Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction
Researchers have developed an embedding-based federated learning system for predicting iron deficiency using routine blood count data. This system was deployed across two distinct clinical environments, the Amsterdam University Medical Centre (AUMC) and NHS Blood and Transplant (NHSBT), to address challenges with non-IID data and communication overhead. The approach utilizes a pre-trained foundation model, DeepCBC, for local representation extraction and a personalized aggregation method called FedMAP, which improved prediction accuracy compared to standard aggregation and local-only training. AI
IMPACT Demonstrates a novel federated learning approach for clinical data, potentially improving diagnostic accuracy and privacy in healthcare AI.