Wireless sensing method for quantitative evaluation of obstruction degree of upper airway

Abstract

The invention discloses a wireless sensing method for upper airway obstruction degree quantitative evaluation, and belongs to the technical field of biomedical signal processing. Comprising the following steps that thoracic and abdominal movement and breath sound signals are collected, physiological signals and labels are obtained, and the thoracic and abdominal movement and breath sound signals are processed; performing feature extraction on the processed thoracoabdominal movement and breath sound signals through a recurrent neural network; and performing iterative training on the hypergraph neural network to obtain a trained breathing parameter quantification model, fusing the extracted features and inputting the fused features into the breathing parameter quantification model, and outputting a ventilation driving and tidal volume prediction result by operating the breathing parameter quantification model so as to obtain an upper airway obstruction degree result. Compared with the prior art, the upper airway obstruction degree quantification method based on bimodal signal fusion of the hypergraph has the advantages that deep learning features are extracted in combination with the LSTM and the self-attention mechanism, and the integrity and accuracy of an evaluation result are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and more specifically to a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction. Background Art

[0002] The detection process of quantitative assessment of the degree of upper airway obstruction aims to accurately capture and quantify the degree of upper airway collapse during sleep through scientific methods, so as to provide objective basis for the research and monitoring of obstructive sleep apnea (OSA). Obstructive sleep apnea (OSA) is a common sleep disorder characterized by repeated complete or partial collapse of the upper airway during sleep. Complete collapse of the upper airway leads to apnea, and partial collapse of the upper airway triggers hypopnea events. The total number of these events per hour, the apnea-hypopnea index (AHI), is widely used to assess the severity of OSA.

[0003] However, existing measurement methods such as apnea-hypopnea index (AHI) and esophageal manometry (EPM) have obvious limitations. With the deepening of research, the limitations of AHI as an evaluation standard have gradually emerged. It cannot fully reflect the true status of OSA between different individuals. Even if the AHI values ​​are similar, the severity of respiratory obstruction and health impact of individuals may be very different. Therefore, AHI cannot fully and accurately reflect the evaluation results of OSA. In order to more accurately assess the severity of upper airway obstruction, esophageal manometry (EPM) came into being. It quantifies the degree of upper airway obstruction by measuring the pressure changes in the esophagus. However, although EPM is regarded as an effective method for assessing the severity of obstruction, the invasive nature of EPM requires the subject to wear an esophageal catheter during sleep during the detection and evaluation process. This will not only cause discomfort to the subject, but also affect the accuracy of the evaluation results obtained by measuring the sleep quality of the subject.

[0004] At present, most of the non-invasive alternatives developed in view of the limitations of EPM are complex and lack sufficient verification, and most detection and evaluation methods still use contact measurement, and the accuracy and reliability of these methods are questionable. The literature "Mann DL, Terrill PI, Azarbarzin A, et al. Quantifying the magnitude of pharyngeal obstruction during sleep using airflow shape [J]. European Respiratory Journal, 2019, 54 (1)" proposed an evaluation method that non-invasively quantifies the severity of upper airway obstruction by analyzing the temporal shape characteristics of individual airflow in nighttime sleep studies, combining the assessment of airflow and ventilatory drive (measured by calibrated intraesophageal diaphragmatic electromyography), and using multivariate regression analysis and airflow shape characteristics (such as inspiration / expiration duration, flatness, and scooping degree) to estimate the ratio of respiratory flow to ventilatory drive, thereby evaluating the severity of upper airway obstruction. Although this method can theoretically realize the automation and non-invasive quantification of the degree of upper airway obstruction, it still relies on contact equipment (such as oral and nasal masks and intraesophageal diaphragmatic electromyography electrodes) in practical applications, which limits its non-invasive advantage in the actual detection process. Therefore, despite various attempts and improvements in the prior art, there is still a lack of a truly effective, simple, non-contact method that can accurately assess the degree of upper airway obstruction, so as to achieve accurate measurement and assessment of the degree of upper airway obstruction without disturbing normal sleep. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In response to the problem in the prior art of how to improve the accuracy of assessment results in the process of realizing non-contact upper airway obstruction degree assessment, the present invention provides a wireless sensing method for quantitative assessment of upper airway obstruction degree, which can realize the assessment of upper airway obstruction degree by wireless sensing, a method for quantifying upper airway obstruction degree based on bimodal signal fusion of hypergraph, and combining LSTM with self-attention mechanism to extract deep learning features, which can effectively integrate signal data from different modalities, thereby improving the completeness and accuracy of the assessment results.

[0007] 2. Technical solution

[0008] The purpose of the present invention is achieved through the following technical solutions.

[0009] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0010] Some embodiments of the present application propose a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction to solve the technical problems mentioned in the above background technology section.

[0011] As the first aspect of the present application, some embodiments of the present application provide a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction, including the following steps: collecting chest and abdominal movement and respiratory sound signals, and obtaining physiological signals and labels to process the chest and abdominal movement and respiratory sound signals; extracting features of the processed chest and abdominal movement and respiratory sound signals through a recurrent neural network; iteratively training the hypergraph neural network to obtain a trained respiratory parameter quantification model, fusing the extracted features and inputting them into the respiratory parameter quantification model, and outputting ventilation drive and tidal volume prediction results by running the respiratory parameter quantification model to obtain the upper airway obstruction degree results.

[0012] Furthermore, chest and abdominal movement and respiratory sound signals are collected synchronously; by acquiring physiological signals and labels, and dividing the physiological signals into data in time periods according to fixed durations, ventilation drive and tidal volume data are acquired based on the physiological signals and labels in each time period.

[0013] Furthermore, the specific process of obtaining ventilation drive and tidal volume data is as follows: the physiological signal is divided into time periods of fixed duration in units of minutes, and the ventilation drive and tidal volume are calculated based on the airflow signal and label of each time period; the ventilation drive of each time period is fused into ventilation drive time series data, and the tidal volume of each time period is fused into tidal volume time series data to obtain ventilation drive and tidal volume data.

[0014] Furthermore, the ventilation drive and tidal volume data are segmented into fixed respiratory cycles, with each fixed respiratory cycle as a segment, and the start time and end time of each segment are recorded to obtain segmented ventilation drive and tidal volume data; according to the start time and end time of each segment, the processed chest and abdominal movement and respiratory sound signals are segmented in time synchronization, and the length of the chest and abdominal movement and respiratory sound signals in each segment is ensured to be consistent.

[0015] Furthermore, the collected breathing sound signal is subjected to noise reduction processing, including framing the collected breathing sound signal and performing short-time Fourier transform on each frame of the breathing sound signal after framing.

[0016] Furthermore, the expression for framing the breathing sound signal is as follows:

[0017] x k (n) = x(n)·w(n-kM);

[0018] In the formula, x k (n) represents the breathing sound signal of the kth frame after framing, x(n) represents the breathing sound signal, k represents the current frame number, k is a natural number; n is the nth sampling point in the kth frame; w() is the windowing function; M represents the frame shift.

[0019] Furthermore, the expression of short-time Fourier transform of the framed breathing sound signal is as follows:

[0020] X k (f) = F{x k (n)};

[0021] Where, X k (f) represents the frequency domain of the kth frame, f represents the frequency, F{} represents the process of short-time Fourier transform, x k (n) represents the breathing sound signal of the kth frame after framing, and x(n) represents the breathing sound signal.

[0022] Furthermore, for X k (f), the noise power spectrum estimate of the kth frame is expressed as The expression is as follows:

[0023]

[0024] Where α is the smoothing factor, α = 0.9; d represents the type of noise;

[0025] The gain function G(k,f) is used to suppress noise and retain the clarity of the breathing signal. The gain function G(k,f) is expressed as follows:

[0026]

[0027] Where p(u|γ(k,f)) is the probability density function of the posterior signal-to-noise ratio; u represents the distribution of noise power; ξ(k,f) is used to characterize the relationship between the noise estimate and the energy of the current signal; and γ(k,f) represents the ratio of the current signal power to the noise power.

[0028] Furthermore, the relationship between the noise estimate and the energy of the current signal ξ(k,f) is expressed as follows:

[0029]

[0030] The expression of the ratio of the current signal power to the noise power γ(k,f) is as follows:

[0031]

[0032] As the second aspect of the present application, some embodiments of the present application provide a processing system for the wireless sensing method for quantitatively evaluating the degree of upper airway obstruction based on the above-mentioned method, including a processing module: collecting chest and abdominal movement and respiratory sound signals, and obtaining physiological signals and labels to process the chest and abdominal movement and respiratory sound signals; an extraction module: performing feature extraction on the processed chest and abdominal movement and respiratory sound signals through a recurrent neural network; an output module: iteratively training the hypergraph neural network to obtain a trained respiratory parameter quantification model, fusing the extracted features and inputting them into the respiratory parameter quantification model, and outputting the ventilation drive and tidal volume prediction results by running the respiratory parameter quantification model, thereby obtaining the upper airway obstruction degree results.

[0033] 3. Beneficial effects

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] (1) This scheme integrates chest and abdominal movement and respiratory sound signals to form a dual-modal monitoring system. Through its dual-modal characteristics, the system can obtain physiological information of the evaluation object from multiple aspects, realizing information complementarity and comprehensive analysis. Compared with the single-modal monitoring method, the dual-modal system can provide a more comprehensive and accurate physiological state assessment and enhance the ability to analyze complex physiological phenomena. In addition, the integration of dual-modal sensing can also improve the robustness of the data, effectively deal with interference in different environments and conditions, and improve the reliability and accuracy of monitoring.

[0036] (2) This scheme uses a hypergraph-based bimodal signal fusion method for quantifying the degree of upper airway obstruction in OSA, and combines LSTM with self-attention mechanism to extract deep learning features. It can effectively integrate signal data from different modalities, thereby improving the integrity and accuracy of information. The LSTM network can capture long-term dependencies in time series signals, while the self-attention mechanism can dynamically adjust the importance of features, so that the respiratory parameter quantification model pays more attention to features that are critical to identifying the degree of upper airway obstruction. The structure of the hypergraph allows the respiratory parameter quantification model to learn features at a higher level, capture more complex patterns and relationships, and improve the prediction accuracy of ventilation drive and tidal volume, thereby improving the accuracy of the evaluation results. In addition, compared with traditional contact upper airway obstruction degree monitoring technology, this scheme avoids direct contact with the subject through wireless sensing of radio frequency sensors and audio sensors, thereby reducing the risk of infection for the evaluated subject during the detection process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1It is a flow chart of a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction in one embodiment of the present invention;

[0038] Figure 2 Schematic diagram of the structure of a respiratory parameter quantization model based on multimodal fusion in one embodiment of the present invention;

[0039] Figure 3 A schematic diagram of input signals of a respiratory parameter quantification model of chest and abdominal movements based on radio frequency sensing in one embodiment of the present invention;

[0040] Figure 4 A schematic diagram of an input signal of a breathing parameter quantization model of a breathing sound signal based on audio sensing in one embodiment of the present invention;

[0041] Figure 5 A schematic diagram of the output results of the respiratory parameter quantification model and the degree of upper airway obstruction in one embodiment of the present invention;

[0042] Figure 6 It is a schematic diagram of the consistency of the degree of obstruction output by the respiratory parameter quantification model in one embodiment of the present invention;

[0043] Figure 7 FIG. 1 is a Bland-Altman diagram of the output of the respiratory parameter quantification model in one embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction of the present invention includes the following steps: collecting chest and abdominal movement and respiratory sound signals, and obtaining physiological signals and labels to process the chest and abdominal movement and respiratory sound signals; extracting features of the processed chest and abdominal movement and respiratory sound signals through a recurrent neural network; iteratively training the hypergraph neural network to obtain a trained respiratory parameter quantification model, fusing the extracted features and inputting them into the respiratory parameter quantification model, and outputting ventilation drive and tidal volume prediction results by running the respiratory parameter quantification model, thereby obtaining the upper airway obstruction degree result.

[0046] Specifically, the specific process of the wireless sensing method for quantitatively evaluating the degree of upper airway obstruction is as follows:

[0047] S1. Collect and process data

[0048] Acquire and process physiological signals, labels, chest and abdominal movement and respiratory sound signals.

[0049] Specifically, data is collected during the complete sleep of the subject being evaluated. Radio frequency sensors and audio sensors are used to synchronously collect chest and abdominal movements and respiratory sound signals of the subject during sleep; polysomnography (PSG) monitoring is performed to obtain physiological signals and labels; based on the physiological signals and labels obtained by PSG, the ventilation drive and tidal volume are calculated breath by breath.

[0050] The radio frequency sensor is used to capture tiny movements of the chest and abdomen of the subject, while the audio sensor is used to record breathing sound signals.

[0051] More specifically, physiological signals include airflow signals and EEG signals. Based on the airflow signals and labels obtained by PSG, the start and end of each respiratory cycle are identified, and then the ventilation drive and tidal volume are calculated breath by breath. Among them, ventilation drive refers to the airflow generation force required for breathing, and tidal volume refers to the amount of gas inhaled or exhaled per breath. The calculation process of ventilation drive is calculated by fitting a ventilation control model (including the effects of chemicals and arousal substances on ventilation drive) to the ventilation mode of sleep apnea; the calculation process of tidal volume is to calculate the average airflow volume of each breath, that is, tidal volume, according to the amplitude of the airflow signal.

[0052] In this embodiment, the acquired airflow signal is divided into data in time periods according to a fixed time length, and analyzed to ensure the continuity and stability of data analysis. In each time period, the ventilation drive and tidal volume are calculated based on the airflow signal and the label.

[0053] In a specific embodiment, the fixed time division process is to divide every 7 minutes as a time period, calculate the ventilation drive and tidal volume based on the airflow signal of each time period, splice the ventilation drive of each time period, and merge it into a complete ventilation drive time series data for the whole night; splice the tidal volume of each time period, and merge it into a complete tidal volume time series data for the whole night, so as to obtain the ventilation drive and tidal volume data for each breath of the whole night.

[0054] A label refers to a mark associated with the physiological signal of the subject being evaluated and used to indicate a specific physiological state or event. The label is used to identify the sleep breathing disorder of the subject being evaluated, that is, a respiratory event related to upper airway obstruction. In this embodiment, the label includes a sleep stage label, a respiratory event label, and other physiological state labels. The sleep stage label is used to identify sleep stages, such as non-rapid eye movement (NREM) and rapid eye movement (REM). These labels are helpful in analyzing the impact of different sleep stages on the degree of upper airway obstruction; the respiratory event label is used to identify respiratory events, such as apnea and hypopnea; other physiological state labels include other physiological state labels related to breathing or sleep, such as heart rate variability and blood oxygen saturation.

[0055] The labels are obtained through the polysomnography (PSG) monitoring system. PSG is a comprehensive sleep monitoring technology that can simultaneously record multiple physiological signals of the subject and automatically or semi-automatically generate corresponding labels based on these signals. Obtaining labels provides an important reference for subsequent data analysis and processing.

[0056] In a specific embodiment, the process of acquiring physiological signals and labels is as follows: using a PSG monitoring system to synchronously collect multiple physiological signals including airflow signals, electroencephalogram (EEG), eye movement signals (EOG), mandibular electromyography (EMG), and electrocardiogram (ECG) during sleep; based on the PSG monitoring results, the sleep stages (such as non-rapid eye movement NREM and rapid eye movement REM) and respiratory events (such as apnea and hypopnea) are marked as labels for subsequent analysis. The process of collecting chest and abdominal movement and respiratory sound signals is as follows: using a radio frequency sensor (such as RFID) to monitor chest and abdominal movement; at the same time, using an audio sensor (such as a microphone array) to capture respiratory sound signals. During the collection process, ensure that the sensor and the monitoring device work synchronously to obtain signal data with consistent time.

[0057] Specifically, if Figure 2 As shown, the noise reduction process is performed on the collected breathing sound signal to reduce the interference of background noise.

[0058] like Figure 3 As shown, it is a schematic diagram of the input signal of the respiratory parameter quantization model of chest and abdominal movement based on radio frequency sensing in this embodiment, wherein the ordinate represents the normalized chest and abdominal movement amplitude. The input respiratory sound signal is subjected to short-time Fourier transform (STFT) by the OMLSA method to obtain frequency domain information, and the optimal gain function is obtained by calculating the noise power spectrum and the respiratory sound signal power spectrum and their ratio, and the spectrum is scaled according to the gain coefficient to reduce the noise component. After processing, the respiratory sound signal is restored to the time domain by inverse short-time Fourier transform (ISTFT). Combined with Figure 2It can be seen that such a noise reduction process can effectively suppress background noise and improve the clarity of the breathing sound signal.

[0059] In a specific embodiment, first, the collected breathing sound signal x(n) is processed into frames, and short-time Fourier transform (STFT) is performed on each frame of the breathing sound signal.

[0060] First, the expression for framing the breathing sound signal x(n) is as follows:

[0061] x k (n) = x(n)·w(n-kM);

[0062] In the formula, x k (n) represents the breathing sound signal of the kth frame after framing, k represents the current frame number, and k is a natural number; n represents the nth sampling point in the kth frame; w() is the windowing function; and M represents the frame shift.

[0063] Secondly, the expression of short-time Fourier transform of the framed breathing sound signal is as follows:

[0064] X k (f) = F{x k (n)};

[0065] Where, X k (f) represents the frequency domain of the kth frame, f represents the frequency, and F{} represents the process of short-time Fourier transform.

[0066] For X k (f) An adaptive noise estimation algorithm is used to dynamically estimate the noise power spectrum.

[0067] Specifically, the noise power spectrum estimate of the kth frame is expressed as Its expression is as follows:

[0068]

[0069] In the formula, α is a smoothing factor, and in this embodiment, α is set to 0.9; d represents the type of noise, and in this embodiment, d is background noise.

[0070] According to x k (n) and The result value is further used to suppress noise and retain the clarity of the breathing sound signal through the gain function G(k,f). The process expression of the gain function G(k,f) is as follows:

[0071]

[0072] Where p(u|γ(k,f)) is the probability density function of the posterior signal-to-noise ratio; u represents the distribution of noise power;

[0073] ξ(k,f) is used to characterize the relationship between the noise estimate and the energy of the current signal, and its expression is as follows:

[0074]

[0075] γ(k,f) represents the ratio of the current signal power to the noise power, and its expression is as follows:

[0076]

[0077] The gain function is used to enhance the breathing sound spectrum. The expression of the enhancement process is as follows:

[0078] |Y k (f)|=G(k,f)·|X k (f)|;

[0079] Where Y k (f) shows the enhanced frequency domain signal.

[0080] For the enhanced frequency domain signal Y k (f) Perform inverse short-time Fourier transform, overlap and add the obtained time domain frame signals, and reconstruct the de-noised breathing sound signal in the time domain The expression is as follows:

[0081]

[0082] Through the above steps, the processed chest and abdominal movement and breathing sound signals are obtained.

[0083] Subsequently, the obtained breath-by-breath ventilation drive and tidal volume data for the whole night are segmented. Every 10 respiratory cycles are taken as a segment, and the start time and end time of each segment are recorded to obtain the segmented ventilation drive and tidal volume data; according to the recorded start time and end time of each segment, the processed chest and abdominal movement and respiratory sound signals are segmented in time synchronization, and the length of each segment of chest and abdominal movement and respiratory sound signals is ensured to be consistent. Figure 5 As shown, the chest and abdominal movements are synchronized with the respiratory sound signals and have the same duration.

[0084] According to the segmented ventilation drive and tidal volume data, P j is the ventilation drive of the jth segment, V j is the tidal volume data of the jth segment, where j is a natural number. Each segment of ventilation drive contains 10 ventilation drive values, and each segment of tidal volume data contains 10 tidal volume values. j and Vj Start time j and end time End k Respectively expressed as:

[0085] Start k =t 10j+1 ;

[0086] End j =t 10j+10 ;

[0087] Where t represents time. For the start time and end time obtained, in the time range [Start j ,End j ], the chest and abdomen movement and breathing sound signals are intercepted, and the process expressions of intercepting the chest and abdomen movement and breathing sound signals are as follows:

[0088] R j (t)={R(t)t∈[Start j ,End j ]};

[0089] A j (t)={A(t)t∈[Start j ,End j ]};

[0090] Among them, R j (t) is the chest and abdominal movement of the jth segment, A j (t) is the breathing sound signal of the jth segment.

[0091] In a specific embodiment, after obtaining the segmented chest and abdominal motion and the respiratory sound signal after noise reduction through the above process, the segmented chest and abdominal motion and the respiratory sound signal are interpolated or trimmed so that the segmented chest and abdominal motion and the respiratory sound signal are the same length, thereby ensuring consistency between different signals in the time dimension.

[0092] Among them, interpolation is to insert known data points between two known points to estimate unknown data points. In signal processing, interpolation can construct a continuous function or an approximately continuous function from known discrete signal points to estimate the values ​​between or outside the signal points. Through interpolation, the signal can be refined on the time axis so that different signals have the same length at the same time resolution. Clipping is to directly delete some data points from the signal to make the signal reach the required length. Through clipping, the signal can be clipped to the same length as other signals for subsequent analysis or comparison.

[0093] Signal interpolation or trimming of the segmented chest and abdominal motion and respiratory sound signals to the same length is to ensure the consistency and comparability of these signals in subsequent analysis, comparison or fusion.

[0094] Specifically, since an adult's breathing time is 3 to 5 seconds, in this embodiment, the segmented chest and abdomen movement and the breathing sound signal are interpolated or trimmed to a length of 40 seconds. With breathing signal The final expression is:

[0095]

[0096] in, is the chest and abdomen movement after interpolation, Interpolate() represents the interpolation operation, and length() represents the length of the chest and abdomen movement or breathing sound signal without interpolation or trimming. It is the chest and abdomen movement after cutting. is the breathing signal after interpolation. It is the breathing signal after trimming.

[0097] S2. Feature extraction

[0098] The processed chest and abdominal movement and respiratory sound signals are then subjected to feature extraction through a recurrent neural network.

[0099] Specifically, the chest and abdominal motion and breathing sound signals with the same length obtained in step S1 are input into the long short-term memory network (LSTM) combined with the self-attention mechanism to extract the deep features f thorax-abdomen With f sound ;.

[0100] Long short-term memory network (LSTM) is a special type of recurrent neural network (RNN) used to process and predict long-term dependencies in sequence data. LSTM effectively controls the inflow, forgetting and output of information by introducing memory units and gating mechanisms (including input gates, forget gates and output gates), thereby overcoming the common gradient vanishing and explosion problems of traditional RNN in long sequence training. This makes LSTM particularly suitable for processing signals with time order. In this embodiment, LSTM is used to process chest and abdominal movement and respiratory sound signals to extract key features and maintain contextual information.

[0101] The self-attention mechanism dynamically focuses on the relationship between different parts of the sequence when processing sequence data. By calculating the correlation between each element in the input sequence and other elements, the self-attention mechanism can assign different weights to each element, thereby focusing on the most important part for the current task. Combining LSTM and self-attention mechanism can effectively extract deep features in chest and abdominal movement and breathing sound signals. thorax-abdomen With f sound This allows for a more comprehensive understanding of the information in complex signals.

[0102] In a specific embodiment, for segmented chest and abdominal movements and breathing sound signals Normalize to obtain normalized chest and abdominal motion and respiratory sound signals Its expression is:

[0103]

[0104] In the formula, μ R Mean value of chest and abdominal motion, σ R is the standard deviation of chest and abdominal motion, μ A is the mean of the breathing sound signal, σ A is the standard deviation of the respiratory sound signal, and [] is used to represent the normalized results of chest and abdominal movement and respiratory sound signals.

[0105] Will Input into LSTM for processing, Divide into multiple time series samples, define the time step as t 1 , then the input of each time step is x t1 , the calculation process in LSTM is as follows:

[0106] Input Gate:

[0107] Forget Gate:

[0108] Candidate memory cells:

[0109] Update memory unit:

[0110] Output Gate:

[0111] Final output:

[0112] Among them, h represents the hidden state, represents the hidden state at the current time step..., Represents the hidden state of the previous time step, that is, at time step t 1 Output of -1; represents the cell state at the previous time step, that is, at time step t 1 -1 memory information; tanh() is the activation function; Represents the current cell state, combining the information of the forget gate and the input gate; W i , W f , W c , W o is the weight matrix used to linearly transform input data and hidden states; b i 、b f 、b C 、b o All represent bias vectors; i t represents the activation value of the input gate; f t Represents the activation value of the forget gate; o t Represents the activation value of the output gate; σ represents the Sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents the element-wise multiplication operation.

[0113] After processing all time steps, the hidden state output of chest and abdomen movement and breathing sound signal can be obtained. The hidden state output H of chest and abdomen movement is thorax-abdomen And the hidden state output H of the breathing sound signal sound .

[0114] The self-attention mechanism is applied to the hidden state outputs of the obtained chest and abdominal motion and breathing sound signals. First, a linear transformation is used, that is, the weight matrix W is used Q , W K , W V , the hidden state output of chest and abdominal motion and breathing sound signals is mapped to query (Q), key (K) and value (V). The process expressions are as follows:

[0115] Q=(H thorax-abdomen / H sound )W Q ;

[0116] K=(H thorax-abdomen / H sound )W K ;

[0117] V=(H thorax-abdomen / H sound )W V ;

[0118] Secondly, calculate the attention score Attention(Q,K,V):

[0119]

[0120] Among them, d k is the dimension of the key and softmax() is the activation function.

[0121] Thus, we can obtain the deep features f in the chest and abdomen motion after attention calculation thorax-abdomen And the deep features f in the breathing sound signal after attention calculation siund , expressed as:

[0122] f thorax-abdomen =Attention(Q,K,V);

[0123] f sound =Attention(Q,K,V);

[0124] Thus, features extracted from chest and abdominal motion and respiratory sound signals are obtained.

[0125] S3. Calculate and output the results

[0126] The hypergraph neural network is iteratively trained to obtain a trained respiratory parameter quantification model. The extracted features are fused and input into the respiratory parameter quantification model. The respiratory parameter quantification model is run to output the ventilation drive and tidal volume prediction results, thereby obtaining the upper airway obstruction degree results.

[0127] Specifically, the f obtained in step S2 thorax-abdomen With f sound The splicing is performed and the trained respiratory parameter quantification model in the hypergraph neural network is input to obtain the ventilation driving force and tidal volume output by the respiratory parameter quantification model, thereby obtaining the upper airway obstruction degree result.

[0128] Hypergraph Neural Network (HGNN) is a network structure that is extended from traditional graph neural network (GNN) and is designed to process data with hyperedges structure more efficiently. Compared with traditional graphs where each edge only connects two nodes, the degree of the edge in the hypergraph neural network is an arbitrary non-negative integer, and the same hyperedge can connect different vertices to indicate the correlation between the connected vertices. This enables better modeling of complex relationships between data.

[0129] Specifically, the specific process of using hypergraph neural network to predict ventilation driving force and tidal volume is as follows:

[0130] Through the hypergraph neural network, the correlation of bimodal data is used to construct the corresponding hyperedge group H 1 and H 2 , and concatenate the hyperedge groups to obtain the association matrix H.

[0131] The correlation matrix H is a matrix representation describing a graph, which is used to represent the correlation between vertices and hyperedges in a hypergraph neural network, and is expressed as follows:

[0132] H=[H 1 ;H 2 ];

[0133] The association matrix H consists of a set of nodes and a set of hyperedges, where each hyperedge e can connect multiple nodes v. For each node in the hypergraph, it is initialized to the concatenated feature h, expressed as:

[0134]

[0135] Where: v represents a hypergraph vertex; e represents a hyperedge. If a vertex is included in a hyperedge, the corresponding position in the association matrix is ​​recorded as 1; if a vertex is not included in a hyperedge, the corresponding position in the association matrix is ​​recorded as 0.

[0136] Then, the association matrix in the hypergraph and the node features contained in the dataset are fed into the hyperedge convolution layer for convolution operation to learn the implicit features of the high-dimensional data, where the expression of the convolution operation is as follows:

[0137] X (0) =X∈R N×C ;

[0138]

[0139] Where: X (0) represents the input after the fusion of bimodal branch features; X (l+1) represents the feature matrix at the l+1 layer; X (l) represents the feature matrix at layer l; σ is the activation function; D v and D e represents normalization; θ (l) represents the filtering matrix, H represents the association matrix, W represents the diagonal matrix of the weights of each hyperedge, N is the number of nodes, and C is the feature dimension of each node.

[0140] The nodes in the last layer of the hypergraph neural network are used to generate the predicted values ​​for each segment. The predicted results are ventilation drive and tidal volume Prediction results The expression is as follows:

[0141]

[0142] The prediction results are constrained by using the loss function, and then the hypergraph neural network is iteratively trained to obtain a trained respiratory parameter quantization model. The loss function is expressed as follows:

[0143]

[0144] Where: L MSE-p represents the loss function of the ventilation drive prediction task; L MSE-v represents the loss function of the tidal volume prediction task; N represents the number of samples of input and output ventilation drive and tidal volume. In this embodiment, N=10.

[0145] Finally, the chest and abdominal movements and respiratory sound signals to be tested are placed in the trained respiratory parameter quantification model, and the ventilation drive and tidal volume prediction results are finally output to accurately assess the degree of upper airway obstruction of the evaluated subject.

[0146] Calculate the ratio of tidal volume to ventilatory drive in each respiratory cycle using the following expression:

[0147]

[0148] Specifically, the ratio of tidal volume to ventilatory drive in each respiratory cycle obtained by the wireless sensing method for quantitative assessment of upper airway obstruction can be used as a key indicator for assessing the degree of upper airway obstruction in obstructive sleep apnea (OSA). The degree of upper airway obstruction is assessed by judging the ratio of tidal volume to ventilatory drive in each respiratory cycle; if the ratio of tidal volume to ventilatory drive is >90%, the result of the upper airway obstruction is normal; if the ratio of tidal volume to ventilatory drive is 70% to 90%, the result of the upper airway obstruction is mild; if the ratio of tidal volume to ventilatory drive is 50% to 70%, the result of the upper airway obstruction is moderate; if the ratio of tidal volume to ventilatory drive is 30% to 50%, the result of the upper airway obstruction is severe; if the ratio of tidal volume to ventilatory drive is <30%, the result of the upper airway obstruction is extremely severe.

[0149] This solution combines chest and abdominal movement and respiratory sound signals to form dual-modal monitoring. Through its dual-modal characteristics, this solution can obtain physiological information from multiple aspects and achieve information complementarity and comprehensive analysis. Compared with the single-modal monitoring method, this solution uses dual-modality to provide a more comprehensive and accurate physiological state assessment and enhance the ability to analyze complex physiological phenomena; in addition, the integration of dual-modal sensing can also improve the robustness of the data, effectively deal with interference in different environments and conditions, and improve the reliability and accuracy of the assessment results.

[0150] In a specific embodiment, when the ratio of tidal volume to ventilatory drive is >90%, it indicates that the ventilatory drive is sufficient and there is no obvious airway obstruction, so the evaluation result is output as normal; when the ratio of tidal volume to ventilatory drive is 70% to 90%, it indicates that the airway is mildly obstructed, so the evaluation result is output as mild; when the ratio of tidal volume to ventilatory drive is 50% to 70%, it indicates that the airway obstruction is obvious, so the evaluation result is output as moderate; when the ratio of tidal volume to ventilatory drive is 30% to 50%, it indicates that the airway obstruction is severe and breathing is significantly affected, so the evaluation result is output as severe; when the ratio of tidal volume to ventilatory drive is <30%, it indicates that the airway obstruction is extremely severe and there is an extremely high risk of breathing being restricted, so the evaluation result is output as extremely severe.

[0151] In a specific embodiment, Figure 4 As shown, it is a schematic diagram of the input signal of the respiratory parameter quantification model of the respiratory sound signal based on audio sensing in this embodiment, wherein the ordinate is the amplitude of the respiratory sound signal after normalization; it shows the results of the ventilation drive and tidal volume and reflects the degree of upper airway obstruction of the evaluated object in the corresponding time period, that is, the ratio of ventilation drive to tidal volume. Secondly, it can be seen that the evaluated object has experienced a process from upper airway patency to obstruction, and the degree of obstruction is mild. In addition, during the non-obstructive period, the ventilation drive level is high, which indicates that the evaluated object is experiencing obstructive apnea or hypopnea.

[0152] like Figures 3 to 5 As shown, a respiratory parameter quantification model designed according to the wireless sensing method for quantitatively evaluating the degree of upper airway obstruction of the present invention inputs a signal into the respiratory parameter quantification model, and runs the respiratory parameter quantification model to output a result, and the output result of the respiratory parameter quantification model reflects the degree of upper airway obstruction.

[0153] like Figure 6 to Figure 7 The figure shows the consistency of obstruction degree and Bland-Altman results obtained by running the respiratory parameter quantification model. Figure 7 The Bland-Altman diagram output by the respiratory parameter quantification model in one embodiment of the present invention. The Bland-Altman diagram is a consistency analysis diagram, which is a graphical method that intuitively reflects the consistency limit of two sets of measurement results. The basic idea is to evaluate the consistency of the two sets of data by calculating the difference between the two sets of measurement results and their average values, and using images to intuitively display the relationship between these differences and the average values.

[0154] In a specific embodiment, a processing system of a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction includes:

[0155] Processing module: collects chest and abdominal movement and breathing sound signals, and obtains physiological signals and labels to process chest and abdominal movement and breathing sound signals;

[0156] Extraction module: extract features from the processed chest and abdominal motion and respiratory sound signals through a recurrent neural network;

[0157] Output module: Iteratively train the hypergraph neural network to obtain a trained respiratory parameter quantification model, fuse the extracted features and input them into the respiratory parameter quantification model, and output the ventilation drive and tidal volume prediction results by running the respiratory parameter quantification model to obtain the upper airway obstruction degree results.

[0158] This scheme uses a hypergraph-based dual-modal signal fusion upper airway obstruction degree quantification method, combined with LSTM and self-attention mechanism to extract deep learning features, which can effectively integrate signal data from different modalities, thereby improving the integrity and accuracy of information. Secondly, the LSTM network can capture long-term dependencies in time series signals, while the self-attention mechanism can dynamically adjust the importance of features, so that the respiratory parameter quantification model pays more attention to features that are critical to identifying the degree of upper airway obstruction; the structure of the hypergraph allows the respiratory parameter quantification model to learn features at a higher level, which can capture more complex patterns and relationships, improve the prediction accuracy of ventilation drive and tidal volume, and thus improve the accuracy of the evaluation results.

[0159] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited to this, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of this patent. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A wireless sensing method for quantitatively evaluating the degree of upper airway obstruction, comprising the following steps: Collect chest and abdominal motion and respiratory sound signals, and obtain physiological signals and labels to process chest and abdominal motion and respiratory sound signals; The processed chest and abdominal motion and respiratory sound signals are subjected to feature extraction through a recurrent neural network; The hypergraph neural network is iteratively trained to obtain a trained respiratory parameter quantification model. The extracted features are fused and input into the respiratory parameter quantification model. The respiratory parameter quantification model is run to output the ventilation drive and tidal volume prediction results, thereby obtaining the upper airway obstruction degree results.

2. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 1, characterized in that: The chest and abdominal movement and respiratory sound signals are collected synchronously; by obtaining physiological signals and labels, and dividing the physiological signals into data in time periods according to a fixed duration, the ventilation drive and tidal volume data are obtained based on the physiological signals and labels in each time period.

3. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 2, characterized in that: The specific process of obtaining ventilation drive and tidal volume data is as follows: divide the physiological signal into time periods with a fixed duration in units of minutes, and calculate the ventilation drive and tidal volume based on the airflow signal and label of each time period; The ventilation drive of each time period is fused into ventilation drive time series data, and the tidal volume of each time period is fused into tidal volume time series data to obtain ventilation drive and tidal volume data.

4. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 3, characterized in that: The ventilation drive and tidal volume data are segmented into fixed respiratory cycles, with each fixed respiratory cycle as a segment, and the start time and end time of each segment are recorded to obtain the segmented ventilation drive and tidal volume data; The processed chest and abdominal motion and respiratory sound signals are segmented in time synchronization according to the start time and end time of each segment, and the length of the chest and abdominal motion and respiratory sound signals of each segment is ensured to be consistent.

5. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 1, characterized in that: The collected breathing sound signal is subjected to noise reduction processing, including framing the collected breathing sound signal and performing short-time Fourier transform on each frame of the breathing sound signal after framing.

6. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 5, characterized in that: The expression for framing the breathing sound signal is as follows: x k (n)=x(n)·w(n-kM); In the formula, x k (n) represents the breathing sound signal of the kth frame after framing, x(n) represents the breathing sound signal, k represents the current frame number, k is a natural number; n is the nth sampling point in the kth frame; w() is the windowing function; M represents the frame shift.

7. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 5, characterized in that: The expression of short-time Fourier transform of the framed breathing sound signal is as follows: X k (f)=F{x k (n)}; Where, X k (f) represents the frequency domain of the kth frame, f represents the frequency, F{} represents the process of short-time Fourier transform, x k (n) represents the breathing sound signal of the kth frame after framing, and x(n) represents the breathing sound signal.

8. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 7, characterized in that: For X k (f), the noise power spectrum estimate of the kth frame is expressed as The expression is as follows: Where α is the smoothing factor, α = 0.9; d represents the type of noise; The gain function G(k,f) is used to suppress noise and retain the clarity of the breathing signal. The gain function G(k,f) is expressed as follows: Where p(u|γ(k,f)) is the probability density function of the posterior signal-to-noise ratio; u represents the distribution of noise power; ξ(k,f) is used to characterize the relationship between the noise estimate and the energy of the current signal; and γ(k,f) represents the ratio of the current signal power to the noise power.

9. The wireless sensing method for quantitatively evaluating the degree of upper airway obstruction according to claim 8, characterized in that: The relationship between the noise estimate and the energy of the current signal ξ(k,f) is expressed as follows: The expression of the ratio of the current signal power to the noise power γ(k,f) is as follows:

10. A processing system for a wireless sensing method for quantitatively evaluating the degree of upper airway obstruction based on any one of claims 1 to 9, characterized in that: It includes a processing module: collecting chest and abdominal movement and breathing sound signals, and obtaining physiological signals and labels to process the chest and abdominal movement and breathing sound signals; Extraction module: extract features from the processed chest and abdominal motion and respiratory sound signals through a recurrent neural network; Output module: Iteratively train the hypergraph neural network to obtain a trained respiratory parameter quantification model, fuse the extracted features and input them into the respiratory parameter quantification model, and output the ventilation drive and tidal volume prediction results by running the respiratory parameter quantification model to obtain the upper airway obstruction degree results.

Images