FL4Health: Private Federated Clinical AI

Recent advancements in generative AI and large language models rely heavily on vast, diverse datasets for optimal performance. However, despite the fact that more data typically leads to improved performance, industries like healthcare are constrained by stringent privacy regulations that prohibit centralized data sharing. Federated learning (FL) addresses this challenge by keeping data local, prioritizing privacy, and securely enabling collaboration through the exchange of model updates rather than raw data. FL operates iteratively, starting with local models initialized from a shared server model. These models are trained on local data and send weight updates back to the server for aggregation, enabling global model improvement while keeping sensitive data secure.
FL4Health is an open-source FL engine developed by members of the AI Engineering team at the Vector Institute. It is designed to streamline cross-silo FL research and deployment, with a particular focus on health applications. In cross-silo federated learning, training participants typically include a small number of reliable institutions with sufficient computational resources, making it ideal for collaborative efforts in domains like healthcare and health applications. FL4Health implements a straightforward user-friendly approach with composable modules. Its core components include a range of distributed optimization strategies, advanced privacy-preserving methods and efficient communication protocols to ease researchers’ use and implementation of state-of-the-art techniques for distributed training with minimal need for configuration, speeding up the experimental phase.
On top of the FL4Health engine, we are also developing a high-level interface called FLorist to simplify federated training and deployment for clinical practitioners, who may have limited coding experience. With FLorist, users can initiate model training by simply providing configurations to the web-service, making it easier for institutions to adopt federated learning effectively and seamlessly in real-world clinical settings. Such capabilities make FL4Health a valuable tool for advancing healthcare AI in a secure, scalable, and collaborative manner.