Navigating Fairness in AI-based Prediction Models: Theoretical Constructs and Practical Applications.

Area of Science:

• Health Informatics
• Artificial Intelligence Ethics
• Medical AI

Background:

• Artificial Intelligence (AI) prediction models are increasingly used in healthcare.
• Ensuring AI fairness is vital to combat health disparities and achieve equitable patient outcomes.
• Conflicting definitions of fairness pose challenges to practical AI implementation.

Purpose of the Study:

• To structure the transition of AI fairness from theory to practice.
• To identify appropriate fairness metrics for medical AI applications.
• To assess the relationship between fairness definitions, intended use, decision types, and distributive justice.

Main Methods:

• Reviewed 27 definitions of AI fairness from recent literature.
• Assessed the relation of each definition with intended use, decision influence, and ethical principles.
• Evaluated clinical utility, performance metrics (AUC), calibration, and statistical parity for medical applications.
• Demonstrated applicability through two use cases.

Main Results:

• Clinical utility, performance-based metrics (AUC), calibration, and statistical parity are recommended as the most relevant group-based fairness metrics for medical AI.
• Different fairness metrics are applicable based on the specific intended use and ethical framework.
• The study provides a foundation for assessing AI fairness and bias mitigation.

Conclusions:

• A structured approach to AI fairness metrics is necessary for equitable healthcare implementations.
• Selecting appropriate fairness metrics is context-dependent on clinical utility and ethical considerations.
• This work promotes more equitable AI development and deployment in healthcare.