Multi-Adversarial Debiasing in Clinical Artificial Intelligence.

Area of Science:

• Machine Learning
• Clinical Informatics
• Algorithmic Fairness

Background:

• Clinical machine learning models can exhibit biases, impacting equitable healthcare outcomes.
• Current debiasing methods often focus on optimizing a single fairness metric, potentially overlooking other bias types.

Purpose of the Study:

• To introduce and evaluate a novel multi-adversarial debiasing framework for clinical machine learning.
• To jointly optimize multiple fairness definitions, specifically demographic parity (DP) and disparate mistreatment (DM).

Main Methods:

• Developed a multi-adversarial debiasing framework extending adversarial debiasing.
• Employed two adversaries representing DP and DM for joint optimization.
• Evaluated the framework on two clinical datasets (UCI Heart Disease, Parkinson's Disease) and two benchmark datasets (COMPAS, Adult Income).

Main Results:

• The multi-adversarial approach successfully reduced DP by 0.03-0.22 and DM by 0.02-0.12 across datasets.
• F1 scores were maintained within 0-16% of baseline models, indicating minimal performance compromise.
• Effectiveness was highest in datasets with balanced representation across protected attributes.

Conclusions:

• Multi-adversarial debiasing offers a more comprehensive approach to mitigating bias in clinical ML than single-metric optimization.
• The framework demonstrates potential for enhancing fairness in healthcare AI applications.
• Dataset characteristics, particularly label representation across protected attributes, influence the efficacy of adversarial debiasing.