Medical knowledge embedding based on recursive neural network for multi-disease diagnosis.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Linguistics

Background:

• Symbolic knowledge representation (first-order logic) is rich but difficult for machine learning.
• Knowledge embedding offers a feasible approach for complex reasoning with quantifiable relationships.
• Integrating symbolic knowledge with machine learning is crucial for advanced AI applications.

Purpose of the Study:

• To propose a Recursive Neural Knowledge Network (RNKN) for multi-disease diagnosis.
• To combine first-order logic medical knowledge with recursive neural networks.
• To improve diagnostic accuracy by leveraging structured medical knowledge and electronic health records.

Main Methods:

• Developed a Recursive Neural Knowledge Network (RNKN).
• Trained the RNKN using manually annotated Chinese Electronic Medical Records (CEMRs).
• Learned diagnosis-oriented knowledge embeddings and weight matrices.

Main Results:

• The RNKN achieved superior diagnostic accuracy compared to four machine learning models, four classical neural networks, and a Markov logic network.
• Performance improved with more explicit evidence extracted from CEMRs.
• Knowledge embedding interpretation within the RNKN became clearer with increased training epochs.

Conclusions:

• The RNKN effectively integrates symbolic medical knowledge with deep learning for enhanced multi-disease diagnosis.
• Recursive neural networks combined with knowledge embeddings provide a powerful framework for medical AI.
• The model's interpretability and performance are positively correlated with data explicitness and training duration.