Constructing Dynamic Brain Functional Networks via Hyper-Graph Manifold Regularization for Mild Cognitive Impairment...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Brain functional networks (BFNs) are crucial for diagnosing brain diseases, but traditional methods using manifold regularization (MR) only capture pairwise relationships.
• Existing MR techniques struggle to represent complex, high-order functional interactions between multiple brain regions.

Purpose of the Study:

• To develop a novel method for constructing dynamic BFNs (DBFNs) using hyper-graph MR (HMR) to better capture multi-regional interactions.
• To apply the proposed HMR-based DBFNs for improved classification of mild cognitive impairment (MCI) subjects.

Main Methods:

• Constructed DBFNs using Pearson's correlation (PC) and formulated it as an optimization problem.
• Utilized k-nearest neighbor (KNN) to build a hyper-graph and derive a hyper-graph manifold regularizer.
• Integrated the hyper-graph manifold regularizer and L1-norm regularizer into the PC-based model to obtain sparse DBFNs (SDBFNs).

Main Results:

• The proposed SDBFNs method achieved superior classification performance for MCI subjects compared to existing state-of-the-art methods.
• Achieved high classification accuracy (ACC) of 82.49%, sensitivity (SEN) of 77.25%, specificity (SPE) of 87.74%, and AUC of 0.9021.
• Demonstrated the biological significance and effectiveness of the expanded MR method and DBFNs.

Conclusions:

• The HMR-based DBFNs method offers a more biologically significant approach to modeling brain functional interactions.
• This technique significantly enhances the classification performance for MCI, providing valuable insights for Alzheimer's disease research and auxiliary diagnosis.