Leveraging Representation Learning for the Construction and Application of a Knowledge Graph for Traditional Chinese...

Area of Science:

• Artificial Intelligence in Medicine
• Knowledge Representation and Reasoning
• Traditional Chinese Medicine Research

Background:

• Knowledge discovery from Traditional Chinese Medicine (TCM) physician data presents significant challenges for artificial intelligence (AI) applications.
• Integrating diverse TCM treatment records into AI models requires advanced knowledge representation techniques.

Purpose of the Study:

• To construct a comprehensive Traditional Chinese Medicine knowledge graph (TCM KG) from physician data.
• To apply the constructed TCM KG to enhance decision-making in TCM diagnosis and treatment.
• To develop and validate a novel framework for TCM KG construction and application.

Main Methods:

• A new framework utilizing representation learning for TCM KG construction and application was designed.
• A transformer-based Contextualized Knowledge Graph Embedding (CoKE) model was employed for KG representation learning and knowledge distillation.
• The framework integrated automatic identification and expansion of multihop relations, resulting in a TCM KG with 59,882 entities and 604,700 triples.

Main Results:

• The framework demonstrated superior performance over baseline models in a link prediction task, evaluated using mean reciprocal rank (MRR) and Hits@N metrics.
• Knowledge graph embedding (KGE) multitagged TCM discriminative diagnosis metrics confirmed the framework's improvements compared to baseline approaches.
• Validation through a link prediction task confirmed the efficacy of the constructed TCM KG.

Conclusions:

• The developed clinical knowledge graph (KG) representation learning and application framework is effective for knowledge discovery and decision-making assistance in TCM diagnosis and treatment.
• The framework exhibits strong potential for applications including KG-fused multimodal diagnosis, KGE-based text classification, and knowledge inference for medical question answering.