Synthetic Data-Driven Approaches for Chinese Medical Abstract Sentence Classification: Computational Study

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Medical abstract sentence classification is vital for medical research and database searches.
• A lack of suitable datasets hinders Chinese medical abstract classification research.
• Precise classification of Chinese medical abstracts, including traditional Chinese medicine, is crucial for global medical advancement.

Purpose of the Study:

• To address the scarcity of labeled Chinese medical abstract datasets.
• To develop accurate text classification algorithms for Chinese medical abstracts.
• To create new training datasets without manual annotation.

Main Methods:

• Generated three training datasets (15,000 sentences each) using translated PubMed data and GPT-3.5 with pseudolabeling or category alignment.
• Created a test dataset of 87,000 sentences from 20,000 abstracts.
• Employed SBERT embeddings for semantic analysis and evaluated models using clustering (SBERT-DocSCAN) and supervised methods (SBERT-MEC).

Main Results:

• Models trained on synthetic datasets outperformed baseline metrics.
• SBERT-DocSCAN achieved up to 91.30% accuracy and F1-score on the test dataset when trained on dataset #3.
• SBERT-MEC also demonstrated robust performance, reaching 90.39% accuracy and 90.35% F1-score.

Conclusions:

• The study successfully generated novel datasets for Chinese medical abstract classification.
• The SBERT-DocSCAN and SBERT-MEC models significantly enhance classification precision, even with synthetic data.
• The findings address data limitations and improve the accuracy of Chinese medical abstract classification.