Detecting mental and physical disorders using multi-task learning equipped with knowledge graph attention network.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Psychosomatic Medicine

Background:

• Mental and physical disorders (MPD) are interconnected, yet current research often analyzes them separately.
• Existing medical informatics approaches lack a concurrent focus on both mental and physical health aspects.
• This gap hinders comprehensive understanding and detection of psychosomatic diseases and co-morbidities.

Purpose of the Study:

• To propose and validate a joint modeling approach for the concurrent detection of mental and physical disorders (MPD).
• To develop a comprehensive MPD knowledge ontology and knowledge graph from online medical consultation records.
• To construct a fine-grained MPD corpus and a multi-task learning model for severity detection.

Main Methods:

• Crawled online medical consultation records to build an MPD knowledge ontology and a knowledge graph (12,673 nodes, 82,195 relations).
• Created an MPD corpus (8909 records) with severity levels (None to Dangerous) via expert-guided annotation.
• Designed a multi-task learning model incorporating a Knowledge Graph Attention Network (KGAT) for MPD severity detection.

Main Results:

• Demonstrated the effectiveness of the proposed joint modeling approach in detecting MPD severity.
• Utilized ontology-based and centrality-based methods to infer additional knowledge, enhancing KGAT's prediction performance and interpretability.
• The developed MPD dataset has been made publicly available for further research.

Conclusions:

• The joint modeling approach effectively addresses the concurrent detection of mental and physical disorders (MPD).
• The created MPD knowledge graph and corpus serve as valuable resources for medical informatics research.
• The findings contribute to improved understanding and detection in psychosomatic medicine, psychiatry, and physical co-morbidity.