Detecting sleep spindles in EEGs using wavelet fourier analysis and statistical features

One of the more difficult tasks in sleep stage scoring is the detection of sleep spindles. Developing an effective method to identify these transitions in sleep electroencephalogram (EEG) recordings is an ongoing challenge, as there are typically hundreds of such transitions in each recording. This paper proposes a statistical model and a method based on wavelet Fourier analysis to detect sleep spindles. In this work, spindle detection is achieved in two phases: a training phase and a testing phase. An EEG signal is first divided into segments, using a sliding window technique. The size of the window is 0.5 s, with an overlap of 0.4 s. Then, each EEG segment is decomposed using a discrete wavelet transform into different levels of decompositions. The wavelet detail coefficient at level 3 (D3) is selected from these parameters, and this is passed through a fast Fourier transform to identify the desired frequency bands {α, β, θ, δ, γ}. Ten statistical characteristics are extracted from each band. Nonparametric Kruskal-Wallis one-way analysis of variance is used to select the important features, representing each of the 0.5 s EEG segments. To detect all possible occurrences of sleep spindles in the original EEG signals, four different window sizes of 0.25, 1.0, 1.5 and 2.0 s are also tested. Finally, the extracted features are used as the input to four classifiers to detect the sleep spindles: a least-squares support vector machine (LS-SVM), K-nearest neighbours, a K-means algorithm and a C4.5 decision tree. The obtained results demonstrate that the proposed method yields optimal results with a window size of 0.5 s. The maximum averages of accuracy, sensitivity and specificity are 97.9%, 98.5% and 97.8%, respectively. This method can efficiently detect spindles in EEG signals, and can assist sleep experts in analysing EEG signals