DyGraphTrans: A temporal graph representation learning framework for modeling disese progression from Electronic...

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Electronic Health Records (EHRs) offer rich longitudinal patient data for disease prediction.
• Existing computational methods for EHR analysis often suffer from high memory usage, computational cost, and lack of interpretability.
• Efficiently processing large-scale EHR data while maintaining accuracy and interpretability is a significant challenge.

Purpose of the Study:

• To introduce DyGraphTrans, a dynamic graph representation learning framework for patient EHR data.
• To address the limitations of existing methods in terms of memory consumption, computational cost, and interpretability.
• To enable accurate and interpretable early disease prediction from EHRs.

Main Methods:

• Representing patient EHR data as a sequence of temporal graphs.
• Utilizing nodes for patients, node features for temporal clinical attributes, and edges for patient similarity.
• Employing a sliding-window mechanism to reduce memory consumption while preserving temporal context.
• Jointly capturing patient similarity and temporal evolution in a memory-efficient and interpretable manner.

Main Results:

• DyGraphTrans demonstrated strong predictive performance on Alzheimer's Disease Neuroimaging Initiative (ADNI), National Alzheimer's Coordinating Center (NACC), and Medical Information Mart for Intensive Care (MIMIC-IV) datasets.
• The model achieved accurate early mortality prediction and disease progression prediction.
• Interpretability analysis showed alignment with known clinical risk factors.

Conclusions:

• DyGraphTrans provides an efficient and interpretable solution for leveraging EHR data for disease prediction.
• The framework successfully models local temporal dependencies and long-range global trends in patient data.
• DyGraphTrans offers a promising approach for advancing computational methods in clinical informatics and precision medicine.