Study Of Sleep Staging Algorithm Based On Feature Selection

Sleep research is of great significance to people’s physical and mental health in daily work.Sleep staging is an important criterion for understanding sleep stage and evaluating sleep quality.Sleep staging is a difficult and time-consuming clinical classification process.Researchers mainly use intelligent algorithms to classify sleep stages automatically and accurately.This thesis mainlyconsisted of four parts: the pre-processing,feature extraction,feature selection and classification algorithms.Firstly,it used heuristic threshold and wavelet denoising function to denoise sleep EEG,thus completing the signal preprocessing.The signal-to-noise was 14 db,which proved that the algorithm can effectively reduce the noise of EEG signal,and it also can protect useful signal spikes and mutations.Then 15 feature algorithms were extracted from three domain: time domain,time-frequency domain and non-linear domain,from which 30 feature parameters wereobtained.They were hjorth parameters,energy of δ/α,lz complexity,fractal dimension,maximum Lyapunov exponent,Hurst exponent,kurtosis,skewness,Tsallis entropy,permutation entropy,fuzzy entropy,sample entropy,standard deviation(α β θ δ),maximum value(α β θ δ)and wavelet energy of(α β θ δ).At the same time,the parameter selection of feature algorithm for sleep EEG signal was studied.Then fisher score(FS),sequential floating forward selection(SFFS),fast correlation-basdfilter solution(FCBF)were used for feature selection.A two-level learning algorithm based on staking model is proposed to classify the six stages of sleep.The first learning classifiers for this algorithm werek-nearest neighbor(KNN),random forest(RF),extremely randomized trees(ERT),multilayerperceptron(MLP),extreme gradient boosting and XGBoost,the second layer was a Logistic regression.It was compared with RF,gradient boosting decision tree(GBDT)and XGBoost.Finally,the Sleep EEG signal in Sleep-EDF database was used to verify the algorithm.The results showd that the maximum classification accuracy(0.9667)can be achieved by using FS feature selection algorithm in Stacking.The kapper coefficient of the algorithm was 0.96.The results showd that proposed method can accurately sleep staging using single-channel EEG and had a high ability to predict sleep staging.