Clinical research text summarization method based on fusion of domain knowledge

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Research

Background:

• Clinical research generates vast amounts of text data.
• Summarizing this data effectively is crucial for knowledge dissemination.
• Existing summarization models may struggle to capture domain-specific nuances.

Purpose of the Study:

• To integrate domain knowledge, specifically PICO (Population, Intervention, Comparison, Outcome) elements, into clinical research text summarization.
• To improve the comprehension of biomedical texts by AI models.
• To enhance the quality of generated summaries for end-users.

Main Methods:

• A novel method, DKGE-PEGASUS (Domain-Knowledge and Graph Convolutional Enhanced PEGASUS), was developed.
• It incorporates a PICO label prediction module and Graph Convolutional Neural Networks (GCN) to enrich text information.
• A pre-trained summarization model's encoder is reinforced with domain knowledge and GCN features.

Main Results:

• Experiments on PubMed and CDSR datasets showed significant effectiveness.
• Achieved Rouge-1 scores of 42.64 (PubMed) and 38.57 (CDSR).
• Demonstrated superior summarization quality compared to baseline models.

Conclusions:

• The proposed DKGE-PEGASUS method effectively identifies critical elements in clinical research texts.
• This approach leads to the generation of higher-quality summaries.
• Integrating domain knowledge significantly enhances clinical text summarization performance.