A survey of recent methods for addressing AI fairness and bias in biomedicine

Bias in medical AI can occur at multiple stages: The paper emphasizes that biases based on race, gender, and other factors can be introduced before, during, or after AI model development, requiring vigilance throughout the entire development lifecycle rather than just at one point.

Debiasing requires active intervention: Rather than assuming AI systems are neutral, the survey identifies specific debiasing methods in biomedical natural language processing and computer vision that developers can implement to reduce bias—suggesting that fairness requires deliberate technical choices.

The problem is widespread but addressable: The authors reviewed 55 peer-reviewed articles on debiasing methods published between 2018-2023 from a pool of 10,041 initial results, indicating this is an active area of research with multiple proposed solutions available to the biomedical AI community.

Clinical stakes make fairness urgent: Because AI systems are being applied to high-stakes decisions like diagnosis and surgery, biased systems could perpetuate existing healthcare inequities and harm vulnerable populations, making bias mitigation a clinical safety issue rather than merely an ethical concern.

General AI debiasing methods can transfer to medicine: The survey suggests that debiasing techniques developed in broader AI domains may be adaptable to biomedical applications, offering additional tools beyond domain-specific approaches.