Enhancing fairness in AI-enabled medical systems with the attribute neutral framework

Area of Science:

• Artificial Intelligence in Medicine
• Healthcare Equity
• Machine Learning for Healthcare

Background:

• Artificial intelligence (AI) in healthcare faces challenges with unfairness and inequity.
• Biased AI models may learn undesirable correlations between sensitive attributes and health data.
• Unequal AI performance across protected groups hinders equitable healthcare deployment.

Purpose of the Study:

• To introduce a novel framework for disentangling and neutralizing biased attributes in AI healthcare models.
• To develop a method that improves representation and fairness across diverse patient subgroups.
• To mitigate unfairness in AI diagnostic models without compromising performance.

Main Methods:

• Introduction of the Attribute Neutral Framework and the Attribute Neutralizer (AttrNzr).
• Generation of neutralized data where protected attributes are not easily predictable.
• Training of a disease diagnosis model (DDM) using neutralized data.

Main Results:

• AttrNzr effectively reduces unfairness in the DDM compared to other mitigation algorithms.
• The Attribute Neutral Framework maintains the overall disease diagnosis performance of the DDM.
• AttrNzr demonstrates utility in neutralizing multiple attributes and during the training phase only.

Conclusions:

• The Attribute Neutral Framework offers a data-centered, model-independent solution to AI fairness challenges in medicine.
• AttrNzr addresses the root cause of unfairness by neutralizing biased data.
• This approach has significant potential for creating more equitable AI-enabled medical systems.