Clinical Note Owns its Hierarchy: Multi-Level Hypergraph Neural Networks for Patient-Level Representation Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5559–5573 July 9-14, 2023 ©2023 Association for Computational Linguistics Clinical Note Owns its Hierarchy: Multi-Level Hypergraph Neural Networks for Patient-Level Representation Learning Nayeon Kim1*, Yinhua Piao2*, and Sun Kim1,2,3,4 1 Interdisciplinary Program in Artificial Intelligence, Seoul National University 2 Department of Computer Science and Engineering, Seoul National University 3 Institute of Computer Technology, Seoul National University 4 AIGENDRUG Co., Ltd. {ny_1031, 2018-27910, sunkim.bioinfo}@snu.ac.kr Abstract Leveraging knowledge from electronic health records (EHRs) to predict a patient’s condition is essential to the effective delivery of appropri- ate care. Clinical notes of patient EHRs contain valuable information from healthcare profes- sionals, but have been underused due to their difficult contents and complex hierarchies. Re- cently, hypergraph-based methods have been proposed for document classifications. Directly adopting existing hypergraph methods on clin- ical notes cannot sufficiently utilize the hier- archy information of the patient, which can degrade clinical semantic information by (1) frequent neutral words and (2) hierarchies with imbalanced distribution. Thus, we propose a taxonomy-aware multi-level hypergraph neural network (TM-HGNN), where multi-level hy- pergraphs assemble useful neutral words with rare keywords via note and taxonomy level hy- peredges to retain the clinical semantic infor- mation. The constructed patient hypergraphs are fed into hierarchical message passing layers for learning more balanced multi-level knowl- edge at the note and taxonomy levels. We val- idate the effectiveness of TM-HGNN by con- ducting extensive experiments with MIMIC-III dataset on benchmark in-hospital-mortality pre- diction.1 1 Introduction With improvement in healthcare technologies, elec- tronic health records (EHRs) are being used to mon- itor intensive care units (ICUs) in hospitals. Since it is crucial to schedule appropriate treatments for patients in ICUs, there are many prognostic models that use EHRs to address related tasks, such as in- hospital mortality prediction. EHRs consist of three types of data; structured, semi-structured, and un- structured. Clinical notes, which are unstructured data, contain valuable comments or summary of the *These authors contributed equally to this work. 1Our codes and models are publicly available at: https://github.com/ny1031/TM-HGNN … resp: pt arrived intubated ---pt intub by Latent space of our graph 𝓅1 𝓅2 𝓃1 𝓉1 𝓃2 𝓉1 𝓃3 𝓉2 𝓃1 𝓉1 𝓃2 𝓉2 𝓃3 𝓉2 𝓃1 𝓉1 𝓃2 𝓉1 𝓃3 𝓉2 𝓃1 𝓉1 𝓃2 𝓉1 𝓃3 𝓉2 Taxonomy Word node Note hyperedge Taxonomy hyperedge (𝑎���) (𝑎) Latent space of hypergraph rhythm fibrillation benadryl benadryl rhythm fibrillation rhythm fibrillation benadryl rhythm fibrillation benadryl Figure 1: (a) Examples of patient clinical notes with difficult contents (e.g. jargons and abbreviations) and complex structures. Patient p1 owns notes of radiol- ogy taxonomy (pink) and nursing taxonomy (blue). (b) Differences between existing hypergraphs and our pro- posed multi-level hypergraphs. patient’s condition written by medical professionals (doctors, nurses, etc.). However, compared to struc- tured data, clinical notes have been underutilized in previous studies due to the difficult-to-understand contents and the complex hierarchies (Figure 1(a)). Transformer-based (Vaswani et al., 2017) methods like ClinicalBERT (Alsentzer et al., 2019; Huang …

Content truncated to 500 words