Exploring trade-offs in equitable stroke risk prediction with parity-constrained and race-free models

Matthew Engelhard, Daniel Wojdyla, Haoyuan Wang, Michael J. Pencina, Ricardo Henao

Abstract: Highlights•Common stroke risk prediction models do not perform equally well in Black and White subgroups.•We propose a method to address this disparity and compare it to the more common practice of removing race as a predictor.•Results show that there is a trade-off between intra-group calibration and inter-group discrimination performance.•Consequently, the choice of model must depend on the potential benefits and harms associated with intended clinical use.•Our results provide a roadmap for development of equitable clinical risk prediction models and highlight pros and cons of a race-free approach.