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

Electronic health record (EHR) datasets have increasingly been harnessed by artificial intelligence (AI) for predictive modeling, yet the ethnicity fairness of these models remains underexplored. To address this issue, we propose FairCare, a novel deep learning framework for ethnically fair EHR representation. FairCare introduces an ethnicity-heterogeneous graph neural network, enhanced with an attention mechanism to correct biases towards predominant nodes, ensuring that minority groups are fairly represented. Two benchmark datasets are collected from the MIMIC-III database, consists of 21,139 samples (i.e., records) and 41,602 patients (3,431,622 EHR records) for the downstream prediction tasks of mortality and decompensation, respectively. The adversarial learning architecture is fine-tuned with fair representation constraints, and demonstrates significant improvements in fairness metrics, with a 3.025-fold increase in demographic parity ratio (DPR) and reductions to 0.352 and 0.087 in disparate impact (DP) and equality of opportunity (EO), respectively. FairCare also outperforms all comparable methods on mortality and decompensation prediction tasks. Specifically, FairCare achieves AUROC scores of 0.9021 and 0.9217 for these tasks, surpassing the second-best methods by margins of 0.0319 (ConCare) and 0.0213 (AdaCare) in AUROC. We believe that the FairCare framework will attract a broad interest of both computational and medical researchers on medical artificial intelligence (AI). The source code and additional resources are available at http://www.healthinformaticslab.org/supp/resources.php.