Evaluating algorithmic fairness in the performance of the AHA PREVENT equations versus pooled cohort equations among...

The performance of AHA PREVENT equation among multiracial individuals has not been assessed.

To assess model performance and fairness of the AHA PREVENT ASCVD (atherosclerotic cardiovascular disease) Equation versus Pooled Cohort Equations (PCE) among single-race and disaggregated multiracial individuals.

This is a retrospective cohort data analysis using electronic health record from a 2010-2023 primary care population in a large integrated healthcare system in California. Patient population was restricted to those who are eligible for both the AHA PREVENT base equation and PCE. The primary outcomes were incident ASCVD based on diagnosis codes. We compared measures of model performance as well as algorithmic fairness between PCE and AHA PREVENT, using Disparate Impact (DI), Average Odds Difference (AOD), and Equal Opportunity Difference (EOD).

A total of 298,276 patients aged 40-79 years were included, with an average follow-up time of 8 years (SD=3), of which 9% (N=26,761) self-identified as multiracial. In aggregate, 3.9% of multiracial patients developed ASCVD, ranging from 2.9% (Multiple Asian) to 5.4% (Multiracial American Indian/Alaska Native), compared to 4.4% of all single-race patients. Discriminatory performance of the AHA PREVENT ASCVD Equation was good for aggregated multiracial groups (AUC=0.747), ranging from 0.692 (Multiracial Black) to 0.793 (Multiracial Pacific Islander), and was comparable to single-race groups (AUC=0.753). Calibration was generally good among most multiracial groups. However, the AHA PREVENT ASCVD Equation was fairer in its predictive ability when compared to the PCE for all three fairness metrics.

With good discriminatory ability and calibration for single race and multiracial groups, the AHA PREVENT ASCVD Equation provides both more accurate and fairer estimates of ASCVD risk compared to PCE.