :
For the prediction of CKM stages 0-3, adding retinal biomarkers to a model that initially consisted of age, sex, BMI, and systolic blood pressure improved multi-class classification accuracy from 0.664 to 0.700 for men and 0.791 to 0.813 for women.
For binary classification of CKM stage 0-2 vs stage 3 using the combined retinal DL model, we obtained sensitivity/specificity of 82.7%/84.8% for men and 65.5%/97.4% for women.
The UK Biobank dataset had an overrepresentation of individuals with CKM stage 2 (42.8%) and an underrepresentation of women with CKM stage 3 (10.9%).
Conclusions :
Early results show that DL models using retinal images may be valuable for the rapid classification of the new CKM Syndrome. However, key limitations were also identified, including limited proportions of several CKM stage subgroups. Further validation on diverse external datasets is required.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.