Navigating fairness in artificial intelligence-based prediction models: theoretical constructs and practical...

Area of Science:

• Health Informatics
• Medical AI
• AI Ethics

Background:

• Artificial intelligence (AI) is increasingly used in healthcare for prediction and decision support.
• Ensuring AI fairness is crucial to prevent health disparities and guarantee equitable patient outcomes.
• Conflicting definitions of AI fairness hinder practical implementation in clinical settings.

Purpose of the Study:

• To bridge the gap between AI fairness theory and practice by identifying appropriate fairness metrics.
• To analyze 27 fairness definitions in relation to intended use, decision type, and distributive justice principles.
• To provide practical guidance for developing and assessing fair AI in healthcare.

Main Methods:

• Literature review of 27 AI fairness definitions.
• Assessment of definitions against intended AI use, decision influence, and ethical principles.
• Analysis of two use cases to demonstrate metric applicability.

Main Results:

• Certain fairness notions have limitations for medical applications.
• Clinical utility, performance metrics (e.g., AUC), calibration, and statistical parity are deemed most relevant group-based metrics.
• The choice of fairness metrics depends on the specific use case and ethical framework.

Conclusions:

• Practical guidance is offered for fair AI development and assessment.
• Understanding the impact of bias mitigation strategies is essential.
• This work supports the equitable implementation of AI-based healthcare solutions.