Deep sparse multi-task learning for feature selection in Alzheimer's disease diagnosis.

Area of Science:

• Neuroimaging analysis
• Machine learning for medical diagnosis
• Computational neuroscience

Background:

• Alzheimer's disease (AD) and mild cognitive impairment (MCI) diagnosis increasingly relies on neuroimaging.
• High-dimensional neuroimaging data with limited samples pose challenges for robust computer-aided diagnosis.
• Sparse regression is a promising machine learning technique but existing methods have limitations in feature selection.

Purpose of the Study:

• To develop a novel deep architecture for recursively discarding uninformative features in neuroimaging data.
• To improve the accuracy of computer-aided diagnosis systems for AD and MCI.
• To leverage feature importance from regression coefficients for enhanced classification.

Main Methods:

• Proposed a novel deep architecture for hierarchical sparse multi-task learning to discard uninformative features.
• Utilized optimal regression coefficients as feature weighting factors in subsequent hierarchical learning stages.
• Incorporated clustering-induced subclass labels as target response values, considering sample class distributions.

Main Results:

• Demonstrated the superiority of the proposed weighted sparse multi-task learning method on the ADNI cohort.
• Achieved high accuracy in both binary and multi-class classification tasks for AD/MCI diagnosis.
• Outperformed existing state-of-the-art methods in neuroimaging-based AD/MCI diagnosis.

Conclusions:

• The proposed hierarchical deep learning approach effectively addresses the challenges of high-dimensional neuroimaging data for AD/MCI diagnosis.
• The method enhances feature selection by recursively discarding irrelevant information and weighting important features.
• This technique offers a robust and superior alternative for computer-aided diagnosis of neurodegenerative diseases.