A sleep spindle detection algorithm based on SVM and WT

Sleep spindles play an important role in assessing sleep pathphysiological. Recently, many researches also indicated some differences in spindle characteristics for different neurological disorders. However, it is a tough task to detect sleep spindles because of the unclear definition of sleep spindle characteristics. The detection is usually done through visual inspection of electroencephalogram(EEG) signals by experts, which has the weaknesses of time-consuming and great subjectivity. This paper proposes a novel sleep spindle detection algorithm, in which we use wavelet transform(WT) to extract the spectrum information of EEG in range from 8 to 25 Hz. Gaussian function is used to normalize the power spectrum and preserve the significant spectrum information. The probability of having a spindle at a given sample is calculated by summing the power spectrum of frequency corresponding to spindle range. After using a smoothing algorithm on the probabilities, we apply Support Vector Machine(SVM) in order to obtain the threshold of probability of spindle candidate. The results of experiment on the public DREAMS sleep spindle database demonstrate the efficiency of the proposed algorithm.