Hybrid multimodal fusion for graph learning in disease prediction

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Graph neural networks (GNNs) are increasingly used for disease prediction by modeling patient similarities.
• Current GNN methods often rely on latent embeddings, which may not fully capture real-world patient relationships.
• Raw patient data (demographics, lab results) contains valuable information for similarity assessment.

Purpose of the Study:

• To develop an improved GNN framework for disease prediction.
• To enhance graph construction by incorporating both latent embeddings and raw patient data.
• To refine graph structures through edge pruning and sparsification for better performance.

Main Methods:

• Constructed adaptive graphs using both patient latent representations and raw data.
• Merged these graphs using weighted summation.
• Applied degree-sensitive edge pruning and kNN sparsification to refine graph connectivity.

Main Results:

• The proposed method demonstrated superior performance on two diagnostic prediction datasets.
• Integrating raw data alongside latent embeddings improved graph structure accuracy.
• Edge pruning and sparsification techniques effectively reduced noise and extraneous connections.

Conclusions:

• Combining latent embeddings and raw data in GNNs offers a more robust approach to disease prediction.
• The proposed graph construction and refinement strategy significantly outperforms existing state-of-the-art methods.
• This work highlights the importance of leveraging diverse data sources for accurate patient similarity modeling in GNNs.