Sleep respiratory status monitoring method, sleep monitoring system, device and medium

Abstract

The application provides a sleep respiratory state monitoring method, a sleep monitoring system, equipment and a medium, and the method comprises the steps of: generating a time-frequency spectrum diagram of a normal mixed heart impact signal; screening out a region associated with a heart rate in the time-frequency spectrum diagram based on a mask matrix to obtain time-frequency spectrum information; decomposing an intrinsic mode function of the normal mixed heart impact signal according to information of the trained mask; extracting time-domain waveform information in the intrinsic mode function; and fusing and analyzing the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result. The sleep respiratory state monitoring method provided by the application obtains high-quality time-frequency spectrum information with little noise, also obtains time-domain waveform information that will not be distorted, further realizes comprehensive analysis of a sleep apnea event through fusion and analysis of the time-frequency spectrum information and the time-domain waveform information, generates an accurate sleep apnea classification result, and provides strong technical support for diagnosis and treatment of sleep disorders.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sleep monitoring, and in particular to a sleep respiration state monitoring method, a sleep monitoring system, a device and a medium. BACKGROUND

[0002] Obstructive sleep apnea (OSA) is a common sleep disorder characterized by upper airway obstruction or narrowing during sleep. OSA can lead to daytime sleepiness, cognitive impairment, and cardiovascular disease, and in severe cases can even cause fatal sleep-related events. Therefore, continuous monitoring and analysis of an individual's respiratory movements during sleep, especially monitoring and analysis of sleep apnea, can help prevent health risks caused by sleep apnea syndrome and other diseases.

[0003] Portable sleep monitoring devices provide an economical and efficient solution for sleep respiration state monitoring in a home environment. Portable sleep monitoring devices usually use a mattress sensor to collect heart rate information, which has good adaptability and convenience. Although the mattress sensor can obtain heart rate information, the signal quality is poor, and the peak value of the signal may be weakened or shifted by a filter in the subsequent signal processing process, or even a false peak value may be generated, which seriously affects the identification of the true signal, thereby limiting the application of the mattress device in sleep apnea monitoring. SUMMARY

[0004] The present application provides a sleep respiration state monitoring method, a sleep monitoring system, a device and a medium to solve the problem of inaccurate sleep apnea monitoring results.

[0005] In a first aspect, the present application provides a sleep respiration state monitoring method, comprising:

[0006] generating a time-frequency spectrum of a normal mixed cardiac impact signal;

[0007] based on a mask matrix, filtering out a region associated with heart rate in the time-frequency spectrum to obtain time-frequency spectrum information;

[0008] decomposing an intrinsic mode function of the normal mixed cardiac impact signal;

[0009] extracting time-domain waveform information in the intrinsic mode function;

[0010] fusing and analyzing the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result.

[0011] In one embodiment, the normal mixed cardiac impact signal is obtained by:

[0012] physiological signals of the chest and abdomen are collected through piezoelectric ceramic sensors;

[0013] raw ballistocardiogram signals are extracted from the physiological signals;

[0014] normal signals in the raw ballistocardiogram signals are screened out by detecting nonlinear features of the raw ballistocardiogram signals, and the normal mixed ballistocardiogram signals are obtained.

[0015] In one embodiment, the generation of the time-frequency spectrogram of the normal mixed ballistocardiogram signal comprises:

[0016] a plurality of different window lengths are determined based on resolution requirements in the time domain and the frequency domain;

[0017] discrete short-time Fourier transforms are performed on the normal mixed ballistocardiogram signals based on the window lengths, and time-frequency spectrograms of multiple scales are obtained.

[0018] In one embodiment, when the region associated with the heart rate in the time-frequency spectrogram is screened out based on the mask matrix, time-frequency spectrogram information is obtained, and the following steps are performed for each time-frequency spectrogram:

[0019] the time-frequency spectrogram is input into a ViT model to obtain a probability matrix output by the ViT model; each probability value in the probability matrix reflects the confidence degree of an image block corresponding to the probability value in the time-frequency spectrogram being associated with the heart rate;

[0020] the mask of the image block corresponding to the probability value greater than or equal to a preset threshold in the probability matrix is determined as 1;

[0021] the mask of the image block corresponding to the probability value less than the preset threshold in the probability matrix is determined as 0;

[0022] a mask matrix is generated according to the masks of the image blocks in the time-frequency spectrogram;

[0023] the mask matrix and the time-frequency spectrogram are combined to generate a region associated with the heart rate;

[0024] the region associated with the heart rate is determined as time-frequency spectrogram information.

[0025] In one embodiment, the decomposition of the intrinsic mode function of the normal mixed ballistocardiogram signal comprises:

[0026] the optimization problem of the variational mode decomposition algorithm is defined as minimizing the sum of the bandwidths of all modes;

[0027] constraint conditions of the optimization problem are determined based on the time-frequency spectrogram information;

[0028] Under the constraint condition, the normal mixed cardiac impact signal is subjected to variational modal decomposition, and a group of intrinsic modal functions is obtained by solving the to-be-optimized problem.

[0029] In one embodiment, the time-domain waveform information in the intrinsic modal functions is extracted, including:

[0030] Target modal signals close to heartbeat frequency, respiration frequency and IJKL wave are screened out from the group of intrinsic modal functions.

[0031] Time-domain waveform information is extracted from each target modal signal respectively.

[0032] In one embodiment, the time-frequency spectrum information and the time-domain waveform information are fused and analyzed to generate a sleep apnea classification result, including:

[0033] Each time-domain waveform information is input into a one-dimensional convolution network to obtain time-domain features of each time-domain waveform information output by the one-dimensional convolution network.

[0034] Each time-frequency spectrum information is input into a two-dimensional convolution network to obtain time-frequency spectrum features of each time-frequency spectrum graph output by the two-dimensional convolution network.

[0035] Each time-domain feature and each time-frequency spectrum feature are controlled on the same time dimension to obtain a plurality of target time-domain features and a plurality of target time-frequency spectrum features.

[0036] The plurality of target time-domain features and the plurality of target time-frequency spectrum features are stacked to obtain fusion features.

[0037] Based on the fusion features, a sleep apnea classification result is determined.

[0038] In a second aspect, the present application further provides a sleep monitoring system, including:

[0039] A time-frequency spectrum graph generation module is configured to generate a time-frequency spectrum graph of a normal mixed cardiac impact signal.

[0040] A time-frequency spectrum information generation module is configured to screen out a region associated with heart rate in the time-frequency spectrum graph based on a mask matrix to obtain time-frequency spectrum information.

[0041] A modal decomposition module is configured to decompose intrinsic modal functions of the normal mixed cardiac impact signal.

[0042] A time-domain waveform information generation module is configured to extract time-domain waveform information in the intrinsic modal functions.

[0043] The sleep apnea analysis module is configured to perform fusion analysis on the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result.

[0044] In a third aspect, the present application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the sleep respiratory state monitoring method according to any one of the above aspects when executing the program.

[0045] In a fourth aspect, the present application further provides a computer-readable storage medium, which comprises a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the sleep respiratory state monitoring method according to any one of the above aspects when executed by a processor.

[0046] The sleep respiratory state monitoring method, the sleep monitoring system, the device, and the medium provided by the present application can ensure that the frequency band directly related to the heart rate fluctuation is highlighted in the time-frequency spectrum graph by generating the time-frequency spectrum graph of the normal mixed ballistocardiogram and screening the key region associated with the heart rate in the time-frequency spectrum graph by using the mask matrix, and high-quality time-frequency spectrum information with little noise can be obtained, in addition, the resolution of the time-frequency spectrum information is limited, and it is difficult to capture the peak amplitude and variability of the heartbeat, and the intrinsic mode function of the normal mixed ballistocardiogram is decomposed, and the time-domain waveform information is extracted therefrom, so that distortion can be avoided, and further, the comprehensive analysis of the sleep apnea event is realized by fusion analysis of the time-frequency spectrum information and the time-domain waveform information, and the accurate sleep apnea classification result is generated, which provides strong technical support for the diagnosis and treatment of sleep disorders. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0048] Figure 1 is a flowchart of the sleep respiratory state monitoring method provided by the present application.

[0049] Figure 2 is a whole flowchart of the sleep respiratory state monitoring method provided by the present application.

[0050] Figure 3 is a DSTFT variation process of the normal mixed ballistocardiogram and a target frequency band graph provided by the present application.

[0051] Figure 4is a schematic diagram of image segmentation training of a ViT model provided by the present application.

[0052] Figure 5 is a schematic diagram of a mask generation process used in model training provided by the present application.

[0053] Figure 6 is a schematic diagram of multiple modalities obtained by variational modal decomposition provided by the present application.

[0054] Figure 7 is a schematic diagram of the overall architecture of a sleep respiratory state monitoring method provided by the present application.

[0055] Figure 8 is a schematic diagram of the structure of a sleep monitoring system provided by the present application.

[0056] Figure 9 is a schematic diagram of the structure of an electronic device provided by the present application. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0058] The terms "first", "second", and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein.

[0059] The embodiments of the present application will be described below in conjunction with Figures 1-9 The sleep respiratory state monitoring method, sleep monitoring system, device and medium provided by the present application are described.

[0060] It should be noted that the sleep respiration state monitoring method provided by the embodiment of the present application is realized based on a sleep monitoring system. The sleep respiration state monitoring method provided by the embodiment of the present application adopts a pressure collection device to monitor a ballistocardiogram (BCG) signal of a chest and abdomen of a human body during sleep. In order to extract more accurate frequency spectrum information, a discrete short-time Fourier transform (DSTFT) is adopted to convert a one-dimensional mixed BCG signal into a high-resolution three-dimensional dynamic frequency spectrum graph. In combination with a heart rate variability (HRV) frequency spectrum and real-time heart rate information, a dynamic time-frequency spectrum mask is designed to highlight a target feature associated with a heart rate, so that a complex dynamic frequency spectrum graph can be processed by using an image segmentation technology such as a ViT (Vision Transformer) and a U-Net, a target region is screened out, and a time-frequency spectrum feature is enhanced. In order to further improve the performance of sleep respiration state monitoring, based on key parameters extracted in a feature enhancement stage, a variational mode decomposition (VDM) technology is applied to extract time domain features such as amplitude, interval, and kurtosis from the mixed BCG signal. The time domain features and the time-frequency spectrum features can comprehensively describe signal features. Multi-modal feature fusion is realized by using a Multi-CNN architecture, which uses a one-dimensional convolutional neural network and a two-dimensional convolutional neural network to process time features and space features respectively. In the processing process, the time-frequency spectrum features and the time domain features are processed in a unified dimension, and are fused in the Multi-CNN. Finally, the model classifies sleep apnea through a full connection layer, fully utilizes and fuses the multi-modal features, and effectively improves the monitoring performance.

[0061] The embodiment of the present application takes a sleep monitoring system as an execution subject to describe the sleep respiration state monitoring method.

[0062] In combination Figure 1 , Figure 1 is a flowchart of the sleep respiration state monitoring method provided by the present application.

[0063] As Figure 1 indicated, the method comprises the following steps:

[0064] Step 101, generating a time-frequency spectrum graph of a normal mixed ballistocardiogram signal;

[0065] Step 102, screening out a region associated with a heart rate in the time-frequency spectrum graph based on a mask matrix to obtain time-frequency spectrum information;

[0066] Step 103, decompose the intrinsic mode function of the normal mixed heart impact signal;

[0067] Step 104, extract time domain waveform information in the intrinsic mode function;

[0068] Step 105, fuse the time-frequency spectrum information and the time domain waveform information for analysis, and generate a sleep apnea classification result.

[0069] The following content can be combined Figure 2 , Figure 2 is the overall flowchart of the sleep respiration state monitoring method provided by the application.

[0070] Specifically, the sleep monitoring system collects physiological signals of the human chest and abdomen through a piezoelectric ceramic sensor, but the physiological signals obtained through the piezoelectric ceramic sensor are actually mixed signals, which mix respiratory, heartbeat, body movement and other physiological signals, so it is necessary to screen out normal mixed BCG signals from the mixed signals.

[0071] Further, the DSTFT algorithm is used to perform discrete short-time Fourier transform on the normal mixed BCG signal to obtain a time-frequency spectrum graph of the normal mixed BCG signal.

[0072] Further, the time-frequency spectrum graph of the normal mixed BCG signal is obtained by the real heart rate and input to the ViT model, and the ViT model is used to perform image blocking, feature encoding and feature decoding, and finally predict the probability value associated with the heart rate of each image block, so that the ViT generates a probability matrix according to the position of the image block according to the probability value associated with the heart rate of all image blocks, and outputs the probability matrix.

[0073] Further, the mask corresponding to each image block is generated by comparing each probability value in the probability matrix with a preset threshold, and the mask matrix is generated according to the mask corresponding to each image block.

[0074] Further, based on the mask matrix, the region associated with the heart rate in the time-frequency spectrum graph is screened out to obtain time-frequency spectrum information, and it can be understood that the mask matrix and the time-frequency spectrum graph are combined to screen out the region associated with the heart rate in the time-frequency spectrum graph, and the region is determined as the time-frequency spectrum information. The time-frequency spectrum information can highlight the target region related to the heart rate and enhance the time-frequency spectrum features in the region.

[0075] In addition, due to the limitation of time and frequency resolution, it is difficult to accurately capture the peak amplitude of each heartbeat and its heart rate variability details by the DSTFT method, in order to solve this problem, the VDM technology is used to decompose the intrinsic mode function (Intrinsic Mode Function, IMF) of the normal mixed BCG signal. The IMF can be considered as a mode decomposed from the signal.

[0076] Further, the time domain waveform information is extracted from the intrinsic mode function IMF, and the time domain characteristics in the original BCG signal are enhanced.

[0077] After the time-frequency spectrum information in the time-frequency spectrum and the time domain waveform information in the intrinsic mode function IMF are extracted, the time-frequency spectrum information and the time domain waveform information are fused and analyzed to generate the sleep apnea classification result.

[0078] Specifically, the time domain waveform information can be processed by a multi-layer one-dimensional convolution network to extract time domain features, and the time-frequency spectrum information can be processed by a multi-layer two-dimensional convolution network to extract time-frequency spectrum features. Further, the time-frequency spectrum features and the time domain features are processed in a unified dimension, and the time-frequency spectrum features and the time domain features in the unified dimension are fused to obtain fusion features. The fusion features are classified and predicted by a multi-layer perceptron (MLP) to obtain the sleep apnea classification result.

[0079] The sleep breathing state monitoring method provided by the application generates a time-frequency spectrum of a normal mixed heart impact signal, and uses a mask matrix to screen a key area in the time-frequency spectrum associated with a heart rate, to ensure that the time-frequency spectrum highlights a frequency band directly related to heart rate fluctuations, to obtain high-quality time-frequency spectrum information with little noise. In addition, considering that the resolution of the time-frequency spectrum information is limited and it is difficult to capture the peak amplitude of the heartbeat and its variability, the intrinsic mode function of the normal mixed heart impact signal is decomposed, and the time domain waveform information is extracted therefrom, which can avoid distortion. Further, through the fusion analysis of the time-frequency spectrum information and the time domain waveform information, comprehensive analysis of the sleep apnea event is realized, and an accurate sleep apnea classification result is generated, which provides strong technical support for the diagnosis and treatment of sleep disorders.

[0080] In some embodiments, the normal mixed heart impact signal is obtained by:

[0081] The physiological signals of the chest and abdomen are collected by a piezoelectric ceramic sensor.

[0082] The original heart impact signal is extracted from the physiological signals.

[0083] The normal signals in the original heart impact signal are screened by detecting the nonlinear characteristics of the original heart impact signal, and the normal mixed heart impact signal is obtained.

[0084] Specifically, the sleep monitoring system uses a hardware design based on a piezoelectric ceramic sensor array, which can collect physiological signals of the human chest and abdomen. The sensor array contains 32 sensors arranged in an 8x4 matrix, with an overall size of 58 cm x 28 cm, covering the adult chest and abdominal area, ensuring effective monitoring of physiological signals such as heartbeat and respiration. Each sensor can be composed of a 3 cm diameter circular piezoelectric element and a brass substrate, mounted on a 3 cm x 5 cm printed circuit board. Through the piezoelectric effect, when a pressure change is applied, the sensor generates a weak current, which is amplified by an analog front-end circuit, then processed by low-pass filtering and analog-to-digital conversion. This design effectively reduces the interference of non-target activities such as body movement on the signal, ensuring high-precision physiological signal capture and providing reliable data for sleep apnea detection.

[0085] The mixed signal obtained by the piezoelectric ceramic sensor usually contains multiple physiological signals such as respiration and heartbeat. In order to accurately analyze these signals, the sleep monitoring system uses a filter to separate the cardiopulmonary signal from the mixed signal, especially the heart rate information related to sleep apnea. A Chebyshev I type finite impulse response (FIR) band-pass filter can be used to separate the mixed signal into a respiratory signal (frequency range 0.1-0.8 Hz) and a BCG signal (frequency range 0.8-15 Hz), as subsequent interference signals need to be filtered out, this can be considered as the original BCG signal.

[0086] Among them, the BCG signal is composed of envelope waveform and IJKL wave, and the IJKL wave is closely related to heart contraction and aortic blood flow. The J peak corresponds to the maximum amplitude of the heartbeat period, and the interval between J peaks directly reflects the heart rate change.

[0087] For non-periodic motion, such as interference signals caused by turning over and body movement, by detecting the nonlinear characteristics of the original BCG signal, such as energy and approximate entropy (ApEn), the interference signal can be distinguished from the normal signal, that is, the normal signal in the original BCG signal can be screened out to obtain the normal mixed ballistocardiogram signal for subsequent processing.

[0088] The embodiment of the present application acquires physiological signals during periodic sleep activities through piezoelectric ceramic sensors, ensuring high-precision physiological signal capture, further extracting the BCG signal with the largest correlation with heart rate from the physiological signal, and screening out the normal signal in the BCG signal by monitoring the nonlinear characteristics of the BCG signal, ensuring that the sample contains accurate BCG signal details, which can be used for subsequent accurate sleep apnea detection.

[0089] In some embodiments, based on step 101, the generating the time-frequency spectrogram of the mixed heart impact signal comprises:

[0090] determining a plurality of different window lengths based on resolution requirements in time domain and frequency domain;

[0091] performing discrete short-time Fourier transform on the normal mixed heart impact signal based on each of the window lengths to obtain time-frequency spectrograms of multiple scales.

[0092] It should be noted that the DSTFT technology is used to extract accurate time-frequency spectrum information from the normal mixed BCG signal, while avoiding the common problems of peak weakening and pseudo-peak in traditional signal processing. HRV is a physiological indicator closely related to sleep apnea. HRV represents the change in the time interval between heartbeats, which reflects the regulation function of the autonomic nervous system. The autonomic nervous system is disturbed during sleep due to sleep apnea events. Specifically, the transient hypoxemia and arousal response caused by sleep apnea events will cause changes in HRV. Therefore, by analyzing the change pattern of HRV, sleep apnea events can be effectively identified.

[0093] Many active health devices identify sleep apnea by analyzing HRV, but this method requires high accuracy of the HRV signal, and bed mat sensor devices face challenges in obtaining heart rate. Because widely used filters can weaken or shift the peaks of the signal, or even introduce pseudo-peak, leading to inaccurate identification of the true signal, which in turn limits the application of bed mat devices in sleep apnea detection. Therefore, in order to avoid this traditional problem, the normal mixed BCG signal is processed by DSTFT to enhance the accuracy of the signal and the effect of feature information extraction.

[0094] Specifically, the DSTFT technology is used to perform multi-resolution DSTFT transform on the normal mixed BCG signal, converting one-dimensional signal into three-dimensional time-frequency spectrogram to capture the frequency characteristics of heart rate changes and other physiological activities.

[0095] The core purpose of DSTFT is to analyze the frequency changes of the signal through short-time windows, so that we can capture the frequency components of the signal at different time points. The formula is as follows:

[0096]

[0097] wherein, is the time-domain discrete normal mixed BCG signal, is the frequency-domain signal after DSTFT, is a time window function, which analyzes the frequency components of the signal by sliding the window in time. The length L of the window directly determines the balance between frequency resolution and time resolution.

[0098] The choice of window length directly affects the representation of the time-frequency spectrum. A longer window helps to improve the frequency resolution, enabling more accurate identification of frequency components, but this leads to a decrease in time resolution, making it difficult to accurately capture the transient changes in the signal. A shorter window, on the other hand, can improve the time resolution, better tracking the rapid changes in the signal, but the frequency resolution will be reduced, as shown in Figure 3 Figure 3 is the DSTFT change process of the normal mixed heart impact signal provided by the present application and the target frequency band diagram, where the shaded part is the spectral range of the target heartbeat activity. This process helps to better capture the signal characteristics related to sleep apnea, especially when there are noise or motion artifacts in the signal. DSTFT can preserve the integrity of the signal, avoiding the peak attenuation or offset problems that may be caused by conventional filtering.

[0099] To deal with signal distortion when the heart rate fluctuates greatly, different window lengths are used, such as 1.3 seconds, 2.6 seconds and 5.1 seconds. Based on each window length, the discrete short-time Fourier transform is performed on the normal mixed BCG signal to obtain time-frequency spectra of multiple scales. According to the time-frequency spectrum of each scale, the corresponding time-frequency spectrum diagram is generated, and multiple scale time-frequency spectrum diagrams can be obtained. For example, a 240x48 multi-resolution image is generated, covering a frequency range of 0.4Hz to 2.8Hz.

[0100] In the process of discrete short-time Fourier transform, the signal is usually divided into multiple overlapping or non-overlapping short-time windows. A window function is applied to each window, and a discrete Fourier transform is performed on each signal segment processed by the window function to obtain the frequency spectrum of the window. The discrete Fourier transform results of each window are further combined to form a time-frequency spectrum. The time-frequency spectrum is a three-dimensional data set, where two dimensions represent time and frequency, and the third dimension represents the signal intensity or energy at that time point and frequency. Finally, the time-frequency spectrum data is visualized to generate a time-frequency spectrum diagram. This process usually involves mapping the intensity values in the time-frequency spectrum to colors to display the relationship between time and frequency on a two-dimensional plane.

[0101] The embodiments of the present application filter out the normal mixed heart impact signal and apply discrete short-time Fourier transform with multiple different window lengths to achieve multi-scale time-frequency analysis of the heart impact signal, thereby generating time-frequency spectrum diagrams of multiple scales. This not only improves the time-frequency resolution of signal analysis, but also enhances the recognition of signal characteristics, enabling a more comprehensive and detailed understanding of the dynamic characteristics of the heart impact signal, providing powerful feature information for the detection of sleep apnea.

[0102] ​In some embodiments, when the region associated with the heart rate in the time-frequency spectrum is screened based on the mask matrix to obtain the time-frequency spectrum information, the following steps are performed for each time-frequency spectrum:

[0103] inputting the time-frequency spectrum into a ViT model to obtain a probability matrix output by the ViT model; each probability value in the probability matrix is used to reflect the confidence degree of the image block corresponding to the position of the probability value in the time-frequency spectrum being associated with the heart rate;

[0104] determining the mask of the image block corresponding to the probability value greater than or equal to the preset threshold in the probability matrix as 1;

[0105] determining the mask of the image block corresponding to the probability value less than the preset threshold in the probability matrix as 0;

[0106] generating a mask matrix according to the masks of the image blocks in the time-frequency spectrum;

[0107] combining the mask matrix and the time-frequency spectrum to generate a region associated with the heart rate;

[0108] determining the region associated with the heart rate as the time-frequency spectrum information.

[0109] It should be noted that the generated time-frequency spectrum can cover a frequency range of 0.4 Hz to 2.8 Hz, although this frequency band can capture heart activity, however, the frequency variation range of heart rate is wide, if all frequency bands are directly processed, it is easy to be affected by spectrum leakage and sidelobe effect. Therefore, by combining HRV information, HRV is used as an important basis for generating a time-frequency spectrum mask to ensure that the time-frequency spectrum highlights the frequency band directly related to heart rate fluctuations, effectively separates the target frequency band from the interference frequency band, and thus improves the detection accuracy of sleep apnea events.

[0110] In the mask generation process, a ViT model is selected for time-frequency spectrum segmentation, and then the segmented time-frequency spectrum is block predicted to predict whether it is associated with the heart rate, and the mask is further generated according to the prediction result.

[0111] The ViT model is a deep learning model for image recognition, and the ability to segment the time-frequency spectrogram is based on its advantages in global information capture and multi-scale feature processing. Traditional convolutional neural networks usually rely on local perception fields, which can cause details to be lost or unable to effectively capture global features in complex time-frequency signal processing. The ViT can more effectively integrate the global information of the frequency domain and the time domain through the self-attention mechanism, especially in the feature extraction process related to HRV, ensuring the preservation of signal details and the integration of overall spectral information. Therefore, the DSTFT time-frequency spectrogram of multiple resolutions can be processed by new image segmentation training methods such as ViT to accurately capture the features of heart activity frequencies.

[0112] Before applying the ViT model, it needs to be trained for image segmentation. The training data and its label data can be combined Figure 4 , Figure 4 is a schematic diagram of the image segmentation training of the ViT model provided by the present application. The training process of the ViT model needs to obtain training data and its label data. The training data is a time-frequency spectrogram sample, and the label data is a mask data generated by the true heart rate.

[0113] In combination with Figure 5 Figure 5 is a schematic diagram of the mask generation process used in the model training. First, the frequency spectrogram sample of the training data is obtained as the model training data, and the mask data corresponding to the true heart rate is obtained as the label. The specific training mask generation process is based on the average heart rate of 3 seconds, rather than the instantaneous heart rate, to determine the center of the target frequency band. Therefore, the true heart rate of the time-frequency spectrogram sample is determined, the true heart rate is taken as the center frequency, and the range of the center frequency ± the preset frequency is further taken as the main frequency band of the mask. The preset frequency can be 0.2 Hz, so the range of the center frequency ± 0.2 Hz is taken as the main frequency band of the mask. This frequency band can cover the changes of normal heart rate fluctuations (usually about ± 0.1 Hz), and thus the mask data of the training time-frequency spectrogram sample can be generated. In the time-frequency spectrogram sample, the mask significantly highlights the target frequency region and suppresses other frequency bands related to non-cardiac activity, thereby ensuring more accurate analysis of the frequency spectrum. Finally, the generated mask can maintain the stability of the frequency spectrum even in the case of large heart rate fluctuations, ensuring the accuracy of the detection of sleep apnea events. Further, the ViT model is trained based on the training data and its label mask data.

[0114] In the image segmentation training process of the ViT, the ViT model first divides the input time-frequency spectrogram sample P into a plurality of image blocks patch with a size of 3x3. Each patch is flattened into a vector , and position information is added to each patch. Position encoding is used to assign position information to each patch, so that the model can perceive its position in the time-frequency spectrogram sample.

[0115] Each patch is flattened into a vector , the formula is:

[0116] , ;

[0117] where, is the vector representation of each patch, and N represents the number of patches.

[0118] Further, the position encoding is generated by the sine and cosine functions, which provides position information for each patch, and the formula is:

[0119]

[0120]

[0121] where pos represents the position of the patch, i is the dimension index, is the dimension of the embedding. In this way, the ViT model can perceive the position of each patch in the time-frequency spectrum sample.

[0122] After completing the position encoding, the ViT model calculates the information interaction between each patch through the self-attention mechanism. The formula of the self-attention mechanism is as follows:

[0123]

[0124] where Q is the query vector (Query), K is the key vector (Key), V is the value vector (Value), and is the dimension of the vector. Through this mechanism, the relationship between different patches is captured, ensuring that the model can focus on local details while also capturing global features.

[0125] After the self-attention mechanism, the encoder in the ViT model generates the weight of each patch, which determines the importance of each patch in the time-frequency spectrum. Further, the model processes these weights through the Softmax layer to convert them into a probability distribution, representing the probability of each patch belonging to the heart rate target area. The output of this process is a probability matrix, representing whether each patch is related to heart rate.

[0126] After repeating the above training process, the performance of the model is evaluated on the validation set, and indicators such as accuracy, recall, and F1 score are used to measure the effectiveness of the model. After the model reaches the preset performance effect, the training is ended.

[0127] Deploy the trained ViT model to an actual application environment. For the time-frequency spectrum obtained in the actual application process, input it into the ViT model to obtain a probability matrix output by the ViT model, wherein the probability matrix includes a plurality of probability values, and each probability value is used to reflect the confidence degree of the image block corresponding to the position of the probability value in the time-frequency spectrum and associated with the heart rate.

[0128] Compare each probability value in the probability matrix with a preset threshold, and according to the comparison result, determine the mask of the image block corresponding to the probability value greater than or equal to the preset threshold in the probability matrix as 1, and determine the mask of the image block corresponding to the probability value less than the preset threshold in the probability matrix as 0, wherein the preset threshold is set according to the actual situation.

[0129] Generate a mask matrix according to the mask of each image block in the time-frequency spectrum.

[0130] Combine the mask matrix and the time-frequency spectrum to generate a region associated with the heart rate, and determine the region as the time-frequency spectrum information.

[0131] Wherein, the combination of the mask matrix and the time-frequency spectrum can be calculated by direct multiplication, and the non-zero elements (usually 1) in the mask matrix will retain the corresponding values in the time-frequency spectrum, and the zero elements will make the corresponding values in the time-frequency spectrum become zero.

[0132] The embodiment of the application can effectively utilize the high-dimensional feature extraction capability of the ViT model, filter out the region closely related to the heart rate from the time-frequency spectrum, and generate an accurate mask matrix, and then accurately extract the time-frequency spectrum information corresponding to the heart rate signal through the combination of the mask matrix and the time-frequency spectrum, thereby providing more fine and reliable data support for the detection of sleep apnea events.

[0133] In some embodiments, based on step 103, the decomposing the intrinsic mode function of the normal mixed heart impact signal comprises:

[0134] Define the optimization problem of the variational mode decomposition algorithm as minimizing the sum of the bandwidths of all modes;

[0135] Based on each time-frequency spectrum information, determine the constraint condition of the optimization problem;

[0136] Under the constraint condition, the normal mixed heart impact signal is subjected to variational mode decomposition, and a group of intrinsic mode functions is obtained by solving the optimization problem.

[0137] It should be noted that the DSTFT method, due to limitations in time and frequency resolution, struggles to accurately capture the peak amplitude and heart rate variability of each heartbeat. To address this issue, VDM (Variable Mode Decomposition) technology is introduced. This is a nonlinear adaptive signal decomposition method capable of accurately extracting target mode components from the original signal without distortion. VMD not only effectively distinguishes the components of the target signal but also provides a clearer periodic signal for subsequent time-domain feature extraction.

[0138] The purpose of VDM is to convert a given signal Decomposed into several band-limited modes Each mode has its center frequency. The goal of this decomposition is to minimize the bandwidth of these modes while ensuring that their summation approximates the original signal as closely as possible. .

[0139] Specifically, the optimization problem of the Variational Mode Decomposition (VMD) algorithm is defined as minimizing the sum of the bandwidths of all modes. Furthermore, based on the time-spectral information obtained from the previous segmentation task, constraints are determined in the process of solving the optimization problem, such as the target frequency and decomposition limitations. To achieve this optimization objective, VMD employs the Lagrange multiplier method to transform the constrained optimization problem into an unconstrained one.

[0140] Its optimization objective can be expressed as:

[0141]

[0142]

[0143] Let be the Dirac Delta function, whose value at t=0 approaches infinity and whose integral value is 1. This formula describes how to obtain the modal components... and center frequency Optimization is performed to decompose the signal and retain its main information.

[0144] Under constraints, variational mode decomposition (IMF) is performed on a normal mixed BCG signal. By solving the optimization problem, a set of intrinsic mode functions (IMFs) is obtained, each representing a mode decomposed from the signal. In one embodiment, based on the distribution of the center frequency, the original signal is decomposed into five IMFs, such as... Figure 6 As shown, Figure 6 This is a schematic diagram of multiple modes obtained by variational mode decomposition provided by the present invention.

[0145] The embodiment of the present application defines an optimization problem of a variational mode decomposition algorithm, accurately performs variational mode decomposition on normal mixed ballistocardiogram signals under the constraint condition of determined time-frequency spectrum information, successfully extracts intrinsic mode functions reflecting heart rate characteristics, realizes efficient decomposition of the ballistocardiogram signals, improves the accuracy of signal processing, and provides more fine and reliable data support for detection of sleep apnea events.

[0146] In some embodiments, based on step 104, the extracting time-domain waveform information in the intrinsic mode functions comprises:

[0147] Filtering a plurality of target mode signals close to heartbeat frequency, breathing frequency and IJKL wave from the set of intrinsic mode functions;

[0148] Extracting time-domain waveform information from each of the target mode signals, respectively.

[0149] Specifically, a plurality of target mode signals close to heartbeat, breathing and IJKL wave frequency are filtered from a set of intrinsic mode functions IMF, the target mode signals can clearly observe the periodicity of heartbeat, breathing and pulse, the amplitude change of each J peak can be extracted, and rich time-domain waveform information such as I-A interval, HRV, peak-valley value difference and J peak value can be extracted from each target mode signal, respectively.

[0150] The embodiment of the present application effectively extracts time-domain waveform information by filtering a plurality of target mode signals close to heartbeat, breathing and IJKL wave frequency and extracting time-domain waveform information therefrom, can avoid distortion of the original signal, and provides more fine and reliable data support for detection of sleep apnea events.

[0151] In some embodiments, based on step 105, the fusion analysis of the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result comprises:

[0152] Inputting each time-domain waveform information into a one-dimensional convolution network to obtain time-domain features of each time-domain waveform information output by the one-dimensional convolution network;

[0153] Inputting each time-frequency spectrum information into a two-dimensional convolution network to obtain time-frequency spectrum features of each time-frequency spectrum graph output by the two-dimensional convolution network;

[0154] Controlling each time-domain feature and each time-frequency spectrum feature on the same time dimension to obtain a plurality of target time-domain features and a plurality of target time-frequency spectrum features;

[0155] Stacking the plurality of target time-domain features and the plurality of target time-frequency spectrum features to obtain fusion features;

[0156] determine a sleep apnea classification result based on the fused features.

[0157] It should be noted that since the time-frequency spectrum information and the time-domain waveform information come from different modalities, an effective fusion strategy is needed. Therefore, a multi-modal convolutional neural network (Multi-CNN) is used to efficiently fuse and process the time-frequency spectrum information and the time-domain waveform information.

[0158] Specifically, on the one hand, the time-domain features are derived from the extraction of time-domain waveform information such as heartbeat intervals and respiratory peak values. These one-dimensional time-domain information is processed by a one-dimensional convolutional network (Conv1D) to extract relevant time-domain patterns.

[0159] The time-domain waveform information is input into a one-dimensional convolutional network for processing, and the calculation formula is:

[0160]

[0161] wherein, is the i-th input time-domain waveform information, is the parameter of one-dimensional convolution, is the time-domain feature extracted from the i-th time-domain waveform information.

[0162] On the other hand, the time-frequency spectrum features are derived from the extraction of time-frequency spectrum graphs after the normal mixed BCG signal is subjected to DSTFT. In order to capture the frequency and time information in the signal, local features are extracted by a two-dimensional convolutional network (Conv2D), and important time-frequency features are gradually extracted by convolution operation.

[0163] The time-frequency spectrum information is input into a two-dimensional convolutional network for processing, and the calculation formula is:

[0164]

[0165] wherein, is the j-th input time-frequency spectrum information, j = 1, 2, 3,..., n, corresponding to time-frequency spectrum information of different window sizes, is the parameter of two-dimensional convolution, is the time-frequency spectrum feature extracted from the j-th time-frequency spectrum information.

[0166] The features processed by the Multi-CNN need to be controlled in the same dimension, facilitating the fusion of different modalities, that is, the time domain features and the time frequency spectrum features are controlled in the same time dimension, so that under the alignment of the time dimension, the fusion features can be obtained by stacking the multiple target time domain features and the multiple target time frequency spectrum features in the feature dimension. The multiple target time domain features and the multiple target time frequency spectrum features obtained after the unified time dimension are stacked to obtain the fusion features, that is, the multiple target time domain features are stacked to obtain a time domain feature matrix, and at the same time, the multiple target time frequency spectrum features are stacked to obtain a time frequency spectrum feature matrix, and then the time frequency spectrum feature matrix and the time domain feature matrix are stacked to obtain the fusion features, as shown in the following formula:

[0167] .

[0168] Based on the fusion features, the full connection layer and the Softmax layer are responsible for the final feature fusion and classification, and output the classification results of sleep apnea:

[0169]

[0170] wherein, is the parameter of the full connection layer, G is the classification result, and S is the fusion feature.

[0171] In addition, in order to verify the effectiveness of the classification model training, the Grad-CAM heat map visualization technology can be used. Grad-CAM can generate a heat map by backtracking the gradient of the model, showing the feature area that the model pays most attention to when making a decision. Through the visualization result of Grad-CAM, it can be seen that the model accurately captures the signals related to heartbeat and breathing when processing the time frequency spectrum features, verifying the effectiveness and reliability of the model.

[0172] The embodiment of the present application inputs the time domain waveform information and the time frequency spectrum information into one-dimensional and two-dimensional convolution networks respectively, extracts time domain and time frequency spectrum features, and stacks these features after aligning in the time dimension, finally generates fusion features, realizes in-depth analysis and high-precision classification of sleep apnea phenomenon, improves the performance of the sleep monitoring system, and provides strong technical support for the diagnosis and treatment of sleep disorders.

[0173] All the above processes can be combined Figure 7 , Figure 7 is the overall architecture diagram of the sleep breathing state monitoring method provided by the present application.

[0174] The sleep monitoring system provided by the present application is described below, and the sleep monitoring system described below can be correspondingly referred to the sleep breathing state monitoring method described above.

[0175] Referring to Figure 8 , Figure 8 Figure 1 is a structural schematic diagram of a sleep monitoring system provided by the application.

[0176] The sleep monitoring system comprises:

[0177] A time-frequency spectrum graph generation module 810 is configured to generate a time-frequency spectrum graph of a normal mixed heart impact signal.

[0178] A time-frequency spectrum information generation module 820 is configured to filter out a region associated with a heart rate in the time-frequency spectrum graph based on a mask matrix to obtain time-frequency spectrum information.

[0179] A modal decomposition module 830 is configured to decompose an intrinsic modal function of the normal mixed heart impact signal.

[0180] A time-domain waveform information generation module 840 is configured to extract time-domain waveform information in the intrinsic modal function.

[0181] A sleep apnea analysis module 850 is configured to perform fusion analysis on the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result.

[0182] The sleep monitoring system provided by the application generates a time-frequency spectrum graph of a normal mixed heart impact signal, and filters out a key region associated with a heart rate in the time-frequency spectrum graph by using a mask matrix, so as to ensure that a frequency band directly related to heart rate fluctuation is highlighted in the time-frequency spectrum graph, and high-quality time-frequency spectrum information with little noise is obtained. In addition, considering that the resolution of the time-frequency spectrum information is limited and it is difficult to capture the peak amplitude of a heartbeat and its variability, the intrinsic modal function of the normal mixed heart impact signal is decomposed, and time-domain waveform information is extracted therefrom, so that distortion can be avoided. Furthermore, through fusion analysis of the time-frequency spectrum information and the time-domain waveform information, comprehensive analysis of a sleep apnea event is realized, and an accurate sleep apnea classification result is generated, thereby providing strong technical support for diagnosis and treatment of sleep disorders.

[0183] Further, the sleep monitoring system is further configured to:

[0184] acquire physiological signals of a chest and abdomen by using a piezoelectric ceramic sensor;

[0185] extract an original heart impact signal from the physiological signals;

[0186] filter out normal signals in the original heart impact signal by detecting nonlinear characteristics of the original heart impact signal to obtain the normal mixed heart impact signal.

[0187] Further, the time-frequency spectrum graph generation module 810 is further configured to:

[0188] determine a plurality of different window lengths based on resolution requirements in time domain and frequency domain;

[0189] perform discrete short-time Fourier transform on the normal mixed ballistocardiogram based on each of the window lengths to obtain a plurality of scale time-frequency spectrograms.

[0190] Further, the time-frequency spectrum information generation module 820 is further configured to:

[0191] input the time-frequency spectrograms into a ViT model to obtain a probability matrix output by the ViT model; each probability value in the probability matrix is used to reflect a confidence degree of an image block corresponding to the probability value in the time-frequency spectrogram being associated with a heart rate;

[0192] determine a mask of the image block corresponding to the probability value greater than or equal to a preset threshold in the probability matrix as 1;

[0193] determine a mask of the image block corresponding to the probability value less than the preset threshold in the probability matrix as 0;

[0194] generate a mask matrix according to the masks of the image blocks in the time-frequency spectrograms;

[0195] combine the mask matrix and the time-frequency spectrograms to generate a region associated with a heart rate;

[0196] determine the region associated with the heart rate as time-frequency spectrum information.

[0197] Further, the modal decomposition module 830 is further configured to:

[0198] define an optimization problem of a variational modal decomposition algorithm as minimizing a sum of bandwidths of all modes;

[0199] determine a constraint condition of the optimization problem based on each time-frequency spectrum information;

[0200] perform variational modal decomposition on the normal mixed ballistocardiogram under the constraint condition, and obtain a group of intrinsic modal functions by solving the optimization problem.

[0201] Further, the time-domain waveform information generation module 840 is further configured to:

[0202] select a plurality of target modal signals close to a heartbeat frequency, a breathing frequency, and an IJKL wave from the group of intrinsic modal functions;

[0203] extract time-domain waveform information from each of the target modal signals, respectively.

[0204] Further, the sleep apnea analysis module 850 is further configured to:

[0205] input each time domain waveform information into a one-dimensional convolution network to obtain time domain features of each time domain waveform information output by the one-dimensional convolution network;

[0206] input each time-frequency spectrum information into a two-dimensional convolution network to obtain time-frequency spectrum features of each time-frequency spectrum output by the two-dimensional convolution network;

[0207] control each time domain feature and each time-frequency spectrum feature on the same time dimension to obtain a plurality of target time domain features and a plurality of target time-frequency spectrum features;

[0208] stack the plurality of target time domain features and the plurality of target time-frequency spectrum features to obtain fusion features;

[0209] determine a sleep apnea classification result based on the fusion features.

[0210] It should be noted that the sleep monitoring system provided by the present application can execute the sleep respiratory state monitoring method described in any of the above embodiments during specific operation, and this embodiment will not be described here.

[0211] Figure 9 is a structural schematic diagram of an electronic device provided by the present application, as shown in Figure 9 the electronic device can include a processor 910, a communication interface 920, a memory 930 and a communication bus 940, wherein the processor 910, the communication interface 920 and the memory 930 complete mutual communication through the communication bus 940. The processor 910 can call the logical instructions in the memory 930 to execute the sleep respiratory state monitoring method, which includes: generating a time-frequency spectrum of a normal mixed ballistocardiogram; based on a mask matrix, screening out a region associated with heart rate in the time-frequency spectrum to obtain time-frequency spectrum information; decomposing an intrinsic mode function of the normal mixed ballistocardiogram; extracting time domain waveform information in the intrinsic mode function; and performing fusion analysis on the time-frequency spectrum information and the time domain waveform information to generate a sleep apnea classification result.

[0212] In addition, the logic instructions in the memory 930 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0213] In another aspect, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, when the program instructions are executed by a computer, the computer can execute the sleep respiratory state monitoring method provided by the above-mentioned embodiments, and the method comprises: generating a time-frequency spectrum of a normal mixed ballistocardiogram; screening out a region associated with a heart rate in the time-frequency spectrum based on a mask matrix to obtain time-frequency spectrum information; decomposing an intrinsic mode function of the normal mixed ballistocardiogram; extracting time-domain waveform information in the intrinsic mode function; and performing fusion analysis on the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result.

[0214] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the sleep respiratory state monitoring method provided by the above-mentioned embodiments, and the method comprises: generating a time-frequency spectrum of a normal mixed ballistocardiogram; screening out a region associated with a heart rate in the time-frequency spectrum based on a mask matrix to obtain time-frequency spectrum information; decomposing an intrinsic mode function of the normal mixed ballistocardiogram; extracting time-domain waveform information in the intrinsic mode function; and performing fusion analysis on the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result.

[0215] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement it without creative labor.

[0216] Those skilled in the art can clearly understand the implementation of the various embodiments by means of software and necessary general hardware platforms through the above description of the embodiments, and of course, the embodiments can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in the various embodiments or some parts of the embodiments.

[0217] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A sleep respiratory status monitoring method, characterized by, The application is applied to a sleep monitoring system, comprising: generating a time-frequency spectrum of a normal mixed ballistocardiogram; screening a region associated with a heart rate in the time-frequency spectrum based on a mask matrix to obtain time-frequency spectrum information; decomposing an intrinsic mode function of the normal mixed ballistocardiogram; extracting time-domain waveform information in the intrinsic mode function; fusing and analyzing the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result; when screening a region associated with a heart rate in the time-frequency spectrum based on a mask matrix to obtain time-frequency spectrum information, the following steps are performed: inputting the time-frequency spectrum into a ViT model to obtain a probability matrix output by the ViT model; each probability value in the probability matrix is used to reflect a confidence degree of an image block corresponding to the probability value in the time-frequency spectrum being associated with a heart rate; determining a mask of the image block corresponding to the probability value greater than or equal to a preset threshold as 1; determining a mask of the image block corresponding to the probability value less than the preset threshold as 0; generating a mask matrix according to the mask of each image block in the time-frequency spectrum; combining the mask matrix and the time-frequency spectrum to generate a region associated with a heart rate; determining the region associated with a heart rate as time-frequency spectrum information.

2. The sleep disordered breathing monitoring method of claim 1, wherein, The normal mixed ballistocardiogram is obtained by the following way: collecting physiological signals of a chest and abdomen through a piezoelectric ceramic sensor; extracting an original ballistocardiogram from the physiological signals; screening a normal signal in the original ballistocardiogram by detecting a nonlinear feature of the original ballistocardiogram to obtain the normal mixed ballistocardiogram.

3. The sleep disordered breathing monitoring method of claim 1, wherein, The time-frequency spectrum of the normal mixed ballistocardiogram comprises: determining a plurality of different window lengths based on resolution requirements in a time domain and a frequency domain; performing discrete short-time Fourier transform on the normal mixed ballistocardiogram based on each window length to obtain time-frequency spectrum graphs of multiple scales.

4. The sleep disordered breathing monitoring method of claim 3, wherein, The intrinsic mode function of the normal mixed ballistocardiogram comprises: defining an optimization problem of a variational mode decomposition algorithm as minimizing a sum of bandwidths of all modes; determining a constraint condition of the optimization problem based on each time-frequency spectrum information; performing variational mode decomposition on the normal mixed ballistocardiogram under the constraint condition to obtain a group of intrinsic mode functions by solving the optimization problem.

5. The sleep disordered breathing monitoring method of claim 4, wherein, The time-domain waveform information in the intrinsic mode function comprises: screening a plurality of target mode signals close to a heartbeat frequency, a breathing frequency and an IJKL wave from the group of intrinsic mode functions; respectively extracting time-domain waveform information from each target mode signal.

6. The sleep disordered breathing monitoring method of claim 5, wherein, The fusion and analysis of the time-frequency spectrum information and the time-domain waveform information to generate a sleep apnea classification result comprises: inputting each time-domain waveform information into a one-dimensional convolution network to obtain time-domain features of each time-domain waveform information output by the one-dimensional convolution network; inputting each time-frequency spectrum information into a two-dimensional convolution network to obtain time-frequency spectrum features of each time-frequency spectrum graph output by the two-dimensional convolution network; The time domain features and the time-frequency spectrum features are controlled on the same time dimension to obtain target time domain features and target time-frequency spectrum features; The target time domain features and the target time-frequency spectrum features are stacked to obtain fusion features; Based on the fusion features, a sleep apnea classification result is determined.

7. A sleep monitoring system characterized by, The method comprises the following steps: The time-frequency spectrum graph generation module is configured to generate a time-frequency spectrum graph of a normal mixed ballistocardiogram; The time-frequency spectrum information generation module is configured to filter out a region associated with a heart rate in the time-frequency spectrum graph based on a mask matrix to obtain time-frequency spectrum information; The modal decomposition module is configured to decompose an intrinsic modal function of the normal mixed ballistocardiogram; The time domain waveform information generation module is configured to extract time domain waveform information in the intrinsic modal function; The sleep apnea analysis module is configured to perform fusion analysis on the time-frequency spectrum information and the time domain waveform information to generate a sleep apnea classification result; When the time-frequency spectrum information is obtained by filtering out a region associated with a heart rate in the time-frequency spectrum graph based on a mask matrix, the following steps are performed: The time-frequency spectrum graph is input into a ViT model to obtain a probability matrix output by the ViT model; each probability value in the probability matrix reflects a confidence degree of an image block corresponding to the probability value in the time-frequency spectrum graph being associated with a heart rate; The mask of the image block corresponding to the probability value greater than or equal to a preset threshold in the probability matrix is determined as 1; The mask of the image block corresponding to the probability value less than the preset threshold in the probability matrix is determined as 0; The mask matrix is generated according to the masks of the image blocks in the time-frequency spectrum graph; The mask matrix and the time-frequency spectrum graph are combined to generate a region associated with a heart rate; The region associated with a heart rate is determined as time-frequency spectrum information.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor executes the computer program to implement the steps of the sleep respiration state monitoring method according to any one of claims 1 to 6.

9. A computer-readable storage medium, the computer-readable storage medium comprising a non-transitory computer-readable storage medium, having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the sleep respiration state monitoring method according to any one of claims 1 to 6.

Images