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

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Health & Wellbeing — Clinical Care & Medical Practice, Medical Devices & Health Technology, Health & Medical Research · Depth: Expert, extended

Summary

An embedding-based Federated Learning (FL) pipeline for iron deficiency prediction from full blood count (FBC) data was developed and deployed across Amsterdam University Medical Centre (AUMC) and NHS Blood and Transplant (NHSBT). The system utilizes a frozen domain-specific haematology foundation model, DeepCBC, for site-local representation extraction, restricting federated training to a compact downstream classifier and significantly reducing communication overhead. The two clinical datasets are structurally non-IID due to distinct population differences. Runtime governance is enforced by FLA3, a healthcare-oriented FL platform. While standard FedAvg reduced ROC-AUC at both sites, FedMAP, a personalized aggregation method, raised ROC-AUC from 0.9470 to 0.9594 at AUMC and from 0.8558 to 0.8671 at NHSBT relative to local-only training, achieving the highest macro ROC-AUC of 0.9133 and best macro balanced accuracy overall.

Key takeaway

For AI Scientists and MLOps Engineers deploying federated learning in healthcare, you should prioritize personalized aggregation methods like FedMAP when dealing with structurally heterogeneous clinical data, as sample-size weighting can degrade performance. Additionally, integrate runtime governance platforms such as FLA3 to ensure policy-based authorization and auditable logging, which are critical for operational deployment across diverse regulatory environments like AUMC and NHSBT.

Key insights

Personalized aggregation and runtime governance are crucial for effective, deployable healthcare federated learning with heterogeneous data.

Principles

• Structural heterogeneity is a clinical property, not just statistical noise.
• Embedding-based FL reduces communication and simplifies deployment.
• Personalized aggregation benefits federations with divergent client relevance.

Method

A two-stage FL design uses a frozen DeepCBC foundation model for local FBC data embedding, then federatively trains a compact MLP classifier using FedMAP for personalized aggregation, all governed by FLA3.

In practice

• Use DeepCBC for FBC data embedding extraction.
• Deploy FLA3 for policy-driven runtime governance.
• Implement FedMAP for non-IID clinical federations.

Topics

• Federated Learning
• Healthcare AI
• Iron Deficiency Prediction
• DeepCBC
• FLA3
• Personalized Aggregation
• Runtime Governance

Code references

• CambridgeCIA/FedMAP
• bloodcounts/FLAAA

Best for: Research Scientist, AI Scientist, MLOps Engineer, AI Engineer

Related on AIssential

• Federated Learning for Multi-Center Sepsis Early Prediction with Privacy-Preserving — Federated learning enables accurate multi-center sepsis prediction while robustly preserving patient data privacy against reconstruction attacks.
• When To Adapt? Adapting the Model or Data in Federated Medical Imaging — Adaptation strategy effectiveness in federated medical imaging depends on the type and magnitude of domain shift.
• dsLassoCov: a federated Lasso approach incorporating covariate control — dsLassoCov efficiently controls covariates in federated learning for high-dimensional biomedical data, enabling privacy-preserving analysis.
• A Reproducible Log-Driven AutoML Framework for Interpretable Pipeline Optimization in Healthcare Risk Prediction — Reproducible, log-driven AutoML reveals healthcare risk prediction pipeline component importance and search space redundancy.
• Federated orthogonal learning for detection of liver lesions from multi-phase contrast-enhanced CT images — FedOG enhances federated learning for medical image segmentation by using orthogonal gradient decomposition to counter data heterogeneity and incomplete phases.
• IEEE Transactions on Artificial Intelligence, Volume 7, Issue 6, June 2026 — AI research in June 2026 emphasizes robustness, privacy, efficiency, and specialized applications across diverse domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.