Classification of Alzheimer's Disease from structural MRI using sparse logistic regression with optional spatial...

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Diagnostics

Background:

• Classifying Alzheimer's Disease (AD) from structural MRI presents challenges due to high-dimensional voxel data.
• Standard logistic regression is unsuitable for datasets with more features (voxels) than samples.
• Sparse Logistic Regression (SLR) offers a solution by incorporating sparsity penalties for automatic feature selection.

Purpose of the Study:

• To apply and compare two formulations of Sparse Logistic Regression (SLR) for classifying Alzheimer's Disease.
• To evaluate the classification accuracy of SLR against Penalized Logistic Regression (PLR) and Maximum uncertainty Linear Discriminant Analysis (MLDA).
• To assess the spatial properties of classifiers generated by different SLR methods.

Main Methods:

• Utilized voxel-wise grey matter volumes from structural MRI scans of 69 Alzheimer's Disease patients and 60 healthy controls.
• Applied the original Sparse Logistic Regression (SLR) formulation, enforcing similar weights for correlated voxels.
• Implemented a spatially regularized Sparse Logistic Regression (SRSLR) formulation to promote smoothness in the discriminating vector.

Main Results:

• Both SLR and SRSLR demonstrated comparable and superior classification accuracies compared to PLR and MLDA.
• Cross-validation confirmed the effectiveness of the sparse logistic regression approaches.
• SRSLR generated classifiers with enhanced spatial smoothness, potentially improving biological interpretability.

Conclusions:

• Sparse Logistic Regression (SLR) is a highly effective method for classifying Alzheimer's Disease using neuroimaging data.
• The spatially regularized formulation (SRSLR) offers advantages in classifier smoothness without compromising accuracy.
• These findings suggest SLR methods are valuable tools for neurodegenerative disease classification and analysis.