Post-Processing Fairness Evaluation of Federated Models: An Unsupervised Approach in Healthcare.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning
• Deep Learning

Background:

• Federated Learning (FL) is crucial for privacy-preserving AI in healthcare.
• Distributed data and learning challenges can lead to inadequate local model training.
• Poorly trained models pose significant risks in critical healthcare applications.

Purpose of the Study:

• To address the inadequacy of local FL model training in healthcare.
• To introduce a post-processing pipeline for improving FL model fairness and performance.
• To enhance the reliability and accuracy of AI models in medical systems.

Main Methods:

• Developed a post-processing pipeline for Federated Learning models.
• Implemented a method to rank models based on fairness by inspecting micro-Manifolds.
• Utilized an unsupervised, model- and data-agnostic methodology for fairness discovery.

Main Results:

• Achieved an average 8.75% increase in Federated model accuracy.
• Demonstrated the effectiveness of the methodology across various Deep Learning architectures.
• Validated the approach in a Federated Learning environment.

Conclusions:

• The proposed post-processing pipeline effectively enhances Federated Learning model fairness and accuracy in healthcare.
• The unsupervised methodology offers a generalizable approach for discovering model fairness.
• This work contributes to more reliable and accurate AI in critical medical applications.