Knowledge graph assisted end-to-end medical dialog generation.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• Medical dialog systems can improve healthcare access, quality, and cost-efficiency.
• Current systems often generate generic responses, leading to unengaging interactions.
• Integrating large-scale medical knowledge is crucial for advanced dialog capabilities.

Purpose of the Study:

• To develop a knowledge-grounded conversation generation model for medical dialog systems.
• To leverage medical knowledge graphs for improved language comprehension and generation.
• To create clinically accurate and human-like medical conversations.

Main Methods:

• Combined pre-trained language models with the Unified Medical Language System (UMLS) medical knowledge base.
• Utilized a medical knowledge graph containing disease, symptom, and laboratory test information.
• Employed MedFact attention for reasoning over knowledge graph triples and a policy network to inject relevant entities.

Main Results:

• The model generated clinically correct and human-like medical conversations on the MedDialog-EN dataset.
• Demonstrated significant performance improvements over state-of-the-art methods in automatic and human evaluations.
• Showcased the effectiveness of transfer learning using a Covid-19 related dialog corpus.

Conclusions:

• Knowledge-grounded models significantly enhance medical dialog systems.
• Integrating medical knowledge graphs improves response generation accuracy and engagement.
• The proposed approach offers a promising direction for advancing e-medicine through AI.