Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction

In three linesReal-world deployment of federated learning pipeline for iron deficiency prediction from full blood count data. Uses DeepCBC (frozen haematology foundation model) + FedMAP (personalised aggregation). Tested across two clinical sites (AUMC, NHSBT) with non-IID data. FedMAP improves ROC-AUC from 0.947→0.959 (AUMC) and 0.856→0.867 (NHSBT) versus local-only training.

Summary generated by Claude — human-verified