Title: Secure Multi-Center Sepsis Early Warning via Federated Learning

Original: arXiv:2606.04338v1 Announce Type: new

Abstract: The decentralized nature and privacy concerns associated with multi-center medical data present significant hurdles to the creation of accurate, centralized models for early sepsis detection. Federated learning (FL) has emerged as a compelling approach for collaborative model development, enabling various institutions to jointly train predictive algorithms without the need to share or aggregate raw patient information. However, the actual efficacy, resilience, and privacy safeguards of this method have not been thoroughly assessed using authentic clinical data. To address this deficiency, our research provides a comprehensive evaluation of federated learning in the context of multi-center sepsis forecasting. The study utilizes a dataset comprising 648 clinically vetted samples sourced from three major tertiary hospitals in China, adhering to strict inclusion and exclusion protocols. We first established a centralized training model as a benchmark for performance comparison, subsequently deploying a horizontal federated learning architecture for distributed collaborative modeling. Our extensive experiments reveal that the FL-based model delivers prediction accuracy nearly on par with the centralized approach, all while effectively preventing privacy breaches. Additionally, security analyses confirm that the transmitted model parameters are resistant to reconstruction attacks, as malicious actors cannot recover original patient records from them. This study not only confirms the viability and security of employing federated learning for clinical sepsis prediction but also offers a robust and practical framework for privacy-conscious collaboration across multiple medical centers.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC