Sleep Staging Algorithm Based On Night Audio Signal

People are under various pressures in contemporary society,which leads to the continuous decline of sleep quality,and the low quality of sleep can have a huge negative impact on life.Therefore,studying and analyzing sleep status can not only improve the quality of sleep,but also help to diagnose sleep-related diseases,which is of great significance to people's daily life and work as well as personal health status.The non-contact sleep monitoring system based on audio sensor can realize long-term sleep monitoring without any burden.Because of it's simplicity and portability,non-contact sleep monitoring system based on audio sensor can become an effective means of household sleep monitoring and medical sleep disease detection.This paper presents a sleep stage algorithm based on night-time audio signals with multiple types of features,which can effectively divide the sleep stage.This algorithm based on audio sensors to collect night-time audio signals,and the audio of respiratory events were get after pre-processing.Analyzing the acoustic characteristics of respiratory event audio to get feature set,then combine with classifiers to achieve sleep staging.The pre-processing algorithm flow consist of pre-emphasis,windowing and framing,noise reduction and endpoint detection algorithm of event,to detecting and extracting the audio of respiration event from nocturnal audio signal.In this paper,the acoustic characteristics of respiratory audio are analyzed from three aspects:time domain,linearity and non-linearity respectively.After acoustic characteristics analysis,the acoustic feature set representing respiratory events is obtained.Then the staging models based on SVM,K nearest neighbor and decision tree are constructed according to the acoustic feature set,and those staging models are used to perform sleep staging.In this paper,the experimental means and experimental environment are designed.The nocturnal sleep audio signal of 13 experimenters are collected.After pre-processing and acoustic analysis,the respiratory acoustic set is obtained.The effects of different classifiers,different integration algorithms,timing information,and feature selection algorithm on the sleep stage results in four stages(REM/deep sleep/light sleep/awake)were analyzed.Experimental results show:Sleep staging accuracy of simple tree classifier is 61.58%,higher than linear SVM 's 60.6%and fine KNN 's 56.65%,so this paper choose the decision tree to achieve sleep stage classification;Sleep staging accuracy of bagged tree classifier was 66.74%,higher than 61.58%of simple tree,58.49%of medium tree,56.72%of complex tree and 64%of boosted tree,indicating that bagging can enhance decision tree classifier's performance;After adding the time information,the sleep staging accuracy of bagged tree increased from 66.74%to 74.52%,indicating that the addition of time information can improve the staging model's accuracy;After adding feature selection algorithm,the accuracyof bagged tree is increased from 74.52%to 76.15%,which shows that feature selection can improve the staging model accuracy too.