Gaussian discriminative component analysis for early detection of Alzheimer's disease: A supervised dimensionality...

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarker Discovery

Background:

• Alzheimer's disease (AD) research utilizes multimodal neuroimaging for pattern characterization.
• Current machine learning approaches often focus on binary classification, with limited exploration of multiclass classification for AD stages.

Purpose of the Study:

• Introduce a supervised Gaussian discriminative component analysis (GDCA) algorithm for improved AD classification.
• Delineate subtle changes in early mild cognitive impairment (EMCI) compared to cognitively normal (CN) individuals.

Main Methods:

• Applied GDCA to multimodal MRI and PET data from 251 CN, 297 EMCI, 196 late MCI (LMCI), and 162 AD subjects (ADNI cohort).
• Utilized a dimensionality reduction technique considering interclass information to define an optimal eigenspace for maximizing eigenvector discriminability.

Main Results:

• Achieved 79.25% accuracy in distinguishing EMCI from CN using 38.97% of GDCA components.
• Attained 67.69% overall accuracy in multiclass AD classification, and 75.28% accuracy in discriminating MCI and AD from CN using 48.90% of components.

Conclusions:

• The GDCA method demonstrates superior performance over state-of-the-art techniques in AD-related multiclass classification.
• GDCA shows stability and reliability in identifying relevant features for optimal classification, with high potential for clinical diagnosis systems.