Hypergraph Foundation Model for Brain Disease Diagnosis.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Existing brain disease diagnosis methods often use graph-based approaches, focusing on low-order correlations between brain regions.
• These methods overlook complex, high-order correlations among different brain diseases and patient data.

Purpose of the Study:

• To propose a Hypergraph Foundation Model (HGFM) for brain disease diagnosis.
• To leverage high-order correlation patterns for improved diagnostic accuracy, particularly in low-data scenarios.

Main Methods:

• Developed an HGFM using self-supervised pretraining on high-order correlation structures.
• Implemented multidimensional pretraining tasks, including functional network link prediction and group interaction network link prediction.
• Utilized few-shot learning for fine-tuning on downstream brain disease diagnosis tasks.

Main Results:

• The HGFM effectively captures latent cross-dimensional high-order correlation patterns.
• Achieved superior performance in predicting four different brain diseases using functional magnetic resonance imaging (fMRI) data.
• Outperformed state-of-the-art methods across all diagnostic tasks.

Conclusions:

• HGFM offers a powerful, high-order correlation-driven approach for brain disease diagnosis.
• Demonstrates significant potential for clinical applications, especially in data-scarce environments.
• Highlights the value of hypergraph computational paradigms in medical diagnostics.