Modeling Disease Progression via Fused Sparse Group Lasso

Area of Science:

• Neuroscience
• Computational Biology
• Biostatistics

Background:

• Alzheimer's Disease (AD) is a leading cause of age-related neurodegeneration.
• Accurate prediction of AD progression and identification of pathological biomarkers are crucial for clinical diagnosis and prognosis.
• Existing methods may not fully capture the complex temporal dynamics of disease progression.

Purpose of the Study:

• To develop novel multi-task learning techniques for predicting Alzheimer's Disease progression using cognitive scores.
• To identify biomarkers predictive of disease progression over time.
• To incorporate temporal smoothness into disease progression models.

Main Methods:

• Proposed a novel convex fused sparse group Lasso (cFSGL) formulation for simultaneous biomarker selection across multiple time points and temporal smoothness.
• Developed efficient computational methods using proximal operators and accelerated gradient descent for the cFSGL formulation.
• Introduced two non-convex formulations using difference of convex (DC) programming to mitigate shrinkage bias.

Main Results:

• Demonstrated the effectiveness of the proposed progression models using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).
• The novel methods showed improved performance compared to existing approaches for predicting disease progression.
• Longitudinal stability selection identified temporal patterns of biomarkers relevant to Alzheimer's Disease progression.

Conclusions:

• The developed multi-task learning techniques provide effective tools for predicting Alzheimer's Disease progression.
• The proposed methods enable the identification of key biomarkers associated with disease advancement.
• These findings contribute to better clinical diagnosis, prognosis, and understanding of Alzheimer's Disease.