High-order graph matching based feature selection for Alzheimer's disease identification

Area of Science:

• Neuroimaging analysis
• Machine learning in medicine
• Biomedical data science

Background:

• Traditional l1-norm feature selection methods analyze samples individually, neglecting inter-sample geometric relationships.
• Geometric relationships among target vectors in training data can offer valuable insights for predictive modeling.

Purpose of the Study:

• To develop a novel feature selection method that incorporates geometric relationships among samples.
• To improve classification accuracy for neurodegenerative diseases by leveraging graph-matching techniques.

Main Methods:

• Formulated feature selection as a graph-matching problem between predicted and target graphs.
• Nodes in the predicted graph represent regional gray matter volume or cortical thickness features.
• Nodes in the target graph represent class labels and clinical scores.
• Devised novel regularization terms in sparse representation for high-order graph matching.
• Fused selected features in kernel space for classification.

Main Results:

• Achieved high classification accuracies on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• Obtained 92.17% accuracy for Alzheimer's disease (AD) classification.
• Achieved 81.57% accuracy for mild cognitive impairment (MCI) classification.

Conclusions:

• The proposed graph-matching feature selection method effectively utilizes geometric relationships for improved classification.
• This approach offers a promising advancement in the analysis of neuroimaging data for disease diagnosis.