Related Experiment Video
Updated: Jan 11, 2026
10:26
Author Spotlight: A 3D Digital Model for the Diagnosis and Treatment of Pulmonary Nodules
Published on: May 19, 2023
2.5K
Named Entity Recognition for Chinese Cancer Electronic Health Records-Development and Evaluation of a Domain-Specific
Junbai Chen1,2, Butian Zhao1, Xiaohan Tian1
1School of Management, Beijing University of Chinese Medicine, No. 11, North Third Ring Road East, Chaoyang District, Beijing, 100029, China, 86 13811833948.
JMIR Medical Informatics
|November 14, 2025
Summary
This study introduces ChCancerBERT, a specialized Named Entity Recognition (NER) model for Chinese cancer electronic health records. It significantly improves the extraction of crucial medical information from breast cancer patient data.
Area of Science:
- Natural Language Processing (NLP) in Healthcare
- Medical Informatics
- Cancer Research
Background:
- Unstructured Chinese cancer electronic health records (EHRs) contain valuable clinical expertise.
- Accurate medical entity recognition (NER) is vital for developing clinical decision support systems.
- Existing NER models often fail to adequately address the nuances of Chinese cancer EHRs.
Purpose of the Study:
- To develop a specialized NER model tailored for Chinese cancer EHRs.
- To enhance the recognition of medical entities within these specialized records.
- To improve the utility of cancer EHR data for research and clinical applications.
Main Methods:
- Developed the ChCancerBERT model by pretraining on a Chinese cancer corpus, building upon the MC Bidirectional Encoder Representations from Transformers (BERT) foundation.
- Integrated dilated-gated convolutional neural networks, bidirectional long short-term memory, multihead attention, and a conditional random field for a multimodel, multilevel NER approach.
- Utilized desensitized inpatient EHRs related to breast cancer from a Beijing hospital for model training and evaluation.
Main Results:
- The ChCancerBERT model achieved high performance in extracting medical entities (symptoms, signs, tests, treatments, time) from Chinese breast cancer EHRs.
- The model surpassed baseline and other comparative models, reaching an F1-score of 86.93% on internal data and 87.26% on the CCKS2019 dataset.
- Demonstrated superior precision (87.24% internal, 87.26% CCKS2019) and recall (86.61% internal, 87.27% CCKS2019) compared to existing methods.
Conclusions:
- The proposed ChCancerBERT approach shows excellent performance for NER in breast cancer EHRs, advancing clinical decision support.
- Incorporating domain-specific corpora significantly enhances BERT model performance in specialized clinical NER tasks.
- This work facilitates better utilization of Chinese cancer EHR data for improved cancer treatment and research.
