Fair-DSP: Fair Dynamic Survival Prediction on Longitudinal Electronic Health Record

Abstract

Scarce medical resources and highly transmissible diseases may overwhelm healthcare infrastructure. Fair allocation based on disease progression and fair distribution among all demographic groups is demanded by society. Surprisingly, there is little work quantifying and ensuring fairness in the context of dynamic survival prediction to equally allocate medical resources. In this study, we formulate individual and group fairness metrics in the context of dynamic survival analysis with time-dependent covariates, in order to provide the necessary foundations to quantitatively analyze the fairness in dynamic survival analysis. We further develop a fairness-aware learner (Fair-DSP) that is generic and can be applied to a dynamic survival prediction model. The proposed learner specifically accounts for time-dependent covariates to ensure accurate predictions while maintaining fairness on the individual or group level. We conduct quantitative experiments and sensitivity studies on the real-world clinical PBC dataset. The results demonstrate that the proposed fairness notations and debiasing algorithm are capable of guaranteeing fairness in the presence of accurate prediction.