Sleep spindle detection using modified extreme learning machine generalized radial basis function method

Abstract

Spindles are the main indicators of stage two sleep. Manual detection of spindles is tedious and time-consuming, hence many attempts have been performed to automatically detect them. Scientists have found that spindles are related to diseases such as Alzheimer's and Schizophrenia. Therefore, designing algorithms to extract certain spindle features to diagnose them with high accuracy is valuable. In this study, new chaotic features, along with a new training and classification algorithm has been used to detect spindles: an algorithm called MELM-GRBF (Modified Extreme Learning Machine). GBRF (Generalized Radial Basis Function) can update the radial basis function (RBF) using a parameter (τ), and the training algorithm is MELM. GRBF centers are selected randomly from the training data. The width and the parameter τ of GRBF are achieved considering two limitations: locality and coverage. In order to have the most similarity between spindle and non-spindle segments and to demonstrate the capability of these (chaotic features and classifier) algorithms, non-spindles are extracted immediately before spindles occur for one second in sleep stage two. After presenting chaotic (Higuchi Fractal Dimension, Katz Fractal Dimension, Sevcik Fractal Dimension, Largest Lyapunov and Entropy) and time series features, the ELM-RBF and MELM-GRBF statistical results are compared. The Best average results of MELM-GRBF classifier after 15 training trials were, 93.10%, 90.34% and 95.90% for accuracy, sensitivity and specificity, respectively and the variances were 0.78, 1.61 and 1.16, respectively. While the ELM-RBF results for accuracy, sensitivity and specificity were 91.06%, 85.83% and 96.32% with 1.64, 2.39 and 1.18 variance, respectively.