Leveraging a Joint learning Model to Extract Mixture Symptom Mentions from Traditional Chinese Medicine Clinical...

Area of Science:

• Computational linguistics
• Medical informatics
• Traditional Chinese Medicine (TCM)

Background:

• The structuring of Traditional Chinese Medicine (TCM) data presents challenges, particularly the 'mixture symptom mention' problem.
• Existing methods may struggle with the complexity and overlapping nature of symptom descriptions in TCM clinical notes.

Purpose of the Study:

• To develop and evaluate an automated method for disassembling mixture symptom mentions in TCM.
• To improve the accuracy and efficiency of extracting entity relations from TCM clinical text.

Main Methods:

• Annotation of over 2,200 clinical notes to create a training dataset.
• Development of an end-to-end joint learning model incorporating a multi-head mechanism for relation extraction.
• Utilization of a pretrained transformer encoder for enhanced context information capture.

Main Results:

• The proposed joint learning model achieved superior performance in recall (0.822), precision (0.825), and F1-score (0.818).
• Performance was 14% higher compared to traditional entity extraction pipelines.
• The model demonstrated efficient automatic feature extraction without external Natural Language Processing (NLP) tools.

Conclusions:

• The joint learning model effectively addresses the mixture symptom mention problem in TCM.
• Its ability to handle overlapping relations facilitates the downstream reassembly of symptom entities.
• This approach offers an efficient and accurate solution for structuring TCM clinical data.