A method for neonatal sleep staging based on compressed EEG
By filtering and compressing the EEG signal of neonatal and combining with deep learning models, a fully automatic sleep staging network is built, which solves the problems of data redundancy and artificial dependence in sleep staging in neonatal, and realizes efficient automatic staging and accurate diagnosis .
Patent Information
- Application Number
- CN202210477229.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-03
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2042-05-03
AI Technical Summary
The prior art has data redundancy and artificial dependence in neonatal sleep staging, which leads to inconvenient data storage, transmission and calculation and difficulty in obtaining common features, limiting the popularity of sleep abnormal diagnosis.
The method based on compressed EEG signals is adopted to extract the overall characteristics of EEG signals through filtering and compression transformation, and combined with deep learning models, a fully automatic neonatal sleep staging network is built to realize automatic staging.
It improves the convenience and readability of EEG signals, reduces the data specificity between different subjects, reduces the time and workload of manual staging, and improves the accuracy and popularity of staging.
Smart Images
Figure CN115067875B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of biomedical technology, and in particular relates to a sleep staging method for newborns. Technical Background
[0002] In clinical practice, electroencephalogram (EEG) signals are usually used to diagnose and stage the sleep process of newborns, and in most cases, sleep staging is manually performed by multiple medical experts. The information contained in the original EEG signals is extremely rich, which also leads to a large amount of redundant information in the original EEG signals that is not necessary for sleep staging, causing inconvenience in data storage, transmission, and calculation. In addition, the "broad but not refined" characteristics of the original EEG signals also lead to strong signal specificity between different subjects, making it difficult to obtain the universal characteristics of sleep from the EEG data of many subjects.
[0003] The interpretation and use of raw EEG signals also require experienced neurophysiologists or pediatric neurologists with expertise in neonatal EEG. The difficulty in understanding EEG signals also limits the popularization of sleep abnormality diagnosis knowledge among medical staff.
[0004] In recent years, deep learning technology has begun to attract much attention in medical diagnosis, especially the application of convolutional neural networks and long short-term memory units in sleep staging, which has greatly improved the accuracy of sleep staging in adults. However, sleep staging is still carried out clinically using manual methods that are cumbersome and time-consuming. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and to fill the gap in the research field of neonatal sleep staging, and to propose a neonatal sleep staging method based on compressed EEG.
[0006] Compared with the existing sleep staging methods, the present invention can greatly improve the convenience of EEG signal data storage, transmission and calculation by compressing the original EEG signal, and the compressed data can well resist the data specificity between different subjects. On the other hand, the present invention uses a deep learning model to realize the automatic staging of the sleep process of newborns, which has a high application value.
[0007] The sleep staging method for newborns proposed by the present invention has the following specific steps.
[0008] (1) Obtain the original EEG signals of N newborns in sleep state, and annotate the sleep stages of each subject's EEG signals; use the stage annotation results of the original EEG signals as the standard stage results.
[0009] (2) Filter the original EEG signal obtained in step (1) to remove background noise, and then compress and transform the original EEG signal to obtain a compressed EEG signal. The specific steps are as follows:
[0010] (2-1) For the EEG signals annotated with sleep stages, the first filter, i.e., Butterworth bandpass filter, is used to filter out background noise and attenuate signal activities less than 1.6 Hz and greater than 70 Hz. Then, the second filter, i.e., finite impulse response (FIR) filter, is used to remove artifacts caused by sweating, movement, cardiac activity, and electrode interference of the subjects, and the signal is enhanced at a slope of 12 dB / decade in the 2-15 Hz frequency band.
[0011] (2-2) Full-wave rectification is performed on the filtered EEG signal, that is, the negative voltage is converted into a positive voltage;
[0012] (2-3) Using a third filter, i.e., a square low-pass filter, to perform envelope detection on the rectified signal to obtain the changing trend of the EEG signal amplitude;
[0013] (2-4) dividing the envelope obtained in step (2-3) into segments according to the compression ratio, where the compression ratio is the ratio of the original EEG signal sampling rate to the compressed pre-acquired signal sampling rate;
[0014] (2-5) compress each segment of the EEG signal, retaining the maximum and minimum values in each segment of the envelope as the upper and lower endpoints of the compressed signal;
[0015] (2-6) Finally, the compression results are displayed in a semi-logarithmic manner; the voltage amplitude of 0-10μV is displayed in a linear coordinate system, and the voltage amplitude of 10-100μV is displayed in a logarithmic coordinate system; in order to prevent the activity changes of small signals from being annihilated by large signals.
[0016] (3) Align the compressed transformed EEG signal obtained in step (2) with the stage annotation of the original EEG in step (1) to obtain labeled sleep stage training data.
[0017] (4) Use sleep staging training data to train a fully automatic neonatal sleep staging network model;
[0018] The fully automatic neonatal sleep staging network model is composed of a convolutional neural network and a bidirectional recurrent long short-term memory unit network model; the training of the model includes a pre-training stage and a sequence training stage;
[0019] (4-1) The specific process of the pre-training stage is as follows:
[0020] (4-1-1) Calculate the number of sleep stages in the sleep data;
[0021] (4-1-2) Compare the numbers of each sleep stage. If the ratio between the sleep stages is within the range of 0.8-1.2, proceed to step (4-1-4), otherwise proceed to step (4-1-3);
[0022] (4-1-3) According to the amount of data of each sleep stage, up-sample the sleep stage with a smaller proportion in the data or down-sample the sleep stage with a larger proportion in the data to balance the data;
[0023] (4-1-4) Inputting the data-balanced sleep data into a coding module mainly composed of a convolutional neural network to extract data features;
[0024] (4-2) The specific process of the sequence training phase is as follows:
[0025] (4-2-1) Input the data features extracted in the pre-training phase into the bidirectional recurrent long short-term memory unit network;
[0026] (4-2-2) The characteristics of each sleep stage are decoded through a bidirectional recurrent long short-term memory unit network. At the same time, the network learns the dependencies between multiple sleep stages to obtain the sequence characteristics of sleep data;
[0027] (4-2-3) combining the sleep stage data features obtained in step (4-1) with the sequence features obtained in step (4-2-2) through skip connections, integrating the information, and predicting the sleep stage;
[0028] During training, the input of the fully automatic neonatal sleep staging network is the compressed EEG signal of each subject, and loss functions with different weights are set according to the proportion of different sleep stages in the total sleep time and the difficulty of identifying different sleep stages to achieve more accurate sleep staging results; the output is the staging training data corresponding to the EEG signal; after training, a fully automatic neonatal sleep staging network is obtained.
[0029] (5) Using the fully automatic neonatal sleep staging network trained in step (4) to perform sleep staging on the EEG signals to be staged; the specific steps are as follows:
[0030] (5-1) Obtaining an EEG signal of a newborn during sleep, and performing compression changes on the EEG signal in step (2), wherein the EEG signal to be staged has similar characteristics to the EEG signal obtained in step (1);
[0031] (5-2) Using the trained neonatal sleep staging network to stage the sleep staging signal;
[0032] The EEG signal to be classified is input into the sleep staging network trained in step (4), and the network outputs the sleep staging result of the signal to be classified.
[0033] In the present invention, the sleep stages are: wakefulness (W), active sleep (AS) and quiet sleep (QS).
[0034] In the present invention, the first filter uses a Butterworth bandpass filter for filtering, the second filter uses a finite impulse response (FIR) filter, and the third filter uses a square low-pass filter for envelope detection.
[0035] In the present invention, the up-sampling is to interpolate the original signal sequence, that is, inserting a new sleep stage between certain sleep stages based on the original signal sampling to supplement the sleep stages with a smaller number.
[0036] In the present invention, the down sampling is to extract the original signal sequence, that is, to randomly extract a certain sleep stage based on the original signal sampling to reduce the number of sleep stages.
[0037] The characteristics and beneficial effects of the present invention are:
[0038] The present invention can greatly improve the convenience of EEG signal data storage, transmission and calculation by compressing the original EEG signal, and the compressed data can well resist the data specificity between different subjects.
[0039] The present invention compresses and transforms the EEG signal, extracts the overall characteristics of the original EEG signal changes, makes the signal characteristics easier to read and understand, and is conducive to the wider popularization of EEG signal interpretation knowledge. The present invention improves the staging accuracy by performing data balancing and weight setting on newborn data.
[0040] The present invention proposes a method of combining compressed EEG signals with a deep learning model. The present invention solves the problems of training data skew and different levels of feature learning difficulty by upsampling and weight configuration. After the training of the neonatal sleep staging network is completed, users can use the compressed EEG signals to automatically stage the neonatal sleep process, which has high practical value. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is an overall flow chart of the sleep staging method for neonates of the present invention.
[0042] Figure 2 and Figure 3 This is a comparison diagram of the original 500 Hz EEG signal compressed to 1 Hz by the compression method of the present invention.
[0043] Figure 4 It is the confusion matrix obtained by using compressed EEG signals through the fully automatic neonatal sleep staging network. DETAILED DESCRIPTION
[0044] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0045] The sleep staging method for newborns proposed by the present invention has the following overall process: Figure 1 As shown, the following steps are included.
[0046] (1) Obtain the original EEG signals of at least 20 newborns during their sleep, with the signal duration of each newborn being at least about 2 hours. Experts will annotate the sleep stages of each newborn's EEG signal to obtain the labeled sleep stage training data corresponding to the signal;
[0047] In this embodiment, the original EEG signals collected by the medical EEG acquisition device do not contain the classification labels of sleep stages. In order to enable the computer to learn the classification characteristics of these sleep stages, medical sleep experts need to manually mark the corresponding sleep stages in each signal segment. Specifically, the experts need to mark the sleep stages contained in each signal segment according to the rules issued by the American Academy of Sleep Medicine (AASM). These marked data can be used to train the deep learning model.
[0048] The American Academy of Sleep Medicine divides the sleep process of adults into five stages: wakefulness (W), rapid eye movement (REM), non-rapid eye movement stage 1 (N1), non-rapid eye movement stage 2 (N2) and non-rapid eye movement stage 3 (N3). The corresponding stages for newborns are wakefulness (W), active sleep (AS) and quiet sleep (QS).
[0049] (2) filtering the original EEG signal obtained in (1), and compressing the original EEG signal after filtering out the background noise to obtain a compressed EEG signal;
[0050] The original EEG signal is filtered through a Butterworth bandpass filter to remove the large background noise and obtain a relatively pure EEG signal; the pre-processed EEG signal is compressed and transformed. The specific steps are as follows:
[0051] (2-1) The preprocessed EEG signal is passed through an asymmetric bandpass filter to attenuate signal activity below 2 Hz and above 15 Hz. The filter is used to remove artifacts caused by sweating, movement, ECG activity, and electrode interference, and to enhance the signal within the 2-15 Hz frequency band with a slope of 12 dB / decade;
[0052] In this embodiment, the asymmetric bandpass filter is designed using the Parks-McClellan algorithm;
[0053] (2-2) Full-wave rectification is performed on the filtered EEG signal, that is, the negative voltage is converted into a positive voltage;
[0054] (2-3) Perform envelope detection on the rectified signal and use a square low-pass filter to obtain the change trend of the EEG signal amplitude;
[0055] (2-4) Dividing the envelope obtained in (2-3) into segments according to the compression ratio, where the compression ratio is the ratio of the original EEG signal sampling rate to the compressed pre-acquired signal sampling rate;
[0056] (2-5) compress each segment of the EEG signal, retaining the maximum and minimum values in each segment of the envelope as the upper and lower endpoints of the compressed signal;
[0057] (2-6) Finally, the compression results are displayed in a semi-logarithmic manner; the voltage amplitude of 0-10μV is displayed in a linear coordinate system, and 10-100μV is displayed in a logarithmic coordinate system. This prevents the activity changes of small signals from being annihilated by large signals;
[0058] Figure 2 and Figure 3 The comparison chart of the EEG signal compressed from the original 500Hz to 1Hz is shown. It can be seen that the compression result retains the characteristics of the overall signal activity of the original EEG well, removes other redundant features, and greatly reduces the data volume.
[0059] (3) In this embodiment, the sleep stage annotation of the original EEG signal has been obtained in step (1). At this time, it is only necessary to align the compressed transformed EEG signal obtained in step 2) with the stage annotation of the original EEG in step (1) to obtain the labeled sleep stage training data.
[0060] (4) Use sleep staging training data to train a fully automatic neonatal sleep staging network;
[0061] In this embodiment, the fully automatic neonatal sleep staging network is composed of a pre-trained encoding module and a sequence decoding module, such as a convolutional neural network (CNN), a gated recurrent unit (GRU), and a long short-term memory unit (LSTM). This embodiment uses a convolutional neural network and a long short-term memory unit network model;
[0062] In general, since the proportion of different sleep stages in the total sleep duration of a newborn is different, in order to prevent the influence of data skew on the model training effect, in this embodiment, the training data is balanced by upsampling to obtain training data in which each sleep stage has the same proportion in the total signal duration. During training, the input of the fully automatic newborn sleep staging network is the EEG signal of each subject after compression and transformation, and the output is the stage training data corresponding to the EEG signal; after training, a fully automatic newborn sleep staging network is obtained;
[0063] The trained sleep staging network obtained in this step takes as input the unlabeled EEG signal of the neonatal sleep process, and outputs the automatic sleep staging result corresponding to the input signal predicted by the network.
[0064] (5) Using the fully automatic neonatal sleep staging network trained in step (4) to perform sleep staging on the EEG signals of the neonates to be staged; the specific steps are as follows:
[0065] (5-1) Obtaining an EEG signal of a newborn during sleep, and performing compression changes on the EEG signal in step (2), wherein the EEG signal to be staged has similar characteristics to the EEG signal obtained in step (1);
[0066] (5-2) In this embodiment, the fully automatic neonatal sleep staging network obtained in step (4) can accept a labeled EEG signal after compression transformation as input, and the output network predicts the sleep staging result corresponding to the EEG signal;
[0067] The compressed and transformed EEG signal to be classified is input into the sleep classification network trained in step (4), and the network outputs the sleep classification result of the signal to be classified.
[0068] Figure 4 The confusion matrix obtained by using compressed EEG signals through a fully automatic neonatal sleep staging network is shown in FIG. 1 . It can be seen that the method of the present invention can effectively realize the automatic sleep staging of the neonatal sleep process.
[0069] Compared with the existing sleep staging methods, the method of the present invention can improve the convenience and readability of EEG signals and reduce the specificity between EEG signals of different subjects while effectively realizing sleep staging. In addition, the present invention adopts a deep learning model method, which can greatly reduce the time and workload of manual sleep staging.
Claims
1. A method for neonatal sleep staging based on compressed EEG, characterized in that: The specific steps are as follows: (1) Obtain the original EEG signals of N newborns in the sleeping state, and annotate the sleep stages of each subject's EEG signals; use the stage annotation results of the original EEG signals as the standard stage results; (2) filtering the original EEG signal obtained in step (1) to remove background noise, and then compressing the original EEG signal to obtain a compressed EEG signal; (3) Align the compressed transformed EEG signal obtained in step (2) with the stage annotation of the original EEG in step (1) to obtain labeled sleep stage training data; (4) Use sleep staging training data to train a fully automatic neonatal sleep staging network model; The fully automatic neonatal sleep staging network model is composed of a convolutional neural network and a bidirectional recurrent long short-term memory unit network model; the training of the model includes 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) Calculate the number of sleep stages in the sleep data; (4-1-2) Compare the numbers of each sleep stage. If the ratio between the sleep stages is within the range of 0.8-1.2, proceed to step (4-1-4), otherwise proceed to step (4-1-3); (4-1-3) According to the amount of data of each sleep stage, up-sample the sleep stage with a smaller proportion in the data or down-sample the sleep stage with a larger proportion in the data to balance the data; (4-1-4) Inputting the data-balanced sleep data into an encoding module to extract data features; (4-2) The specific process of the sequence training phase is as follows: (4-2-1) Input the data features extracted in the pre-training phase into the bidirectional recurrent long short-term memory unit network; (4-2-2) The characteristics of each sleep stage are decoded through a bidirectional recurrent long short-term memory unit network. At the same time, the network learns the dependencies between multiple sleep stages to obtain the sequence characteristics of sleep data; (4-2-3) combining the sleep stage data features obtained in step (4-1) with the sequence features obtained in step (4-2-2) through skip connections, integrating the information, and predicting the sleep stage; During training, the input of the fully automatic neonatal sleep staging network is the compressed EEG signal of each subject, and the loss function with different weights is set according to the proportion of different sleep stages in the total sleep time and the difficulty of identifying different sleep stages, so as to achieve more accurate sleep staging results; The output is the period training data corresponding to the EEG signal; After the training is completed, a fully-automatic neonatal sleep staging network is obtained; (5) Use the fully automatic neonatal sleep staging network trained in step (4) to perform sleep staging on the EEG signals to be staged; the specific steps are as follows: (5-1) Obtaining an EEG signal of a newborn during sleep, and performing compression changes on the EEG signal in step (2), wherein the EEG signal to be staged has similar characteristics to the EEG signal obtained in step (1); (5-2) Using the trained neonatal sleep staging network to stage the sleep staging signal; The EEG signal to be classified is input into the sleep staging network trained in step (4), and the network outputs the sleep staging result of the signal to be classified.
2. The method for neonatal sleep staging according to claim 1, characterized in that: The sleep stages described in step (1) are divided into wakefulness (W), active sleep (AS) and quiet sleep (QS).
3. The method for neonatal sleep staging according to claim 2, characterized in that: The specific process of step (2) is as follows: (2-1) For the EEG signals annotated with sleep stages, the first filter is used to filter out background noise and attenuate signal activities less than 1.6 Hz and greater than 70 Hz. Then, the second filter is used to remove artifacts caused by sweating, movement, ECG activity and electrode interference of the subjects, and the signal is enhanced at a slope of 12 dB / decade in the 2-15 Hz frequency band. (2-2) Full-wave rectification is performed on the filtered EEG signal, that is, the negative voltage is converted into a positive voltage; (2-3) Using a third filter to perform envelope detection on the rectified signal to obtain the changing trend of the EEG signal amplitude; (2-4) dividing the envelope obtained in step (2-3) into segments according to the compression ratio, where the compression ratio is the ratio of the original EEG signal sampling rate to the compressed pre-acquired signal sampling rate; (2-5) compress each segment of the EEG signal, retaining the maximum and minimum values in each segment of the envelope as the upper and lower endpoints of the compressed signal; (2-6) Finally, the compression results are displayed in a semi-logarithmic manner; the voltage amplitude of 0-10μV is displayed in a linear coordinate system, and the voltage amplitude of 10-100μV is displayed in a logarithmic coordinate system to prevent the activity changes of small signals from being annihilated by large signals.
4. The method for neonatal sleep staging according to claim 3, characterized in that: The first filter uses a Butterworth bandpass filter for filtering; the second filter uses a finite length unit impulse response filter; and the third filter uses a square low-pass filter.
5. The method for neonatal sleep staging according to claim 1, characterized in that: The upsampling in step (4) is to interpolate the original signal sequence, that is, inserting a new sleep stage between certain sleep stages based on the original signal sampling to supplement the smaller number of sleep stages.
6. The method for neonatal sleep staging according to claim 1, characterized in that: The downsampling in step (4) is to extract the original signal sequence, that is, to randomly extract a certain sleep stage based on the original signal sampling to reduce the number of sleep stages.