Multi-Gate Mixture of Multi-View Graph Contrastive Learning on Electronic Health Record

Area of Science:

• Medical Informatics
• Machine Learning
• Graph Neural Networks

Background:

• Electronic Health Records (EHRs) contain valuable patient data for predictive modeling.
• Graph Neural Networks (GNNs) show promise for EHR representation learning but require manual relationship labeling and often focus on single tasks.
• Existing methods may not fully leverage the rich, interconnected information within EHR data.

Purpose of the Study:

• To develop an advanced EHR representation learning method that overcomes limitations of current GNN approaches.
• To improve the accuracy of patient-specific prediction tasks by exploiting multi-task learning and self-contrastive learning on graph structures.
• To enhance the utility of EHR data for clinical decision support and patient outcome prediction.

Main Methods:

• Proposed a multi-gate mixture of multi-view graph contrastive learning (MMMGCL) framework for EHR representation.
• Represented patient visits as graphs with a hierarchical structure and pre-trained node features using GloVe with ontology knowledge.
• Employed a joint learning strategy to optimize both prediction task losses and contrastive losses for improved representations.

Main Results:

• MMMGCL demonstrated improved performance on patient readmission, mortality, and length of stay prediction tasks.
• The method outperformed straightforward graph-based approaches on two large-scale medical datasets (MIMIC-III and eICU).
• Enhanced EHR representations led to more accurate downstream prediction outcomes.

Conclusions:

• The proposed MMMGCL method offers a more effective approach to EHR representation learning.
• Multi-task learning and self-contrastive learning on graph structures significantly improve predictive accuracy.
• This work provides a robust framework for leveraging complex EHR data for clinical predictions.