Medical Graph Diffusion: Hybrid Graph Diffusion With Heterogeneous Graph Convolutional Networks for Medical Text...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Medical Informatics

Background:

• Text classification is crucial for extracting knowledge from medical texts.
• Existing methods may not fully capture the complex structural relationships within medical documents.

Purpose of the Study:

• To propose a novel Medical Graph Diffusion (MGD) model for enhanced medical text classification.
• To model intricate word-level, sentence-level, and word-sentence-level relationships within medical documents.

Main Methods:

• Constructed a text heterogeneous graph to represent multi-level structural relationships.
• Employed graph diffusion convolution to reconstruct the graph, overcoming direct neighbor limitations.
• Utilized a heterogeneous graph convolutional network and multilayer perceptron for classification.

Main Results:

• The MGD model achieved notable improvements in Accuracy and F1 scores across various benchmarks (long text, short text, medical text).
• Demonstrated superior performance compared to representative baseline methods.
• Ablation studies confirmed the significant impact of heterogeneous graph construction and diffusion graph convolution.

Conclusions:

• The MGD model effectively captures complex structural information in medical texts.
• Graph diffusion convolution and heterogeneous graph construction are key to the model's performance.
• The MGD model offers a robust and effective solution for medical text classification.