A continuum model for cardiovascular risk using OS-CMR biomarkers

Background Known Cardiovascular risk factors have a weak to moderate predictive value only. Coronary vascular dysfunction is the earliest abnormality associated with atherosclerosis. Recently, Oxygenation-Sensitive Cardiovascular Magnetic Resonance (OS-CMR) was shown to provide markers of coronary vascular function. We hypothesized that markers automatically extracted from OS-CMR images are associated with cardiovascular risk. Purpose To identify clusters of coronary vascular function phenotypes related to established cardiovascular risk factors. Methods We analyzed CMR data from consecutive subjects of a large local cardiovascular cohort study with a wide range of cardiovascular risk or disease. The extracted data comprised 116 experimental metrics derived from normalized myocardial signal intensity curves across the cardiac cycle, including dynamics over time, intensity distributions, and tissue heterogeneity. Subjects were stratified into 3 risk groups based on INTERHEART risk factor scores: controls (n=280, score 0-8), mild risk (n=160, score 9-15), and high risk (n=94, score ≥16). From the initial 118 features (116 experimental biomarkers plus age and sex), univariate feature selection using one-way ANOVA F-statistics identified the 30 most discriminative features between risk groups. All features were standardized to zero mean and unit variance prior to dimensionality reduction. PCA computed the first two principal components, which explained 88.8% and 3.2% of the total variance, respectively. PC1 was dominated by global statistical OCG intensity biomarkers, while PC2 was primarily driven by myocardial heterogeneity, sex, and end-diastolic OCG signal intensity metrics. Results Images from 534 OS-CMR scans were included. The PCA visualization revealed a clear disease continuum along PCI, with group centroids showing progressive separation: controls (PC1: 0.97), mild risk (PC1: -0.41), and high risk (PC1: -2.19), see Fig. 1. The sample distribution shows a continuous pattern with substantial overlap between adjacent risk groups (53.4% of samples in transition zones), supporting a spectrum model rather than discrete disease states. PC1, primarily loaded with vascular function biomarkers derived from the signal intensity curves, provided strong discrimination (88.8% variance explained). Conclusion Our results indicate that coronary vascular function markers automatically extracted from OS-CMR images follow a continuous relationship with the presence and number of markers for cardiovascular risk, consistent with a spectrum model rather than discrete states of metabolically determined coronary vascular function. The results confirm the robustness of OS-CMR to identify clinically relevant markers of coronary vascular function.