Weighted Heterogeneous Graph-Based Incremental Automatic Disease Diagnosis Method.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Knowledge Representation

Background:

• Automatic diagnosis models face challenges in classifying numerous diseases and collecting extensive disease-symptom datasets.
• Existing models struggle with scalability and data acquisition for comprehensive diagnostic capabilities.

Purpose of the Study:

• To develop an experience-infused knowledge model for multi-department symptom-based automatic diagnosis.
• To address data collection limitations and large-scale multi-classification issues using knowledge graphs and incremental learning.

Main Methods:

• Constructed a heterogeneous knowledge graph via graph fusion and entity linking.
• Employed incremental learning to continuously update the knowledge graph with experiential knowledge from data.
• Developed adaptive neural network models for each dataset, integrating learned parameters back into the knowledge graph.

Main Results:

• The proposed model demonstrated improved diagnostic accuracy across three public datasets, with average improvements of 5%, 2%, and 15%.
• The model exhibited a strong ability to resist forgetting, maintaining performance on historical data after class increments.
• Incremental learning effectively addressed data collection challenges and enabled scalable multi-classification.

Conclusions:

• The experience-infused knowledge model offers a robust solution for automatic diagnosis, overcoming limitations of traditional approaches.
• Incremental learning and knowledge graph integration are effective strategies for building adaptive and accurate diagnostic systems.
• The model shows significant potential for enhancing clinical decision support systems through continuous learning and knowledge integration.