Improving Fairness in AI Models on Electronic Health Records: The Case for Federated Learning Methods.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Prioritizing prediction performance in health AI often overlooks potential biases.
• Federated learning (FL) is a popular paradigm in healthcare, but fairness-focused FL methods require further investigation.
• Existing FL approaches for healthcare lack comprehensive bias mitigation strategies.

Purpose of the Study:

• To propose and evaluate a novel federated learning approach for mitigating bias in healthcare AI models.
• To address the underinvestigated aspects of fairness-aware FL models in the healthcare domain.
• To offer a practical solution for bias concerns in electronic health record (EHR) data analysis.

Main Methods:

• Developed a comprehensive federated learning (FL) framework incorporating adversarial debiasing.
• Implemented a fair aggregation method compatible with various fairness metrics.
• Applied the approach to healthcare data, specifically electronic health records (EHRs).

Main Results:

• The proposed FL approach demonstrated superior fairness performance compared to baseline methods.
• The method achieved bias mitigation with minimal impact on overall model accuracy.
• Empirical results simulated large-scale, real-world healthcare scenarios, validating the approach's effectiveness.

Conclusions:

• The developed FL approach effectively mitigates bias in healthcare AI models.
• Federated learning offers implicit benefits for data imbalance and size in healthcare applications.
• This method provides a practical and promising solution for fair and accurate health AI.