WO2022134242A1 - 一种自动睡眠分期的建立方法及其应用 - Google Patents

一种自动睡眠分期的建立方法及其应用 Download PDF

Info

Publication number
WO2022134242A1
WO2022134242A1 PCT/CN2021/071979 CN2021071979W WO2022134242A1 WO 2022134242 A1 WO2022134242 A1 WO 2022134242A1 CN 2021071979 W CN2021071979 W CN 2021071979W WO 2022134242 A1 WO2022134242 A1 WO 2022134242A1
Authority
WO
WIPO (PCT)
Prior art keywords
scale
sleep
entropy
sampling
time series
Prior art date
Application number
PCT/CN2021/071979
Other languages
English (en)
French (fr)
Inventor
黄锷
Original Assignee
江苏爱谛科技研究院有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 江苏爱谛科技研究院有限公司 filed Critical 江苏爱谛科技研究院有限公司
Priority to US17/618,502 priority Critical patent/US20220323000A1/en
Publication of WO2022134242A1 publication Critical patent/WO2022134242A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4806—Sleep evaluation
    • A61B5/4815—Sleep quality
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4806—Sleep evaluation
    • A61B5/4812—Detecting sleep stages or cycles
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4806—Sleep evaluation
    • A61B5/4809—Sleep detection, i.e. determining whether a subject is asleep or not
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4863—Measuring or inducing nystagmus
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/369—Electroencephalography [EEG]
    • A61B5/372—Analysis of electroencephalograms
    • A61B5/374—Detecting the frequency distribution of signals, e.g. detecting delta, theta, alpha, beta or gamma waves
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7246—Details of waveform analysis using correlation, e.g. template matching or determination of similarity
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems

Definitions

  • the invention relates to the technical field of data processing, in particular to a method for establishing automatic sleep staging and its application.
  • REM rapid eye movement
  • NREM non-rapid eye movement
  • iMSE Intrinsic Multiscale Entropy, iMSE
  • MSE Multiscale Entropy
  • entropy is defined as the "unpredictable" physical measurement of coarse-grained nonlinear time series on multiple time scales, and can be defined using sample entropy or approximate entropy when calculating entropy.
  • the number of non-overlapping samples merged into one sample is defined as the time scale of coarse graining in MSE.
  • timescale 1 represents the original time of the measurement and is digitized at the original sampling rate.
  • the time scale n represents the coarse-grained time series, the sampling interval is n times the original data, and the sampling rate is 1/n of the original sampling rate. Therefore, the coarse-grained time scale represents the length of time in the number of original sampling intervals.
  • MSE Intrinsic Multiscale Entropy
  • the empirical mode decomposition algorithm can perform two preprocessing steps of denoising and detrending in order to extract the required information from nonlinear and non-stationary real signals.
  • EMD can act as an adaptive binary filter bank to decompose complex time series into a set of eigenmode functions (IMFs).
  • IMFs eigenmode functions
  • Each eigenmode function is characterized by a small bandwidth and zero mean, so the eigenmode function is stationary.
  • the filtered time series can be reconstructed using different combinations of eigenmode functions, and the reconstructed time series can filter out high frequency noise components or low frequency trends from the original time series.
  • iMSE method as a new signal analysis method.
  • iMSE quantizes the entropy of the quantized components at multiple coarse-grained time scales (ie, sampling scales). Meanwhile, the filter band represents the second filter scale in iMSE. Then, the entropy values are displayed in a two-dimensional matrix on two axes of sampling scale and filtering scale, which greatly enhances the power of the original MSE.
  • this method only needs to calculate the entropy value of the optimal sampling scale and the filtering scale, that is, automatic sleep staging can be performed through the entropy value.
  • This method will greatly reduce the computational complexity of sleep staging with multi-scale entropy, thereby increasing the speed of automatic sleep staging.
  • the present invention provides a method for establishing automatic sleep staging, comprising the following steps: acquiring several groups of PSG signals and artificial sleep marker information of the PSG signals; The original time series is decomposed into a set of eigenmode functions or quasi eigenmode functions; the eigenmode functions or quasi eigenmode functions are combined to obtain m groups of time series sets; multi-scale entropy analysis, Use n sampling scales to calculate the entropy value of m groups of time series sets to obtain an entropy matrix with m ⁇ n elements; establish a correlation coefficient matrix between the consciousness level and the elements of the entropy matrix, and find the largest positive correlation in the correlation coefficient matrix
  • a mode decomposition method is adopted, and the mode decomposition method is one of the following methods: empirical mode decomposition method , ensemble mode decomposition method, adaptive binary mask empirical mode decomposition method.
  • a set of high-pass filters are used, and the cut-off frequencies of the high-pass filters are 32Hz, 16Hz, 8Hz, 4Hz, 2Hz and 1Hz.
  • the PSG signal comprises at least one of the following EEG signals: Fp4-A1, F4-A1, C4-A1, P4-A1, O2-A1.
  • the level of consciousness is determined according to artificial sleep marker information, and the level of consciousness is used to reflect the degree of wakefulness in sleep, wherein the wakefulness stage is quantified as 6, the rapid eye movement stage is quantified as 5, and the NREM1 stage is quantified as 4 , the NREM2 stage is quantized to 3, the NREM3 stage is quantized to 2, and the NREM4 stage is quantized to 1.
  • the correlation coefficient matrix between the level of consciousness and the elements of the entropy matrix is established based on the Pearson coefficient.
  • an artificial intelligence method is used to calculate the threshold between different sleep states.
  • the present invention also provides an automatic sleep staging method, which is characterized by comprising the following steps: acquiring the PSG signal of the person to be tested; decomposing the PSG signal of the person to be tested into the original time series of several stages; obtaining the original time of one stage sequence, decompose the original time series into a set of eigenmode functions or similar eigenmode functions; obtain the entropy value at this scale according to the sampling scale and filtering scale of the largest positive correlation element or the largest negative correlation element; The entropy value determines the sleep state of the person to be tested at this stage.
  • the PSG signal comprises at least one of the following EEG signals: Fp4-A1, F4-A1, C4-A1, P4-A1, O2-A1.
  • the time of each stage is 30 seconds.
  • the automatic sleep staging establishment method in the present invention we can establish an automatic sleep staging method, which only needs to measure the entropy value of the patient to be tested at the optimal sampling scale and the filtering scale, that is, automatic sleep can be performed through the entropy value. Staging. This method will greatly reduce the computational complexity of sleep staging with multi-scale entropy, thereby increasing the speed of automatic sleep staging.
  • Fig. 1 is a flow chart of a method for establishing automatic sleep staging according to the present invention.
  • Figure 2 shows five-channel EEG images of six different sleep stages.
  • Figure 3 is a typical IMF-like set of 6 sleep states (channels C4-A1).
  • FIG. 4 is a flowchart of a method for decomposing and recombining original time series to obtain a time series set in the present invention.
  • Fig. 5 shows the entropy value calculation of sampling scale from 1 to n for each group of time series sets.
  • FIG. 6 is an example diagram of a two-dimensional entropy matrix for six different sleep stages.
  • Figure 7 shows the discrete level of consciousness and the correlation coefficient matrix between the elements of each entropy matrix in the five entropy matrices of the five EEG recording channels.
  • Figure 8a shows the time series and trends of manually calibrated sleep stages, PEDCL and NEDCL in the F4-A1 channel
  • Figure 8b shows the time series and trends of manually calibrated sleep stages, PEDCL and NEDCL in the C4-A1 channel.
  • Figure 9a shows the intra-subject comparison results between the PEDCL values of the six sleep states of five channels
  • Figure 9b shows the intra-subject comparison results of the NEDCL values of the six sleep states of the five channels.
  • Figure 10a is the result of the inter-subject comparison between the PEDCL values of the six sleep states of five channels
  • Figure 10b is the result of the inter-subject comparison of the NEDCL values of the six sleep states of the five channels.
  • FIG. 11 is a flow chart of the automatic sleep staging method in the present invention.
  • step 110 several groups of polysomnographic (PSG) signals and artificial sleep marker information of the polysomnographic are acquired.
  • Polysomnography equipment can be used to measure and record one or more physiological signals, such as F3-A2 brainwave signal, F4-A1 brainwave signal, C3-A2 brainwave signal, C4-A1 brainwave signal, P3 -A2 brain wave signal, P4-A1 brain wave signal, left eye movement signal, right eye movement signal, etc. Since sleep is a brain wave activity, the physiological signals closer to the brain can better reflect the sleep state. Generally, brain wave signals are used for sleep staging.
  • the cyclic alternating pattern (CAP) in sleep is a periodic brain wave change that occurs during non-REM sleep, and this component reflects the microstructure of sleep.
  • Polysomnography recordings in cyclic alternating patterns in sleep can be downloaded from the Phsionet website.
  • the database provides a total of 108 PSG records covering 8 different pathological conditions. In the present invention, we only select three pathological conditions of normality, insomnia and narcolepsy for research. In order to reduce the variables of the automatic sleep staging method of the present invention, we also require the electroencephalogram (Electroencephalogram, EEG) channel and sampling frequency must be the same.
  • EEG electroencephalogram
  • Figure 2 is a typical five-channel brainwave atlas in sleep state, which includes awake stage, non-rapid eye movement stage (including the first stage - the fourth stage, denoted as NREM1, NREM2, NREM3 and NREM4 respectively) and rapid eye movement (rapid eye movement, REM) stage.
  • NREM1, NREM2, NREM3 and NREM4 the first stage - the fourth stage
  • REM rapid eye movement
  • Six examples are shown in Figure 2, corresponding to the five-channel EEG images of the six different sleep stages described above.
  • low-frequency oscillations are included in the EEG signals of NREM3 and NREM4 representing deep sleep stages. Therefore, NREM3 and NREM4 are also known as slow wave sleep.
  • Step 120 decompose the original time series of several stages in the PSG signal into a set of eigenmode functions or quasi-eigenmode functions respectively. Because in the PSG recording, sleep is divided into a stage according to 30s, and the sleep state is analyzed. Therefore, when we establish an automatic sleep staging method in the present invention, we also analyze 30s as a stage, and decompose the original time series of several stages in the PSG signal into a set of eigenmode functions or quasi-eigenmode functions. . The essence of pre-analysis is to decompose the original time series into a set of independent narrow-bandwidth and detrended zero-mean eigenmode functions or eigenmode-like functions with binary frequency bands.
  • This step is critical because entropy is calculated from the probability density function of the data, but probability density can only be calculated on data without a trend.
  • the modal decomposition method can be used to decompose. Because the empirical mode decomposition is an ideal binary filter bank, nonlinear time series can be adaptively decomposed into a set of eigenmode functions.
  • the modal decomposition method refers to any modal decomposition method that can obtain the eigenmode function components in the present invention, such as the empirical mode decomposition method (Empirical Mode Decomposition, EMD), the ensemble mode decomposition method (Ensemble Empirical Mode Decomposition, EEMD), or Adaptive Binary Mask Empirical Mode Decomposition (Conjugate Adaptive Dyadic Masking Empirical Mode Decomposition, CADM-EMD).
  • EMD empirical mode decomposition method
  • EEMD Ensemble Mode decomposition method
  • CADM-EMD Adaptive Binary Mask Empirical Mode Decomposition
  • the modal decomposition method for sleep staging has the following drawbacks: (1) Unless an expensive computation and well-designed masking method is used in the augmentation algorithm, it is difficult to completely solve the modal aliasing problem, and the resulting eigenmode function ( IMF) may have different IMF intrinsic frequency band mismatches in different sleep stages. For example, the distribution of the instantaneous frequency of IMF1 from the EEG signal recorded in the sleep stage of NREM1 is different from the distribution of the instantaneous frequency of the EEG signal recorded in the sleep stage of NREM4.
  • the commonly defined delta band is 0.5-4Hz
  • theta band is 4-8Hz
  • the alpha band is 8-16Hz
  • the beta band is 16-30Hz
  • the gamma band is 30-60Hz.
  • These frequency bands are similar, but not identical.
  • high frequency noise is first removed from the time series using a low pass filter with a cutoff frequency of 64 Hz.
  • the frequency bands of the first 6 IMF classes decomposed by the alternative method are 32-64Hz (similar to the gamma band), 16-32Hz (beta band), 8-16Hz (alpha band), 4-8Hz (theta band) ), 2-4Hz (delta frequency band) and 1-2Hz (low delta frequency band), the corresponding relationship is shown in Table 1.
  • class IMF1 32-64Hz Gamma band (gamma band) class IMF2 16-32Hz Beta band class IMF3 8-16Hz Alpha band class IMF4 4-8Hz
  • Theta band class IMF5 2-4Hz Delta band (delta band) class IMF6 1-2Hz Low delta band (L ⁇ band)
  • FIG. 3 is a typical IMF-like set of six sleep states obtained by this method, and the selected EEG channel in this figure is the C4-A1 channel.
  • the filter frequency bands of IMF1-6 are respectively ⁇ frequency band (32-64Hz), ⁇ frequency band (16-32 Hz), ⁇ frequency band (8-16 Hz), ⁇ frequency band (4-8 Hz), ⁇ frequency band (2-4 Hz) and Low delta band (1-2Hz).
  • Step 121 combine the eigenmode functions or quasi-eigenmode functions obtained by decomposition to obtain m groups of time series sets.
  • This set of filtered time series can be reconstituted using various combinations of eigenmode functions or eigenmode-like functions into a new set of m sets of detrended zero-mean time series showing different aspects of the original data. information (such as high frequencies only) or any specific selected frequency band. As shown in Fig.
  • a set of complex time series is decomposed through modal decomposition or its alternative method to obtain a set of eigenmode functions or quasi-eigenmode functions, and then Combining the set of eigenmode functions or eigenmode-like functions, the recombined filtered time series will cover all possible frequency segment representations of the original data.
  • the above 6 class eigenmode functions as an example, from these initial 6 class eigenmode functions, 14 additional filtered time series can be reconstructed for further analysis.
  • Step 130 multi-scale entropy analysis, using n sampling scales to calculate the entropy values of the m groups of time series sets obtained in step 121 respectively, to obtain an entropy matrix with m ⁇ n elements.
  • the entropy value calculation of the sampling scale from 1 to n is performed for each group of time series sets, thereby obtaining an entropy matrix with m ⁇ n elements.
  • Multiscale entropy analysis evaluates the complexity of time series based on entropy values corresponding to multiple different time scales.
  • multiple sampling scales can be used to estimate the entropy value of each filtered time series containing one or more eigenmode functions or combinations of eigenmode-like functions to obtain a row vector of entropy.
  • the sampling scale is defined as the number of consecutive samples that are non-overlappingly merged into a coarse-grained time series from the original time series.
  • the length of the sampled time series is 1/n the length of the original time series, where n is the sampling scale. Only time series with a sampling scale of 1 are the original time series.
  • Any entropy definition method can be used to calculate the entropy value in coarse-grained time series, such as approximate entropy, sample entropy, etc.
  • FIG. 6 which is the result of calculating the entropy value from 60 sampling scales ranging from 2 to 120 with a step size of 2 for the time series of 20 different filtering scales in the present invention.
  • a two-dimensional mean entropy matrix with 20 ⁇ 60 elements is computed for 20 filtered time series at 60 different sampling scales.
  • Figure 6 represents 6 typical average entropy matrices for 6 different sleep stages. In these subplots, the entropy values are shown in different colors, the X-axis labels the sampling scale from 2 to 120 with a stride of 2, and the Y-axis labels the filter scale in the class eigenmode 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 clear differences between different sleep stages.
  • Step 140 establishing a correlation coefficient matrix between the consciousness level and the elements of the entropy matrix, and finding out the sampling scale and the filtering scale of the largest positive correlation element or the largest negative correlation element in the correlation coefficient matrix.
  • the level of consciousness was determined according to the artificial sleep signature of the PSG signal.
  • the discrete consciousness level (DCL) in sleep was first defined according to manually classified sleep stages, which was used to reflect the degree of wakefulness in sleep.
  • the waking stage which represents the highest level of consciousness, is quantified as 6, the REM 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, which represents the lowest level of consciousness is quantized to 1.
  • each subject had nearly 1000 stages in a night's sleep, which formed a time series with amplitudes as defined above, known as the DCL series.
  • Figure 7 shows the DCL and the matrix of correlation coefficients between elements in 5 entropy matrices for 5 EEG recording channels.
  • the first five subplots represent the correlation coefficient matrices derived from five different EEG channels, and the sixth subplot shows the mean matrix of the first five subplots.
  • the filter scale of the element with the largest positive correlation is IMFs1-3 ( ⁇ - ⁇ frequency band), and the sampling scale is 2 (ie 2/512 seconds at a sampling rate of 512Hz).
  • Sampling scale its correlation coefficient is 0.64869; while the filter scale of the element with the largest negative correlation is IMF1 (gamma frequency band), the sampling scale is 102 (that is, the sampling scale of 102/512 seconds at the sampling rate of 512Hz), its correlation The coefficient is -0.73802.
  • the location is close to all six subjects and all five electrodes, we can choose the sampling scale and filtering scale as the best sampling scale and filtering scale for the patient to be tested, and then according to the sampling scale and filtering scale
  • the entropy value under the scale is used to automatically perform sleep staging for the patient to be tested. This conclusion is correct, as shown by the average matrix in the figure above.
  • PEDCL Positive Entropy for DCL
  • NEDCL Negative Entropy for DCL
  • Figure 8 shows manually labelled sleep states, PEDCL and NEDCL, for two EEG channels of F4-A1 ( Figure 8a) and C4-A1 ( Figure 8b).
  • the PEDCL and NEDCL values for the different stages and their trends are shown in the graph, where the trend refers to the actual stage readings filtered by a digital low-pass filter with a cutoff frequency of 1 cycle per hour.
  • the sampling rate for PEDCL and NEDCL is 120 cycles per hour using a cycle of 30s to mark sleep.
  • trends in PEDCL and manual sleep stages were similar, whereas NEDCL was negatively correlated with manual sleep stage trends.
  • Both results obtained from the F4-A1 and C4-A1 channels perfectly matched the manual sleep cycles of 6 different subjects. Therefore, the automatic sleep staging method provided in the present invention can well stage the sleep state.
  • Step 150 calculate the entropy value of the subject under the sampling scale and the filtering scale, and judge the sleep state of the patient according to the entropy value.
  • the entropy values of several sampling scales and filter scale positions near the maximum positive correlation element or the maximum negative correlation element can also be collected to increase the adaptability of the automatic sleep staging method.
  • DCL discrete level of consciousness
  • FIG 9 presents the results of a within-subject statistical comparison of the six sleep stages, in which the PEDCL mean and standard deviation values for all subjects and five EEG channels are shown, where Figure 9a is for six of the five channels Results of intra-subject comparisons between PEDCL values of sleep states; Figure 9b is the results of intra-subject comparisons of NEDCL values of six sleep states of five channels.
  • the results support the following order of PEDCL values: awake > NREM1 > NREM2 > NREM3 > NREM4.
  • NREM4 is abbreviated as "N4"; NREM3 is abbreviated as “N3”; NREM2 is abbreviated as “N2”; NREM1 is abbreviated as “N1”; REM is abbreviated as “R”; W”.
  • NREM4 is abbreviated as "N4"; NREM3 is abbreviated as “N3”; NREM2 is abbreviated as “N2”; NREM1 is abbreviated as “N1”; REM is abbreviated as “R”; W”.
  • Figure 10 presents the mean and standard deviation of the between-subject comparison for PEDCL and NEDCL values across six subjects, where Figure 10a is for six sleeps of five channels Results of the between-subject comparison of PEDCL values between states; Figure 10b is the results of the between-subject comparison of NEDCL values between the six sleep states of five channels. None of the individual subject values were significantly different (statistical significance by Kolmogorov-Smirnov test, P ⁇ 0.05). These results suggest that PEDCL and NEDCL can be used as objective quantitative methods to reflect the fluctuation pattern of sleep cycles based on the dynamic range of PEDCL and NEDCL. For purposes of automatically classifying sleep stages, dynamic range may be considered to determine thresholds for classification between sleep stages.
  • the automatic sleep staging establishment method in the present invention we can establish an automatic sleep staging method, which only needs to measure the entropy value of the patient to be tested at the optimal sampling scale and the filtering scale, that is, automatic sleep can be performed through the entropy value. Staging. This method will greatly reduce the computational complexity of sleep staging with multi-scale entropy, thereby increasing the speed of automatic sleep staging.
  • the present invention also provides an automatic sleep staging method.
  • the automatic sleep staging method includes step 210 , acquiring the PSG signal of the test subject, and using the PSG signal to stage sleep.
  • Step 220 decompose the PSG signal of the subject to be tested into the original time series of several stages, the time of each stage can be consistent with the establishment method of automatic sleep staging, in the current manual staging, 30 seconds are used as a stage, The present invention is not limited to this, and the time of each stage can also be shortened to better examine the influence of the sleep state on other aspects.
  • Step 230 Obtain the original time series of a stage, and decompose the original time series into a set of eigenmode functions or quasi-eigenmode functions.
  • the decomposition method of the original time series is the same as step 120 in the automatic sleep stage establishment method.
  • Step 240 analyze the eigenmode function or the quasi-eigenmode function according to the optimal filtering scale and the sampling scale, and obtain the entropy value at the scale.
  • the optimal filtering scale and sampling scale refer to the sampling scale and filtering scale of the largest positive correlation element or the largest negative correlation element in the correlation coefficient matrix in the automatic sleep stage establishment method.
  • Fig. 8 we have seen that the entropy values at the optimal filtering scale and the sampling scale are consistent with the results of artificial sleep staging, and the entropy values at this scale are very suitable for artificial sleep staging.
  • Step 250 determine the sleep state of the test subject at this stage.
  • Step 260 it is judged whether all the stages have been judged. If the judgment is not completed, return to step 230 to obtain the original time series of the next stage; if all stages are judged, output the sleep staging result of the subject to be tested.
  • automatic sleep staging method in the present invention it is only necessary to measure the entropy value of the patient to be tested at the optimal sampling scale and the filtering scale, that is, automatic sleep staging can be performed by the entropy value. This method will greatly reduce the computational complexity of sleep staging with multi-scale entropy, thereby increasing the speed of automatic sleep staging.
  • the present invention also provides an artificial intelligence method for assisting sleep staging.
  • This artificial intelligence method uses a two-layer feedforward pattern recognition neural network model in the Matlab toolbox.
  • a total of 200 entropy values were selected as input to the neural network model from five different EEG channels of five entropy matrices, and four different sleep stages were defined as slow-wave sleep (SWS, including NREM3 and NREM4), light sleep (NREM1 and NREM2), the rapid eye movement (REM) and awake phases were used as training targets for the model.
  • SWS slow-wave sleep
  • NREM1 and NREM2 light sleep
  • REM rapid eye movement
  • awake phases were used as training targets for the model.
  • the performance of automatic sleep classification can be shown in a confusion matrix, as shown in Table 2.
  • the corrected percentages for the four categories that are diagonal elements in the confusion matrix are 88.6%, 85.8%, 84.2%, and 81.8%, respectively.
  • the consistency between the above four state judgment classification and target classification is greater than 80%. Therefore, the automatic sleep staging method provided by the present invention has a good accuracy rate, and the output result is highly matched with the manually calibrated sleep state.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Artificial Intelligence (AREA)
  • Psychiatry (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physiology (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Anesthesiology (AREA)
  • Psychology (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

一种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（110）；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（120）；将类本征模态函数进行组合，得到m组时间序列集合（121）；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（130）；建立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（140）；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，计算待测者在采样尺度和滤波尺度的熵值，根据熵值判断待测者的睡眠状态（150）。

Description

一种自动睡眠分期的建立方法及其应用 技术领域
本发明涉及数据处理技术领域，特别涉及一种自动睡眠分期的建立方法及其应用。
背景技术
我们一生中三分之一的时间是在睡眠中度过的，睡眠是我们生活中至关重要的一部分。睡眠质量不仅影响了我们的日常健康，也会影响我们的经济生产力。尽管睡眠的功能还不完全清楚，但是最近的研究发现，除了传统认为的增强记忆以外，神经胶质细胞的收缩和脑脊液冲洗的增加在废物清除中也起着关键的作用。研究还表明，睡眠不是一个统一的状态，从一个状态到另一个状态的切换是高度非线性且不稳定的过程。睡眠障碍会带来严重的后果，包括生活质量下降、并发症、早期死亡等，并对经济社会成本造成巨大的影响。这些因素使得对睡眠的研究至关重要。传统上，根据R&K(Rechtschaffen and Kales)标准，正常的睡眠分为五个阶段：快速眼动(REM)和其他四个非快速眼动(NREM)阶段(1-4)。根据多导睡眠图记录的特征(包括脑电图，肌电图和眼电图以及其他参数)来估计睡眠状态，在临床实践中，睡眠技术人员以视觉方式确定每个30s的睡眠阶段。但是，手动评分是一个耗时的过程，并且由于不同阶段之间具有相似性，不同的人可能会得出不一致的结果。因此，通过多导睡眠图评估睡眠质量是一项劳动强度大，耗时长且容易出错的过程。除此之外，随着睡眠监测的患者数量的增多，有限的睡眠分析师越来越难以满足日渐增多的睡眠图分析需求。
因此，目前亟需一种准确且客观的方法来对睡眠阶段进行自动分期并评估睡眠质量。
发明内容
为了解决上述问题，在本发明中，我们提出了本征多尺度熵iMSE(intrinsic Multiscale Entropy,iMSE)作为一种新的信号分析方法。首先，使用多尺度熵 MSE(Multiscale Entropy,MSE)来研究睡眠状态，旨在利用多个时间尺度上的熵求和来量化复杂度。在多尺度熵中，熵定义为对多个时间尺度上的粗粒化非线性时间序列进行“不可预测”的物理测量，在计算熵时，可以使用样本熵或者近似熵来定义。合并为一个样本的不重叠样本的数量被定义为MSE中粗粒化的时间尺度。在进行数字信号处理时，时间尺度1表示测量的原始时间，并按原始采样率进行数字化。时间尺度n表示粗粒化的时间序列，采样间隔为原始数据的n倍，采样率为原始采样率的1/n。因此，粗粒化的时间尺度以原始采样间隔的数量表示时间长度。MSE的方法反映了一个观点，即熵是一种取决于采样间隔时间尺度的度量。其次，由于睡眠过程不是稳定的过程，因此，我们提出本征多尺度熵(iMSE)，即通过结合经验模态分解(Empirical Mode Decomposition，EMD)与多尺度熵来解决非稳态的限制。经验模态分解算法能够执行去噪声和去趋势两个预处理步骤，以便从非线性和非稳态的真实信号中提取所需要的信息。EMD可以作为自适应的二元滤波器组，将复杂的时间序列分解为一组本征模态函数(IMFs)。每一个本征模态函数具有带宽较小且零均值的特征，因此，本征模态函数是稳态的。可以使用本征模态函数的不同组合来重建滤波后的时间序列，重建后的时间序列可以从原始的时间序列中滤除高频噪声成分或者低频趋势。作为一种功能组合(functional combination)，我们提出了iMSE方法作为一种新的信号分析方法。
此外，为了避免高计算成本和EMD的其他细节问题，我们引入一种简单的基于滤波器的伪EMD方法，该方法模仿了EMD的功能，避免模态混叠的问题，从而从时间序列中***地提取了滤波后的分量。iMSE在多个粗粒度时标(即采样尺度)上对量化后的分量的熵进行量化。同时，滤波频带代表iMSE中的第二滤波尺度。然后，将熵值显示在采样尺度和滤波尺度的两个轴上的二维矩阵中，极大地增强了原始MSE的功能。
在本发明中，我们提供一种自动睡眠分期的建立方法，通过对采样尺度和滤波尺度的两个轴上的熵矩阵进行分析，找出适用于自动睡眠分期的最佳采样尺度和滤波尺度。在分析患者的睡眠状态时，采用本方法只需要计算最佳采样尺度和滤波尺度的熵值，即可以通过该熵值进行自动睡眠分期。该方法将大大减少用多尺度熵进行睡眠分期的计算量，进而提高了自动睡眠分期的速度。
为了实现上述发明目的，本发明提供一种自动睡眠分期的建立方法，包括 以下步骤：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，用于将PSG信号中的每一阶段的原始时间序列分解为一组本征模态函数或者类本征模态函数；将所述本征模态函数或者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；建立意识水平与熵矩阵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的采样尺度和滤波尺度；所述采样尺度为所述粗粒度尺度；所述滤波尺度为时间序列集合；根据最大正相关元素或者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。
优选地，将PSG信号中的每一阶段的原始时间序列分解为一组本征模态函数时，采用模态分解方法，所述模态分解方法为下列方法其中之一：经验模态分解法，集合经模态分解法，自适应性二进位遮罩经验模态分解法。
优选地，将PSG信号中的每一阶段的原始时间序列分解为一组类本征模态函数时，采用一组高通滤波器，所述高通滤波器的截止频率分别为32Hz、16Hz、8Hz、4Hz、2Hz和1Hz。
优选地，PSG信号至少包含以下脑电信号其中之一：Fp4-A1，F4-A1，C4-A1，P4-A1，O2-A1。
优选地，意识水平根据人工睡眠标记信息而定，所述意识水平用于反映睡眠中的清醒程度，其中，清醒阶段被量化为6，快速眼动阶段被量化为5，NREM1阶段被量化为4，NREM2阶段被量化为3，NREM3阶段被量化为2，以及NREM4阶段被量化为1。
优选地，建立意识水平与熵矩阵元素之间的相关系数矩阵时，基于Pearson系数。
优选地，根据待测者在最大正相关元素或者最大负相关元素的采样尺度和滤波尺度的熵值判断患者的睡眠状态时，采用人工智能方法计算不同睡眠状态之间的阈值。
本发明还提供一种自动睡眠分期方法，其特征在于，包含以下步骤：获取待测试者的PSG信号；将待测试者的PSG信号分解为若干个阶段的原始时间序列；取得一个阶段的原始时间序列，将该原始时间序列分解为一组本征模态函数或者类本征模态函数；根据最大正相关元素或者最大负相关元素的采样尺度 和滤波尺度，获得在该尺度的熵值；根据所述熵值，判断待测试者在该阶段的睡眠状态。
优选地，PSG信号至少包含以下脑电信号其中之一：Fp4-A1，F4-A1，C4-A1，P4-A1，O2-A1。
优选地，将待测试者的PSG信号分解为若干个阶段的原始时间序列时，每一阶段的时间为30秒。
通过本发明中的自动睡眠分期建立方法，我们可以建立一种自动睡眠分期方法，该方法只需要测量待测患者在最佳采样尺度和滤波尺度的熵值，即可以通过该熵值进行自动睡眠分期。该方法将大大减少用多尺度熵进行睡眠分期的计算量，进而提高了自动睡眠分期的速度。
附图说明
图1本发明自动睡眠分期的建立方法流程图。
图2为六种不同睡眠阶段的五通道EEG图像。
图3是6种睡眠状态的典型的类IMF集合(通道为C4-A1)。
图4为本发明中进行原始时间序列分解及重新组合得到时间序列集合的方法流程图。
图5为对于每一组时间序列集合均进行从1到n的采样尺度的熵值计算。
图6为六种不同睡眠阶段的二维熵矩阵的示例图。
图7为5个EEG记录通道的5个熵矩阵中离散意识水平和各个熵矩阵元素之间的相关系数矩阵。
图8a为F4-A1通道中手动标定睡眠阶段、PEDCL和NEDCL的时间序列和趋势；图8b为C4-A1通道中手动标定睡眠阶段、PEDCL和NEDCL的时间序列和趋势。
图9a为五通道的六个睡眠状态的PEDCL值之间的受试者内部比较结果；图9b为五通道的六个睡眠状态的NEDCL值之间的受试者内部比较结果。
图10a为五通道的六个睡眠状态的PEDCL值之间的受试者之间比较结果；图10b为五通道的六个睡眠状态的NEDCL值之间的受试者之间比较结果。
图11为本发明中自动睡眠分期方法的流程图。
具体实施方式
以下配合附图及本发明的较佳实施例，进一步阐述本发明为达成预定发明目的所采取的技术手段。
请参照图1所示，为本发明自动睡眠分期的建立方法的详细实施方式。在步骤110中，获取若干组多导睡眠图(polysomnographic，PSG)信号以及多导睡眠图的人工睡眠标记信息。多导睡眠图仪器可被用来测量并记录一种或多种生理信号，如F3-A2脑波信号，F4-A1脑波信号，C3-A2脑波信号，C4-A1脑波信号，P3-A2脑波信号，P4-A1脑波信号，左眼眼动信号，右眼眼动信号等。由于睡眠属于脑波活动，因此离脑部越近的生理信号越能反映睡眠状态，一般选用脑波信号来进行睡眠分期。睡眠中的循环交替模式(cyclic alternating pattern，CAP)是在非快速眼动睡眠期出现的一种呈周期性的脑波改变，该成分反映了睡眠的微观结构。睡眠中的循环交替模式中的多导睡眠图记录可以从Phsionet网站上下载。该数据库提供了一共108种PSG记录，其中包含了8种不同的病理情况。在本发明中，我们仅选择正常、失眠和嗜睡症三种病理情况进行研究，为了减少本发明自动睡眠分期建立方法的变量，我们还要求这些多导睡眠图记录中记录的脑电波(Electroencephalogram，EEG)通道和采样频率必须一致。基于这些标准，我们以512Hz的采样率采样的Fp4，F4，C4，P4和O2这五个EEG通道采集的PSG记录。六个正常对照受试者(n1，n2，n3，n5，n10和n11)，五个失眠受试者(ins2，ins4，ins6，ins7和ins8)和五个嗜睡症受试者(narco1，narco2，narco3，narco4和narco5)用于本发明中的研究。对于每个PSG记录，文本文件中都包含事件时间，睡眠阶段和CAP注释信息。图2为睡眠状态典型的五通道脑电波图谱，睡眠状态中包括清醒阶段、非快速眼动阶段(包括第一阶段-第四阶段，分别记为NREM1，NREM2，NREM3和NREM4)以及快速眼动(rapid eye movement，REM)阶段。图2中显示了六个示例，分别对应于上述六种不同睡眠阶段的五通道EEG图像。从图2中可以看出，表示深度睡眠阶段的NREM3和NREM4的EEG信号中包含了低频振荡。因此，NREM3和NREM4也被称作为慢波睡眠。
步骤120，对上述PSG信号中若干阶段的原始时间序列分别分解为一组本征模态函数或者类本征模态函数。因为PSG记录中，将睡眠按照30s分为一个阶段，进行睡眠状态的分析。因此，在本发明中我们建立自动睡眠分期方法时， 也将30s作为一个阶段进行分析，将PSG信号中若干阶段的原始时间序列分别分解为一组本征模态函数或者类本征模态函数。预分析的本质是将原始时间序列分解为一组独立的窄带宽和具有二进频带的去趋势零均值的本征模态函数或者类本征模态函数。该步骤至关重要，因为熵是根据数据的概率密度函数计算得出的，但是概率密度只能在没有趋势的数据上进行计算。在对原始时间序列进行分解时，可以采用模态分解的方法进行分解。因为经验模态分解是理想的二元滤波器组，可以将非线性时间序列自适应地分解为一组本征模态函数。模态分解方法指本发明中利用任意一种可以取得本征模态函数分量的模态分解方法，例如经验模态分解法(Empirical Mode Decomposition,EMD)，集合经模态分解法(Ensemble Empirical Mode Decomposition,EEMD),或者自适应性二进位遮罩经验模态分解法(Conjugate Adaptive Dyadic Masking Empirical Mode Decomposition,CADM-EMD)。
在本发明中，我们还提供一种替代方法用于克服模态分解方法用于睡眠分期的不足。模态分解方法用于睡眠分期存在以下缺陷：(1)除非在增强算法中使用昂贵的计算和精心设计的屏蔽方法，否则很难完全解决模态混叠问题，所得的本征模态函数(IMF)可能在不同的睡眠阶段具有不同的IMF固有频段的失配。例如，来自记录在NREM1的睡眠阶段的EEG信号的IMF1的瞬时频率的分布不同于记录在NREM4的睡眠阶段的EEG信号的瞬时频率的分布。(2)将来自IMF的结果频段与EEG信号的通用频段对齐并不容易。例如，通常定义的δ频段是0.5-4Hz，θ频段为4-8Hz，α频段为8-16Hz，β频段为16-30Hz，γ频段为30-60Hz。这些频段是相似的，但是并非完全相同。为了便于与通常定义的频段进行比较，我们可以模拟一个预定的(但是不是自适应的)滤波器组，以另一种方法从EEG信号中提取一组类IMF函数。(3)对于那些不熟悉模态分解方法的人，使用滤波器方法更容易实现。在本发明提供的替代方法中，使用截止频率为64Hz的低通滤波器首先从时间序列中去除高频噪声。然后，使用一组截止频率依次为32Hz、16Hz、8Hz、4Hz、2Hz和1Hz的高阶高通滤波器来提取前6个类IMF。从理论上来讲，通过替代方法分解的前6个类IMF的频段为32-64Hz(类似于γ频段)、16-32Hz(β频段)、8-16Hz(α频段)、4-8Hz(θ频段)、2-4Hz(δ频段)和1-2Hz(低δ频段)，其对应关系如表1中所示。
滤波后的组分 频率范围 频段
类IMF1 32-64Hz Gamma band(γ频段)
类IMF2 16-32Hz Beta band(β频段)
类IMF3 8-16Hz Alpha band(α频段)
类IMF4 4-8Hz Theta band(θ频段)
类IMF5 2-4Hz Delta band(δ频段)
类IMF6 1-2Hz Low delta band(Lδ频段)
表1
采用滤波器方法取得的类IMF函数，可以很好地解决来自IMF的结果频段与EEG通用频段不易对齐的问题。图3是通过该方法得到的6种睡眠状态的典型的类IMF集合，该附图中选取的EEG通道为C4-A1通道。其中类IMF1-6的滤波频带分别为γ频段(32-64Hz)、β频段(16-32Hz)、α频段(8-16Hz)、θ频段(4-8Hz)、δ频段(2-4Hz)和低δ频段(1-2Hz)。
步骤121，将分解得到的本征模态函数或者类本征模态函数进行组合，得到m组时间序列集合。可以使用本征模态函数或者类本征模态函数的各种组合将这组滤波后的时间序列重组为m组去趋势零均值时间序列的新集合，该新集合显示了原始数据不同方面的信息(如仅高频)或者任何特定的选定频段。如图4所示，显示了步骤120-步骤121的过程，将一组复杂的时间序列通过模态分解或者其替代方法，分解得到一组本征模态函数或者类本征模态函数，再将该组本征模态函数或者类本征模态函数进行组合，重新组合的滤波时间序列将覆盖原始数据的所有可能的频率分段表示。以上述的6个类本征模态函数为例，从这最初的6个类本征模态函数中，可以重建14个额外的滤波时间序列，以进行进一步的分析。20个经过过滤的时间序列包括仅IMF1，IMFs 1-2，IMFs 1-3,IMFs 1-4,IMFs 1-5,IMFs 1-6；仅IMF2，IMFs2-3,IMFs2-4,IMFs2-5,IMFs 2-6；仅IMF3,IMFs3-4,IMFs3-5,IMFs 3-6；仅IMF 4,IMFs4-5,IMFs 4-6；仅IMF5和IMF5-6。
步骤130，多尺度熵分析，使用n个采样尺度对步骤121中得到的m组时间序列集合分别进行熵值计算，得到具有m×n个元素的熵矩阵。如图5中所示，对于每一组时间序列集合均进行从1到n的采样尺度的熵值计算，从而得到一 个具有m×n个元素的熵矩阵。多尺度熵分析是根据对应于多个不同时间尺度的熵值来评估时间序列的复杂度。在iMSE中，可以使用多个采样尺度来估计每个包含一个或多个本征模态函数或者类本征模态函数的组合后的滤波时间序列的熵值，以获得熵的行向量。采样尺度定义为从原始时间序列中非重叠地合并到一个粗粒度时间序列的连续样本的数量。采样时间序列的长度是原始时间序列长度的1/n，其中n是采样尺度。只有采样尺度为1的时间序列才是原始时间序列。可以采用任何熵定义方法来计算粗粒度时间序列中的熵值，如近似熵，样本熵等。在本发明中，我们选择近似熵(approximate entropy，ApEn)来计算多尺度熵分析中经过滤波的时间序列的熵向量。为了减小个例以及不同睡眠阶段对熵值的影响，客观地显示各个不同睡眠状态的熵值情况，我们可以选择多个睡眠阶段进行研究，分别按照步骤120至步骤130，求出各个睡眠阶段的熵矩阵，然后针对具有相同人工标记的睡眠阶段的熵矩阵求出平均值，得到平均熵矩阵。请参照图6所示，为本发明中20种不同的滤波尺度的时间序列从2-120的60个采样尺度以2为步长进行熵值计算的结果。在60个不同的采样尺度上，针对20个滤波后的时间序列，计算了具有20×60个元素的二维平均熵矩阵。图6代表了6种不同睡眠阶段的6个典型的平均熵矩阵。在这些子图中，熵值以不同的颜色显示，X轴标注从2到120的采样尺度，步长为2，Y轴标注在类本征模态函数中的滤波尺度，从1到20。每个熵矩阵代表在滤波和采样的多个控制条件下针对同一阶段的EEG信号的熵度量，不同睡眠阶段之间存在明显的差异。
步骤140，建立意识水平与熵矩阵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最大负相关元素的采样尺度和滤波尺度。其中，意识水平根据PSG信号的人工睡眠标记而定。为了建立意识水平与熵测值之间的关系，首先根据手动分类的睡眠阶段定义睡眠中的离散意识水平(discrete consciousness level,DCL)，意识水平用于反映睡眠中的清醒程度。代表最高意识水平的清醒阶段被量化为6，快速眼动阶段被量化为5，NREM1阶段被量化为4，NREM2阶段被量化为3，NREM3阶段被量化为2，以及代表最低意识水平的NREM4阶段被量化为1。平均而言，每个受试者在一夜的睡眠中有将近1000个阶段，这形成了具有上述定义的有幅值的时间序列，成为DCL序列。接下来，我们将检查DCL序列与每个iMSE矩阵元素的时间序列之间的相关性。我们发 现，基于Pearson相关系数，一些元素与睡眠中的DCL呈正相关，另一些元素则呈负相关。图7显示了针对5个EEG记录通道的5个熵矩阵中DCL和各个元素之间的相关系数矩阵。前五个子图表示从五个不同的EEG通道得出的相关系数矩阵，第六个子图显示前五个子图的均值矩阵。然后，我们从每个相关系数矩阵中各选取一个与DCL序列正相关性和负相关性最大的一个元素。例如，通道Fp2-A1(子图1)的矩阵中，最大正相关的元素的滤波尺度为IMFs1-3(α-γ频段)，采样尺度为2(即在512Hz的采样率下2/512秒的采样尺度)，其相关系数为0.64869；而最大负相关的元素的滤波尺度为IMF1(γ频段)，采样尺度为102(即在512Hz的采样率下102/512秒的采样尺度)，其相关系数为-0.73802。该位置对于所有的六个受试者和所有的五个电极都是接近的，我们可以选择该采样尺度和滤波尺度作为待测患者的最佳采样尺度和滤波尺度，进而根据该采样尺度和滤波尺度下的熵值对待测患者自动进行睡眠分期。正如上图中的平均矩阵所示，这个结论是正确的。因此，我们决定根据整体结果选择这两个熵元素，如子图6所示，以代表所有的受试者。在本发明中，除了使用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位置的熵值之外，还可以采集最大正相关元素或最大负相关元素附近的几个采样尺度和滤波尺度位置的熵值，增加该自动睡眠分期方法的适应性。请参照图8所示，我们研究了本征熵与离散意识水平之间的正负关系。与DCL序列正相关的熵元素记为PEDCL(Positive Entropy for DCL)，与DCL序列负相关的熵元素记为NEDCL(Negative Entropy for DCL)。图8显示了F4-A1(图8a)和C4-A1(图8b)的两个EEG通道的手动标记睡眠状态，PEDCL和NEDCL。不同阶段的PEDCL和NEDCL值以及其趋势均在图中显示，其中趋势是指截止频率为每小时1周期的数字低通滤波器过滤的实际各个阶段的读数。手动睡眠标记，PEDCL和NEDCL的采样率是使用30s一个周期每小时120个周期对睡眠进行标记。如图8中所示，PEDCL和手动睡眠阶段的趋势是相似的，而NEDCL与手动睡眠阶段趋势呈负相关。从F4-A1和C4-A1通道获得的两个结果都与6个不同受试者的手动睡眠周期完全匹配。因此，本发明中提供的自动睡眠分期方法将可以很好地对睡眠状态进行分期。
步骤150，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下的熵值，根据该熵值判断患者的睡眠状 态。在本发明中，还可以采集最大正相关元素或者最大负相关元素附近的几个采样尺度和滤波尺度位置的熵值，增加该自动睡眠分期方法的适应性。
为了验证离散意识水平的取值是否具有合理性，我们研究了本征多尺度熵中进行的复杂性度量是否与离散意识水平的取值一致。应当指出的是，根据手动睡眠阶段，前述离散的意识水平(DCL)的值为1-6并非是线性的。但是，可以肯定的是，NREM3的意识水平在理论上要高于NREM4，但是NREM4和NREM3的意识水平之间并不存在线性关系。将清醒状态的意识定义为睡眠周期中的最高级别是符合逻辑的，并且四个NREM睡眠阶段的意识级别应按照以下顺序排列：NREM1>NREM2>NREM3>NREM4。因此，PEDCL被认为是一种使用的熵测量，与睡眠中的意识水平呈正相关。现在，至关重要的是要验证通过iMSE进行的复杂性度量是否与此意识水平的取值一致。图9给出了六个睡眠阶段的受试者内部统计比较的结果，其中显示了所有受试者和五个EEG通道的PEDCL平均值和标准差值，其中，图9a为五通道的六个睡眠状态的PEDCL值之间的受试者内部比较结果；图9b为五通道的六个睡眠状态的NEDCL值之间的受试者内部比较结果。通常，结果支持PEDCL值按以下顺序排列：清醒>NREM1>NREM2>NREM3>NREM4。在图中，NREM4的缩写为“N4”；NREM3的缩写为“N3”；NREM2的缩写为“N2”；NREM1的缩写为“N1”；REM的缩写为“R”；清醒状态的缩写为“W”。与PEDCL的结果相反，NEDCL值的顺序是清醒<NREM1<NREM2<NREM3<NREM4。唯一例外的是REM。尽管REM睡眠状态是在50多年前发现的，但是对REM和非REM睡眠之间的神经回路转换仍然知之甚少。因此，REM阶段称为反常睡眠阶段。目前已经有研究提出了在REM和非REM睡眠阶段之间切换的脑干触发器控制。重要的是，我们的结果表明REM的PEDCL和NEDCL值更加接近于NREM2阶段的值：对于PEDCL，REM阶段的值低于NREM1和清醒阶段；对于NEDCL，REM阶段的值高于NREM1和清醒阶段。但是，这种独特的特征足以让我们仅根据EEG记录对睡眠阶段进行分类。结合眼电位描记法(EOG)和肌张力数据，我们可以消除任何不确定性，并使得这种分类很容易确定。接下来，我们将检查不同睡眠阶段的受试者间的比较。结果在图10中给出，其中给出了在六个受试者之间针对PEDCL和NEDCL值的受试者之间比较的平均值和标准偏差，其中，图10a为五通道的六个睡眠状态的PEDCL值之间的受试者之间比较结果；图10b为五通 道的六个睡眠状态的NEDCL值之间的受试者之间比较结果。个体受试者的值均无显著差异(统计学显著性通过Kolmogorov-Smirnov检验，P<0.05)。这些结果表明，PEDCL和NEDCL可以用作客观定量方法，以基于PEDCL和NEDCL的动态范围反映睡眠周期的波动模式。出于自动对睡眠阶段进行分类的目的，可以考虑动态范围以确定睡眠阶段之间分类的阈值。
通过本发明中的自动睡眠分期建立方法，我们可以建立一种自动睡眠分期方法，该方法只需要测量待测患者在最佳采样尺度和滤波尺度的熵值，即可以通过该熵值进行自动睡眠分期。该方法将大大减少用多尺度熵进行睡眠分期的计算量，进而提高了自动睡眠分期的速度。
请参照图11所示，本发明还提供一种自动睡眠分期方法。该自动睡眠分期方法包括步骤210，获取待测试者的PSG信号，用此PSG信号对睡眠进行分期。步骤220，将待测试者的PSG信号分解为若干个阶段的原始时间序列，每个阶段的时间可以与自动睡眠分期的建立方法一致，目前的手动分期中，均是以30秒为一个阶段，本发明并不以此为限，也可以缩短每个阶段的时间以更好地检查睡眠状态对其他方面的影响。步骤230，取得一个阶段的原始时间序列，将该原始时间序列分解为一组本征模态函数或者类本征模态函数。在该步骤中，原始时间序列的分解方法与自动睡眠分期建立方法中的步骤120相同。步骤240，根据最佳滤波尺度和采样尺度对本征模态函数或类本征模态函数进行分析，获得在该尺度的熵值。此处，最佳滤波尺度和采样尺度是指在自动睡眠分期建立方法中，相关系数矩阵中最大正相关元素或者最大负相关元素的采样尺度和滤波尺度。在图8中，我们已经看到在最佳滤波尺度和采样尺度的熵值与人工睡眠分期的结果趋势一致，该尺度下的熵值非常适用于人工睡眠分期。步骤250，根据最佳滤波尺度和采样尺度的熵值，判断待测试者在该阶段的睡眠状态。步骤260，判断是否所有的阶段均判断完毕。如未判断完毕，则返回步骤230，获取下一个阶段的原始时间序列；若所有阶段均判断完毕，则输出待测试者的睡眠分期结果。采用本发明中的自动睡眠分期方法，只需要测量待测患者在最佳采样尺度和滤波尺度的熵值，即可以通过该熵值进行自动睡眠分期。该方法将大大减少用多尺度熵进行睡眠分期的计算量，进而提高了自动睡眠分期的速度。
为了克服个体差异以及不同睡眠状态之间的阈值问题，本发明中还提供一种人工智能方法，用于辅助进行睡眠分期。该人工智能方法使用Matlab工具箱 中的两层前馈模式识别神经网络模型。从五个熵矩阵的五个不同EEG通道中总共选择了200个熵值作为神经网络模型的输入，并且将四个不同的睡眠阶段定义为慢波睡眠(SWS，包括NREM3和NREM4)，轻睡眠(NREM1和NREM2)，快速眼动阶段(REM)和清醒阶段作为模型的训练目标。自动睡眠分级的性能可以在混淆矩阵中显示，如表2所示。作为混淆矩阵中对角线元素的四个类别的校正百分比分别为88.6％、85.8％、84.2％和81.8％。以上四个状态判定分类与目标分类的一致性均大于80％。因此，本发明提供的自动睡眠分期方法具有较好的正确率，其输出结果与手动标定的睡眠状态高度匹配。
Figure PCTCN2021071979-appb-000001
表2
以上所述仅是本发明的优选实施例而已，并非对本发明做任何形式上的限制，虽然本发明已以优选实施例揭露如上，然而并非用以限定本发明，任何熟悉本专业的技术人员，在不脱离本发明技术方案的范围内，当可利用上述揭示的技术内容作出些许更动或修饰为等同变化的等效实施例，但凡是未脱离本发明技术方案的内容，依据本实用发明的技术实质对以上实施例所作的任何简单修改、等同变化与修饰，均仍属于本发明技术方案的范围内。

Claims (10)

  1. 一种自动睡眠分期的建立方法，其特征在于，包括以下步骤：
    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，用于将PSG信号中的每一阶段的原始时间序列分解为一组本征模态函数或者类本征模态函数；将所述本征模态函数或者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；
    意识水平根据人工睡眠标记信息而定；
    建立所述意识水平与所述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；所述采样尺度为所述粗粒度尺度；
    根据最大正相关元素或者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。
  2. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，将PSG信号中的每一阶段的原始时间序列分解为一组本征模态函数时，采用模态分解方法，所述模态分解方法为下列方法其中之一：经验模态分解法，集合经模态分解法，自适应性二进位遮罩经验模态分解法。
  3. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，将PSG信号中的每一阶段的原始时间序列分解为一组类本征模态函数时，采用一组高通滤波器，所述高通滤波器的截止频率分别为32Hz、16Hz、8Hz、4Hz、2Hz和1Hz。
  4. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，所述PSG信号至少包含以下脑电信号其中之一：Fp4-A1，F4-A1，C4-A1，P4-A1，O2-A1。
  5. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，所述意识水平根据人工睡眠标记信息而定，所述意识水平用于反映睡眠中的清醒程度，其中，清醒阶段被量化为6，快速眼动阶段被量化为5，NREM1阶段被量化为4，NREM2阶段被量化为3，NREM3阶段被量化为2，以及NREM4阶段被量化为1。
  6. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，建立意识水平与熵矩阵元素之间的相关系数矩阵时，基于Pearson系数。
  7. 根据权利要求1中的自动睡眠分期的建立方法，其特征在于，根据待测者在最大正相关元素或者最大负相关元素的采样尺度和滤波尺度的熵值判断患者的睡眠状态时，采用人工智能方法计算不同睡眠状态之间的阈值。
  8. 一种权利要求1的方法的应用，其特征在于，包含以下步骤：
    获取待测试者的PSG信号；
    将待测试者的PSG信号分解为若干个阶段的原始时间序列；
    取得一个阶段的原始时间序列，将该原始时间序列分解为一组本征模态函数或者类本征模态函数；
    根据权利要求1中的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，计算待测试者在所述采样尺度和滤波尺度的熵值；
    根据所述熵值，判断待测试者在该阶段的睡眠状态。
  9. 根据权利要求8中的自动睡眠分期方法，其特征在于：所述PSG信号至少包含以下脑电信号其中之一：Fp4-A1，F4-A1，C4-A1，P4-A1，O2-A1。
  10. 根据权利要求8中的自动睡眠分期方法，其特征在于，将待测试者的PSG信号分解为若干个阶段的原始时间序列时，每一阶段的时间为30秒。
PCT/CN2021/071979 2020-12-24 2021-01-15 一种自动睡眠分期的建立方法及其应用 WO2022134242A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/618,502 US20220323000A1 (en) 2020-12-24 2021-01-15 A Construction Method for Automatic Sleep Staging and Use Thereof

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011548920.9A CN112690759B (zh) 2020-12-24 2020-12-24 一种自动睡眠分期的建立方法及其应用
CN202011548920.9 2020-12-24

Publications (1)

Publication Number Publication Date
WO2022134242A1 true WO2022134242A1 (zh) 2022-06-30

Family

ID=75509682

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/071979 WO2022134242A1 (zh) 2020-12-24 2021-01-15 一种自动睡眠分期的建立方法及其应用

Country Status (3)

Country Link
US (1) US20220323000A1 (zh)
CN (1) CN112690759B (zh)
WO (1) WO2022134242A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7088153B2 (ja) * 2019-09-19 2022-06-21 カシオ計算機株式会社 Ｃａｐ（周期性脳波活動）検出装置、ｃａｐ（周期性脳波活動）検出方法及びプログラム
CN115422984B (zh) * 2022-11-04 2023-01-24 北京理工大学 一种基于时间尺度信号分解及熵特征的信号分类方法
CN116509336B (zh) * 2023-06-27 2024-05-03 安徽星辰智跃科技有限责任公司 基于波形分析的睡眠周期性检测及调节方法、***和装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006255134A (ja) * 2005-03-17 2006-09-28 Ikeda Denshi Kogaku Kenkyusho:Kk 脳波計測表示方法及び装置
CN106473705A (zh) * 2016-09-21 2017-03-08 广州视源电子科技股份有限公司 用于睡眠状态监测的脑电信号处理方法和***
CN108742517A (zh) * 2018-03-27 2018-11-06 重庆邮电大学 一种基于Stacking单导联脑电睡眠自动分期方法
CN109602417A (zh) * 2018-11-23 2019-04-12 杭州妞诺科技有限公司 基于随机森林的睡眠分期方法及***
CN110897639A (zh) * 2020-01-02 2020-03-24 清华大学深圳国际研究生院 一种基于深度卷积神经网络的脑电睡眠分期方法
CN111493822A (zh) * 2020-03-23 2020-08-07 济南国科医工科技发展有限公司 一种基于睡眠脑电的快速眼动期睡眠行为障碍分类方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9186083B2 (en) * 2009-02-24 2015-11-17 Natus Medical Incorporated Method and system for sleep stage determination
CN102274022B (zh) * 2011-05-10 2013-02-27 浙江大学 一种基于脑电信号的睡眠状态监测方法
US11039784B2 (en) * 2014-12-05 2021-06-22 Agency For Science, Technology And Research Sleep profiling system with feature generation and auto-mapping
EP3649933A4 (en) * 2017-07-07 2021-07-14 Osaka University PAIN DETERMINATION USING TREND ANALYSIS, MEDICAL DEVICE WITH MACHINE LEARNING, ECONOMIC DISCRIMINATION MODEL AND IOT, TAILORED MACHINE LEARNING AND NOVEL BRAIN WAVE CHARACTERISTICS DETERMINATION OF PAIN
CN108968915B (zh) * 2018-06-12 2021-07-02 山东大学 基于熵特征及支持向量机的睡眠状态分类方法及***

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006255134A (ja) * 2005-03-17 2006-09-28 Ikeda Denshi Kogaku Kenkyusho:Kk 脳波計測表示方法及び装置
CN106473705A (zh) * 2016-09-21 2017-03-08 广州视源电子科技股份有限公司 用于睡眠状态监测的脑电信号处理方法和***
CN108742517A (zh) * 2018-03-27 2018-11-06 重庆邮电大学 一种基于Stacking单导联脑电睡眠自动分期方法
CN109602417A (zh) * 2018-11-23 2019-04-12 杭州妞诺科技有限公司 基于随机森林的睡眠分期方法及***
CN110897639A (zh) * 2020-01-02 2020-03-24 清华大学深圳国际研究生院 一种基于深度卷积神经网络的脑电睡眠分期方法
CN111493822A (zh) * 2020-03-23 2020-08-07 济南国科医工科技发展有限公司 一种基于睡眠脑电的快速眼动期睡眠行为障碍分类方法

Also Published As

Publication number Publication date
CN112690759A (zh) 2021-04-23
CN112690759B (zh) 2022-04-19
US20220323000A1 (en) 2022-10-13

Similar Documents

Publication Publication Date Title
WO2022134242A1 (zh) 一种自动睡眠分期的建立方法及其应用
Srinivasan et al. Artificial neural network based epileptic detection using time-domain and frequency-domain features
Peker An efficient sleep scoring system based on EEG signal using complex-valued machine learning algorithms
KR101535352B1 (ko) 전두엽 뇌파를 이용한 우울증 심도 측정 방법
CN106413541B (zh) 用于诊断睡眠的***和方法
CN110135285B (zh) 一种使用单导设备的脑电静息态身份认证方法及装置
Hussein et al. Robust detection of epileptic seizures using deep neural networks
Decat et al. Beyond traditional sleep scoring: Massive feature extraction and data-driven clustering of sleep time series
CN112842279B (zh) 一种基于多维度特征参数的睡眠质量评估方法及装置
CN116509338B (zh) 基于模态分析的睡眠周期性检测及调节方法、***和装置
Si et al. An expert system for EEG monitoring in the pediatric intensive care unit
TWI288875B (en) Multiple long term auto-processing system and method thereof
Khosropanah et al. A hybrid unsupervised approach toward EEG epileptic spikes detection
Jiang et al. A multi-scale parallel convolutional neural network for automatic sleep apnea detection using single-channel EEG signals
Ganesan et al. Binary state prediction of sleep or wakefulness using EEG and EOG features
Kloosterman et al. Boosts in brain signal variability track liberal shifts in decision bias
Zhang et al. Automatic Sleep Staging using Multi‐dimensional Feature Extraction and Multi‐kernel Fuzzy Support Vector Machine
CN116525063B (zh) 基于时频分析的睡眠周期性检测及调节方法、***和装置
CN116211308A (zh) 一种高强度运动下机体疲劳评估方法
CN115969330A (zh) 一种睡眠情绪活跃水平检测量化的方法、***和装置
CN115966308A (zh) 一种睡眠记忆活跃水平检测量化的方法、***和装置
Ranjan et al. Automatic detection of mental health status using alpha subband of EEG data
Yamunarani et al. EEG–Based Bipolar Disorder Deduction Using Machine Learning
Rakhonde et al. Sleep stage classification for prediction of human sleep disorders by using machine learning approach
Saastamoinen et al. Systematic performance evaluation of a continuous-scale sleep depth measure

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21908261

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21908261

Country of ref document: EP

Kind code of ref document: A1