Multi-layer information fusion based on graph convolutional network for knowledge-driven herb recommendation.

Area of Science:

• Integrative Medicine
• Computational Biology
• Pharmacology

Background:

• Traditional Chinese Medicine (TCM) prescriptions are a valuable resource, but understanding complex herb-symptom relationships is challenging.
• Existing artificial intelligence (AI) herb recommendation models often overlook crucial herb property information and rely on basic statistical methods.
• This limitation hinders the accurate perception of intricate correlations between symptoms and herbs, impacting clinical applications and new prescription discovery.

Purpose of the Study:

• To develop an advanced AI model for Traditional Chinese Medicine (TCM) herb recommendation.
• To integrate herb property information using a knowledge graph to enhance model performance.
• To improve the understanding of symptom-herb correlations for clinical applications and novel prescription discovery.

Main Methods:

• Constructed a herb knowledge graph to incorporate herb properties as auxiliary information.
• Proposed a graph convolution model with multi-layer information fusion.
• Developed methods to obtain rich and less noisy symptom and herb feature representations.

Main Results:

• The proposed graph convolution model demonstrated superior performance compared to baseline models on a TCM prescription dataset.
• Achieved significant improvements in Precision@5 (+6.2%), Recall@5 (+16.0%), and F1-Score@5 (+12.0%).
• The model effectively captures complex correlations between symptoms and herbs by leveraging herb properties.

Conclusions:

• Integrating herb properties through knowledge graphs and graph convolution models enhances TCM herb recommendation.
• The developed AI model offers a more sophisticated approach to understanding symptom-herb relationships in TCM.
• This advancement holds significant potential for improving TCM clinical applications and facilitating the discovery of new prescriptions.