Related Experiment Video
Updated: Jul 31, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
A general approach for improving deep learning-based medical relation extraction using a pre-trained model and
Tao Chen1, Mingfen Wu1, Hexi Li1
1Department of Computer Science and Engineering, Faculty of Intelligent Manufacturing, Wuyi University, No.22, Dongcheng village, Pengjiang district, Jiangmen City, Guangdong Province, 529020, China.
Database : the Journal of Biological Databases and Curation
|December 5, 2019
Summary
This study introduces a novel deep learning method combining Bidirectional Encoder Representations from Transformers (BERT) and 1D convolutional neural networks (1D-CNN) for medical relation extraction, achieving state-of-the-art results without extensive human labeling.
Area of Science:
- Biomedical Informatics
- Natural Language Processing
- Machine Learning
Background:
- Automatic extraction of biomedical relations is vital for applications using literature and clinical data.
- Current deep learning models for medical relation extraction often require extensive labeled data, risking overfitting.
- Human labeling for training data is time-consuming and resource-intensive.
Purpose of the Study:
- To develop an improved medical relation extraction approach that minimizes the need for large-scale human-labeled datasets.
- To leverage pre-trained language models and fine-tuning techniques for enhanced relation extraction performance.
- To achieve state-of-the-art results in extracting meaningful relations from biomedical text.
Main Methods:
- Utilized Bidirectional Encoder Representations from Transformers (BERT) for pre-training on large unstructured biomedical text.
- Combined the pre-trained BERT model with a one-dimensional convolutional neural network (1D-CNN) for fine-tuning.
- Evaluated the proposed approach on three diverse biomedical relation extraction datasets: BioCreative V, TCM, and i2b2 2012.
Main Results:
- The proposed BERT-CNN model achieved state-of-the-art performance across all three tested datasets.
- Demonstrated significant relative improvements in F1 score: 22.2% (chemical-disease), 7.77% (TCM), and 38.5% (temporal relations).
- The fine-tuning approach effectively improved relation extraction accuracy without additional manual data annotation.
Conclusions:
- The integration of pre-trained BERT with 1D-CNN offers a powerful and efficient method for medical relation extraction.
- This approach effectively overcomes the limitations of data scarcity and overfitting in deep learning models for biomedical text.
- The developed model provides a robust solution for advancing biomedical informatics and clinical data analysis.