FairFML: Fair Federated Machine Learning with a Case Study on Reducing Gender Disparities in Cardiac Arrest Outcome Prediction

Federated learning can reduce bias while protecting privacy: FairFML demonstrates that healthcare institutions can collaborate on machine learning models without sharing raw patient data, while simultaneously addressing algorithmic fairness—solving two critical challenges in healthcare AI at once.

Significant fairness improvements are achievable: The framework improved model fairness by up to 65% compared to centralized models in the cardiac arrest prediction case study, showing that fairness enhancements are not merely theoretical but measurable and substantial in real clinical applications.

Fairness doesn't require sacrificing accuracy: A key finding is that FairFML maintained predictive performance comparable to both local and centralized models while improving fairness, addressing a common concern that fair AI systems must trade off accuracy for equity.

The solution is flexible and adaptable: FairFML is model-agnostic, meaning it can work with different types of machine learning approaches—from traditional statistical methods to deep learning—making it practical for diverse healthcare institutions with different technical infrastructures.

Gender disparities in medical AI are addressable: The cardiac arrest outcome prediction case study specifically demonstrates that gender-based algorithmic disparities in healthcare can be identified and reduced through this framework, with direct implications for equitable patient care.