Grouping attributes zero-shot learning for tongue constitution recognition.

Area of Science:

• Integrative Medicine
• Artificial Intelligence
• Medical Diagnostics

Background:

• Traditional Chinese Medicine (TCM) posits that individual constitutions influence disease susceptibility and treatment outcomes.
• Machine learning (ML) shows promise in recognizing TCM constitutions via tongue image analysis.
• Current ML methods face challenges with limited tongue image datasets and lack of TCM domain knowledge, leading to imbalanced constitution categories.

Purpose of the Study:

• To propose a novel constitution recognition method for Traditional Chinese Medicine (TCM) using zero-shot learning and TCM domain knowledge.
• To address the limitations of existing ML approaches, including small datasets and imbalanced categories.

Main Methods:

• Developed a new zero-shot learning (ZSL) framework incorporating TCM domain knowledge.
• Implemented attribute grouping and discriminant latent feature learning within the ZSL method.
• Constructed new tongue image databases for experimental validation.

Main Results:

• The proposed ZSL method effectively recognizes TCM constitutions from tongue images.
• Attribute grouping and discriminant latent feature learning significantly improved performance.
• The approach successfully addressed the issue of imbalanced constitution categories.

Conclusions:

• The integration of TCM knowledge with zero-shot learning offers a promising solution for constitution recognition.
• The developed method enhances ML performance in analyzing TCM constitutions, overcoming data scarcity and imbalance.
• This approach has the potential to improve personalized disease prediction and treatment strategies in TCM.