Chinese Clinical Named Entity Recognition With Segmentation Synonym Sentence Synthesis Mechanism: Algorithm Development and Validation

Jian Tang ¹, Zikun Huang ², Hongzhen Xu ¹

⁰Department of Pharmacy, People's Hospital of Guilin, 12 Wenming Road, Guilin, 541000, China, 86 18978320258.

Jmir Medical Informatics +

|

December 2, 2024

View abstract on PubMed

Summary

This study introduces a novel dataset augmentation algorithm for clinical named entity recognition (CNER) that enhances model performance by overcoming data scarcity and labeling challenges. The Segmentation Synonym Sentence Synthesis (SSSS) method improves CNER model accuracy without needing extensive manual annotation.

Area Of Science

Natural Language Processing
Machine Learning
Bioinformatics

Background

Clinical Named Entity Recognition (CNER) is crucial for extracting information from electronic medical records.
Deep learning models are prevalent in CNER but often require large annotated datasets and extensive dictionaries.
Existing CNER methods face challenges due to text complexity, entity diversity, and boundary ambiguity.

Purpose Of The Study

To address data scarcity and labeling difficulties in CNER tasks.
To propose a dataset augmentation algorithm that leverages existing knowledge without manual dictionary expansion.
To improve the generalization performance of CNER models.

Main Methods

Developed the Segmentation Synonym Sentence Synthesis (SSSS) algorithm using proximity word calculation and lexical segmentation.
Recombined synonymous vocabulary from natural language data to expand the dataset.
Applied the SSSS algorithm to RoBERTa + CRF and RoBERTa + BiLSTM + CRF models, evaluating on CCKS 2017 and 2019 datasets.

Main Results

SSSS + RoBERTa + CRF achieved an F1-score of 91.30% on the CCKS-2017 dataset.
SSSS + RoBERTa + BiLSTM + CRF achieved an F1-score of 91.35% on the CCKS-2017 dataset.
Both models achieved F1-scores above 83% on the CCKS-2019 dataset.

Conclusions

The SSSS algorithm effectively expands CNER datasets.
The proposed method significantly improves CNER model performance.
This approach successfully mitigates challenges related to data acquisition, annotation, and model generalization.

Keywords:

AI Chinese electronic medical records EMR RoBERTa algorithm artificial intelligence clinical named entity recognition dataset dataset augmentation entity entity recognition natural language processing segmentation word embedding