GSAformer: Group sparse attention transformer for functional brain network analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional brain network (FBN) analysis using fMRI is crucial for classifying neurological and mental disorders.
• Traditional methods often separate network construction from classification, leading to suboptimal diagnostic models.
• Existing transformer-based approaches overlook shared patterns among subjects, limiting their effectiveness.

Purpose of the Study:

• To propose GSAformer, a group sparse attention-based model for enhanced brain disorder diagnosis.
• To integrate population-level prior knowledge and inter-subject relationships into a unified framework.
• To improve the generalization and interpretability of brain disorder classification models.

Main Methods:

• Constructing brain connectivity matrices using Pearson's correlation for individual subjects.
• Incorporating a group sparse prior into the transformer architecture to model inter-subject relationships.
• Applying group sparsity across attention matrices and using a maximum mean discrepancy (MMD) constraint for consistency.

Main Results:

• GSAformer demonstrated improved classification performance across three public datasets, with accuracy increases of 3.8%, 4.1%, and 14.7% compared to standard transformers.
• The model effectively captures group sparse topological structures within populations.
• The framework supports end-to-end adaptive learning with computational complexity comparable to standard transformers.

Conclusions:

• GSAformer offers a significant advancement in brain disorder classification by effectively modeling inter-subject relationships and population-level prior knowledge.
• The proposed method enhances classification accuracy and model interpretability.
• This approach represents a promising direction for leveraging advanced machine learning techniques in clinical neuroscience.