A quantitative pipeline for bias detection and mitigation in medical imaging AI models: a prototype using chest radiography

Abstract

Prior research has demonstrated that AI systems can demonstrate statistical bias, highlighting the importance for methods to detect and mitigate bias prior to deployment. This study introduces a prototype software tool designed to detect and mitigate bias in an AI system at the post-processing stage (i.e., after classifier training) without retraining. We describe a quantitative pipeline tailored to binary data attributes and illustrate its utility for a classifier trained to predict ICU admission for COVID-19 on chest X-ray (CXR). The test dataset included 1048 patients with a 14.0% prevalence of ICU admission within 24 hours after CXR exam. For bias detection in the model’s predictions between the subgroups, we developed a Python function that computes seven different group-level fairness metrics that measure bias, using patient sex as an example attribute. To mitigate bias at the post-processing stage without requiring retraining of the underlying model, the tool was developed to perform group thresholding to improve fairness. The treatment equality ratio was 0.535 before bias mitigation, deviating the most from ideal fairness (1.0) of all fairness metrics evaluated. After mitigation, treatment equality improved to 0.996 while the model maintained 95% sensitivity, demonstrating that statistical bias could be reduced without compromising sensitivity and avoiding retraining. We also developed radar plot functions to visualize the metrics before and after mitigation. Our pipeline successfully quantifies statistical bias across multiple fairness metrics and includes actionable mitigation steps. Future work will evaluate additional AI models across various attributes and extend mitigation strategies across other model development stages to further refine and validate the pipeline.

Ähnliche Arbeiten