Multi-Adversarial Debiasing in Clinical Artificial Intelligence

While multiple types of biases can occur in clinical machine learning, the status quo in algorithmic debiasing is to optimize a single fairness metric in the training procedure. We propose a multi-adversarial debiasing framework that builds on the established technique of adversarial debiasing to jointly optimize two or more fairness definitions. Our experiments use two adversaries corresponding to demographic parity (DP) and disparate mistreatment (DM). Evaluating four datasets, including two clinical datasets (UCI Heart Disease and a Parkinson's Disease digital health dataset) and two algorithmic fairness benchmarks (COMPAS and Adult Income), we find that our multi-adversarial approach reduces DP between 0.03-0.22 and DM between 0.02-0.12 while maintaining the F1 score within 0-16% of the baseline models. Analyzing these performance variations, we find that adversarial debiasing is most effective for datasets with adequate representation of positive and negative labels across protected attribute values, but the effectiveness declines when this is not the case.