MERGE: A Multi-graph Attentive Representation learning framework integrating Group information from similar patients

Area of Science:

• Medical Big Data Analytics
• Machine Learning in Healthcare
• Predictive Health Modeling

Background:

• Electronic Health Records (EHRs) are crucial for predicting future patient health status.
• Existing deep learning models often struggle with sparse and low-quality EHR data, limiting prediction accuracy.
• Individual patient data alone may not capture comprehensive health information.

Purpose of the Study:

• To develop an advanced deep learning framework for medical prediction that overcomes limitations of individual EHR data.
• To improve the accuracy of patient health status prediction by incorporating group information from similar patients.

Main Methods:

• Proposed MERGE (Multi-graph attEntive Representation learning framework integrating Group information from similar patiEnts).
• MERGE utilizes an individual representation learning module for temporal characteristics.
• A group representation learning module integrates data from similar patients to supplement individual information.

Main Results:

• MERGE demonstrated superior performance in predicting in-hospital mortality on the MIMIC-III dataset.
• The framework also showed effectiveness in predicting cardiovascular diseases (CVDs) on the Xiangya dataset.
• Experimental results confirm MERGE outperforms existing state-of-the-art methods.

Conclusions:

• Integrating group information from similar patients significantly enhances the accuracy of medical prediction models.
• MERGE offers a promising approach for leveraging complex EHR data in healthcare.
• The framework effectively addresses data sparsity and quality issues in individual patient records.