Medicine and Health

The limits of fair medical imaging AI in real-world generalization

Y. Yang, H. Zhang, et al.

This study reveals the critical challenges of fairness in medical AI for disease classification across various imaging modalities, highlighting how demographic shortcuts lead to biased predictions. Conducted by Yuzhe Yang, Haoran Zhang, Judy W. Gichoya, Dina Katabi, and Marzyeh Ghassemi, the research uncovers that less demographic attribute encoding in models can yield better performance in diverse clinical settings, emphasizing best practices for equitable AI applications.

00:00

Playback language: English

Index

Abstract

This study investigates the fairness of medical AI in disease classification across various imaging modalities (radiology, dermatology, ophthalmology) and datasets. The research confirms that medical imaging AI leverages demographic shortcuts, leading to biased predictions. While algorithmic correction improves fairness within the original data distribution, this 'local optimality' doesn't generalize to new test settings. Surprisingly, models with less demographic attribute encoding often perform better ('globally optimal') in out-of-distribution evaluations. The study establishes best practices for maintaining performance and fairness in diverse clinical deployments.

Publisher

Nature Medicine

Published On

Oct 14, 2024

Authors

Yuzhe Yang, Haoran Zhang, Judy W. Gichoya, Dina Katabi, Marzyeh Ghassemi

Related Publications

Explore these studies to deepen your understanding of the subject.

Economics

The development of Kondratieff's theory of long waves: the place of the AI economy humanization in the 'competencies-innovations-markets' model

A. E. Tyulin, A. A. Chursin, et al.

Social Work

Peer reviewers' dilemmas: a qualitative exploration of decisional conflict in the evaluation of grant applications in the medical humanities and social sciences

G. Vallée-tourangeau, A. Wheelock, et al.

Psychology

Prevalence of suicidal thoughts and attempts in the transgender population of the world: a systematic review and meta-analysis

P. Kohnepoushi, M. Nikouei, et al.

Social Work

The delayed and combinatorial response of online public opinion to the real world: An inquiry into news texts during the COVID-19 era

Y. Du, H. Cheng, et al.

Listen, Learn & Level Up

Over 10,000 hours of research content in 25+ fields, available in 12+ languages.

No more digging through PDFs, just hit play and absorb the world's latest research in your language, on your time.

listen to research audio papers with researchbunny