Patient-GAT: Sarcopenia Prediction using Multi-modal Data Fusion and Weighted Graph Attention Networks.

Area of Science:

• Medical Informatics
• Machine Learning
• Graph Neural Networks

Background:

• Graph Attention Networks (GAT) are widely used for node classification in graph data.
• Limited research exists on applying GAT to patient similarity networks for healthcare applications.

Purpose of the Study:

• To propose Patient-GAT, a novel method for predicting chronic health conditions using GAT on patient similarity networks.
• To integrate multi-modal data for robust patient vector representation and disease prediction.

Main Methods:

• Multi-modal data fusion to create patient vector representations from imputed lab variables and structured data.
• Construction of patient similarity networks based on fused patient representations.
• Application of Graph Attention Networks (GAT) to patient networks for disease prediction, specifically sarcopenia.

Main Results:

• Patient-GAT demonstrated superior performance in predicting sarcopenia compared to baseline models.
• Analysis of the contribution of temporal lab data representation was conducted.
• Model interpretability was explored through attention coefficient analysis.

Conclusions:

• Patient-GAT offers a promising approach for chronic disease prediction by leveraging patient similarity networks and GAT.
• The model's effectiveness is validated on real-world electronic health records (EHRs).
• The study highlights the potential of GAT in personalized medicine and healthcare analytics.