Federated Patient Hashing

Jie Xu, Zhenxing Xu, Peter B. Walker, Fei WangAAAI 2020 · DOI

Abstract

Federated Patient Hashing (FPH) is a novel framework designed to enable collaborative patient retrieval across multiple medical institutions without compromising data privacy. By integrating federated learning with multi-modal hashing, FPH generates compact binary "fingerprints" from heterogeneous clinical data. This approach allows for fast, accurate similarity matching across distributed networks while ensuring that sensitive patient records remain strictly local to their home institutions.

Introduction

In modern clinical research, identifying similar patient cohorts is essential for trial construction, disease tracking, and personalized medicine. However, two primary obstacles hinder this process: privacy silos and data complexity. Patient data is protected by strict regulations like HIPAA, preventing hospitals from pooling raw data into central repositories. Furthermore, medical records are inherently multi-modal, consisting of structured lab results and unstructured clinical notes, making large-scale similarity searches computationally intensive.

Federated Patient Hashing addresses these challenges by learning to map complex, multi-modal data into short, searchable binary codes within a federated learning environment. This ensures that only model updates—never raw data—are shared between institutions.

Methodology

The FPH framework utilizes a specialized objective function and distributed optimization strategies to train robust retrieval models.

The Objective Function

To ensure the utility of the generated hash codes, the model minimizes two distinct losses:

  • Similarity Preserving Loss ($L_s$): This ensures that patients with similar clinical profiles are assigned similar binary codes, facilitating accurate retrieval.
  • Heterogeneity Digging Loss ($L_h$): This aligns different data modalities (e.g., structured charts and unstructured notes) for the same patient, ensuring they converge on a single, consistent binary representation.

Federated Optimization Strategies

The authors developed two primary architectures for cross-hospital training:

  1. Centralized (Federated SMD): A central server coordinates updates using Stochastic Mirror Descent based on Bregman divergence, aggregating gradients from local hospitals to update a global model.
  2. Decentralized: Institutions update the model parameters sequentially in a peer-to-peer fashion, removing the need for a central coordinator.
Centralized and Decentralized federated learning architectures for hospital networks.

Experimental Evaluation

The framework was evaluated using the MIMIC-III dataset, focusing on the retrieval of patients based on the onset and severity of Acute Kidney Injury (AKI). The study involved over 16,000 ICU stays and utilized both structured data and clinical notes processed via Latent Semantic Analysis.

Key Results

FPH was benchmarked against several state-of-the-art hashing methods, including Linear Cross-modal Hashing (LCH), Semantic Correlation Maximization (SCM), and MLP-based hashing.

Performance comparison of FPH against baseline hashing methods across different bit lengths and retrieval counts.
  • Superior Accuracy: FPH consistently achieved higher Mean Average Precision (MAP) than all baselines. For instance, on the AKI onset task, FPH reached ~53.7% MAP, significantly outperforming traditional methods like LCH (~46.6%).
  • Efficiency and Scalability: The model maintains high performance even with short binary codes (16 to 64 bits), enabling lightning-fast retrieval in large-scale databases.
  • Robustness: The decentralized training strategy performed nearly as well as the centralized one, proving that hospitals can collaborate effectively without a central middleman.

Discussion and Future Work

FPH demonstrates that hashing can serve as an effective privacy layer; since hash codes are non-invertible, they provide an additional safeguard against data reconstruction. While the current results are promising, future research could focus on reducing communication overhead during gradient exchange and integrating more advanced NLP encoders, such as BERT, to better capture the nuances of clinical narratives. Additionally, addressing the requirement for zero-centered data through privacy-preserving global mean calculation remains an area for further refinement.

Summary updated

BibTeX
@inproceedings{xu2020federated,
  author = {Jie Xu and Zhenxing Xu and Peter B. Walker and Fei Wang},
  title = {{Federated Patient Hashing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.6121},
}
Figure