Mitigating bias in chest X-ray disease diagnosis via de-biased disentangled representation learning.

Abstract

Ensuring fairness and reliability in AI-driven healthcare systems is critical, particularly in chest X-ray diagnosis, where models often exhibit subgroup biases. While existing methods primarily address biases from imbalanced data, they overlook those arising from visual bias attributes. To address this, we propose a novel de-biased model based on disentangled representation learning, comprising three core components: a Disentanglement Module to separate visual bias and disease features, a Disease Diagnosis Module for accurate diagnoses, and a Feature Augmentation Module to enhance disease feature extraction through feature-level sample diversification. Evaluated on the MIMIC-CXR and CheXpert datasets, our model achieves higher accuracy and reduced subgroup bias, outperforming baseline and comparative methods. This work presents a promising approach to improving fairness and diagnostic accuracy in chest X-ray analysis. The code is available at https://github.com/WishLyle/FAIR_DRL_CXR.