Sleep Disorder Recognition Method Based On Codec Network

Sleep quality is closely related to human health.Previous studies on sleep mostly focused on sleep disorders of apnea,but few on non apnea.The detection of sleep disorders is usually to score the arousal caused by sleep disorders,which is usually done manually by sleep experts by checking PSG records of several periods.This method has two obvious defects:one is that the task is very cumbersome and time-consuming;the other is that the scoring result depends on the knowledge and experience of the rater.Therefore,a well functioning and automated non apnea/hypopnea wake-up detection method will help health researchers to determine the impact of these events on health and develop more effective treatments to reduce arousal frequency.In this paper,we use a variety of original polysomnograms to establish a deep learning based physiological time series codec network for the analysis of non apnea/hypopnea sleep disorders.In traditional machine learning,it is usually necessary to process the original biological signal and extract features,which is a very time-consuming problem and requires a lot of domain knowledge and experiments.The analysis of sleep disorders depends on time series.Inspired by image segmentation,the model uses an end-to-end method to map a complete sleep sequence to the corresponding class label sequence.Our model can automatically learn the variable interaction between signals and any related time dependence,and automatically extract these wake-up features from rich physiological time series.The main work of this paper is as follows:1.Using Fourier convolution to fuse the time-frequency characteristics of polysomnograms.Polysomnograms is obtained by sampling in the time dimension,which can only reflect the time domain characteristics of the signal,and many information of identifying sleep disorders are contained in the frequency domain characteristics.We combine the time-domain and frequency-domain features of the signal by Fourier convolution to reflect the essential features of the sleep map from two dimensions,which is helpful to model recognition.2.This paper proposes a network based on encoding and decoding structure to deal with the task of labeling sleep disorders.The input of the model is the complete sleep map of the subjects to ensure that the model can capture any dependent features.The algorithm combines the encoding and decoding network with the long-term and short-term memory network,and directly uses the original polysomnogram as the input of the algorithm.Compared with other algorithms on the benchmark data set,the algorithm achieves better recognition effect.3.On the basis of the encoding and decoding algorithm,this paper further proposes an encoding and decoding network based on attention mechanism,which selectively focuses on the relevant input signals.By weighting the output of the encoder,not only the input and output can be aligned,but also more context information of the original physiological signal can be used.Attention model can help to solve some problems existing in traditional codec network,such as the insufficient utilization of information between physiological signals,and the performance of the model may decline with the increase of the duration of input polysomnogram.4.Propose a multi task learning mechanism to improve the generalization ability of the model,and use other related auxiliary tasks to learn more complex features,so as to improve the network performance.According to the proposed learning mechanism,the corresponding data set is constructed.Multi task learning mechanism can not only help the model to better identify the target area,but also provide advantages for auxiliary tasks,and make a great contribution to the generalization ability of the model.