New Findings Show How Positive-Sum Fairness Changes the Performance of Medical AI Models | HackerNoon
Briefly

from Hackernoon1 month ago
M2 improves overall performance as it outperforms M1 in AUROC, leveraging an additional relevant medical feature. Although it shows more disparity in fairness, this does not equate to harm under positive-sum fairness.
Hackernoonhttps://hackernoon.com/new-findings-show-how-positive-sum-fairness-changes-the-performance-of-medical-ai-models
M4 enhances fairness specifically for lung lesions and consolidations by using a gradient reversal layer that eliminates race information, thereby preventing demographic shortcuts from influencing model performance.
Hackernoonhttps://hackernoon.com/new-findings-show-how-positive-sum-fairness-changes-the-performance-of-medical-ai-models
M3 displays inconsistent results, often performing below the baseline except for pneumonia. It appears that its strategy of maximizing race prediction may intensify demographic shortcuts previously exploited.
Hackernoonhttps://hackernoon.com/new-findings-show-how-positive-sum-fairness-changes-the-performance-of-medical-ai-models
Read at Hackernoon
#model-performance #fairness #medical-imaging #disparity #machine-learning
[
|
]