The concept of positive-sum fairness allows for increased performance in medical AI to be achieved through the use of sensitive attributes, such as race, without undermining subgroup performance. Our approach suggests that not all disparities are harmful; rather, some can be acceptable if they contribute positively overall. Explore whether enhancements in performance lead to only insignificant disparities among different groups, thus emphasizing that an increase in an AI model's efficacy might not inherently compromise fairness.
Our analysis demonstrates that when all demographic encodings are removed, the performance gap narrows. However, by selectively using the race attribute during training, we observed an improvement in overall performance, albeit at the cost of increased disparity. This illustrates the balance necessary in medical AI, where the ethical use of sensitive attributes should be aligned with striving for better healthcare outcomes for all demographics.
Collection
[
|
...
]