EHRFL: Federated Learning Framework for Heterogeneous EHRs and Precision-guided Selection of Participating Clients
arxiv(2024)
摘要
In this study, we provide solutions to two practical yet overlooked scenarios
in federated learning for electronic health records (EHRs): firstly, we
introduce EHRFL, a framework that facilitates federated learning across
healthcare institutions with distinct medical coding systems and database
schemas using text-based linearization of EHRs. Secondly, we focus on a
scenario where a single healthcare institution initiates federated learning to
build a model tailored for itself, in which the number of clients must be
optimized in order to reduce expenses incurred by the host. For selecting
participating clients, we present a novel precision-based method, leveraging
data latents to identify suitable participants for the institution. Our
empirical results show that EHRFL effectively enables federated learning across
hospitals with different EHR systems. Furthermore, our results demonstrate the
efficacy of our precision-based method in selecting reduced number of
participating clients without compromising model performance, resulting in
lower operational costs when constructing institution-specific models. We
believe this work lays a foundation for the broader adoption of federated
learning on EHRs.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要