Advanced Sampling Technique in Radiology Free-Text Data for Efficiently Building Text Mining Models by Deep Learning
Wei-Chieh Hung1,2,3, Yih-Lon Lin4, Chi-Wei Lin1,2
1Department of Family and Community Medicine, E-Da Hospital, I-Shou University, Kaohsiung 82445, Taiwan.
Diagnostics (Basel, Switzerland)
|January 22, 2024
Summary
Advanced sampling methods, like vector sum minimization, improve deep learning models for identifying vertebral compression fractures (VCF) in radiology reports. This method efficiently selects critical data, enhancing predictive accuracy.
Area of Science:
- Medical Informatics
- Natural Language Processing
- Machine Learning
Background:
- Accurate identification of vertebral compression fractures (VCF) in radiology reports is crucial for patient care.
- Traditional text mining methods may struggle with the complexity and volume of free-text clinical data.
- Deep learning models offer potential for semantic analysis but require efficient data sampling strategies.
Purpose of the Study:
- To establish and evaluate advanced sampling methods for building efficient semantic text mining models.
- To compare the performance of different sampling techniques in identifying VCF from radiology reports using deep learning.
- To propose an optimized sampling method for critical data selection in free-text analysis.
Main Methods:
- Utilized a dataset of 27,401 free-text radiology reports from spine X-ray examinations.
- Developed supervised long short-term memory (LSTM) networks for VCF identification.
- Compared four sampling methods: vector sum minimization, vector sum maximization, stratified, and simple random sampling at fixed percentages (1/10, 1/20, 1/30, 1/40).
- Evaluated predictive accuracy using the area under the receiver operating characteristics (AUROC) curve.
Main Results:
- Vector sum minimization consistently yielded the highest AUROC values across all sampling ratios.
- AUROC values for vector sum minimization ranged from 0.981 to 0.895 at decreasing sampling ratios.
- Vector sum maximization demonstrated the lowest predictive accuracy.
- The proposed vector sum minimization method efficiently identified critical representative samples.
Conclusions:
- Vector sum minimization is an effective advanced sampling method for free-text data in building semantic text mining models.
- This method significantly enhances the efficiency and predictive accuracy of deep learning models, such as LSTMs, for clinical applications like VCF detection.
- The findings suggest a novel approach for optimizing data selection in medical natural language processing tasks.
