Improved sparse decomposition based on a smoothed L0 norm using a Laplacian kernel to select features from fMRI data.

Area of Science:

• Neuroimaging
• Machine Learning
• Data Science

Background:

• Functional magnetic resonance imaging (fMRI) data presents a
• few samples and large features
• challenge for multivariate classification.
• Sparse representation methods offer promise for fMRI feature selection despite computational limitations.

Purpose of the Study:

• To introduce and evaluate the Laplacian smoothed L0 norm (LSL0) approach for feature selection in fMRI data.
• To assess LSL0's performance in sparse source estimation and classification accuracy.

Main Methods:

• Proposed the Laplacian smoothed L0 norm (LSL0) method for fMRI feature selection.
• Utilized the Laplacian function to approximate the L0 norm within a sparse decomposition framework, building upon the smoothed L0 norm (SL0) method.

Main Results:

• LSL0 demonstrated feasibility and robustness in sparse source estimation and feature selection on both simulated and real fMRI data.
• LSL0 achieved higher classification accuracy compared to SL0, ICA, and t-test, particularly at high noise levels.
• LSL0 and SL0 required less computational time than ICA and t-test for fMRI decoding.

Conclusions:

• LSL0 surpasses SL0 in sparse source estimation accuracy under high noise conditions and in feature selection efficacy.
• Both LSL0 and SL0 methods provide superior feature selection performance compared to Independent Component Analysis (ICA) and t-test for fMRI decoding.