FedMAP: Personalised Federated Learning for Real Large-Scale Healthcare Systems

Fan Zhang, Carlos Esteve, Sören Dittmer, Carola‐Bibiane Schönlieb, Michael S. Roberts, Ip, Samantha, Consortium, BloodCounts!, de Wit, Norbert C. J., Wood, Angela, Rudd, James Hf, Lane, Nicholas, Gleadall, Nicholas S, Schönlieb, Carola-Bibiane, Roberts, Michael

Abstract

Federated learning (FL) promises to enable collaborative machine learning across healthcare sites whilst preserving data privacy. Practical deployment remains limited by statistical heterogeneity arising from differences in patient demographics, treatments, and outcomes, and infrastructure constraints. We introduce FedMAP, a personalised FL (PFL) framework that addresses heterogeneity through local Maximum a Posteriori (MAP) estimation with Input Convex Neural Network priors. These priors represent global knowledge gathered from other sites that guides the model while adapting to local data, and we provide a formal proof of convergence. Unlike many PFL methods that rely on fixed regularisation, FedMAP's prior adaptively learns patterns that capture complex inter-site relationships. We demonstrate improved performance compared to local training, FedAvg, and several PFL methods across three large-scale clinical datasets: 10-year cardiovascular risk prediction (CPRD, 387 general practitioner practices, 258,688 patients), iron deficiency detection (INTERVAL, 4 donor centres, 31,949 blood donors), and mortality prediction (eICU, 150 hospitals, 44,842 patients). FedMAP incorporates a three-tier design that enables participation across healthcare sites with varying infrastructure and technical capabilities, from full federated training to inference-only deployment. Geographical analysis reveals substantial equity improvements, with underperforming regions achieving up to 14.3% performance gains. This framework provides the first practical pathway for large-scale healthcare FL deployment, which ensures clinical sites at all scales can benefit, equity is enhanced, and privacy is retained.

Introduction

Federated Learning (FL) is often considered the optimal solution for healthcare AI, allowing models to learn from sensitive data across multiple institutions without compromising patient privacy. However, a significant gap exists between theoretical research and clinical reality; only approximately 5% of reported FL implementations reach real-world use. This gap is primarily driven by two factors: statistical heterogeneity (non-IID data) across diverse patient populations and significant disparities in technical infrastructure between large teaching hospitals and smaller community clinics.

Methodology: Adaptive Priors and Tiered Deployment

FedMAP addresses these challenges by reframing personalised federated learning (PFL) as a bi-level optimization problem rooted in Maximum a Posteriori (MAP) estimation.

Learnable Priors via ICNNs

Unlike traditional FL methods that use rigid, fixed penalties to anchor local models to a global average, FedMAP utilizes Input Convex Neural Networks (ICNNs) to parameterize a global "prior." This prior adaptively learns complex relationships between sites, acting as a smart regularizer that understands how specific sites deviate from the norm. The mathematical convexity of ICNNs allows for a formal proof of convergence, ensuring stability in complex clinical environments.

FedMAP Overview: The framework consists of a two-step optimization process (A) and a three-tier deployment architecture (B) to accommodate varying site capabilities.

Three-Tier Architecture

To overcome infrastructure barriers, FedMAP introduces a flexible three-tier participation model:

  • Tier 1 (Full FL): Large institutions participate in iterative training of both the global prior and local models.
  • Tier 2 (Local Fine-tuning): Medium-sized sites utilize a pre-trained global prior to perform a one-time local optimization.
  • Tier 3 (Inference-only): Resource-constrained sites with no training capacity deploy the global model directly for predictions.

Experimental Results and Clinical Impact

The framework was validated across three massive clinical datasets: CPRD (258,688 patients across 387 practices), INTERVAL (31,949 donors across 4 centers), and eICU (44,842 patients across 150 hospitals).

FedMAP consistently outperformed standard FedAvg and state-of-the-art PFL methods like pFedME and pFedBayes. In the CPRD cardiovascular risk prediction task, FedMAP achieved a median C-index of 0.717, representing a 3% improvement over the next-best method. Notably, even in Tier 2 scenarios, the framework provided a 6.8% improvement over standard baselines.

The "Robin Hood" Effect and Healthcare Equity

A key finding was the strong negative correlation between a site's standalone performance and its benefit from FedMAP. Smaller sites with limited data gained the most from joining the federation. Geographical analysis in the UK demonstrated that FedMAP significantly reduced regional performance disparities, with underperforming regions like North East England seeing gains of up to 14.3%, effectively leveling the quality of care across the country.

Discussion and Future Work

While FedMAP offers a robust pathway for large-scale deployment, it introduces higher computational overhead compared to simple averaging methods due to bi-level optimization. Future research will focus on improving the interpretability of the ICNN "black box" priors and integrating formal Differential Privacy (DP) guarantees. Additionally, addressing distribution shifts for Tier 3 sites remains a priority to ensure continued model reliability as clinical data evolves.

Summary updated

Abstract

Federated learning (FL) promises to enable collaborative machine learning across healthcare sites whilst preserving data privacy. Practical deployment remains limited by statistical heterogeneity arising from differences in patient demographics, treatments, and outcomes, and infrastructure constraints. We introduce FedMAP, a personalised FL (PFL) framework that addresses heterogeneity through local Maximum a Posteriori (MAP) estimation with Input Convex Neural Network priors. These priors represent global knowledge gathered from other sites that guides the model while adapting to local data, and we provide a formal proof of convergence. Unlike many PFL methods that rely on fixed regularisation, FedMAP's prior adaptively learns patterns that capture complex inter-site relationships. We demonstrate improved performance compared to local training, FedAvg, and several PFL methods across three large-scale clinical datasets: 10-year cardiovascular risk prediction (CPRD, 387 general practitioner practices, 258,688 patients), iron deficiency detection (INTERVAL, 4 donor centres, 31,949 blood donors), and mortality prediction (eICU, 150 hospitals, 44,842 patients). FedMAP incorporates a three-tier design that enables participation across healthcare sites with varying infrastructure and technical capabilities, from full federated training to inference-only deployment. Geographical analysis reveals substantial equity improvements, with underperforming regions achieving up to 14.3% performance gains. This framework provides the first practical pathway for large-scale healthcare FL deployment, which ensures clinical sites at all scales can benefit, equity is enhanced, and privacy is retained.
BibTeX
@misc{zhang2024fedmap,
  author = {Fan Zhang and Carlos Esteve and Sören Dittmer and Carola‐Bibiane Schönlieb and Michael S. Roberts and Ip, Samantha and Consortium, BloodCounts! and de Wit, Norbert C. J. and Wood, Angela and Rudd, James HF and Lane, Nicholas and Gleadall, Nicholas S and Schönlieb, Carola-Bibiane and Roberts, Michael},
  title = {{FedMAP: Personalised Federated Learning for Real Large-Scale Healthcare Systems}},
  year = {2024},
  eprint = {2405.19000},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.19000},
}
Figure