Aligning Federated Learning with Existing Trust Structures in Health Care Systems

Imrana Abdullahi Yari, René Raab, Arne Küderle, Bjoern M. EskofierInternational Journal of Environmental Research and Public Health 2023 · DOI

TL;DR

This paper introduces CareNetFL, a novel federated learning (FL) framework designed for patient-centered health information systems (PHSs) that leverage peer-to-peer (P2P) architectures. CareNetFL addresses the challenges of trust, scalability, and data heterogeneity in decentralized health data by integrating FL with existing healthcare trust structures, specifically by involving medical practitioners as trusted aggregators.

What Problem Does This Paper Address?

The paper tackles the challenge of effectively utilizing machine learning (ML) on decentralized health data stored in patient-centered health information systems (PHSs). While PHSs, often built on peer-to-peer (P2P) networks, offer enhanced data sovereignty and resilience, they present significant hurdles for traditional ML training due to data fragmentation, heterogeneity, and privacy concerns. Existing federated learning (FL) approaches, which allow decentralized training without centralizing raw data, often rely on centralized aggregation servers or random entities, raising concerns about trust, security, and potential malicious actors within the healthcare context. Furthermore, the multi-tier and hybrid architectures common in PHSs are not well-supported by standard FL designs. The core problem is how to implement FL in these P2P PHS environments in a way that aligns with existing healthcare workflows, builds trust among stakeholders, and ensures robust security and performance.

Methods

The authors propose CareNetFL, a conceptual federated learning framework. The key innovation lies in its integration with existing healthcare trust structures. Instead of relying on random or centralized entities for model aggregation, CareNetFL designates entrusted parties, specifically medical practitioners, as local aggregators.

Here's how it works:

  • Hierarchical Aggregation: CareNetFL adopts a hierarchical approach. PHS providers act as global orchestrators, while practitioners serve as local aggregators for their patient groups.
  • Patient-Practitioner Clustering: Patients are grouped and associated with their practitioners. This clustering helps to mitigate issues arising from imbalanced and skewed data distributions, as practitioners have a better understanding of their patients' specific conditions and data.
  • Leveraging Trust Structures: By involving practitioners, the framework capitalizes on the existing trust relationship between patients and their doctors. This makes the aggregation process more secure and trustworthy, as practitioners are already invested in patient care and are subject to professional and ethical standards.
  • Identity-Based Authentication: The framework emphasizes the use of identity-based systems for end-to-end security and accountability throughout the FL lifecycle, ensuring that model updates originate from verified sources.
  • Model Life Cycle: CareNetFL defines a comprehensive model life cycle, including phases for certification, study definition, user participation, pre-training, identification, association, selection, hierarchical aggregation, version control, and deletion.

Datasets

Not discussed in the paper. The paper focuses on a conceptual framework and does not present empirical results from training on specific datasets.

Key Insights and Ideas

  • Practitioners as Trusted Aggregators: The most significant insight is the shift from random or centralized aggregators to trusted medical practitioners. This directly addresses the trust deficit in FL for healthcare by aligning with established patient-doctor relationships.
  • Bridging the Gap Between FL and Healthcare Workflows: CareNetFL explicitly aims to integrate FL into existing healthcare processes, rather than forcing healthcare to adapt to a new, potentially untrusted, FL paradigm. This is crucial for adoption.
  • Addressing P2P PHS Architecture: The framework is designed to be compatible with the multi-tier and hybrid architectures of P2P PHSs, which are often overlooked by traditional FL designs.
  • Mitigating Data Heterogeneity: Patient-practitioner clustering is a clever way to reduce the impact of non-IID (non-independent and identically distributed) data, a common challenge in FL, especially with diverse patient populations.
  • Enhanced Security and Accountability: By leveraging identity-based systems and trusted practitioners, the framework aims to provide stronger security and accountability compared to systems that rely solely on anonymization or less trusted aggregation mechanisms.

Key Results

Not discussed in the paper. As this is a conceptual framework, no quantitative or qualitative results from experiments are presented. The paper outlines the proposed architecture and its potential benefits.

Limitations and Future Work

  • Conceptual Framework: The primary limitation is that CareNetFL is a conceptual design. It has not yet been implemented or empirically validated.
  • Computational Demands: The paper acknowledges that implementing concepts like identity authentication, user-to-user mapping, and model update management at the PHS provider network will require substantial computational power and resources.
  • Privacy Leakage Risk: While the framework aims to enhance privacy, there's a potential risk of privacy leakage when the learning orchestrator accesses patient metadata and training model parameters for user selection.
  • Future Work: The authors suggest implementing CareNetFL with suitable datasets and conducting experiments to evaluate its performance against benchmarks. They also plan to explore various aggregation methods and privacy-preserving techniques for a concrete implementation. Further research will focus on the practical challenges of integrating the proposed concepts into real-world PHS networks.

Summary updated

Abstract

Patient-centered health care information systems (PHSs) on peer-to-peer (P2P) networks (e.g., decentralized personal health records) enable storing data locally at the edge to enhance data sovereignty and resilience to single points of failure. Nonetheless, these systems raise concerns on trust and adoption in medical workflow due to non-alignment to current health care processes and stakeholders' needs. The distributed nature of the data makes it more challenging to train and deploy machine learning models (using traditional methods) at the edge, for instance, for disease prediction. Federated learning (FL) has been proposed as a possible solution to these limitations. However, the P2P PHS architecture challenges current FL solutions because they use centralized engines (or random entities that could pose privacy concerns) for model update aggregation. Consequently, we propose a novel conceptual FL framework, CareNetFL, that is suitable for P2P PHS multi-tier and hybrid architecture and leverages existing trust structures in health care systems to ensure scalability, trust, and security. Entrusted parties (practitioners' nodes) are used in CareNetFL to aggregate local model updates in the network hierarchy for their patients instead of random entities that could actively become malicious. Involving practitioners in their patients' FL model training increases trust and eases access to medical data. The proposed concepts mitigate communication latency and improve FL performance through patient-practitioner clustering, reducing skewed and imbalanced data distributions and system heterogeneity challenges of FL at the edge. The framework also ensures end-to-end security and accountability through leveraging identity-based systems and privacy-preserving techniques that only guarantee security during training.
BibTeX
@article{yari2023aligning,
  author = {Imrana Abdullahi Yari and René Raab and Arne Küderle and Bjoern M. Eskofier},
  title = {{Aligning Federated Learning with Existing Trust Structures in Health Care Systems}},
  journal = {International Journal of Environmental Research and Public Health},
  year = {2023},
  doi = {10.3390/ijerph20075378},
}
Figure