Automatic sleep apnea detection method based on S transformation and MDSNet

Abstract

An automatic sleep apnea detection method based on S transformation and MDSNet comprises the following steps that S1, multiple physiological signals of a patient to be detected are collected, dimensionality reduction is conducted through supervised principal component analysis, and three main components are obtained; s2, respectively processing the three main components based on S transformation to obtain three S-dynamic harmonic spectrum matrixes, wherein the three S-dynamic harmonic spectrum matrixes are in one-to-one correspondence with the three main components; s3, performing single-channel mapping on the S-dynamic harmonic spectrum matrix to obtain a single-channel S-dynamic harmonic spectrum, and constructing a mixed RGB image based on the three single-channel S-dynamic harmonic spectrum; s4, the trained improved MDRNet network model is used for recognizing the mixed RGB image, and a recognition result is output. According to the sleep apnea detection method, the accuracy of the sleep apnea detection result is improved, meanwhile, different types of sleep apnea can be distinguished, and the application range is wide.

Description

Technical Field

[0001] The present invention relates to the technical field of sleep apnea detection, in particular to an automatic sleep apnea detection method based on S transform and MDSNet. Background Art

[0002] Sleep apnea is one of the most common and serious sleep disorders. It refers to the repeated interruption of breathing during sleep, resulting in airflow limitation. Apnea can lead to serious harmful physiological disturbances and may have long-term sequelae, such as hypertension and cardiopulmonary dysfunction, impaired neuropsychological ability, depression and early all-cause mortality. Sleep apnea includes obstructive sleep apnea, central sleep apnea (CSA), mixed sleep apnea (MSA), and hypopnea.

[0003] Traditional sleep apnea detection is mainly based on polysomnography (PSG), which collects a large number of physiological signals from patients during sleep, including electroencephalogram (EEG), and then manually annotates apnea events in the readings. The difference in scores and other people's errors in this process are the main obstacles to the correct diagnosis of apnea. In order to solve this problem, many automatic sleep apnea detection methods have been born in the prior art.

[0004] At present, deep learning methods are mostly used in the prior art to detect sleep apnea. However, only one signal is used in the detection process. For example, Chinese patent CN117338253B discloses a sleep apnea detection method and device based on physiological signals, which collects physiological signals and separates respiratory effort signals. Based on the respiratory effort signals, spatial feature extraction, attention feature extraction and time series feature extraction are performed in sequence to obtain time series feature data, and the time series feature data is input into the deep learning network to obtain the detection results. The use of only one signal leads to a low accuracy rate of the detection results. At the same time, this method is only applicable to the detection of obstructive sleep apnea and has a limited scope of application. Summary of the invention

[0005] In order to solve the problems of low detection accuracy and relatively single detection type in the prior art, the present invention provides an automatic sleep apnea detection method based on S transform and MDSNet, which improves the accuracy of sleep apnea detection results and can distinguish different types of sleep apnea, i.e., has a wide range of applications.

[0006] In order to achieve the above object, the specific scheme adopted by the present invention is: an automatic sleep apnea detection method based on S transform and MDSNet, comprising the following steps:

[0007] S1, collects multiple physiological signals of the patient to be tested, and uses supervised principal component analysis to reduce the dimension to obtain three main components;

[0008] S2, based on S transformation, the three main components are processed respectively to obtain three S-dynamic harmonic spectrum matrices, and the three S-dynamic harmonic spectrum matrices correspond to the three main components one by one;

[0009] S3, performing single-channel mapping on the S-dynamic harmonic spectrum matrix to obtain a single-channel S-dynamic harmonic spectrum map, and constructing a mixed RGB map based on the three single-channel S-dynamic harmonic spectrum maps;

[0010] S4, use the trained improved MDRNet network model to recognize the mixed RGB image and output the recognition result.

[0011] As an optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: S1 includes:

[0012] S11, collecting physiological signals of the patient to be tested, the physiological signals including six EEG signals, two EOG signals, an EMG signal, a breathing signal from the abdomen, a breathing signal from the chest, an airflow signal, an oxygen saturation signal and an ECG signal;

[0013] S12, denoising the physiological signal;

[0014] S13, normalizing the physiological signal after denoising in S12;

[0015] S14, supervised principal component analysis is used to reduce the dimension of the normalized physiological signal in S13 to obtain three main components.

[0016] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: S12 includes:

[0017] S121, using wavelet basis functions to perform multi-layer wavelet decomposition of physiological signals;

[0018] S122, selecting a threshold λ to perform threshold processing on the physiological signal after wavelet decomposition;

[0019] S123, performing inverse wavelet transform on the physiological signal after threshold processing to reconstruct the physiological signal after denoising.

[0020] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: the wavelet basis function is a Biorthogonal function.

[0021] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: the threshold λ is:

[0022]

[0023] Where σ is the standard deviation of the noise and l is the length of the physiological signal.

[0024] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: S14 includes:

[0025] S141, performing normalization processing on the normalized physiological signals so that the mean value of each physiological signal is 0 and the variance is 1;

[0026] S142, mapping the response variables in the supervised principal component analysis to the physiological signals, and calculating the covariance matrix of the mapped physiological signals;

[0027] S143, performing eigenvalue decomposition on the covariance matrix to obtain the direction of the supervised principal component;

[0028] S144, projecting the normalized physiological signal based on the direction of the supervised principal component to obtain three main components.

[0029] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: the S transform in S2 is:

[0030]

[0031] Among them, n is the time index, indicating the sampling point of the physiological signal; N is the total number of sampling points of the physiological signal; k is the time variable, x[n] is the value of the physiological signal at time point n; f is the frequency variable; j is the imaginary unit; and e is the base of the natural logarithm.

[0032] As another optimization scheme of the automatic sleep apnea detection method based on S transform and MDSNet, in S3, the mapping function of the S-dynamic harmonic spectrum matrix is:

[0033]

[0034] Among them, A is the S-dynamic harmonic spectrum matrix, min_val is the minimum value in the S-dynamic harmonic spectrum matrix, max_val is the maximum value of the S-dynamic harmonic spectrum matrix, and scaled_matrix is ​​the mapped S-dynamic harmonic spectrum matrix.

[0035] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: in S4, the backbone network of the improved MDRNet network model is ResNet, and a multi-scale structure is introduced into ResNet.

[0036] As another optimization scheme of the above-mentioned automatic sleep apnea detection method based on S transform and MDSNet: in S4, the backbone network of the improved MDRNet network model is ResNet, and the DANet attention module is introduced into ResNet.

[0037] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides an automatic sleep apnea detection method based on S transform and MDSNet, collects a variety of physiological signals, and reduces the dimensions of the multiple signals to obtain three main components, removes redundant and irrelevant features, retains information in the data that is highly relevant to sleep apnea, and enhances the accuracy of sleep apnea detection results of the features after dimensionality reduction; converts the three main components into RGB images, and uses an improved network model to recognize the RGB images, thereby improving the accuracy of the detection results and being able to distinguish different types of sleep apnea, i.e., having a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 are three S-dynamic harmonic spectra;

[0039] Figure 2 is a mixed RGB image;

[0040] Figure 3 It is to improve the structure of the MDRNet network model;

[0041] Figure 4 This is the DANet module structure diagram. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further elaborated in detail below in conjunction with specific embodiments. The parts of the present invention that are not described and disclosed in detail in the following embodiments should be understood as the prior art known or should be known to those skilled in the art.

[0043] Example 1

[0044] An automatic sleep apnea detection method based on S transform and MDSNet comprises the following steps:

[0045] S1, collects multiple physiological signals of the patient to be tested, and uses supervised principal component analysis to reduce the dimension to obtain three main components, specifically:

[0046] S11, collect physiological signals of the patient to be tested, the physiological signals include six EEG signals (electroencephalogram channels), two EOG signals (electrooculogram), EMG signal (electromyogram of the chin), respiratory signal from the abdomen, respiratory signal from the chest, airflow signal, oxygen saturation signal (SaO2) and ECG signal, these physiological signals are collected through a multi-lead sleep monitor (Polysomnography, PSG); the six EEG signals are F3-M2, F4-M1, C3-M2, C4-M1, O1-M and O2-M1, the two EOG signals are E1-M2 and E2-M2, all physiological signals are measured in microvolts, the sampling frequency is 200Hz, and the oxygen saturation signal is upsampled to 200Hz through sampling technology to synchronize with other physiological signals.

[0047] EEG and EOG signals are used to monitor and analyze sleep stages and wakefulness; EMG signals record jaw muscle activity; respiratory signals monitor breathing patterns and respiratory events, airflow signals and oxygen saturation signals are used to evaluate respiratory function and oxygen levels, and ECG signals are used to record cardiac activity. The synchronous recording of these signals helps to fully understand the patient's sleep status and identify and classify various sleep-related events, such as awakenings, apnea, hypopnea, and periodic limb movements.

[0048] S12, denoising the physiological signal. In the process of collecting physiological signal data, high-frequency noise generated by the equipment is inevitable. Therefore, the collected physiological signal needs to be denoised. In this embodiment, the time-frequency localization characteristics of wavelet transform are used to decompose the physiological signal at multiple scales, so as to process the physiological signal and noise at different scales respectively, remove the noise in the physiological signal, and retain the main features and details of the physiological signal. Specifically:

[0049] S121, wavelet decomposition: using wavelet basis function, multi-layer wavelet decomposition is performed on the physiological signal. In this embodiment, Biorthogonal (Bior) is selected as the wavelet basis function, the decomposition layer number is 5, and multi-layer wavelet decomposition is performed on the physiological signal x(n):

[0050] x(n)=A5+D5+D4+D3+D2+D1,

[0051] Among them, n is the time index, indicating the sampling point of the signal; A5 is the approximate component of the 5th layer; D5, D4, D3, D2, and D1 are the detail components of each layer respectively.

[0052] S122, threshold processing: select a threshold λ to perform threshold processing on the physiological signal after wavelet decomposition. Specifically, select a threshold λ to perform threshold processing on the detail component. The method for selecting the threshold λ is:

[0053]

[0054] Where σ is the standard deviation of the noise and l is the length of the physiological signal.

[0055] S123, wavelet reconstruction: performing inverse wavelet transform on the physiological signal after threshold processing, that is, performing inverse wavelet transform on the detail component and the approximate component after threshold processing, and reconstructing the physiological signal after denoising:

[0056]

[0057] in, It is the detail component after threshold processing.

[0058] S13, normalizing the physiological signal after denoising in S12; in order to unify the scale range of 14 physiological signal data, eliminate the scale difference between different samples or features, and facilitate the subsequent better extraction of physiological signal features, the denoised physiological signal is subjected to Minmax normalization, and the Minmax normalization function is:

[0059]

[0060] Where x is a sample or feature in the physiological signal dataset, max(x) and min(x) represent the maximum and minimum values ​​in the physiological signal dataset, respectively. The normalized physiological signal data is within the specified range of [0,1].

[0061] S14, using supervised principal component analysis to reduce the dimension of the normalized physiological signals in S13, and obtain three main components. This embodiment has 14 physiological signals, and the data shows redundancy and high correlation, which affects the accuracy of the detection results. Therefore, it is necessary to reduce the dimension of the data, which can map high-dimensional data to low-dimensional space, retain important information in the data, and remove redundant and irrelevant features, which not only reduces the consumption of computing resources, but also improves the accuracy of the detection results. In this embodiment, supervised principal component analysis (Supervised Principal Component Analysis, SupervisedPCA) is used to reduce the dimension of physiological signals. Supervised PCA uses label information to know the selection of principal components, so that the retained components can not only explain the maximum variance of the physiological signal data, but also are highly correlated with apnea, thereby enhancing the detection ability of the features after dimensionality reduction for apnea. Specifically:

[0062] S141, standardizing data, standardizing the normalized physiological signals so that the mean of each physiological signal is 0 and the variance is 1. In this embodiment, each signal collected is an array, and 14 signals are 14 arrays. The 14 arrays are combined to form a data matrix X, that is, the data of the physiological signals are defined as the data matrix X;

[0063] S142, mapping the response variable in the supervised principal component analysis to the physiological signal, and calculating the covariance matrix of the mapped physiological signal. Specifically, calculating the projection of the response variable y in the data matrix X: Z = X T y, and map the calculated response variable to the data space to guide the selection of principal components; calculate the projected covariance matrix: S = ZZ T .

[0064] S143, calculate the principal component direction: perform eigenvalue decomposition on the covariance matrix, find the largest eigenvector in the eigenvalues, and obtain the supervised principal component direction, that is, the largest eigenvector is the supervised principal component direction.

[0065] S144, projecting data: projecting the normalized physiological signals based on the supervised principal component direction to obtain three dimensionality reduction signals. Specifically, projecting the data matrix X using the supervised principal component direction W:

[0066] X SPCA =XW

[0067] Get X SPCA It is the data matrix after dimensionality reduction, with three columns, each of which represents a main component.

[0068] S2, based on S transform, processes the three main components respectively to obtain three S-dynamic harmonic spectrum matrices, and the three S-dynamic harmonic spectrum matrices correspond to the three main components one by one; S transform is a time-frequency analysis method that combines the advantages of short-time Fourier transform (STFT) and wavelet transform, and can provide local characteristics of the signal in the time and frequency domains at the same time; it has the ability to provide multi-scale analysis and can provide characteristics of different frequencies of the signal; it is suitable for analyzing non-stationary signals and can effectively capture the instantaneous frequency and amplitude changes of the signal. Specifically:

[0069] A sliding window of 200ms is applied to each of the three main components, and an S-transform is applied to the main component in each processing window. For the signal of each sliding window, after the S-transform, the window provides an S-dynamic harmonic spectrum matrix containing the time-frequency positioning and multi-scale resolution information of the main component. The discrete time signal x[n] is represented by S(n,f) after the S-transform:

[0070]

[0071] Among them, ω(nk, f) is a frequency-dependent discrete Gaussian window function, defined as:

[0072]

[0073] Finally, the S transform formula is obtained:

[0074]

[0075] Among them, n is the time index, indicating the sampling point of the physiological signal; N is the total number of sampling points of the physiological signal; k is the time variable, indicating the center position of the Gaussian window function; x[n] is the value of the physiological signal at time point n; f is the frequency variable; j is the imaginary unit; e is the base of the natural logarithm, which is 2.72.

[0076] S3, single-channel mapping is performed on the S-dynamic harmonic spectrum matrix to obtain a single-channel S-dynamic harmonic spectrum map, and a mixed RGB map is constructed based on the three single-channel S-dynamic harmonic spectrum maps; specifically, each element in the three S-dynamic harmonic spectrum matrices is linearly mapped to the [0,255] interval according to the numerical size, and three single-channel S-dynamic harmonic spectrum maps are obtained, such as Figure 1 As shown, the mapping function is:

[0077]

[0078] Among them, A is the S-dynamic harmonic spectrum matrix, min_val is the minimum value in the S-dynamic harmonic spectrum matrix, max_val is the maximum value of the S-dynamic harmonic spectrum matrix, and scaled_matrix is ​​the mapped S-dynamic harmonic spectrum matrix.

[0079] The three single-channel S-dynamic harmonic spectra are superimposed as a channel of the mixed RGB image along the Z axis direction, and the mixed RGB is displayed using the imshow function in Matlab, as shown in Figure 2 As shown, specifically:

[0080] RGB(i,j)=(R(i,j),G(i,j),B(i,j))

[0081] Among them, RGB(i, j) represents the RGB color value at position (i, j), R(i, j), G(i, j) and B(i, j) represent the pixel values ​​of the red, green and blue channels at position (i, j), respectively.

[0082] S4, use the trained improved MDRNet network model to recognize the mixed RGB image and output the recognition result. In this embodiment, the improved MDRNet network model uses ResNet as the backbone network, and introduces a multi-scale structure into ResNet to improve the network's ability to extract tiny features in the S-dynamic harmonic spectrum; in this embodiment, the 14 physiological signals collected themselves have weak and non-stationary characteristics, and the 5 types of sleep pauses are very similar, and the increase in the depth of the network model will lose some features. The shallow feature map has a small receptive field and more small target detail information, but lacks rich semantic information. The deep feature map has a large receptive field and rich semantic information, but contains less small target information. In order to prevent the network model from ignoring some weak but meaningful information during feature extraction, a multi-scale structure is introduced into ResNet, such as Figure 3 As shown in the figure, the first output result of the ResNet network is branched and passed through the Avg Pool layer and the Flatten layer, and the feature fusion is output through the Concat layer. The output is passed through the fully connected layer to obtain the final classification result. The use of a multi-scale structure can effectively extract tiny features in the mixed RGB image and improve the network recognition accuracy.

[0083] In this embodiment, the mixed RGB image is formed by superimposing the data after dimensionality reduction, which helps to improve the network model to capture the global information about sleep breathing disorders of the patient to be detected. However, the feature information provided by the data of different channels contributes to different degrees to the final recognition effect. Therefore, the DANet attention module is introduced in ResNet to assign different weights to the feature information of different channels, enhance the feature expression ability of the network, and enable the model to better learn the complex features in the image. Specifically, the DANet module includes a channel attention module (CAM) and a position attention module (PAM), wherein the channel attention module assigns an optimal weight to the feature map of each channel in the mixed RGB image to ensure that the time domain information and frequency domain information in the image are better utilized, and the PAM assigns a weight to the feature information of each position in each layer of the channel feature map of the mixed RGB image to ensure that the model network focuses on the useful feature information in each layer of the channel feature map.

[0084] Figure 4The structure of the DANet module is shown in the figure. It can be seen from the figure that the CAM and PAM of DANet are parallel structures, which can optimize the features of an image at the same time, and finally summarize and output two sets of optimized features. In this embodiment, the DANet module is added to each Res Block. Based on the framework of Conv Block and Identity Block, an additional branch is added to the result output part to enter DANet to obtain the channel weight w and the weight is dot-producted with the original output to obtain the final result. The structure of Improved Conv Block and Improved Identity Block is shown in the figure. Figure 3 This strategy enables the model to adaptively adjust the weights of the channels at each stage, thereby improving the ability to perceive the features of each stage.

[0085] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An automatic sleep apnea detection method based on S transform and MDSNet, characterized in that: The following steps are involved: S1, collects multiple physiological signals of the patient to be tested, and uses supervised principal component analysis to reduce the dimension to obtain three main components; S2, based on S transformation, the three main components are processed respectively to obtain three S-dynamic harmonic spectrum matrices, and the three S-dynamic harmonic spectrum matrices correspond to the three main components one by one; S3, performing single-channel mapping on the S-dynamic harmonic spectrum matrix to obtain a single-channel S-dynamic harmonic spectrum map, and constructing a mixed RGB map based on the three single-channel S-dynamic harmonic spectrum maps; S4, use the trained improved MDRNet network model to recognize the mixed RGB image and output the recognition result.

2. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 1, characterized in that: The S1 includes: S11, collecting physiological signals of the patient to be tested, the physiological signals including six EEG signals, two EOG signals, an EMG signal, a breathing signal from the abdomen, a breathing signal from the chest, an airflow signal, an oxygen saturation signal and an ECG signal; S12, denoising the physiological signal; S13, normalizing the physiological signal after denoising in S12; S14, supervised principal component analysis is used to reduce the dimension of the normalized physiological signal in S13 to obtain three main components.

3. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 2, characterized in that: The S12 includes: S121, using wavelet basis functions to perform multi-layer wavelet decomposition of physiological signals; S122, selecting a threshold λ to perform threshold processing on the physiological signal after wavelet decomposition; S123, performing inverse wavelet transform on the physiological signal after threshold processing to reconstruct the physiological signal after denoising.

4. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 3, characterized in that: The wavelet basis function is a Biorthogonal function.

5. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 3, characterized in that: The threshold λ is: Where σ is the standard deviation of the noise and l is the length of the physiological signal.

6. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 2, characterized in that: The S14 includes: S141, performing normalization processing on the normalized physiological signals so that the mean value of each physiological signal is 0 and the variance is 1; S142, mapping the response variables in the supervised principal component analysis to the physiological signals, and calculating the covariance matrix of the mapped physiological signals; S143, performing eigenvalue decomposition on the covariance matrix to obtain the direction of the supervised principal component; S144, projecting the normalized physiological signal based on the direction of the supervised principal component to obtain three main components.

7. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 1, characterized in that: The S in S2 is transformed into: Among them, n is the time index, indicating the sampling point of the physiological signal; N is the total number of sampling points of the physiological signal; k is the time variable, x[n] is the value of the physiological signal at time point n; f is the frequency variable; j is the imaginary unit; and e is the base of the natural logarithm.

8. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 1, characterized in that: In S3, the mapping function of the S-dynamic harmonic spectrum matrix is: Among them, A is the S-dynamic harmonic spectrum matrix, min_val is the minimum value in the S-dynamic harmonic spectrum matrix, max_val is the maximum value of the S-dynamic harmonic spectrum matrix, and scaled_matrix is ​​the mapped S-dynamic harmonic spectrum matrix.

9. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 1, characterized in that: In S4, the backbone network of the improved MDRNet network model is ResNet, and a multi-scale structure is introduced into ResNet.

10. The automatic sleep apnea detection method based on S transform and MDSNet as claimed in claim 1, characterized in that: In S4, the backbone network of the improved MDRNet network model is ResNet, and the DANet attention module is introduced into ResNet.

Images