A Weighted Heterogeneous Graph-Based Dialog System.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Linguistics

Background:

• Traditional knowledge graphs (KGs) struggle to represent complex symptom-disease relationships due to unweighted edges.
• Existing disease diagnosis dialog systems often rely on data-driven methods, lacking deep relational understanding.
• Accurate representation of symptom-symptom and symptom-disease connections is crucial for effective diagnosis.

Purpose of the Study:

• To develop a weighted heterogeneous graph-based dialog system for improved disease diagnosis.
• To enhance the comprehension of symptom-disease and symptom-symptom relationships in diagnostic systems.
• To create a more efficient and accurate dialog system for identifying diseases.

Main Methods:

• Construction of a weighted heterogeneous graph utilizing symptom co-occurrence and a novel symptom frequency-inverse disease frequency metric.
• Implementation of a graph-based deep Q-network (graph-DQN) for dialog management.
• Integration of graph convolutional network (GCN) with DQN to learn embeddings from graph structure and attributes.

Main Results:

• The proposed graph-DQN system demonstrates comparable performance to state-of-the-art models.
• The system achieves disease diagnosis with significantly fewer dialog turns.
• Enhanced capability in distinguishing between diseases with similar symptoms is observed.

Conclusions:

• The weighted heterogeneous graph approach effectively models complex symptom relationships for disease diagnosis.
• Graph-DQN offers a powerful framework for dialog management in medical applications.
• This system represents a significant advancement in intelligent disease diagnosis dialog systems.