Named entity recognition in the medical domain

• Main Author: Kusalavan, Kirubhaharini
• Other Authors: Chng Eng Siong
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Data
• Online Access: https://hdl.handle.net/10356/165225
• Institution: Nanyang Technological University

This report presents a project that aims to develop Named Entity Recognition (NER) models for two datasets in the medical domain: emergency hotline data and N2C2 clinical notes. The project objectives include reviewing existing models and architectures, training on general datasets to find the best model and architecture, and creating a pipeline to train and deploy NER models for different domains. The NER models will be used to auto-fill forms during emergencies by detecting entities in call transcripts and to identify essential entities from clinical notes. The paper also discusses the sampling method used to derive subsets from the datasets, the backend and frontend of the NER Flask application, and presents the results and discussions. For the GMB dataset, RoBERTa outperformed BERT by 0.19% and DistilBERT by 1.58%. RoBERTa and BERT showed similar results for the CoNLL-2003 dataset, with the latter scoring 0.02% higher and 0.85% better than DistilBERT. MedBERT was the best model for the N2C2 dataset, performing 1.71% better than BERT. However, the implementation of augmentation techniques for the GMB and N2C2 datasets did not yield significant improvements in the results of the NER models. Lastly, The emergency hotline dataset models showed similar results, with BioClinical BERT scoring the highest. These models can be deployed using the Flask application introduced in this report to receive useful outputs.