Regularized multi-task learning with individual-feature-based task correlations for Alzheimer's cognitive score...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Alzheimer's disease (AD) prediction is crucial for early intervention and management.
• Multi-task sparse learning aids in predicting multiple cognitive scores and identifying biomarkers.
• Existing methods struggle with inaccurate task correlations, hindering prediction performance.

Purpose of the Study:

• To develop a novel multi-task learning framework for Alzheimer's disease (AD) prediction.
• To capture task correlations at a fine-grained, feature-specific level.
• To improve the prediction of multiple cognitive scores and identify AD biomarkers.

Main Methods:

• Proposed the individual-feature-based task correlation matrices guided multi-task learning (IFTMTL) method.
• Constructed a non-smooth convex objective function for joint regression of multiple cognitive scores.
• Integrated feature-level task correlations and Pearson coefficients, optimized via an iterative algorithm.

Main Results:

• IFTMTL significantly outperformed 11 competitive methods in normalized mean squared error (nMSE) and correlation coefficient (CC) metrics.
• Achieved a 4.09% improvement in nMSE and 1.68% in CC compared to Multi-Task Feature Learning.
• Identified key affected brain regions in AD: left hippocampus, left middle temporal, and right entorhinal.

Conclusions:

• IFTMTL enhances AD prediction by incorporating feature-level task correlations, improving knowledge transfer.
• The method surpasses existing approaches in cognitive score prediction and biomarker identification.
• Identified hippocampus and middle temporal regions as crucial for AD prediction and clinical analysis.