Research On Sleep State Assessment Technology Based On Brain Waves

Sleep is very important.Unfortunately,nearly 25% of adults have sleep disorders.Clinically,EEG is the first choice for doctors to analyze sleep problems.However,the amount of EEG data is huge,the interference is serious,and the interpretation criteria are complicated.The interpretation between experts would have much difference.In addition,the collection of medical EEG is complicated.The current artificial sleep analysis consumes huge personal and financial resources,and the sleep problems faced by the public urgently need convenient,stable,and automated analysis methods.In view of this,this thesis had mainly done three work on the research of sleep state assessment technology based on brain waves.First,this thesis studied the detection technology of the human body’s moments of falling asleep and awakening from the perspective of traditional signal analysis.This part proposed a specific-duration and high-energy sine wave model,which extracts the subsignals that reflect the human sleep state from the brain waves according to the frequency attributes of the sub-signals.This thesis had shown that specific-duration and high-energy sine waves could effectively reflect the moment of falling asleep and awakening of the human.In the Dreams database,the Kappa consistency between the specific-duration and high-energy sine wave model and the expert was 0.58.The Kappa consistency also got0.59 on the Sleep-EDF when the model was applied to the database directly.Second,this thesis adopted the idea of feature extraction combined with machine learning to study the automatic sleep staging algorithm.This part proposed the sine wave filtering theory and the widely applicable second-order feature generation scheme One Hot random forest cotyledons and nearest neighbor features.Analysis of variance and chart analysis were adopted in the feature analysis.The classifier used the Stack model integration theory,which combined the prediction results of four machine learning algorithms based on different principles: support vector machine,neural network,nearest neighbor algorithm and XGBoost.In the Sleep-EDF database,the classification accuracy of 2 to 6 sleep stages were 99%,95%,93%,91%,and 90% respectively,which were the same as the published works,and some exceeded.The accuracy in the Dreams database were 96%,90%,86%,83%,and 81%,which exceeded the published results.The model had good stability,and the overall performance of the same model in two different databases was better than the published works.Third,in order to test the theoretical research results of the subject in practice,this thesis designed and constructed an experimental platform for brainwave acquisition and measurement and control.In the short-term sleep monitoring experiments,29 valid experiment records were obtained from 7 subjects.The accuracy of the falling asleep and awakening moment detected by the specific-duration and highenergy sine wave model was 88%.The automatic sleep staging algorithm had an average detection accuracy of 85% for the subjects’ sleep status,and the accuracy of falling asleep and awakening moment based on the sleep staging results of the algorithm was 92%.The two theoretical research results of this thesis had shown good practical results under a simple experimental device.It could be initially applied to the assessment of human sleep status.