Cortical surface biomarkers for predicting cognitive outcomes using group l2,1 norm

Area of Science:

• Neuroscience
• Biomarker Discovery
• Machine Learning in Medicine

Background:

• Traditional regression models for Alzheimer's disease (AD) often overlook interrelationships within neuroimaging data and between cognitive outcomes.
• This limitation can reduce the predictive power of neuroimaging measures as biomarkers for cognitive decline in AD.

Purpose of the Study:

• To develop and validate a novel sparse multitask learning model, Group-Sparse Multi-task Regression and Feature Selection (G-SMuRFS).
• To enhance the prediction of cognitive outcomes using detailed cortical thickness measures in an Alzheimer's disease cohort.

Main Methods:

• G-SMuRFS employs a group-level l2,1-norm strategy to group related features anatomically.
• The model incorporates prior knowledge of feature groupings and accounts for correlations among multiple cognitive outcomes.
• The approach was tested on a large cohort examining the predictive power of cortical thickness for three cognitive scores.

Main Results:

• G-SMuRFS demonstrated superior predictive performance compared to traditional regression methods.
• The model successfully identified a concise set of biologically meaningful cortical surface markers.
• The approach effectively leverages interrelationships within neuroimaging data and between cognitive outcomes.

Conclusions:

• G-SMuRFS offers a more powerful and biologically interpretable approach for predicting cognitive outcomes in Alzheimer's disease using neuroimaging data.
• The method advances the use of sparse multitask learning for biomarker discovery in neurological disorders.
• This study highlights the importance of considering data structure for optimal predictive modeling in AD research.