An Attention Model With Transfer Embeddings to Classify Pneumonia-Related Bilingual Imaging Reports: Algorithm
Hyung Park1, Min Song2, Eun Byul Lee2
1Department of Pulmonary and Critical Care Medicine, Asan Medical Center, Seoul, Republic of Korea.
JMIR Medical Informatics
|April 6, 2021
Summary
This study developed a deep learning model to accurately classify pneumonia in bilingual radiology reports, achieving high accuracy and improving electronic health data analysis.
Area of Science:
- Medical Informatics
- Artificial Intelligence in Healthcare
- Radiology Report Analysis
Background:
- Accurate labeling of outcomes in electronic health records is crucial.
- Deep learning has been applied to classify radiologic reports.
- Pneumonia classification in bilingual reports remains an underexplored area.
Purpose of the Study:
- To develop and evaluate a deep learning method for classifying pneumonia in bilingual radiologic reports.
- To compare the performance of a long short-term memory (LSTM)-Attention model against other machine learning and deep learning methods.
Main Methods:
- Retrospective analysis of 5450 radiologic reports (chest CT and X-ray) from Asan Medical Center (2008-2018).
- Development and testing of a long short-term memory (LSTM)-Attention model.
- Comparison of model performance using metrics like accuracy, AUROC, and F1 score.
Main Results:
- The proposed LSTM-Attention model achieved 91.01% accuracy on the test set (n=1090), with high AUROCs for negative (0.98), positive (0.97), and obscure (0.90) classifications.
- Top-performing models utilized FastText or LSTM architectures.
- Convolutional neural network (CNN) models showed lower accuracy (73.03%).
Conclusions:
- The developed deep learning method demonstrates excellent performance in classifying pneumonia from bilingual radiologic reports.
- This approach can enhance pneumonia research by enabling precise outcome extraction from electronic health data.
