GraphEHR: Heterogeneous Graph Neural Network for Electronic Health Records
Track: Work in Progress
Categories: Hospitals / Emergency Services
Keywords: Heterogeneous Graph Neural Networks, Electronic Health Records, Hospitals
TL;DR: GraphEHR employs Heterogeneous Graph Neural Networks to predict multiple diseases from Electronic Health Records, enhancing hospital decision-making by considering all patient characteristics.
Abstract: Electronic Health Records (EHR) are valuable for patient analysis and disease prediction using deep learning algorithms. However, previous approaches focused on single predictive tasks and feature selection, which may not be optimal in real clinical settings where all patient characteristics, including historical data, treatments, and past illnesses, are important. To address this limitation, we propose \emph{\textbf{GraphEHR}}, which leverages Heterogeneous Graph Neural Networks (HGNNs) for various predictive tasks using EHR data. \emph{\textbf{GraphEHR}} aims to comprehensively predict multiple diseases by modeling all patient characteristics through a novel graph-based patient embedding. This approach effectively captures complex relationships within EHR and enables predictions across a wide range of diseases. In comparative experiments using the MIMIC-III EHR database, encompassing up to 13 predictive tasks,
\emph{\textbf{GraphEHR}} showcased its capability to grasp the complexity and multi-dimensional nature of EHR when compared against multiple baseline models across various tasks. By considering all aspects of a patient's medical history, this holistic modeling approach enhances clinical decision-making and facilitates patient management.
Submission Number: 7
Loading