Wavelet transform has been widely applied in extracting characteristic information in spike sorting.As the wavelet coefficients used to distinguish various spike shapes are often disorganized, there still lacks in effective unsupervised methods to select the most discriminative features.In this paper, we propose an unsupervised feature selection method, employing kernel density estimation to select those wavelet coefficients with bimodal or multimodal distributions.This method is tested on a simulated spike data set, and the average misclassification rate after fuzzy C-means clustering has been greatly reduced, which proves this kernel density estimation based feature selection approach is effective in selecting the most discriminative features.