Structure-aware siamese graph neural networks for encounter-level patient similarity learning

Area of Science:

• Biomedical Informatics
• Machine Learning
• Graph Neural Networks

Background:

• Patient similarity learning is crucial for clinical decision-making.
• Electronic Health Records (EHRs) present challenges due to data sparsity and missing values.
• Existing methods struggle with the complexity of patient relationships and data incompleteness.

Purpose of the Study:

• To propose a novel deep learning framework, SSGNet, for robust encounter-level patient similarity learning.
• To capture intrinsic graph structures within EHR data.
• To mitigate the impact of missing values in patient records.

Main Methods:

• Organized EHRs as a graph, with patient encounters as nodes.
• Employed Siamese Graph Neural Networks (GNNs) to learn node embeddings and similarities simultaneously.
• Utilized a low-rank and contrastive objective for graph structure optimization and enhanced model capacity.

Main Results:

• SSGNet demonstrated significant improvements in Accuracy, Precision, Recall, and F1 score for pairwise similarity classification.
• Achieved superior mean Average Precision (mAP) in similar encounter retrieval tasks.
• Maintained stable performance across varying data missing rates, validating robustness.

Conclusions:

• SSGNet offers a superior approach to patient similarity learning compared to existing methods.
• The framework effectively addresses challenges posed by sparse and incomplete EHR data.
• SSGNet shows potential for enhancing clinical decision support systems.