Hypergraph Transformers for EHR-based Clinical Predictions.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Electronic health records (EHR) contain extensive patient data crucial for digital medicine.
• Extracting meaningful patient visit representations from diverse medical codes in EHR is challenging.
• Existing methods often fail to capture complex inter-code relationships, limiting prediction accuracy.

Purpose of the Study:

• To develop an advanced method for learning patient visit representations from EHR data.
• To improve the accuracy of downstream clinical predictions by modeling complex medical code interactions.
• To leverage hypergraphs and self-attention for enhanced EHR data analysis.

Main Methods:

• Utilized hypergraphs to model multi-way relationships among medical codes within patient visits.
• Implemented a self-attention mechanism to dynamically select relevant medical codes per visit.
• Jointly learned representations for patient visits and medical codes.

Main Results:

• The proposed hypergraph-based method significantly outperformed existing approaches on two EHR datasets.
• Demonstrated superior performance in supporting downstream clinical prediction tasks.
• The model provided interpretable insights into the relationships between medical codes and clinical outcomes.

Conclusions:

• Hypergraph modeling effectively captures complex medical code interactions in EHR data.
• Self-attention enhances the relevance of medical codes for improved predictive generalization.
• This approach offers a powerful tool for advancing digital medicine and clinical decision support.