Effective Non-IID Degree Estimation for Robust Federated Learning in Healthcare Datasets.

Area of Science:

• Machine Learning
• Healthcare Informatics
• Data Science

Background:

• Federated learning (FL) is crucial for unbiased AI in diverse healthcare settings.
• Non-Independent and Not Identically Distributed (non-IID) data across healthcare systems poses a significant challenge for FL.
• Variations in patient demographics and protocols create data distribution differences.

Purpose of the Study:

• To develop a method for estimating the degree of non-IID data between healthcare datasets.
• To introduce metrics for evaluating non-IID degree estimation methods.
• To propose a novel non-IID FL algorithm for improved healthcare AI.

Main Methods:

• Developed a novel index to quantify the non-IID degree between datasets.
• Introduced variability, separability, and computational time as evaluation metrics for non-IID estimation.
• Integrated the non-IID degree index as regularization into existing FL algorithms for AKI prediction.

Main Results:

• The proposed non-IID degree estimation method effectively identifies data distribution differences.
• The method demonstrated consistent estimates across dataset subsamples with reduced computational time and improved interpretability.
• The novel non-IID FL algorithm achieved superior test accuracy for AKI prediction compared to local, concurrent FL, and centralized learning.

Conclusions:

• The developed non-IID degree estimation method is a robust and efficient tool for federated learning in healthcare.
• The novel non-IID FL algorithm enhances the performance of machine learning models in multi-institutional healthcare data.
• Accurate estimation of data distribution differences is key to successful federated learning for clinical applications like AKI prediction.