Federated learning algorithms to overcome statistical and algorithmic bias and privacy concerns in machine learning for health

Project Summary Federated learning has emerged as a promising technique in biomedical research, providing the potential to construct robust common machine learning models with datasets from multiple institutions without having to share data among groups. However, the current implementations of this technique present several challenges that must be addressed before it can be widely adopted in the field of biomedicine. These challenges include issues related to bias and data heterogeneity, privacy and security, and interoperability among institutions. To mitigate these challenges, it is essential to integrate domain-specific knowledge with the theoretical advances in computer science and statistics for bias correction and privacy preservation. In this regard, we propose new mathematical and algorithmic approaches that can make federated learning a practical reality in biomedical research. These methods include developing domain-specific add-on algorithms to traditional federated learning frameworks, and leveraging techniques such as optimal transport and secure multiparty computation to overcome statistical and algorithmic bias, and privacy concerns. To this end, we will first develop a framework to calculate optimal transport cost of the data distributions across sites. This cost will be used to determine whether addition of a site to the federated learning is beneficial to the resulting model. We will then use this cost during the learning process as a regularization in order to take into account different data distributions in the resulting model. We will also use the metadata from the sites in the federated averaging process to avoid bias in the resulting model. Lastly, we will develop federated data pre-processing frameworks using secure multiparty computation to overcome the privacy issues.