阅读说明：本技术 一种自动睡眠分期的建立方法及其应用 (Method for establishing automatic sleep stages and application thereof ) 是由 黄锷 于 2020-12-24 设计创作，主要内容包括：本发明提供一种自动睡眠分期的建立方法及其应用。该自动睡眠分期的建立方法包括：获取若干组PSG信号以及PSG信号的人工睡眠标记信息；预分析,将PSG信号中的原始时间序列分解为一组类本征模态函数；将所述类本征模态函数进行组合,得到m组时间序列集合；多尺度熵分析,使用n个采样尺度对m组时间序列集合进行熵值计算,得到具有m×n个元素的熵矩阵；建立所述意识水平与熵矩阵的元素之间的相关系数矩阵,找出相关系数矩阵中最大正相关元素或者最大负相关元素相对应的采样尺度和滤波尺度；采样尺度为粗粒度尺度；根据最大正相关元素或最大负相关元素的采样尺度和滤波尺度,计算待测者在该采样尺度和滤波尺度的熵值,根据该熵值判断患者的睡眠状态。(The invention provides a method for establishing an automatic sleep stage and application thereof. The method for establishing the automatic sleep stages comprises the following steps: acquiring a plurality of groups of PSG signals and artificial sleep marking information of the PSG signals; pre-analyzing, namely decomposing an original time sequence in the PSG signal into a group of eigenmode functions; combining the similar intrinsic mode functions to obtain m groups of time sequence sets; performing multi-scale entropy analysis, namely performing entropy calculation on the m groups of time sequence sets by using n sampling scales to obtain an entropy matrix with m multiplied by n elements; establishing a correlation coefficient matrix between the consciousness level and elements of the entropy matrix, and finding out a sampling scale and a filtering scale corresponding to the maximum positive correlation element or the maximum negative correlation element in the correlation coefficient matrix; the sampling scale is coarse grain size; and according to the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element, calculating the entropy values of the person to be measured in the sampling scale and the filtering scale, and judging the sleep state of the patient according to the entropy values.)

1. A method for establishing an automatic sleep stage is characterized by comprising the following steps:

acquiring a plurality of groups of PSG signals and artificial sleep mark information of the PSG signals;

pre-analysis for decomposing the original time series of each stage in the PSG signal into a set of eigenmode functions or eigenmode-like functions; combining the intrinsic mode functions or similar intrinsic mode functions to obtain m groups of time sequence sets;

performing multi-scale entropy analysis, namely performing entropy calculation on the m groups of time sequence sets by using n sampling scales to obtain an entropy matrix with m multiplied by n elements;

the consciousness level is determined according to the artificial sleep marking information;

establishing a correlation coefficient matrix between the consciousness level and the elements of the entropy matrix, and finding out a sampling scale and a filtering scale corresponding to the maximum positive correlation element or the maximum negative correlation element in the correlation coefficient matrix; the sampling scale is the coarse-grained scale;

and according to the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element, calculating the entropy value of the person to be measured in the sampling scale and the filtering scale, and judging the sleep state of the patient according to the entropy value.

2. The method for establishing an automatic sleep stage according to claim 1, wherein a modal decomposition method is used when decomposing the original time series of each stage in the PSG signal into a set of intrinsic mode functions, the modal decomposition method being one of the following methods: the empirical mode decomposition method is a mode decomposition method and an adaptive binary mask empirical mode decomposition method.

3. The method of establishing an automatic sleep stage as claimed in claim 1, wherein a set of high pass filters with cut-off frequencies of 32Hz, 16Hz, 8Hz, 4Hz, 2Hz and 1Hz are used in decomposing the original time series of each stage in the PSG signal into a set of eigenmode-like functions.

4. The method of establishing an automatic sleep stage as claimed in claim 1, wherein said PSG signal comprises at least one of the following brain signals: fp4-A1, F4-A1, C4-A1, P4-A1 and O2-A1.

5. The method for establishing an automatic sleep stage according to claim 1, wherein the consciousness level for reflecting the degree of wakefulness in sleep is determined according to the artificial sleep marker information, wherein the wakefulness stage is quantized to 6, the rapid eye movement stage is quantized to 5, the NREM1 stage is quantized to 4, the NREM2 stage is quantized to 3, the NREM3 stage is quantized to 2, and the NREM4 stage is quantized to 1.

6. The method of establishing an automatic sleep stage as claimed in claim 1, wherein the matrix of correlation coefficients between the level of consciousness and the elements of the entropy matrix is established based on Pearson coefficients.

7. The method for establishing an automatic sleep stage according to claim 1, wherein the threshold between different sleep states is calculated by an artificial intelligence method when the sleep state of the patient is judged according to the entropy of the maximum positive correlation element or the maximum negative correlation element of the person under test in the sampling scale and the filtering scale.

8. Use of the method of claim 1, characterized in that it comprises the following steps:

acquiring a PSG signal of a person to be tested;

decomposing a PSG signal of a to-be-tested person into an original time sequence of a plurality of stages;

obtaining an original time sequence of a stage, and decomposing the original time sequence into a group of intrinsic mode functions or quasi-intrinsic mode functions;

calculating the entropy of the testee at the sampling scale and the filtering scale according to the sampling scale and the filtering scale corresponding to the maximum positive correlation element or the maximum negative correlation element in claim 1;

and judging the sleep state of the person to be tested at the stage according to the entropy value.

9. The automatic sleep staging method of claim 8, wherein: the PSG signal at least comprises one of the following brain signals: fp4-A1, F4-A1, C4-A1, P4-A1 and O2-A1.

10. The automatic sleep staging method of claim 8 wherein the PSG signal of the test subject is decomposed into an original time series of stages, each stage having a time of 30 seconds.

Technical Field

The invention relates to the technical field of data processing, in particular to an automatic sleep staging establishing method and application thereof.

Background

One third of our lifetime is spent in sleep, which is a vital part of our lives. Sleep quality not only affects our daily health, but also affects our economic productivity. Although the function of sleep is not fully understood, recent studies have found that, in addition to the traditionally thought of memory enhancement, glial cell contraction and increased cerebrospinal fluid flushing play a critical role in waste clearance. Studies have also shown that sleep is not a uniform state, and that switching from one state to another is a highly non-linear and unstable process. Sleep disorders can have serious consequences, including reduced quality of life, complications, early death, etc., and have a great impact on economic and social costs. These factors make studies on sleep critical. Traditionally, normal sleep is divided into five stages according to the R & k (rechtschaffen and kales) standard: rapid Eye Movement (REM) and four other non-rapid eye movement (NREM) stages (1-4). Sleep states are estimated from features of polysomnography recordings, including electroencephalogram, electromyography, and electrooculogram, among other parameters, and in clinical practice, sleep technicians visually determine sleep stages for each 30 s. However, manual scoring is a time consuming process, and different people may give inconsistent results due to similarities between different stages. Therefore, assessing sleep quality through polysomnography is a labor intensive, time consuming and error prone process. In addition, as the number of patients monitored for sleep increases, it becomes increasingly difficult for limited sleep analysts to meet the increasing demands for sleep map analysis.

Therefore, there is a need for an accurate and objective method for automatically staging sleep stages and assessing sleep quality.

Disclosure of Invention

In order to solve the above problems, in the present invention, we propose intrinsic Multiscale Entropy imae (imase) as a new signal analysis method. First, sleep states were studied using Multiscale Entropy MSE (MSE), which aims to quantify complexity using Entropy summation over multiple time scales. In multi-scale entropy, entropy is defined as a physical measure of "unpredictability" of a coarsely-grained non-linear time series over multiple time scales, which can be defined using sample entropy or approximate entropy when calculating entropy. The number of non-overlapping samples that are combined into one sample is defined as the time scale of coarse graining in MSE. In digital signal processing, time scale 1 represents the raw time of the measurement and is digitized at the raw sampling rate. The time scale n represents the coarsely grained time series with a sampling interval of n times the original data and a sampling rate of 1/n of the original sampling rate. Thus, the time scale of coarse grain represents the length of time in terms of the number of original sampling intervals. The method of MSE reflects the idea that entropy is a measure that depends on the time scale of the sampling interval. Second, since the sleep process is not a stable process, we propose intrinsic multi-scale entropy (imese), i.e. the limitation of non-stationary states is solved by combining Empirical Mode Decomposition (EMD) with multi-scale entropy. Empirical mode decomposition algorithms can perform both denoising and detrending preprocessing steps to extract the required information from the non-linear and non-stationary real signal. EMD can be used as an adaptive binary filter bank to decompose a complex time series into a set of eigenmode functions (IMFs). Each eigenmode function is characterized by a small bandwidth and zero mean, and thus, the eigenmode function is steady-state. Different combinations of eigenmode functions may be used to reconstruct the filtered time series, which may filter out high frequency noise components or low frequency trends from the original time series. As a functional combination, we propose the imase method as a new signal analysis method.

Furthermore, to avoid the high computational cost and other detail problems of EMD, we introduce a simple filter-based pseudo-EMD method that mimics the function of EMD, avoiding the problem of modal aliasing, and thus systematically extracts the filtered components from the time series. The iMSE quantizes the entropy of the quantized components over a number of coarse-grained time scales (i.e., sample scales). Meanwhile, the filter band represents a second filter scale in the imese. The entropy values are then displayed in a two-dimensional matrix on both axes of the sampling scale and the filtering scale, greatly enhancing the function of the original MSE.

In the invention, an establishing method of automatic sleep stages is provided, and the optimal sampling scale and filtering scale suitable for the automatic sleep stages are found out by analyzing entropy matrixes on two axes of the sampling scale and the filtering scale. When the sleep state of the patient is analyzed, the method only needs to calculate the entropy values of the optimal sampling scale and the optimal filtering scale, namely, the automatic sleep staging can be carried out through the entropy values. The method greatly reduces the calculation amount of sleep staging by using multi-scale entropy, thereby improving the speed of automatic sleep staging.

In order to achieve the above object, the present invention provides a method for establishing an automatic sleep stage, comprising the following steps: acquiring a plurality of groups of PSG signals and artificial sleep marking information of the PSG signals; pre-analysis for decomposing the original time series of each stage in the PSG signal into a set of eigenmode functions or eigenmode-like functions; combining the intrinsic mode functions or similar intrinsic mode functions to obtain m groups of time sequence sets; performing multi-scale entropy analysis, namely performing entropy calculation on the m groups of time sequence sets by using n sampling scales to obtain an entropy matrix with m multiplied by n elements; establishing a correlation coefficient matrix between the consciousness level and the entropy matrix elements, and finding out the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element in the correlation coefficient matrix; the sampling scale is the coarse-grained scale; the filtering scale is a time sequence set; and according to the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element, calculating the entropy value of the person to be measured in the sampling scale and the filtering scale, and judging the sleep state of the patient according to the entropy value.

Preferably, when the original time series of each stage in the PSG signal is decomposed into a set of eigenmode functions, a mode decomposition method is used, which is one of the following methods: the empirical mode decomposition method is a mode decomposition method and an adaptive binary mask empirical mode decomposition method.

Preferably, when decomposing the original time series of each stage in the PSG signal into a set of eigenmode-like functions, a set of high-pass filters is used, the cut-off frequencies of which are 32Hz, 16Hz, 8Hz, 4Hz, 2Hz and 1Hz, respectively.

Preferably, the PSG signal comprises at least one of the following brain electrical signals: fp4-A1, F4-A1, C4-A1, P4-A1 and O2-A1.

Preferably, the consciousness level for reflecting the degree of wakefulness in sleep is determined according to the artificial sleep marker information, wherein the wakefulness stage is quantified as 6, the rapid eye movement stage is quantified as 5, the NREM1 stage is quantified as 4, the NREM2 stage is quantified as 3, the NREM3 stage is quantified as 2, and the NREM4 stage is quantified as 1.

Preferably, the matrix of correlation coefficients between the level of consciousness and the elements of the entropy matrix is established based on Pearson coefficients.

Preferably, when the sleep state of the patient is judged according to the entropy of the maximum positive correlation element or the maximum negative correlation element of the testee and the sampling scale and the filtering scale, the threshold value between different sleep states is calculated by adopting an artificial intelligence method.

The invention also provides an automatic sleep staging method, which is characterized by comprising the following steps: acquiring a PSG signal of a person to be tested; decomposing a PSG signal of a to-be-tested person into an original time sequence of a plurality of stages; obtaining an original time sequence of a stage, and decomposing the original time sequence into a group of intrinsic mode functions or quasi-intrinsic mode functions; according to the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element, obtaining an entropy value in the scale; and judging the sleep state of the person to be tested at the stage according to the entropy value.

Preferably, the PSG signal comprises at least one of the following brain electrical signals: fp4-A1, F4-A1, C4-A1, P4-A1 and O2-A1.

Preferably, the PSG signal of the test subject is decomposed into an original time series of several phases, each phase having a time of 30 seconds.

By the automatic sleep stage establishing method, an automatic sleep stage establishing method can be established, and the method only needs to measure the entropy values of the patient to be measured in the optimal sampling scale and the filtering scale, namely, the automatic sleep stage can be carried out through the entropy values. The method greatly reduces the calculation amount of sleep staging by using multi-scale entropy, thereby improving the speed of automatic sleep staging.

Drawings

FIG. 1 is a flow chart of a method for establishing an automatic sleep stage according to the present invention.

Fig. 2 is a five-channel EEG image of six different sleep stages.

FIG. 3 is a typical IMF-like set of 6 sleep states (channel C4-A1).

FIG. 4 is a flowchart of a method for decomposing and recombining original time series to obtain a time series set according to the present invention.

FIG. 5 is an entropy calculation of the sample scale from 1 to n for each set of time series.

Fig. 6 is an exemplary diagram of a two-dimensional entropy matrix for six different sleep stages.

Fig. 7 is a matrix of correlation coefficients between discrete consciousness levels and individual entropy matrix elements in the 5 entropy matrices of the 5 EEG recording channels.

FIG. 8a is a time series and trend of manually calibrating sleep stages, PEDCL and NEDCL in the F4-A1 channel; fig. 8b is a time series and trend of manually demarcating sleep stages, PEDCL and NEDCL in the C4-a1 channel.

Figure 9a is a comparison of internal subject values of PEDCL between six sleep states for five channels; figure 9b is a comparison of the intra-subject comparisons between NEDCL values for six sleep states for five channels.

Figure 10a is a comparison between subjects between PEDCL values for six sleep states for five channels; figure 10b is a comparison between subjects' NEDCL values for six sleep states for five channels.

Fig. 11 is a flowchart of an automatic sleep staging method of the present invention.

Detailed Description

The technical means adopted by the invention to achieve the preset purpose are further described below by combining the accompanying drawings and the preferred embodiments of the invention.

Fig. 1 shows a detailed embodiment of the method for establishing an automatic sleep stage according to the present invention. In step 110, several sets of Polysomnographic (PSG) signals and artificial sleep marker information of the polysomnography are acquired. Polysomnography instruments may be used to measure and record one or more physiological signals, such as F3-A2 brain wave signals, F4-A1 brain wave signals, C3-A2 brain wave signals, C4-A1 brain wave signals, P3-A2 brain wave signals, P4-A1 brain wave signals, left eye movement signals, right eye movement signals, and the like. Because sleep belongs to brain wave activity, the physiological signals closer to the brain can reflect the sleep state, and the brain wave signals are generally selected for sleep staging. A Cyclic Alternating Pattern (CAP) in sleep is a periodic change in brain waves that occurs during non-rapid eye movement sleep, a component that reflects the microstructure of sleep. Polysomnography records in a cyclically alternating pattern in sleep may be downloaded from the website of the Phsionet. The database provides a total of 108 PSG records containing 8 different pathological conditions. In the present invention, we only select three pathological conditions of normal, insomnia and lethargy for study, and we also require that the channels and sampling frequency of the brain waves (EEG) recorded in these polysomnography recordings must be consistent in order to reduce the variables of the inventive automatic sleep stage establishment method. Based on these criteria, we acquired PSG recordings of five EEG channels Fp4, F4, C4, P4 and O2 sampled at a sampling rate of 512 Hz. Six normal control subjects (n1, n2, n3, n5, n10 and n11), five subjects with insomnia (ins2, ins4, ins6, ins7 and ins8) and five subjects with narcolepsy (narco1, narco2, narco3, narco4 and narco5) were used in the study of the invention. For each PSG record, the event time, sleep stage and CAP annotation information is contained in the text file. Fig. 2 is a typical five-channel electroencephalogram of a sleep state including a waking stage, a non-rapid eye movement stage (including first to fourth stages, respectively, NREM1, NREM2, NREM3, and NREM4), and a Rapid Eye Movement (REM) stage. Six examples are shown in fig. 2, corresponding to the five-channel EEG images for the six different sleep stages described above. As can be seen in fig. 2, the EEG signals of NREM3 and NREM4, which represent deep sleep stages, contain low frequency oscillations. Therefore, NREM3 and NREM4 are also referred to as slow wave sleep.

Step 120, decomposing the original time series of several phases in the PSG signal into a set of eigenmode functions or quasi-eigenmode functions, respectively. Since sleep was divided into one stage by 30s in the PSG record, analysis of sleep state was performed. Therefore, when the automatic sleep stage method is established, 30s is also used as a stage for analysis, and the original time sequence of a plurality of stages in the PSG signal is respectively decomposed into a group of eigenmode functions or quasi-eigenmode functions. The essence of the pre-analysis is to decompose the original time series into a set of independent narrow bandwidths and detrended zero-mean eigenmode functions or eigenmode-like functions with dyadic bands. This step is critical because entropy is calculated from the probability density function of the data, but the probability density can only be calculated on data without trends. When the original time series is decomposed, a modal decomposition method can be adopted for decomposition. Since empirical mode decomposition is an ideal binary filter bank, a non-linear time series can be adaptively decomposed into a set of eigenmode functions. The modal Decomposition method refers to any modal Decomposition method capable of obtaining the intrinsic Mode function component, such as an Empirical Mode Decomposition (EMD), an Ensemble Empirical Mode Decomposition (EEMD), or an Adaptive binary mask Empirical Mode Decomposition (CADM-EMD) method.

In the present invention, we also provide an alternative approach to overcome the shortcomings of the modality decomposition approach for sleep staging. The modal decomposition method for sleep staging has the following disadvantages: (1) unless expensive calculations and elaborate masking methods are used in the enhancement algorithm, it is difficult to fully solve the modal aliasing problem and the resulting Intrinsic Mode Functions (IMFs) may have different IMF eigenband mismatches at different sleep stages. For example, the distribution of instantaneous frequencies of IMF1 from EEG signals recorded at sleep stages of NREM1 is different from the distribution of instantaneous frequencies of EEG signals recorded at sleep stages of NREM 4. (2) It is not easy to align the resulting frequency bands from the IMF with the generic frequency bands of the EEG signal. For example, the commonly defined delta band is 0.5-4Hz, theta band is 4-8Hz, alpha band is 8-16Hz, beta band is 16-30Hz, and gamma band is 30-60 Hz. These frequency bands are similar, but not identical. To facilitate comparison with the commonly defined frequency bands, we can model a predetermined (but not adaptive) filter bank to extract a set of IMF-like functions from the EEG signal in another way. (3) For those not familiar with the modal decomposition method, the use of the filter method is easier to implement. In an alternative approach provided by the present invention, high frequency noise is first removed from the time series using a low pass filter with a cutoff frequency of 64 Hz. Then, the first 6 classes of IMFs were extracted using a set of high-order high-pass filters with cut-off frequencies of 32Hz, 16Hz, 8Hz, 4Hz, 2Hz and 1Hz in order. Theoretically, the frequency bands of the first 6 classes of IMFs decomposed by the substitution method are 32-64Hz (similar to the γ frequency band), 16-32Hz (β frequency band), 8-16Hz (α frequency band), 4-8Hz (θ frequency band), 2-4Hz (δ frequency band), and 1-2Hz (low δ frequency band), and the correspondence is shown in table 1.

TABLE 1

The IMF-like function obtained by adopting the filter method can well solve the problem that the result frequency band from the IMF is not easy to align with the EEG universal frequency band. Fig. 3 is a typical IMF-like set of 6 sleep states obtained by this method, the EEG channel selected in the figure being the C4-a1 channel. The filtering frequency bands of the IMF1-6 are respectively a gamma frequency band (32-64Hz), a beta frequency band (16-32Hz), an alpha frequency band (8-16Hz), a theta frequency band (4-8Hz), a delta frequency band (2-4Hz) and a low delta frequency band (1-2 Hz).

And step 121, combining the intrinsic mode functions or quasi-intrinsic mode functions obtained by decomposition to obtain m groups of time sequence sets. The set of filtered time series can be recombined using various combinations of eigenmode functions or eigenmode-like functions into a new set of m detrended zero-mean time series that displays information on different aspects of the original data (e.g., high frequencies only) or any particular selected frequency band. As shown in fig. 4, the process from step 120 to step 121 is shown, a complex time series is decomposed by mode decomposition or its alternative method to obtain a set of intrinsic mode functions or eigenmode-like functions, and then the set of intrinsic mode functions or eigenmode-like functions are combined, and the recombined filtering time series will cover all possible frequency segments of the original data. Taking the above-mentioned 6 eigenmode-like functions as an example, from the first 6 eigenmode-like functions, 14 additional filtered time series can be reconstructed for further analysis. The 20 filtered time series include only IMF1, IMFs 1-2, IMFs1-3, IMFs 1-4, IMFs 1-5, IMFs 1-6; only IMF2, IMFs2-3, IMFs2-4, IMFs2-5, IMFs 2-6; only IMF 3, IMFs3-4, IMFs3-5, IMFs 3-6; IMF 4 only, IMFs4-5, IMFs 4-6; only IMF5 and IMF 5-6.

And step 130, performing multi-scale entropy analysis, and performing entropy calculation on the m groups of time sequence sets obtained in the step 121 by using n sampling scales to obtain an entropy matrix with m × n elements. As shown in fig. 5, entropy calculation of a sampling scale from 1 to n is performed for each set of time series, resulting in an entropy matrix having m × n elements. Multi-scale entropy analysis evaluates the complexity of a time series in terms of entropy values corresponding to a number of different time scales. In imese, a plurality of sampling scales may be used to estimate entropy values of filtered time series each containing one or more eigenmode functions or combinations of eigenmode-like functions to obtain row vectors of entropy. The sampling scale is defined as the number of consecutive samples from the original time series that are non-overlapping combined into one coarse-grained time series. The length of the sample time series is 1/n of the length of the original time series, where n is the sample size. Only the time series with the sampling scale 1 is the original time series. Any entropy definition method may be employed to compute the entropy values in the coarse-grained time series, such as approximate entropy, sample entropy, and the like. In the present invention, we select approximate entropy (ApEn) to compute the entropy vector of the filtered time series in the multi-scale entropy analysis. In order to reduce the influence of each example and different sleep stages on entropy and objectively display the entropy of each sleep state, we can select multiple sleep stages to study, find the entropy matrix of each sleep stage according to steps 120 to 130, and then find the average value of the entropy matrices of sleep stages with the same artificial mark to obtain the average entropy matrix. FIG. 6 shows the result of entropy calculation from 60 sample scales of 2-120 in steps of 2 for a time series of 20 different filter scales according to the present invention. On 60 different sampling scales, a two-dimensional average entropy matrix with 20 × 60 elements was calculated for the 20 filtered time series. Fig. 6 represents 6 typical average entropy matrices for 6 different sleep stages. In these sub-graphs, the entropy values are displayed in different colors, the X-axis marks the sampling scale from 2 to 120, the step size is 2, and the Y-axis marks the filtering scale in the eigenmode-like function, from 1 to 20. Each entropy matrix represents an entropy measure for the EEG signal of the same stage under multiple control conditions of filtering and sampling, with significant differences between different sleep stages.

Step 140, a correlation coefficient matrix between the consciousness level and the entropy matrix elements is established, and the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element in the correlation coefficient matrix are found out. Wherein the level of consciousness is based on an artificial sleep marker of the PSG signal. In order to establish a relationship between the level of consciousness and the entropy measure, a discrete level of consciousness (DCL) in sleep is first defined according to the manually classified sleep stages, the level of consciousness being used to reflect the degree of wakefulness in sleep. The waking phase representing the highest level of consciousness is quantified as 6, the rapid eye movement phase is quantified as 5, the NREM1 phase is quantified as 4, the NREM2 phase is quantified as 3, the NREM3 phase is quantified as 2, and the NREM4 phase representing the lowest level of consciousness is quantified as 1. On average, each subject had nearly 1000 stages in one night of sleep, which formed an amplitude time series with the above definition, a DCL series. Next, we will examine the correlation between the DCL sequence and the time series of each imese matrix element. We found that some elements are positively correlated with DCL in sleep and others are negatively correlated based on Pearson correlation coefficients. Fig. 7 shows a matrix of correlation coefficients between DCL and the respective elements in 5 entropy matrices for 5 EEG recording channels. The first five subgraphs represent the matrix of correlation coefficients derived from five different EEG channels, and the sixth subgraph shows the mean matrix of the first five subgraphs. Then, we select an element with the largest positive correlation and negative correlation with the DCL sequence from each correlation coefficient matrix. For example, in the matrix of the channel Fp2-a1 (sub-diagram 1), the filtering scale of the maximum positively correlated element is IMFs1-3(α - γ frequency band), the sampling scale is 2 (i.e., 2/512 seconds of sampling scale at 512Hz sampling rate), and the correlation coefficient is 0.64869; while the filter scale for the element of maximum negative correlation is IMF1 (gamma band), the sample scale is 102 (i.e., the sample scale of 102/512 seconds at a sampling rate of 512 Hz), and its correlation coefficient is-0.73802. The position is close to all six subjects and all five electrodes, the sampling scale and the filtering scale can be selected as the optimal sampling scale and the filtering scale of the patient to be tested, and then sleep staging is automatically carried out on the patient to be tested according to entropy values under the sampling scale and the filtering scale. This conclusion is correct, as shown by the average matrix in the above figure. Therefore, we decided to select these two entropy elements based on the overall result, as shown in figure 6, to represent all subjects. In the invention, besides using the entropy values of the sampling scale and the filtering scale position of the maximum positive correlation element or the maximum negative correlation element, the entropy values of a plurality of sampling scales and filtering scale positions near the maximum positive correlation element or the maximum negative correlation element can be collected, thereby increasing the adaptability of the automatic sleep staging method. Referring to FIG. 8, we studied the positive and negative relationship between intrinsic entropy and discrete level of consciousness. The Entropy elements positively correlated to the DCL sequence are denoted as PEDCL (Positive Entropy for DCL) and the Entropy elements negatively correlated to the DCL sequence are denoted as NEDCL (negative Entropy for DCL). Fig. 8 shows the manually labeled sleep states, PEDCL and NEDCL, for the two EEG channels F4-a1 (fig. 8a) and C4-a1 (fig. 8 b). The values of PEDCL and NEDCL at the different stages and their trends are shown in the figure, where the trend refers to the actual individual stage readings filtered by a digital low pass filter with a cut-off frequency of 1 cycle per hour. Manual sleep tagging, the sampling rate of PEDCL and NEDCL is 120 cycles per hour using 30s one cycle to tag sleep. As shown in fig. 8, the trends for PEDCL and manual sleep stage are similar, while NEDCL is negatively correlated with manual sleep stage trend. Both results obtained from the F4-a1 and C4-a1 channels matched well with the manual sleep cycles of 6 different subjects. Therefore, the automatic sleep staging method provided by the invention can be used for staging the sleep state well.

And 150, calculating an entropy value of the person to be measured under the sampling scale and the filtering scale according to the sampling scale and the filtering scale of the maximum positive correlation element or the maximum negative correlation element, and judging the sleep state of the patient according to the entropy value. In the invention, entropy values of a plurality of sampling scales and filtering scale positions near the maximum positive correlation element or the maximum negative correlation element can be acquired, and the adaptability of the automatic sleep staging method is improved.

In order to verify whether the value of the discrete consciousness level is reasonable, whether complexity measurement conducted in the intrinsic multi-scale entropy is consistent with the value of the discrete consciousness level is researched. It should be noted that the aforementioned discrete level of consciousness (DCL) values of 1-6 are not linear, depending on the manual sleep stage. However, it is certain that the level of consciousness of NREM3 is theoretically higher than that of NREM4, but there is no linear relationship between the levels of consciousness of NREM4 and NREM 3. Defining awareness of the awake state as the highest level in the sleep cycle is logical, and the levels of awareness for the four NREM sleep stages should be arranged in the following order: NREM1> NREM2> NREM3> NREM 4. Thus, PEDCL is considered a measure of entropy used, with a positive correlation with the level of consciousness in sleep. It is now crucial to verify whether the complexity measure by imese is consistent with the value of this level of consciousness. Figure 9 gives the results of a statistical comparison of the internal subjects for six sleep stages showing the PEDCL mean and standard deviation values for all subjects and five EEG channels, wherein figure 9a is the result of an internal subject comparison between PEDCL values for six sleep states for five channels; figure 9b is a comparison of the intra-subject comparisons between NEDCL values for six sleep states for five channels. Generally, the results support the arrangement of PEDCL values in the following order: sober > NREM1> NREM2> NREM3> NREM 4. In the figure, NREM4 is abbreviated "N4"; the abbreviation "N3" for NREM 3; the abbreviation "N2" for NREM 2; the abbreviation "N1" for NREM 1; REM is abbreviated "R"; the awake state is abbreviated as "W". In contrast to the results for PEDCL, the order of NEDCL values was awake < NREM1< NREM2< NREM3< NREM 4. The only exception is REM. Although REM sleep states were discovered more than 50 years ago, little is known about the neural circuit transitions between REM and non-REM sleep. Thus, the REM stage is called an abnormal sleep stage. There have been studies that propose brainstem trigger control to switch between REM and non-REM sleep stages. Importantly, our results indicate that the PEDCL and NEDCL values for REM are much closer to those of the NREM2 stage: for PEDCL, the REM phase values are lower than NREM1 and awake phase; for NEDCL, the value of REM phase is higher than NREM1 and awake phase. However, this unique feature is sufficient for us to classify sleep stages based on EEG recordings only. In combination with eye potentiometry (EOG) and myotonometry data, we can remove any uncertainty and make this classification easy to determine. Next, we will examine comparisons between subjects at different sleep stages. The results are given in figure 10, where the mean and standard deviation of the comparison between subjects for PEDCL and NEDCL values between six subjects are given, where figure 10a is the result of the comparison between subjects between PEDCL values for six sleep states for five channels; figure 10b is a comparison between subjects' NEDCL values for six sleep states for five channels. No significant difference in the values of individual subjects (statistical significance by Kolmogorov-Smirnov test, P < 0.05). These results indicate that PEDCL and NEDCL can be used as an objective quantification method to reflect the fluctuating pattern of sleep cycles based on the dynamic range of PEDCL and NEDCL. For the purpose of automatically classifying sleep stages, dynamic ranges may be considered to determine thresholds for classification between sleep stages.

By the automatic sleep stage establishing method, an automatic sleep stage establishing method can be established, and the method only needs to measure the entropy values of the patient to be measured in the optimal sampling scale and the filtering scale, namely, the automatic sleep stage can be carried out through the entropy values. The method greatly reduces the calculation amount of sleep staging by using multi-scale entropy, thereby improving the speed of automatic sleep staging.

Referring to fig. 11, the present invention further provides an automatic sleep staging method. The automatic sleep staging method includes step 210 of obtaining a PSG signal of a person to be tested and staging sleep using the PSG signal. Step 220, the PSG signal of the testee is decomposed into an original time sequence of several stages, the time of each stage may be consistent with the establishment method of the automatic sleep stage, and in the current manual stage, 30 seconds is used as one stage. Step 230, obtain an original time series of a phase, and decompose the original time series into a set of eigenmode functions or eigenmode-like functions. In this step, the decomposition method of the original time series is the same as the step 120 in the automatic sleep stage creation method. And 240, analyzing the intrinsic mode function or the eigenmode-like function according to the optimal filtering scale and the sampling scale to obtain an entropy value of the scale. Here, the optimal filter scale and the optimal sampling scale refer to a sampling scale and a filter scale of a maximum positive correlation element or a maximum negative correlation element in a correlation coefficient matrix in the automatic sleep stage establishing method. In fig. 8, we have seen that the entropy values at the optimal filter and sample scales are consistent with the resulting trend of artificial sleep stages, which entropy values at this scale are well suited for artificial sleep stages. And step 250, judging the sleep state of the person to be tested at the stage according to the entropy values of the optimal filtering scale and the sampling scale. And step 260, judging whether all the stages are finished. If not, returning to step 230 to obtain the original time sequence of the next stage; if all stages are judged to be finished, the sleep staging results of the testee are output. By adopting the automatic sleep staging method, only the entropy of the patient to be measured at the optimal sampling scale and the filtering scale needs to be measured, and the automatic sleep staging can be carried out through the entropy. The method greatly reduces the calculation amount of sleep staging by using multi-scale entropy, thereby improving the speed of automatic sleep staging.

In order to overcome the problems of individual difference and threshold values between different sleep states, the invention also provides an artificial intelligence method for assisting sleep staging. The artificial intelligence method uses a two-layer feedforward pattern recognition neural network model in a Matlab toolbox. A total of 200 entropy values were selected from five different EEG channels of five entropy matrices as inputs to the neural network model, and four different sleep stages were defined as slow wave sleep (SWS, including NREM3 and NREM4), light sleep (NREM1 and NREM2), rapid eye movement stage (REM) and waking stage as training targets for the model. The performance of the automatic sleep grading may be displayed in a confusion matrix, as shown in table 2. The correction percentages for the four classes as diagonal elements in the confusion matrix are 88.6%, 85.8%, 84.2% and 81.8%, respectively. The consistency of the four state judgment classifications and the target classification is more than 80%. Therefore, the automatic sleep staging method provided by the invention has better accuracy, and the output result is highly matched with the manually calibrated sleep state.

TABLE 2

Although the present invention has been described with reference to the preferred embodiments, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.