Deep-learning prediction of cardiovascular outcomes from routine retinal images in individuals with type 2 diabetes

Abstract

Background: Prior studies have demonstrated an association between retinal vascular features and cardiovascular disease (CVD), however most studies have only evaluated a few simple parameters at a time. Our aim was to determine whether a deep-learning artificial intelligence (AI) model could be used to predict CVD outcomes from routinely obtained diabetic retinal screening photographs and to compare its performance to a traditional clinical CVD risk score.

Methods: We included 6127 individuals with type 2 diabetes without myocardial infarction or stroke prior to study entry. The cohort was divided into training (70%), validation (10%) and testing (20%) cohorts. Clinical 10-year CVD risk was calculated using the pooled cohort equation (PCE) risk score. A polygenic risk score (PRS) for coronary heart disease was also obtained. Retinal images were analysed using an EfficientNet-B2 network to predict 10-year CVD risk. The primary outcome was time to first major adverse CV event (MACE) including CV death, myocardial infarction or stroke.

Results: 1241 individuals were included in the test cohort (mean PCE 10-year CVD risk 35%). There was a strong correlation between retinal predicted CVD risk and the PCE risk score (r = 0.66) but not the polygenic risk score (r = 0.05). There were 288 MACE events. Higher retina-predicted risk was significantly associated with increased 10-year risk of MACE (HR 1.05 per 1% increase; 95% CI 1.04-1.06, p < 0.001) and remained so after adjustment for the PCE and polygenic risk score (HR 1.03; 95% CI 1.02-1.04, p < 0.001). The retinal risk score had similar performance to the PCE (both AUC 0.697) and when combined with the PCE and polygenic risk score had significantly improved performance compared to the PCE alone (AUC 0.728). An increase in retinal-predicted risk within 3 years was associated with subsequent increased MACE likelihood.

Conclusions: A deep-learning AI model could accurately predict MACE from routine retinal screening photographs with a comparable performance to traditional clinical risk assessment in a diabetic cohort. Combining the AI-derived retinal risk prediction with a coronary heart disease polygenic risk score improved risk prediction. AI retinal assessment might allow a one-stop CVD risk assessment at routine retinal screening.

Keywords: Artificial intelligence; Cardiovascular risk; Diabetes; Retina.

Fig. 1

Example of pre-processing result (top) and Average of Grad-CAM heatmaps for PCE risk score prediction superimposed on a sample left eye and right eye retina (bottom). A Original retinal image; B pre-processed image; C heatmap for left eye retinas; D heatmap for right eye retinas

Fig. 2

Correlations between individual-level retinal predicted 10-year MACE, clinical CVD risk and genetic CHD risk. A Predicted 10-year MACE vs. CHD PRS. B Predicted 10-year MACE vs. PCE risk score. Red dots represent females and blue dots represent males

Fig. 3

Major adverse cardiovascular events stratified by retinal predicted CVD risk. Kaplan-Meier curves showing MACE-free survival at 10 years stratified by tertiles of retinal predicted CVD risk derived from individual-level prediction at baseline

Fig. 4

Major Adverse Cardiovascular Events Based on Progression of Retinal Predicted CVD Risk