Read the complete story in our latest article.

Mitigating algorithmic disparities is a critical challenge in healthcare research, where ensuring equity and fairness is paramount. While large-scale healthcare data exist across multiple institutions, cross-institutional collaborations often face privacy constraints, highlighting the need for privacy-preserving solutions that also promote fairness. In this study, we present Fair Federated Machine Learning (FairFML), a model-agnostic solution designed to reduce algorithmic bias in cross-institutional healthcare collaborations while preserving patient privacy. As a proof of concept, we validated FairFML using a real-world clinical case study on reducing gender disparities in cardiac arrest outcome prediction. We demonstrate that the proposed FairFML framework enhances fairness in federated learning (FL) models without compromising predictive performance. Our findings show that FairFML improves model fairness by up to 65% compared to the centralized model, while maintaining performance comparable to both local and centralized models, as measured by the receiver operating characteristic analysis. FairFML offers a promising and flexible solution for FL collaborations, with its adaptability allowing seamless integration with various FL frameworks and models, from traditional statistical methods to deep learning techniques. This makes FairFML a robust approach for developing fairer models across diverse clinical and biomedical applications.

The implementation was developed and tested with Python 3.11.
Install required Python packages by running

pip install -r requirements.txt

We partitioned Adult dataset from the UCI repository into a total of five clients with gender being the sensitive attribute.

The following commands in this example are to be run in the code directory.

cd code

python baseline_models.py

The central model is trained using pooled training data and is evaluated using testing data at each client. Local models are trained and evaluated independently at each client.

python utils/tune_local_lambda_adult.py

The raw outputs (.txt files) and the summary of results (.csv files) are saved in code/outputs/adult/local.

python utils/tune_FL_gamma_adult.py [FL strategy]

For example, to use personalized FedAvg (PerAvg), run

python utils/tune_FL_gamma_adult.py PerAvg

The raw outputs (.txt files) and the summary of results on validation set (.csv files) are saved in code/outputs/adult/FL/PerAvg. The trained models are saved in code/outputs/adult/models/group/PerAvg.

python utils/evaluate_adult_test.py [FL strategy]

For example, to use PerAvg, run

python utils/evaluate_adult_test.py PerAvg

For PerAvg, both server-side and client-side models can be evaluated. Results will be saved in code/outputs/adult/FL/PerAvg as test_results_server_model.csv and test_results_client_model.csv. Results for each lambda value will also be saved in each directory (e.g. code/outputs/adult/FL/PerAvg/lambda_1/test_result_lambda1_server_model.csv).

Currently, FairFML is only implemented for FedAvg and PerAvg based on PFLlib. Other FL algorithms available in PFLlib can be adapted by incorporating fairness penalty in the loss function. Check out our code for details.

• Siqi Li (Email: [email protected])
• Qiming Wu (Email: [email protected])
• Nan Liu (Email: [email protected])