Patient multi-relational graph structure learning for diabetes clinical assistant diagnosis

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Biology

Background:

• Electronic health records (EHRs) and advanced data analysis are crucial for healthcare research and decision-making.
• Graph neural networks (GNNs) are utilized in smart healthcare but struggle with inaccurate graph topologies from EHR data, hindering disease prediction.
• Existing GNNs assume accurate graph structures, which is often not the case with complex EHR interactions, leading to inefficiencies.

Purpose of the Study:

• To develop a new model, PM-GSL, for improved diabetes clinical assistant diagnosis using patient multi-relational graph structure learning.
• To address the limitations of traditional GNNs in handling noisy or false topologies within EHR-derived graphs.
• To enhance disease prediction accuracy by learning optimal graph structures from heterogeneous patient data.

Main Methods:

• Constructed a patient multi-relational graph incorporating demographics, diagnostics, lab tests, and medication interactions from EHRs.
• Generated three candidate graphs (original subgraph, overall feature graph, higher-order semantic graph) to capture node characteristics and higher-order semantics.
• Fused these graphs into a heterogeneous graph and jointly optimized its structure with GNNs for disease prediction.

Main Results:

• The proposed PM-GSL model demonstrated superior performance in diabetes clinical assistant diagnosis tasks.
• Experimental results indicated that PM-GSL outperforms state-of-the-art models in predicting diabetes.
• The model effectively handles the heterogeneity and complexity of patient multi-relational graphs.

Conclusions:

• PM-GSL offers a significant advancement in leveraging EHR data for clinical decision support in diabetes.
• The approach of learning and optimizing graph structures improves the accuracy of GNN-based disease prediction.
• This method provides a robust framework for smart healthcare applications dealing with complex, multi-relational patient data.