Abstract:
The application of artificial intelligence (AI) in healthcare has tremendous potential
for improving diagnostic precision and optimizing treatment and patient
care. However, increasing dependence on such tools brings up urgent
questions regarding the amplification of existing biases, which may detract
from their ability to improve fair clinical decision-making. Adversarial debiasing,
a method that utilizes fairness measures by contrasting a core predictive
model with an adversarial network to reduce the influence of sensitive features,
has emerged as an effective way of mitigating bias in AI systems. This review
combines findings from 25 studies on several areas, encompassing the technical
elements of adversarial learning and its practical applications in
healthcare. The review offers extensive data and thoroughly assesses technological,
ethical, and practical issues. This study reveals that adversarial debiasing
improves fairness indicators and presents significant trade-offs, including
reduced sensitivity and interpretability. We conclude with recommendations
for future research avenues, encompassing prospective multicenter trials,
adaptive training methodologies, hybrid debiasing strategies, and formulating
standardized regulatory frameworks.
