Empirical Comparison of Post-processing Debiasing Methods for Machine Learning Classifiers in Healthcare

Area of Science:

• Machine Learning
• Healthcare Disparities
• Algorithmic Fairness

Background:

• Machine learning classifiers in healthcare can perpetuate or worsen existing health disparities due to biased training data.
• Addressing these biases is crucial for equitable healthcare delivery.

Purpose of the Study:

• To rigorously compare state-of-the-art post-processing debiasing methods for machine learning models in healthcare.
• To evaluate the trade-offs between predictive performance and various group fairness metrics.

Main Methods:

• Comparison of multiple post-processing debiasing techniques.
• Evaluation across diverse synthetic and real-world healthcare datasets.
• Assessment using various performance and fairness metrics, including impact on untreated attributes.

Main Results:

• Identification of strengths and weaknesses for each compared debiasing method.
• Analysis of the trade-offs between group fairness and predictive performance.
• Examination of trade-offs among different notions of group fairness and impact on untreated attributes.

Conclusions:

• Post-processing methods offer a privacy-preserving way to ensure fairness without retraining.
• Optimal implementation in healthcare requires balancing accuracy and fairness based on specific needs.
• The study provides insights for effectively mitigating bias in healthcare machine learning.