Heterogeneous Competing Causes of Death and Transportability of the PREVENT Equations

The American Heart Association's PREVENT equations represent a major advance in cardiovascular risk prediction, in part because they account for the possibility that patients may die from non-cardiovascular causes before experiencing a cardiovascular event. Despite this advance, recent external validations have found that PREVENT systematically underestimates cardiovascular risk in Black patients, socioeconomically disadvantaged groups, veterans, and those with Medicaid or no insurance. We propose that one underrecognized contributor is how PREVENT handles non-cardiovascular death: by treating all such deaths from cancer, overdose, suicide, injury, and other causes as a single combined category. This aggregation is a problem because the mix of competing causes differs across populations. The direct consequence is that the model's competing-risk adjustment reflects the cause-of-death composition of its derivation cohort rather than that of the population being evaluated, and standard recalibration methods cannot correct this mismatch. The direction of the resulting error is predictable: populations with disproportionately high overdose, suicide, and injury mortality relative to the cancer-dominant pattern of general-population derivation cohorts will have their cardiovascular risk systematically underestimated. This problem is further obscured by a common methodological choice in validation studies: using Kaplan-Meier survival estimates rather than cause-specific competing-risk methods to measure observed event rates. This approach inflates apparent event rates in exactly the subgroups most affected, making miscalibration harder to interpret and potentially masking its true direction. Future cardiovascular risk prediction models should separate non-cardiovascular deaths into cause-specific competing processes. External validations should use cause-specific competing-risk estimation. Health systems considering adoption of PREVENT should validate it locally before implementation, with particular attention to calibration within subgroups defined by competing-mortality burden.