Application of Deep Learning Methods to SNOMED CT Encoding of Clinical Texts: From Data Collection to Extreme Multi-Label Text-Based Classification

Proceedings of Recent Advances in Natural Language Processing, pages 557–565 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_063 557 Application of Deep Learning Methods to SNOMED CT Encoding of Clinical Texts: From Data Collection to Extreme Multi-Label Text-Based Classification Anton Hristov1,2, Aleksandar Tahchiev1, Hristo Papazov3, Nikola Tulechki1, Todor Primov1 and Svetla Boytcheva 1,4 1Sirma AI (Ontotext), Bulgaria anton.hristov@ontotext.com, alexander.tahchiev@ontotext.com, nikola.tulechki@ontotext.com, todor.primov@ontotext.com, svetla.boytcheva@ontotext.com 2 Faculty of Mathematics and Informatics, Sofia University ”St. Kliment Ohridski”, Bulgaria 3 MIT, Mathematics PhD Student, USA h.g.papazov@gmail.com 4 Institute of Information and Communication Technologies, Bulgarian Academy of Sciences, Bulgaria Abstract Concept normalization of clinical texts to stan- dard medical classifications and ontologies is a task with high importance for healthcare and medical research. We attempt to solve this problem through automatic SNOMED CT en- coding, where SNOMED CT is one of the most widely used and comprehensive clini- cal term ontologies. Applying basic Deep Learning models, however, leads to undesir- able results due to the unbalanced nature of the data and the extreme number of classes. We propose a classification procedure that fea- tures a multiple-step workflow consisting of la- bel clustering, multi-cluster classification, and clusters-to-labels mapping. For multi-cluster classification, BioBERT is fine-tuned over our custom dataset. The clusters-to-labels map- ping is carried out by a one-vs-all classifier (SVC) applied to every single cluster. We also present the steps for automatic dataset gen- eration of textual descriptions annotated with SNOMED CT codes based on public data and linked open data. In order to cope with the problem that our dataset is highly unbalanced, some data augmentation methods are applied. The results from the conducted experiments show high accuracy and reliability of our ap- proach for prediction of SNOMED CT codes relevant to a clinical text. 1 Introduction The task of automatic encoding of clinical text with standard medical classifications and ontologies is with high importance for healthcare organizations and medical research. Truly, more than 80% of clinical documents are stored in free-text format. This paper presents a research effort in solving the problem of automatic encoding of textual descrip- tion of medical diagnoses with one of the most widely used (Lee et al., 2013) and comprehensive ontologies – the Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT)1. One of the most important characteristics of SNOMED CT, which makes it significantly different from the rest of the standard medical classifications, is that it is based on compositional grammar2. Another aspect of popularity and importance of SNOMED CT for health information is interoperability, that is discussed in (Peterson and Liu, 2020). SNOMED CT is well known for being one of the most com- prehensive medical ontologies, which makes the task of automatic encoding an extreme scale classi- fication task with more than 360,000 medical codes. Currently, this task has not been solved with suffi- cient accuracy for all possible classes. Usually the developed solutions cover restricted terminology from 10 to a couple of thousands terms (Gaudet- Blavignac et al., 2021). The compositional nature of the SNOMED CT codes gives us the …

Content truncated to 500 words