A federated learning framework with knowledge graph and temporal transformer for early sepsis prediction in multi-center ICUs

Early sepsis prediction: Early prediction of sepsis in ICU patients is important for improving survival rates.

Multi-center data fragmentation: Accurate predictive model development is hampered by fragmented data across healthcare institutions.

Temporal clinical data complexity: Predictive modeling is challenged by the complex temporal nature of medical data.

Healthcare privacy constraints: Multi-institution clinical prediction faces stringent privacy constraints on patient data sharing.

Local distribution adaptation: A global model must rapidly adapt to local hospital data distributions.

Federated KG temporal transformer framework: The paper proposes a framework integrating federated learning with a medical knowledge graph, temporal transformer, and meta-learning for multi-center early sepsis prediction.

Privacy-preserving federated training: The approach trains models collaboratively across multiple hospitals without sharing raw patient data.

Knowledge graph clinical relationships: The model uses a knowledge graph to incorporate structured medical relationships.

Temporal transformer time-series modeling: The model employs a temporal transformer to capture long-range dependencies in clinical time-series data.

MAML local adaptation: The framework incorporates model-agnostic meta-learning to adapt the global model to local data distributions.

High sepsis prediction AUC:

Improves over centralized models:

Improves over standard FL: