Abstract

A recent analysis of common stroke risk prediction models showed that performance differs between Black and White subgroups, and that applying standard machine learning methods does not reduce these disparities. There have been calls in the clinical literature to correct such disparities by removing race as a predictor (i.e., race-free models). Alternatively, a variety of machine learning methods have been proposed to constrain differences in model predictions between racial groups. In this work, we compare these approaches for equitable stroke risk prediction. We begin by proposing a discrete-time, neural network-based time-to-event model that incorporates a parity constraint designed to make predictions more similar between groups. Using harmonized data from Framingham Offspring, MESA, and ARIC studies, we develop both parity-constrained and unconstrained stroke risk prediction models, then compare their performance with race-free models in a held-out test set and a secondary validation set (REGARDS). Our evaluation includes both intra-group and inter-group performance metrics for right-censored time to event outcomes. Results illustrate a fundamental trade-off in which parity-constrained models must sacrifice intra-group calibration to improve inter-group discrimination performance, while the race-free models strike a balance between the two. Consequently, the choice of model must depend on the potential benefits and harms associated with the intended clinical use. All models as well as code implementing our approach are available in a public repository. More broadly, these results provide a roadmap for development of equitable clinical risk prediction models and illustrate both merits and limitations of a race-free approach.