HealthGAT: Node Classifications in Electronic Health Records using Graph Attention Networks

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Graph Neural Networks

Background:

• Electronic Health Records (EHRs) are crucial in healthcare but often used in raw, tabular formats.
• Traditional data pre-processing limits the performance and applicability of EHR data in downstream tasks.
• Existing graph-based methods struggle to capture the complexity of medical relationships within EHRs.

Purpose of the Study:

• To introduce HealthGAT, a novel graph attention network framework for advanced EHR data representation.
• To overcome the limitations of raw EHR data formats and traditional pre-processing techniques.
• To improve the performance of healthcare applications utilizing EHR data.

Main Methods:

• Developed HealthGAT, a hierarchical graph attention network framework.
• Employed an iterative refinement process for medical code embeddings.
• Introduced customized EHR-centric auxiliary pre-training tasks to leverage embedded medical knowledge.

Main Results:

• HealthGAT generates superior embeddings from EHR data compared to traditional graph-based methods.
• The framework demonstrated significant advancements in EHR data analysis and representation.
• Achieved outstanding performance in node classification and downstream tasks like readmission prediction and diagnosis classification.

Conclusions:

• HealthGAT offers a comprehensive approach to analyzing complex medical relationships within EHRs.
• The model represents a significant advancement over standard EHR data representation techniques.
• HealthGAT proves effective across various healthcare scenarios, enhancing predictive capabilities.