MEGACare: knowledge-guided multi-view hypergraph predictive framework for healthcare

• Main Authors: Wu, Jialun, He, Kai, Mao, Rui, Li, Chen, Cambria, Erik
• Other Authors: School of Computer Science and Engineering
• Format: Article
• Language: English
• Published: 2023
• Subjects: Engineering::Computer science and engineering; Electronic Health Record; Healthcare
• Online Access: https://hdl.handle.net/10356/171200
• Institution: Nanyang Technological University

Predicting a patient's future health condition by analyzing their Electronic Health Records (EHRs) is a trending subject in the intelligent medical field, which can help clinicians prescribe safely and effectively, and also make more accurate diagnoses. Benefiting from powerful feature extraction capabilities, graph representation learning can capture complex relationships and achieve promising performance in many clinical prediction tasks. However, existing works either exclusively consider single domain knowledge with an independent task or do not fully capitalize on domain knowledge that can provide more predictive signals in the code encoding stage. Moreover, the heterogeneous and high-dimensional nature of EHR data leads to a deficiency of hardly encoding implicit high-order correlations. To address these limitations, we proposed a knowledge-guided Multi-viEw hyperGrAph predictive framework (MEGACare) for diagnosis prediction and medication recommendation. Our MEGACare leveraged multi-faceted medical knowledge, including ontology structure, code description, and molecular information to enhance medical code presentations. Furthermore, we constructed an EHR hypergraph and a multi-view learning framework to capture the high-order correlation between patient visits and medical codes. Specifically, we propose three perspectives around the pairwise relationship between patient visits and medical codes to comprehensively learn patient representation and enhance the robustness of our framework. We evaluated our MEGACare framework against a set of state-of-the-art methods for two clinical outcome prediction tasks in the public MIMIC-III dataset, and the results showed that our proposed framework was superior to the baseline methods.