A linked mixture model of coronary atherosclerosis

Background & Objectives: Characterizing the progression of coronary atherosclerosis is a critical public health goal. The most common quantitative summary, the CAC score, is modelled by a variety of statistical methods, both as a predictor of coronary events and as an outcome of behavioral and population-specific risk factors. Little attempt has been made, however, to ground these statistical models in the underlying physiology of arterial aging, which would allow us to describe the onset and growth of CAC over a patient's life. Methods: Using a generative growth model for arterial plaque accumulation, we identify severe under-estimation in the age of initial onset and rate of progression (doubling time) of CAC growth with standard ln(CAC + 1) or ln(CAC | CAC > 0) models, and use this growth model to motivate new statistical approaches to CAC using logistic and log-linear mixture regressions. We compare statistical models directly by computing their average parameter biases using 540 growth trajectory simulations (113,760 patients, 268,200 observations). Results: While all models used can successfully estimate the influence of risk factors with minimal bias, we demonstrate substantial improvements in predictive accuracy in the timing of CAC onset and progression with logistic regression and linked hurdle-lognormal mixture regression, compared with standard ln(CAC + 1) or ln(CAC | CAC > 0) models. Conclusions: Using models that can account for patient-specific onset and progression rates, accurate descriptions of CAC trajectories can be made even in cross-sectional (single scan per patient) designs, with substantial clinical and epidemiological utility.