Unsupervised Sparse Multi-Task Learning With Application to Alzheimer's Disease

Area of Science:

• Neuroscience
• Biostatistics
• Medical Imaging

Background:

• Alzheimer's disease (AD) progression is challenging to track using high-dimensional fMRI data.
• Identifying interpretable brain functional connectivity biomarkers is crucial for understanding AD.
• Existing methods struggle with high dimensionality, disease-stage heterogeneity, and collinearity in fMRI data.

Purpose of the Study:

• To propose a unified statistical method, Sparse Multi-task Adaptive Regularization Truncation (SMART), for identifying AD biomarkers.
• To address challenges of high dimensionality, disease-stage heterogeneity, and connection collinearity in fMRI data.
• To enhance the accuracy, stability, and interpretability of biomarker identification for AD progression.

Main Methods:

• Developed SMART, a method incorporating L1-penalty for sparsity, L2-penalty for stable activation patterns, and a truncated L2-penalty (TLP) for adaptive grouping.
• Utilized a DC-ADMM algorithm for efficient computation and convergence to KKT points.
• Validated SMART through comprehensive simulation studies and real-data analysis of AD neuroimaging data.

Main Results:

• SMART demonstrated superior accuracy in identifying functional connectivity biomarkers for AD.
• The method showed enhanced stability in feature selection across different disease stages.
• SMART improved the interpretability of biomarkers for AD cohort studies.

Conclusions:

• SMART provides a robust and interpretable framework for identifying brain functional connectivity biomarkers in Alzheimer's disease.
• The method effectively handles the complexities of high-dimensional fMRI data, offering advantages over existing approaches.
• The accompanying R package facilitates the application of SMART in AD research.