Imaging-derived cardiovascular OrganAge predicts mortality and identifies smoking and diabetes as major accelerators of cardiovascular ageing in UK Biobank

Background

Chronological age inadequately captures the heterogeneity in cardiovascular ageing. We hypothesise that machine learning (ML) models applied to large-scale cardiovascular imaging could quantify biological cardiovascular ageing and highlight major drivers of accelerated decline.

Purpose

To derive an imaging-based estimate of biological cardiovascular (CV) ageing (CVOrganAge) and to quantify the relative contributions of major demographic and cardiometabolic factors to accelerated CV ageing.

Methods

Using cardiac MRI and carotid ultrasound phenotypes from over 33,000 UK Biobank participants, we trained sex-specific gradient-boosted machine-learning models to estimate CVOrganAge from high-dimensional imaging features (80/20 train–test split; 10-fold cross-validation). Imaging features with >50% missingness were excluded. The performance was quantified using root-mean-square error (RMSE), mean absolute error (MAE), and Pearson’s correlation coefficient (r). We evaluated the association of CVOrganAge with subsequent all-cause mortality using Cox models and quantified its associations with established risk factors using multivariable models. We also estimated sex-stratified differences in CVOrganAge gap (CVOrganAge–chronological age) associated with prevalent cardiovascular conditions. Interpretability was assessed using feature importance.

Results

Higher CVOrganAge gap was associated with increased all-cause mortality after adjustment for chronological age, sex, BMI, smoking, and deprivation (hazard ratio per year-increase 1.41, 95% CI 1.27–1.58; P<0.00001). In multivariable analyses, current smoking showed the strongest association with higher CVOrganAge, followed by diabetes, which exceeded BMI as the leading cardiometabolic determinant. Diabetes was associated with the greatest CVOrganAge gap (~1.3–1.5 years older than disease-free individuals), with heart failure also showing a large gap (~1.1–1.4 years) and evidence of sex heterogeneity across several diagnoses. Key model contributors included pericardial fat, chamber volumes and function (e.g., RA minimum volume, LA ejection fraction), LV strain, aortic/arterial stiffness metrics, and carotid intima–media thickness.

Conclusions

In this large UK Biobank imaging cohort, ML-derived CVOrganAge identified individuals with accelerated cardiovascular ageing and was independently associated with subsequent mortality after adjustment for conventional risk factors. Although observational, these results support further evaluation of CVOrganAge as a scalable imaging-derived metric to enhance population risk stratification and more precisely target preventive cardiovascular interventions.

Figure 1: Multivariable association of CVOrganAge with CV risk factors

Figure 2: Interpretable determinants and clinical relevance of CVOrganAge in UK Biobank: a) top imaging contributors, b) disease-associated age gaps, and c) diabetes-onset stratificationFigure 1Figure 2