The Effect of Enforcing Fairness on Reshaping Explanations in Machine Learning Models.

Area of Science:

• Machine Learning in Healthcare
• Artificial Intelligence Ethics
• Clinical Decision Support

Background:

• Trustworthy machine learning in healthcare demands predictive performance, fairness, and explainability.
• Clinicians may distrust models if explanations change after fairness adjustments.
• The impact of fairness improvements on model explainability remains under-explored.

Purpose of the Study:

• To investigate how bias mitigation techniques affect Shapley-based feature rankings in machine learning models.
• To quantify the changes in feature importance after applying fairness constraints.
• To evaluate the stability of Shapley-based rankings across different model classes and datasets.

Main Methods:

• Applied bias mitigation techniques to enhance fairness across racial subgroups.
• Quantified changes in Shapley-based feature importance rankings.
• Evaluated ranking stability on three diverse datasets: UTI risk, anticoagulant bleeding risk, and recidivism risk.
• Assessed multiple machine learning model classes.

Main Results:

• Increasing model fairness significantly altered feature importance rankings.
• Changes in feature rankings sometimes varied across different racial subgroups.
• The stability of Shapley-based rankings differed across model classes.

Conclusions:

• Enhancing fairness in healthcare AI can substantially reshape model explanations.
• Model explainability is sensitive to fairness interventions, posing challenges for clinical trust.
• A holistic approach is needed to evaluate accuracy, fairness, and explainability concurrently.