Introduction: The American Heart Association’s race-agnostic Predicting Risk of cardiovascular EVENTs (PREVENT) framework has emerged to better capture multilevel determinants of health in contemporary population. Whether PREVENT better predicts incident atherosclerotic cardiovascular disease (ASCVD) than the preceding Pooled Cohort Equations (PCE) is unknown in healthcare populations. We evaluated calibration and discrimination of the two equations across three major U.S. healthcare networks. Methods: We retrospectively evaluated electronic health records of 209,879 individuals (Mass General Brigham [MGB]: 118,801; Mount Sinai Healthcare System: 38,834; Penn Medicine: 52,424) aged 40-79 years and without prior ASCVD between 2010-2012. Calibration was assessed using discordance ratio between 10-year observed and predicted risk. Discrimination was evaluated by Harrell’s C-index. Results: Based on PREVENT, the mean [SD] estimated 10-year ASCVD risk ranged from 5.5 [4.7] % in MGB (mean age 56y, 42% female), 6.6 [5.6]% in Mount Sinai (59y, 54% female), and 6.4 [5.2]% in Penn (60y, 55% female). PREVENT substantially underestimated empirical 10-year ASCVD incidence in MGB (16.9%) and Mount Sinai (10.9%), whereas it closely mirrored the observed rate in Penn (5.5%). Both PCE and PREVENT had comparable discrimination C-index [95% CI] across MGB (0.69 [0.69-0.70] vs. 0.69 [0.68-0.69]), Mount Sinai (0.72 [0.71-0.73] vs. 0.70 [0.70-0.71]), and Penn (0.68 [0.67-0.69] vs. 0.67 [0.66-0.68]). Nevertheless, compared to PREVENT, PCE demonstrated better overall calibration in MGB (-70% vs. -42%) and Mount Sinai (-36% vs. +5%), whereas largely overestimated risk in Penn (+3% vs +94%). Calibration varied across demographics with PREVENT consistently underestimating risk in Hispanics, Asians, Black and African Americans across all sites, whereas PCE largely overestimated risk for White or multiracial Penn healthcare patients. Conclusions: The race-agnostic PREVENT model discriminated ASCVD incidence comparably or better than the race-informed PCE across three geographically distinct academic health systems in the Northeast. However, calibration metrics between the two models varied widely across health systems and demographics.
Read full abstract