Knowledge-Driven Graph Representation Learning for Myocardial Infarction Localization.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Electrocardiograms (ECG) are vital for myocardial infarction (MI) localization.
• Current deep learning methods for MI localization are data-driven, limiting performance with rare MI cases.
• Integrating prior medical knowledge can enhance deep learning model accuracy.

Purpose of the Study:

• To develop a knowledge-driven graph representation learning (KD-GRL) framework for improved MI localization.
• To guide deep learning models in identifying key MI localization features using integrated medical knowledge.
• To enhance the detection of rare MI cases through a knowledge-enhanced approach.

Main Methods:

• Constructed a MI localization knowledge graph (KG) integrating ECG leads, morphology, diagnostic rules, and demographics.
• Utilized parallel patient multi-feature extractors to obtain entity embeddings.
• Employed an edge relation projection (ERP) method for KG aggregation.
• Framed MI localization as a link prediction task within the KG.

Main Results:

• Achieved F1-scores of 48.90% on the PTB dataset and 46.06% on the PTBXL dataset.
• Outperformed traditional data-driven methods on both public datasets.
• Demonstrated superior performance in localizing rare MI cases due to the integration of diagnostic knowledge.

Conclusions:

• The KD-GRL framework effectively improves MI localization accuracy by incorporating medical knowledge.
• This knowledge-driven approach enhances the identification of key features for MI diagnosis.
• The method shows significant potential for improving the clinical diagnosis of rare MI subtypes.