Enhancing Cardiovascular Disease (CVD) Risk Prediction Using Metabolic Profiling in the UK Biobank—Pooled Cohort Equations vs. Retinal Imaging Based Dr. Noon CVD

Introduction and Objective: 

Circulating metabolites hold potential for improving cardiovascular disease (CVD) risk prediction. However, their added value to existing risk calculators like the Pooled Cohort Equations (PCE) and retina based artificial intelligence (AI) tool like Dr Noon CVD remains uncertain.

Methods: 

This study used UK Biobank, a prospective cohort of 0.5 million participants aged 40-69 years. 17447 participants with available blood metabolic profiling and no history of CVD at baseline were included. Cox LASSO model identified a set of metabolites possibly associated with incident CVD events over 10-year follow-up in a training set of UK Biobank (50% of the individuals). The predictive performance of the PCE and Dr Noon CVD was assessed using Cox regression before and after adding the metabolites.

Results: 

At 10 year follow-up, 563 participants had CVD events. Among the 168 metabolites – glucose, lactate, albumin, phospholipids in medium HDL, triglycerides in chylomicrons and extremely large VLDL and triglycerides in large LDL were associated with CVD events. For metabolites alone baseline Harrell’s C-index= 0.667 (95% CI: 0.635-0.699). After adding selected metabolites C-index for Dr Noon CVD improved by 2.72% from 0.711 (95% CI: 0.683-0.738) to 0.738 (95% CI: 0.712-0.765). For PCE, the performance increased by 2.93% from 0.710 (95% CI: 0.682-0.738) to 0.739 (95% CI: 0.713-0.766). For individuals with prediabetes and diabetes, the C-index improved by 5.11% (from 0.684 to 0.735) for PCE as compared to 2.86% (from 0.706 to 0.735) for Dr Noon CVD.

Conclusion: 

Integration of metabolic profiling with CVD risk prediction models enhances performance, with similar improvements for Dr. Noon CVD and PCE. In diabetics and prediabetics, the PCE model showed a larger improvement, however, both models achieved comparable overall performance after incorporating metabolites, indicating similar final predictive performance for both approaches.