Low-dimensional feature fusion strategy for overlapping neuron spike sorting

Abstract

Abstract Complex action potentials of the brain neurons from extracellular recordings are usually represented as overlapping superimposed spikes resulting from the neuron spike bursts. In this paper, a novel feature fusion strategy is proposed for the high-precision classification of the neuron spikes. Based on the wavelet coefficients and the principal component analysis (PCA) features of the spikes, the eigenvectors of the adjacency matrices are constructed using the Locality Preserving Projections (LPP) algorithm. The fusion adjacency matrix of different features is constructed using different weighted adjacency matrices method. The feature fusion data based on the wavelet coefficients and PCA features are projected from high-dimensional space to low-dimensional space. The experiment results show that the noise level is 0.4, and the accuracy of this method is increased by 9% on average compared with non-fusion feature classification. According to the experimental results, a high-precision classification is obtained based on the feature fusion method with a low-dimensional feature space. The proposed feature fusion method effectively reduces the dimension number of the features and remedies the missing information.