Multi-modal classification of neurodegenerative disease by progressive graph-based transductive learning

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Graph-based transductive learning (GTL) is utilized when training data is scarce.
• Conventional GTL methods construct fixed graphs based on feature domain similarities, which may be suboptimal for label propagation due to outliers.
• This can limit the accuracy of predicting clinical scores and labels.

Purpose of the Study:

• To introduce a progressive GTL (pGTL) method for improved feature-to-phenotype alignment.
• To enhance classification accuracy in medical imaging studies, particularly for neurodegenerative diseases.
• To extend the pGTL approach to multi-modal imaging data.

Main Methods:

• A novel iterative approach refines inter-subject relationships using label domain representations.
• The intrinsic data representation is updated based on refined relationships.
• The method was extended to multi-modal imaging data for improved classification.
• Validation was performed on Alzheimer's disease and Parkinson's disease datasets.

Main Results:

• The proposed pGTL method demonstrated superior classification accuracy compared to state-of-the-art methods.
• pGTL effectively identified subjects across different stages of Alzheimer's and Parkinson's disease.
• The iterative refinement process led to more accurate feature-to-phenotype alignment.

Conclusions:

• pGTL offers a more robust and accurate approach to classification in scenarios with limited training data.
• The method shows significant potential for clinical applications in diagnosing and monitoring neurodegenerative diseases.
• Extending pGTL to multi-modal data further enhances its diagnostic capabilities.