CN115067875A

CN115067875A - Neonate sleep staging method based on compressed electroencephalogram

Info

Publication number: CN115067875A
Application number: CN202210477229.9A
Authority: CN
Inventors: 张跃慧; 陈炜; 陈晨; 许艳
Original assignee: Fudan University
Current assignee: Fudan University
Priority date: 2022-05-03
Filing date: 2022-05-03
Publication date: 2022-09-20

Abstract

The invention belongs to the technical field of biological medical treatment, and particularly relates to a neonate sleep staging method based on compressed electroencephalogram. The method comprises the following steps: acquiring an original neonatal brain electrical signal data set, carrying out stage labeling on sleep stages, and taking stage labeling results as standard stage results; performing compression conversion on the original electroencephalogram signal after the background noise is filtered to obtain an electroencephalogram signal after the compression conversion, so that the data volume is greatly reduced; aligning the transformed electroencephalogram signal with the stage mark of the original electroencephalogram; training by using sleep staging training data to obtain a full-automatic neonate sleep staging network; sleep staging the sleep process of the neonate using a fully automated neonatal sleep staging network. The invention can compress the electroencephalogram signals to reduce the redundancy of the original data information, and constructs an automatic staging model to stage the sleep process of the neonate by a more portable electroencephalogram signal data source, thereby having wide clinical application prospect.

Description

Neonate sleep staging method based on compressed electroencephalogram

Technical Field

The invention belongs to the technical field of biological medical treatment, and particularly relates to a sleep staging method for a newborn.

Technical Field

Clinically, electroencephalogram (EEG) is commonly used for diagnosis and sleep staging of neonates during their sleep, and sleep staging is most often done manually by a number of medical professionals. The information contained in the original electroencephalogram signal is extremely rich, which causes the problem that the original electroencephalogram signal contains a large amount of redundant information which is not needed by the problem of sleep stage, and causes the problem of inconvenient utilization of data storage, transmission, calculation and the like. In addition, the characteristic of 'boldness but not delicacy' of the original electroencephalogram signals also causes strong signal specificity among different testees, and common sleep characteristics are difficult to acquire from electroencephalogram data of a plurality of testees.

The interpretation and utilization of the original brain electrical signals also need to be performed by experienced neurobiologists or pediatric neurologists with professional knowledge of newborn brain electrograms. The difficulty in understanding the electroencephalogram signals also limits the popularization of the sleep abnormality diagnosis knowledge in medical care personnel.

In recent years, deep learning technology has attracted much attention in medical diagnosis, and particularly, application of a convolutional neural network and a long-short term memory unit to sleep staging problems greatly improves the accuracy of adult sleep staging, but currently, a manual mode which is tedious in process and needs a lot of time is still adopted clinically to perform sleep staging work.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, make up the vacancy in the field of research on sleep staging of newborns, and provide a method for staging sleep of newborns based on compressed electroencephalogram.

Compared with the existing sleep staging method, on one hand, the method can greatly improve the convenience of the electroencephalogram signals in data storage, transmission and calculation by compressing the original electroencephalogram signals, and the compressed data can well resist the data specificity among different testees. On the other hand, the invention realizes automatic staging of the sleep process of the newborn by utilizing the deep learning model, and has higher application value.

The invention provides a sleep staging method for a newborn, which comprises the following specific steps.

(1) Acquiring N original electroencephalograms of the neonate in a sleep state, and labeling the sleep stage of the electroencephalograms of each testee; and taking the stage labeling result of the original electroencephalogram signal as a standard stage result.

(2) Filtering the original electroencephalogram signals obtained in the step (1), filtering background noise, and then performing compression transformation on the original electroencephalogram signals to obtain the electroencephalogram signals after the compression transformation, wherein the specific steps are as follows:

(2-1) filtering background noise of the electroencephalogram signals marked in the sleep stage by using a first filter, namely a Butterworth band-pass filter, and attenuating signal activities of less than 1.6Hz and more than 70 Hz; then, a second filter, namely a finite-length single-bit impulse response (FIR) filter is utilized to remove artifacts caused by sweating, movement, electrocardio-activity, electrode interference and the like of a tested person, and simultaneously, signals are enhanced within a frequency band of 2-15Hz with the slope of 12 dB/decade;

(2-2) carrying out full-wave rectification on the electroencephalogram signals after filtering, namely converting negative voltage into positive voltage;

(2-3) performing envelope detection on the rectified signal by using a third filter, namely a square low-pass filter, so as to obtain the variation trend of the amplitude of the electroencephalogram signal;

(2-4) carrying out segmentation division on the envelope obtained in the step (2-3) according to a compression ratio, wherein the compression ratio is the ratio of the sampling rate of the original electroencephalogram signal to the sampling rate of a signal obtained in advance after compression;

(2-5) compressing each section of electroencephalogram signal, and keeping the maximum value and the minimum value in each section of envelope as the upper end point and the lower end point of the compressed signal;

(2-6) finally, displaying the compression result in a semilogarithmic mode; the voltage amplitude of 0-10 MuV is displayed by a linear coordinate system, and the voltage amplitude of 10-100 MuV is displayed by a logarithmic coordinate system; to prevent activity changes of small signals from being annihilated by large signals.

(3) Aligning the compressed and transformed electroencephalogram signal obtained in the step (2) with the stage mark of the original electroencephalogram in the step (1) to obtain marked sleep stage training data.

(4) Training a full-automatic neonate sleep staging network model by using sleep staging training data;

the full-automatic newborn sleep staging network model consists of a convolutional neural network and a bidirectional cyclic long-short term memory unit network model; the training of the model comprises a pre-training stage and a sequence training stage;

(4-1) the specific process of the pre-training stage is as follows:

(4-1-1) calculating the number of sleep stages in the sleep data;

(4-1-2) comparing the number of sleep stages, if the ratio of the sleep stages is in the range of 0.8-1.2, entering the step (4-1-4), otherwise entering the step (4-1-3);

(4-1-3) up-sampling sleep stages with a smaller proportion in the data or down-sampling sleep stages with a larger proportion in the data according to the data amount of each sleep stage so as to balance the data;

(4-1-4) inputting the sleep data with balanced data into a coding module mainly composed of a convolutional neural network for data feature extraction;

(4-2) the specific process of the sequence training phase is as follows:

(4-2-1) inputting the data features extracted in the pre-training stage into the bidirectional cyclic long and short term memory unit network;

(4-2-2) decoding the characteristic of each sleep stage through the bidirectional cyclic long-short term memory unit network, and at the same time, learning the dependency relationship among a plurality of sleep stages by the network to obtain the sequence characteristic of the sleep data;

(4-2-3) combining the sleep stage data characteristics obtained in the step (4-1) with the sequence characteristics obtained in the step (4-2-2) through jump connection, and synthesizing information to predict the sleep stage;

during training, the input of the full-automatic neonate sleep staging network is an electroencephalogram signal of each tested person after compression and transformation, and loss functions with different weights are set according to the ratio of different sleep stages in the total sleep time and the recognition difficulty of different sleep stages, so that a more accurate sleep staging result is realized; outputting the stage training data corresponding to the electroencephalogram signal; and after the training is finished, obtaining the fully-automatic neonate sleep staging network after the training is finished.

(5) Performing sleep staging on the electroencephalogram signals to be staged by using the full-automatic neonate sleep staging network obtained after the training in the step (4); the method comprises the following specific steps:

(5-1) acquiring an electroencephalogram signal of a neonate in a sleep process, and carrying out compression change on the electroencephalogram signal in the step (2), wherein the electroencephalogram signal to be staged is similar to the electroencephalogram signal acquired in the step (1) in characteristics;

(5-2) staging the staging signal by using the trained neonatal sleep staging network;

inputting the electroencephalogram signals to be staged into the sleep staging network trained in the step (4), and outputting the sleep staging result of the signals to be staged by the network.

In the invention, the sleep stages are as follows: awake (W), Active Sleep (AS) and Quiet Sleep (QS).

In the invention, the first filter is a Butterworth band-pass filter for filtering. The second filter is a finite-length unit impulse response (FIR) filter. And the third filter adopts a square low-pass filter for envelope detection.

In the invention, the up-sampling is to interpolate the original signal sequence, i.e. a new sleep stage is inserted between certain sleep stages on the basis of the original signal sampling so as to supplement the sleep stages with less quantity.

In the invention, the down-sampling is to extract the original signal sequence, namely to randomly extract a certain sleep stage on the basis of the original signal sampling so as to reduce a plurality of sleep stages.

The invention has the characteristics and beneficial effects that:

the invention can greatly improve the convenience of the electroencephalogram signals in data storage, transmission and calculation by compressing the original electroencephalogram signals, and the compressed data can well resist the data specificity among different testees.

The invention extracts the overall characteristics of the original EEG signal change after the EEG signal is compressed and transformed, so that the signal characteristics are easier to read and understand, and the EEG signal interpretation knowledge is more widely popularized. The invention improves staging precision by carrying out data equalization and weight setting on the newborn data.

The invention provides a method for combining compressed and transformed electroencephalogram signals with a deep learning model. The invention solves the problems of different training data skewness and different feature learning difficulty degrees by adopting an upsampling and weight configuration mode. After the network training of the newborn sleep stage is completed, the user can utilize the compressed and transformed electroencephalogram signals to complete the automatic stage of the newborn sleep process, and the method has high practical value.

Drawings

Fig. 1 is an overall flowchart of the sleep staging method for a neonate according to the present invention.

FIGS. 2 and 3 are comparative diagrams of the compression of original 500Hz brain electrical signals to 1Hz by the compression method of the present invention.

Fig. 4 is a confusion matrix obtained by a fully automatic neonatal sleep staging network using compressed brain electrical signals.

Detailed Description

The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

The overall process of the sleep staging method for the newborn is shown in fig. 1 and comprises the following steps.

(1) Acquiring original electroencephalogram signals of at least 20 neonates in a sleep process, wherein the signal duration of each neonate is at least about 2 hours, and carrying out sleep stage labeling on the electroencephalogram signals of each neonate by an expert to obtain labeled sleep stage training data corresponding to the signals;

in this embodiment, the original electroencephalogram signals acquired by the medical electroencephalogram acquisition device do not include the classification labels of the sleep stages, and in order to enable the computer to learn the classification characteristics of the sleep stages, a medical sleep specialist needs to manually label the corresponding sleep stage in each signal. Specifically, the expert should label the sleep stages contained in each signal according to the rules promulgated by the american society for sleep medicine (AASM). These labeled data can be used to train a deep learning model.

The american society of sleep medicine divides the sleep process of adults into five stages: a waking period (W), a rapid eye movement period (REM), a non-rapid eye movement first period (N1), a non-rapid eye movement second period (N2) and a non-rapid eye movement third period (N3). Corresponding to the neonate were respectively: awake (W), Active Sleep (AS) and Quiet Sleep (QS).

(2) Filtering the original electroencephalogram signals obtained in the step (1), and performing compression conversion on the original electroencephalogram signals with the background noise removed to obtain the electroencephalogram signals after the compression conversion;

removing larger background noise of the original electroencephalogram signal through a Butterworth band-pass filter to obtain a purer electroencephalogram signal; carrying out compression transformation on the preprocessed electroencephalogram signals, and specifically comprising the following steps:

(2-1) passing the preprocessed electroencephalogram signal through an asymmetric band-pass filter to attenuate signal activities smaller than 2Hz and larger than 15 Hz. The filter is used for removing artifacts caused by sweating, movement, electrocardio-activity, electrode interference and the like of a tested person, and simultaneously enhancing signals by using a slope of 12dB/decade in a frequency band of 2-15 Hz;

in the embodiment, the asymmetric band-pass filter is designed by adopting a Parks-McClellan algorithm;

(2-3) carrying out envelope detection on the rectified signal, and carrying out envelope detection through a square low-pass filter to obtain the variation trend of the amplitude of the electroencephalogram signal;

(2-6) finally, displaying the compression result in a semilogarithmic mode; the voltage amplitudes of 0-10 μ V are shown in a linear coordinate system and 10-100 μ V in a logarithmic coordinate system. To prevent activity changes of small signals from being annihilated by large signals;

fig. 2 and fig. 3 show a contrast diagram of the electroencephalogram signal compressed from 500Hz original electroencephalogram to 1Hz, so that the compression result well retains the characteristics of the overall signal activity of the original electroencephalogram, other redundant characteristics are removed, and the data volume is greatly reduced.

(3) In this embodiment, the sleep stage annotation of the original electroencephalogram signal is already obtained in step (1), and at this time, only the compressed and transformed electroencephalogram signal obtained in step (2) needs to be aligned with the stage annotation of the original electroencephalogram in step (1), so as to obtain annotated sleep stage training data.

(4) Training by using sleep staging training data to obtain a full-automatic neonate sleep staging network;

in this embodiment, the full-automatic neonatal sleep staging Network is composed of a pre-training encoding module and a sequence decoding module, such as a Convolutional Neural Network (CNN), a Gate controlled loop Unit (GRU), a Long Short Term Memory (LSTM), and the like, and the embodiment uses a Convolutional Neural Network and a Long Short Term Memory Network model;

in general, since the ratios of different sleep stages in the total sleep time duration of the newborn are different, in order to prevent the influence of data skew on the training effect of the model, in this embodiment, the training data is subjected to data equalization by an upsampling method to obtain the training data with the same ratio of each sleep stage in the total signal time duration. During training, the input of the full-automatic neonate sleep staging network is an electroencephalogram signal of each tested person after compression and transformation, and the output is stage training data corresponding to the electroencephalogram signal; after training is finished, obtaining a fully-automatic neonate sleep staging network after training is finished;

the trained sleep stage network obtained in the step is input into the electroencephalogram signals of the unmarked sleep process of the neonate, and the automatic sleep stage results corresponding to the input signals obtained by the network prediction are output.

(5) Performing sleep staging on the neonatal electroencephalogram to be staged by using the full-automatic neonatal sleep staging network obtained after the training in the step (4); the method comprises the following specific steps:

(5-2) in the embodiment, the full-automatic neonate sleep staging network obtained in the step (4) can accept the marked electroencephalogram signal after compression and transformation as an input, and an output network pre-measures a sleep staging result corresponding to the electroencephalogram signal;

and (4) inputting the electroencephalogram signals to be subjected to compression change in a staging way into the sleep staging network trained in the step (4), and outputting the sleep staging result of the signals to be staged by the network.

Fig. 4 shows a confusion matrix obtained by a full-automatic neonate sleep staging network by using a compressed electroencephalogram signal, and it can be seen that the method of the present invention can effectively realize the automatic sleep staging work of the neonate in the sleep process.

Compared with the existing sleep staging method, the method can improve the convenience and readability of the electroencephalogram signals and reduce the specificity among the electroencephalogram signals of different testees on the premise of effectively realizing sleep staging. In addition, the invention adopts a deep learning model method, and can greatly reduce the time and workload of manual sleep staging.

Claims

1. A neonate sleep staging method based on compressed electroencephalogram is characterized by comprising the following specific steps:

(1) acquiring N original electroencephalograms of the neonate in a sleep state, and labeling the sleep stage of the electroencephalograms of each testee; taking a stage labeling result of the original electroencephalogram signal as a standard stage result;

(2) filtering the original electroencephalogram signals obtained in the step (1), filtering background noise, and then performing compression transformation on the original electroencephalogram signals to obtain electroencephalogram signals after the compression transformation;

(3) aligning the compressed and transformed electroencephalogram signal obtained in the step (2) with the stage mark of the original electroencephalogram in the step (1) to obtain marked sleep stage training data;

(4-1) the specific process of the pre-training stage is as follows:

(4-1-1) calculating the number of sleep stages in the sleep data;

(4-1-4) inputting the sleep data with balanced data into a coding module for data feature extraction;

(4-2) the specific process of the sequence training phase is as follows:

during training, the input of the full-automatic neonate sleep staging network is an electroencephalogram signal of each tested person after compression and transformation, and loss functions with different weights are set according to the ratio of different sleep stages in the total sleep time and the recognition difficulty of different sleep stages, so that a more accurate sleep staging result is realized; outputting the stage training data corresponding to the electroencephalogram signal; after training is finished, obtaining a fully-automatic neonate sleep staging network after training is finished;

(5) performing sleep staging on the electroencephalogram signals to be staged by using the full-automatic neonate sleep staging network obtained by training in the step (4); the method comprises the following specific steps:

2. The method of staging of neonatal sleep according to claim 1, wherein in step (1) the sleep stage is divided into: awake (W), Active Sleep (AS) and Quiet Sleep (QS).

3. The method for staging sleep in a newborn infant as claimed in claim 2, wherein the specific process of step (2) is:

(2-1) filtering background noise of the electroencephalogram signals marked in the sleep stage by adopting a first filter, and attenuating signal activities of less than 1.6Hz and more than 70 Hz; then, removing artifacts caused by sweating, movement, electrocardio-activity and electrode interference of a tested person by using a second filter, and simultaneously enhancing signals by using a slope of 12dB/decade in a frequency band of 2-15 Hz;

(2-3) carrying out envelope detection on the rectified signal by adopting a third filter to obtain the variation trend of the amplitude of the electroencephalogram signal;

(2-6) finally, displaying the compression result in a semilogarithmic mode; voltage amplitudes of 0-10 μ V are shown in a linear coordinate system and voltage amplitudes of 10-100 μ V are shown in a logarithmic coordinate system to prevent activity changes of small signals from being annihilated by large signals.

4. The method of claim 3, wherein the first filter is filtered using a Butterworth band-pass filter; the second filter selects a finite-length single-bit impulse response filter; the third filter is a squared low-pass filter.

5. The method according to claim 1, wherein the step (4) of upsampling is to interpolate the original signal sequence by inserting new sleep stages between certain sleep stages based on the original signal samples to supplement a smaller number of sleep stages.

6. The method according to claim 1, wherein the down-sampling in step (4) is to perform decimation on the original signal sequence, that is, to perform random decimation on a sleep stage based on the original signal sampling, so as to reduce a larger number of sleep stages.