Exploring Deep Learning-based Approaches for Predicting Concept Names in SNOMED CT

Abstract

Ontologies or terminologies have been widely used as formal representation of biomedical knowledge. New concepts are constantly added to biomedical ontologies due to the evolving nature of biomedical knowledge. Much progress has been made to identify new concepts in SNOMED CT, the largest clinical healthcare terminology. However, proper naming of new concepts remains challenging and relies on the ontology curators’ manual effort. In this paper, we explore three deep learning-based approaches, given bags of words, to automatically predict concept names that comply with the naming convention of SNOMED CT. These deep learning models are simple neural network, Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN) combined with LSTM. Our experiments showed that LSTM-based approach achieved the best performance: a precision of 65.98%, a recall of 61.04%, and an F1 score of 63.41% for predicting concept names for newly added concepts in the March 2018 Edition of SNOMED CT. It also achieved a precision of 74.58%, a recall of 73.33%, and an F1 score of 73.95% for naming missing concepts identified by our previous work. Further examination of results revealed inconsistencies within SNOMED CT which may be leveraged for quality assurance purpose.