FairCare: Adversarial training of a heterogeneous graph neural network with attention mechanism to learn fair representations of electronic health records

Yan Wang, Ruochi Zhang, Qian Yang, Qiong Zhou, Shengde Zhang, Yusi Fan, Lan Huang, Kewei Li, Fengfeng Zhou

Abstract: Highlights•The proposed FairCare framework uses an ethnicity-heterogeneous graph neural network to ensure demographic parity in clinical prognosis predictions through reduced bias in electronic health records (EHRs).•Adversarial training was employed to minimize learning bias, with the objective of accurate class label prediction while preventing ethnicity-specific bias, using a min-max optimization problem for the loss function.•The experimental data shows that FairCare achieves the state-of-the-art results on two benchmark tasks of EHRs while substantially improves the ethnicity fairness.