BACKGROUND AND AIMS: Cardiovascular disease (CVD) is a leading cause of death worldwide with type 1 and type 2 diabetes as known risk factors. Recent studies have shown deep learning (DL) predictors of CVD risk factors based on fundus images are significantly associated with CVD risk independently of the same risk factors. However, reported increment in CVD risk prediction of the DL models compared to a model using clinical risk factors alone has been small. The objective of this study was to examine whether DL models using retinal images could improve the predictive performance for CVD risk - in people with type 1 diabetes (T1D) and type 2 diabetes (T2D) - when compared with a baseline Poisson model using clinical covariates.
MATERIALS AND METHODS: The cohort was constructed using the Scottish Diabetes Research Network dataset that links all fundus images between 2005-2017 from the Scottish Diabetic Retinopathy Screening (SDRS) programme to eHealth records. We follow 23,103 and 201,957 people with T1D and T2D respectively who were alive any time between 1 January 2008 and 1 January 2018 with no prior CVD event and at least one gradable screening episode prior to follow-up until first incident CVD yielding 166,113.1 and 1,269,142 person-years of follow up respectively. There were 2,003 and 38,569 incident CVD events respectively which included Coronary Heart Disease, stroke, Peripheral Vascular Disease, Cerebrovascular Disease, Atherosclerosis, Coronary Artery Disease, Acute Myocardial Infarction. Using fundus images from the SDRS programme we trained DL networks - for T1D and T2D separately - to yield DL predictors of: CVD, diabetic retinopathy (DR), eGFR, and systolic blood pressure (SBP). Each of the DL predictors is then included in a Poisson regression 10-year CVD risk prediction model containing known CVD risk factors. We compared the predictive performance between Poisson models (a restricted model including age, diabetes duration, and sex as covariates and a full baseline using further known risk factors) and the Poisson models that included the DL predictors. Half of the cohort were used to train all models and half to evaluate performance.
RESULTS: C-statistics and test log-likelihoods are shown in the table. DL contributed no performance improvement to the restricted baseline models, with C-statistics 0.75 and 0.67, nor to the full baseline models, with C-statistics 0.82 and 0.71, for T1D and T2D respectively.
CONCLUSION: There has been much optimism that DL approaches may significantly improve prediction of CVD risk beyond conventional predictive models. However this study suggests the added value of DL predictors for CVD risk prediction may be small.
