Mapping of Disease Names to Disease Codes based on Natural Language Processing Techniques

Anđelka Zečević^1*, Jovana Kovačević², and Radoslav Davidović³

¹Mathematical Institute, Serbian Academy of Sciences and Arts

²Faculty of Mathematics, University of Belgrade

³Institute of Nuclear Sciences Vinča

andjelkaz [at] mi.sanu.ac.rs

Abstract

Information aggregation from various gen, disease, and gen-disease databases such as DisGeNet, COSMIC, HumsaVar, Orphanet, ClinVar, HPO, and Diseases into a unique database would enable researchers to analyze and compare valuable domain findings in a more convenient and systematic way. However, the aggregation poses numerous challenges due to non-uniform information annotation across the databases. In this work, we address the problem of mapping a disease name, when needed, into a standardized disease code (DOID) based on Natural Language Processing text representation techniques. We examine the benefits and limitations of using off-the-shelf embeddings such as Med2vec, and language models such as BioBERT, UmlsBERT, and PubMedBERT in retrieval scenarios with respect to standard full-text search. In addition to qualitative improvements, we elaborate on the technical requirements and computational complexities that come with the embracement of language models and semantic search.