B0898
Title: FedScore: A privacy-preserving framework for federated scoring system development
Authors: Siqi Li - Duke-NUS Medical School (Singapore) [presenting]
Abstract: Federated learning (FL) is gaining popularity in healthcare research for the integration of information from multiple sites while safeguarding data privacy. However, most FL applications are for black-box models, while interpretable models are predominantly developed using single-source data, limiting their applicability to other sites. To address this issue, FedScore is proposed, a first-of-its-kind framework for building federated scoring systems across multiple sites. The FedScore framework comprises five modules: federated variable ranking, federated variable transformation, federated score derivation, federated model selection, and federated model evaluation. For a proof-of-concept, FedScore is applied to develop a hypothetical federated scoring system for predicting mortality within 30 days after an emergency department visit. To accomplish this, the ED data, collected from a tertiary hospital in Singapore, is artificially partitioned into 10 simulated sites. 10 local scoring systems and a pooled scoring system based on centralized data are built, using a pre-existing scoring system for benchmark comparison. The FedScore model exhibited notable accuracy and stability, achieving an average area under the curve closest to the pooled model and a standard deviation lower than most local models. FedScore fills a gap in medical research and could serve as a foundation for creating reliable clinical scoring systems that protect data privacy in a variety of medical contexts.
