Fine-Tuning Bidirectional Encoder Representations From Transformers (BERT)-Based Models on Large-Scale Electronic Health Record Notes: An Empirical Study

Abstract

Background: The bidirectional encoder representations from transformers (BERT) model has achieved great success in many natural language processing (NLP) tasks, such as named entity recognition and question answering. However, little prior work has explored this model to be used for an important task in the biomedical and clinical domains, namely entity normalization.

Objective: We aim to investigate the effectiveness of BERT-based models for biomedical or clinical entity normalization. In addition, our second objective is to investigate whether the domains of training data influence the performances of BERT-based models as well as the degree of influence.

Methods: Our data was comprised of 1.5 million unlabeled electronic health record (EHR) notes. We first fine-tuned BioBERT on this large collection of unlabeled EHR notes. This generated our BERT-based model trained using 1.5 million electronic health record notes (EhrBERT). We then further fine-tuned EhrBERT, BioBERT, and BERT on three annotated corpora for biomedical and clinical entity normalization: the Medication, Indication, and Adverse Drug Events (MADE) 1.0 corpus, the National Center for Biotechnology Information (NCBI) disease corpus, and the Chemical-Disease Relations (CDR) corpus. We compared our models with two state-of-the-art normalization systems, namely MetaMap and disease name normalization (DNorm).

Results: EhrBERT achieved 40.95% F1 in the MADE 1.0 corpus for mapping named entities to the Medical Dictionary for Regulatory Activities and the Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT), which have about 380,000 terms. In this corpus, EhrBERT outperformed MetaMap by 2.36% in F1. For the NCBI disease corpus and CDR corpus, EhrBERT also outperformed DNorm by improving the F1 scores from 88.37% and 89.92% to 90.35% and 93.82%, respectively. Compared with BioBERT and BERT, EhrBERT outperformed them on the MADE 1.0 corpus and the CDR corpus.

Conclusions: Our work shows that BERT-based models have achieved state-of-the-art performance for biomedical and clinical entity normalization. BERT-based models can be readily fine-tuned to normalize any kind of named entities.

Keywords: BERT; deep learning; electronic health record note; entity normalization; natural language processing.

Figure 1

Overview of this paper's methods. Bidirectional encoder representations from transformers (BERT) [11] was trained on Wikipedia text and the BookCorpus dataset. BioBERT [13] was initialized with BERT and fine-tuned using PubMed and (PubMed Central) PMC publications. We initialized the BERT-based model that was trained using 1.5 million electronic health record notes (EhrBERT) with BioBERT and then fine-tuned it using unlabeled electronic health record (EHR) notes. We further fine-tuned EhrBERT using annotated corpora for the entity normalization task. CDR: Chemical-Disease Relations; MADE: Medication, Indication, and Adverse Drug Events; NCBI: National Center for Biotechnology Information.

Figure 2

Model architectures. An example of entity normalization is shown and the named entity “dyspnea on exertion” is normalized to the term “60845006” in the Systematized Nomenclature of Medicine—Clinical Terms (SNOMED-CT) vocabulary (SNOMED International, 2019). The size of classes depends on the vocabularies used in a corpus, which is about 380,000 (Medical Dictionary for Regulatory Activities [MedDRA] and SNOMED-CT) for the Medication, Indication, and Adverse Drug Events (MADE) 1.0 corpus and 11,000 (MErged DIsease voCabulary [MEDIC]) for the National Center for Biotechnology Information (NCBI) Disease and Chemical-Disease Relations (CDR) corpora. BERT: bidirectional encoder representations from transformers; C: d^Trm-dimensional representation; [CLS]: classifier token; E: d^emb-dimensional embedding; T: d^Trm-dimensional vector; Trm: bidirectional transformer.

Figure 3

A case study. The left column shows examples where EhrBERT gave valid predictions. The right column shows examples where EhrBERT failed to give valid predictions. The rectangles denote mentions and weights of the word pieces in these mentions. The darker the color is, the larger the weight is. Split word pieces are denoted with “##.” The text in green and red indicate gold and predicted answers respectively. EhrBERT: bidirectional encoder representations from transformers (BERT)-based model that was trained using 1.5 million electronic health record notes.