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

•We propose the LSL0 method for feature selection of fMRI data.•LSL0 showed better sparse source estimation than SL0 at high noise levels.•LSL0 showed better classification accuracy than SL0 in most cases.•LSL0 and SL0 showed higher classification accuracy than the t-test and ICA method.

BackgroundFeature selection plays an important role in improving the classification accuracy of multivariate classification techniques in the context of fMRI-based decoding due to the “few samples and large features” nature of functional magnetic resonance imaging (fMRI) data. Recently, several sparse representation methods have been applied to the voxel selection of fMRI data. Despite the low computational efficiency of the sparse representation methods, they still displayed promise for applications that select features from fMRI data.New methodIn this study, we proposed the Laplacian smoothed L0 norm (LSL0) approach for feature selection of fMRI data. Based on the fast sparse decomposition using smoothed L0 norm (SL0) (Mohimani, 2007), the LSL0 method used the Laplacian function to approximate the L0 norm of sources.ResultsResults of the simulated and real fMRI data demonstrated the feasibility and robustness of LSL0 for the sparse source estimation and feature selection.Comparison with existing methodsSimulated results indicated that LSL0 produced more accurate source estimation than SL0 at high noise levels. The classification accuracy using voxels that were selected by LSL0 was higher than that by SL0 in both simulated and real fMRI experiment. Moreover, both LSL0 and SL0 showed higher classification accuracy and required less time than ICA and t-test for the fMRI decoding.ConclusionsLSL0 outperformed SL0 in sparse source estimation at high noise level and in feature selection. Moreover, LSL0 and SL0 showed better performance than ICA and t-test for feature selection.