Manifold regularized multitask feature learning for multimodality disease classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Data Analysis

Background:

• Multimodality methods improve Alzheimer's disease (AD) and mild cognitive impairment (MCI) classification.
• Existing methods often overlook modality-specific data distribution, crucial for accurate classification.
• Multitask feature selection aims to jointly select common features across modalities.

Purpose of the Study:

• To propose a manifold regularized multitask feature learning method for AD/MCI classification.
• To preserve intrinsic relatedness among multiple data modalities.
• To retain essential data distribution information within each modality for improved classification.

Main Methods:

• Feature learning is treated as a single task for each modality.
• Group-sparsity regularizer captures relatedness among tasks (modalities) for joint feature selection.
• A manifold-based Laplacian regularizer preserves data distribution information within each task.
• Multikernel support vector machine (SVM) is used for data fusion.
• The method is extended to a semisupervised learning setting.

Main Results:

• The proposed method achieves improved classification performance for AD and MCI.
• It successfully identifies disease-related brain regions crucial for diagnosis.
• Experimental validation uses magnetic resonance imaging (MRI), FDG-PET, and cerebrospinal fluid (CSF) data from the AD neuroimaging initiative database.

Conclusions:

• The manifold regularized multitask feature learning method effectively integrates multimodal data for enhanced AD/MCI classification.
• Preserving modality-specific data distributions alongside inter-modality relatedness is key to improving diagnostic accuracy.
• The method aids in discovering relevant biomarkers for disease diagnosis.