Sleep Spindle Detection From Single-channel Eeg Using CNN-SNN Algorithm

The Sleep Spindle reflects the transient burst behavior of the brain during sleep in the frequency range of 11Hz-16 Hz,and is represented as a spindle waveform on the Electro Encephalogram(EEG),mainly generated by the interaction between the thalamocortical network and the reticular nucleus of the thalamus.By characterizing the sleep spindle,it is possible to study and discriminate the health status of individuals in terms of cognition,memory and related brain diseases.In the process of sleep spindle detection,the widely used manual detection method is still considered as the most accurate gold standard,but it requires a high level of expertise from physicians and the results vary greatly between physicians,so it takes up more clinical resources and has a mediocre performance.In order to solve the problem of automatic recognition of sleep spindle waves,the following studies were conducted in this thesis.1)Validation of traditional signal processing methods.The power,correlation,and root mean square of the sigma(11-16Hz)band signal relative to the broadband signal were thresholded and discriminated using 25 clinical sleep EEG data from Xuanwu Hospital of Capital Medical University.The performance of traditional detection methods was verified by such sleep spindle wave detection algorithms,and it was found that this class of methods is more sensitive to specific waveforms,but has greater limitations due to strict threshold detection,which cannot meet diverse clinical needs.2)Two convolutional neural network sleep spindle detection methods based on single channel EEG data are proposed.Two convolutional neural network structures,1D-CNN and2D-CNN,are constructed with the original waveform detection and the detection method combining time-frequency features,respectively.For the time-frequency characteristics of sleep spindle,this thesis uses single-channel EEG as the basis and obtains the time-frequency domain distribution of EEG by wavelet transform as the input of the convolutional neural network.By changing the input data from channel-time-domain to frequency-time-domain,the network is further improved for the sleep spindle detection problem by ensuring that the channel less data meets the detection requirements.Both the 1D-CNN and 2D-CNN proposed in this thesis outperform existing detection algorithms in terms of training speed and performance scores.3)The method of sleep spindle detection based on spiking neural network is proposed.The conversion of convolutional neural network by CNN-SNN method avoids the problem of difficult direct training of spiking neural network;on the basis of 1D-CNN and 2D-CNN,we get two kinds of spiking neural network with better structure for sleep spindle detection problem,so that spiking neural network can be used in sleep spindle detection problem.This enables the practical use of spiking neural networks for the sleep spindle detection problem and achieves good classification and detection performance,which is conducive to the construction of low-power sleep monitoring devices.