FD-GATDR: A Federated-Decentralized-Learning Graph Attention Network for Doctor Recommendation Using EHR

TL;DR

This paper introduces FD-GATDR, a federated-decentralized graph attention network for doctor recommendation using EHR data, incorporating time embedding and privacy-preserving learning to improve accuracy and data security.

Contribution

It proposes a novel federated-decentralized learning approach with graph attention and time embedding for privacy-aware doctor recommendation from heterogeneous EHR data.

Findings

Improves AUC by up to 6.2% over baseline models.

Achieves convergence rate of O(1/T) in federated learning.

Effectively models heterogeneous EHR data with time embedding.

Abstract

In the past decade, with the development of big data technology, an increasing amount of patient information has been stored as electronic health records (EHRs). Leveraging these data, various doctor recommendation systems have been proposed. Typically, such studies process the EHR data in a flat-structured manner, where each encounter was treated as an unordered set of features. Nevertheless, the heterogeneous structured information such as service sequence stored in claims shall not be ignored. This paper presents a doctor recommendation system with time embedding to reconstruct the potential connections between patients and doctors using heterogeneous graph attention network. Besides, to address the privacy issue of patient data sharing crossing hospitals, a federated decentralized learning method based on a minimization optimization model is also proposed. The graph-based recommendation system has been validated on a EHR dataset. Compared to baseline models, the proposed method improves the AUC by up to 6.2%. And our proposed federated-based algorithm not only yields the fictitious fusion center's performance but also enjoys a convergence rate of O(1/T).