All-Cause Death Prediction Method for CHD Based on Graph Convolutional Networks.

Area of Science:

• Cardiovascular Disease Research
• Machine Learning in Healthcare
• Graph Neural Networks

Background:

• Coronary heart disease (CHD) presents a significant public health challenge due to high morbidity and mortality.
• Current CHD risk prediction models often rely on shallow machine learning, limiting accuracy by overlooking global patient interactions.

Purpose of the Study:

• To propose a novel graph-based approach for CHD risk prediction.
• To enhance the accuracy of CHD prediction for individualized patient management strategies.

Main Methods:

• Framed CHD prediction as a graph node classification task, representing individuals as nodes and their associations as graph edges.
• Utilized an adaptive multi-channel graph convolutional neural network (AM-GCN) for feature extraction from graph topology and node features.
• Incorporated an attention mechanism to learn adaptive importance weights for embeddings and modeled relationships using population and K-nearest neighbor graphs.

Main Results:

• The proposed AM-GCN model demonstrated superior performance on the CHD dataset compared to non-graph models.
• Achieved a 1.3% increase in accuracy, a 5.1% improvement in Area Under the Curve (AUC), and a 4.6% enhancement in F1-score.

Conclusions:

• The graph-based approach, particularly with the AM-GCN model, significantly improves CHD risk prediction.
• This method offers a more effective strategy for individualized patient management in cardiovascular health.