Automatic recognition system of hypopnea and apnea based on graph convolutional neural network
Through the automatic identification system of hypoventilation and apnea based on graph convolutional neural network, snoring data is converted into graph structures. The classification of convolutional neural networks solves the problem of early screening at home for OSAHS patients, and efficient and convenient snoring monitoring is achieved and AHI index calculation meets the needs of home diagnosis.
Patent Information
- Application Number
- CN202111160071.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-30
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-09-30
AI Technical Summary
The existing technology is difficult to achieve efficient and convenient early screening of OSAHS patients at home. Traditional diagnostic methods such as polysomnography require a high usage environment and high cost, and are complex in operation, which cannot meet the daily monitoring needs of home.
The low ventilation and apnea automatic recognition system based on graph convolutional neural network is adopted to convert snoring data into graph structures through visual methods, and the convolutional neural network is used to realize the classification of normal snoring, hypoventilated snoring and apnea-related snoring. The number of hypoventilation and apnea events during sleep throughout the night was counted, and the AHI index was calculated to assist in diagnosis.
It improves the efficiency of automatic snoring recognition, provides convenient early-home screening tools for OSAHS patients, and can run on various mobile platforms such as mobile phones, computers and tablets, reducing diagnostic costs and meeting the needs of home monitoring.
Smart Images

Figure CN113749622B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of sleep snoring detection, and in particular to an automatic detection system for sleep apnea events and hypopnea events based on a visual graph and a graph convolutional neural network model. Background Art
[0002] Obstructive sleep apnea-hypopnea syndrome (OSAHS) is a potentially dangerous source disease. The main cause of the disease is apnea and hypopnea during sleep caused by upper airway stenosis or obstruction. Among them, apnea refers to the complete cessation of oral and nasal airflow for more than 10 seconds during sleep; hypopnea refers to a decrease of more than 50% in the intensity of respiratory airflow compared with the basal level during sleep, accompanied by micro-awakening or a decrease of more than 4% in blood oxygen saturation compared with the basal level. Apnea or hypopnea can easily cause nocturnal hypopnea, which in turn induces chronic damage to multiple organs throughout the body. In addition, its long-term existence can not only lead to or develop respiratory failure at home, but is also a risk factor for diseases such as cerebrovascular accidents, myocardial infarction, and hypertension. Therefore, early diagnosis and treatment of OSAHS can significantly improve the quality of life of patients, prevent the occurrence of various accidents and complications, and reduce the occurrence of sudden death.
[0003] The diagnosis and treatment guidelines for OSAHS provide criteria for its diagnosis: clinically, there is a typical phenomenon of snoring during sleep at night, accompanied by symptoms such as apnea, daytime sleepiness and inattention; physical examinations reveal pharyngeal stenosis, enlarged tonsils, thick uvula, adenoid hyperplasia and other phenomena; the apnea hypopnea index (AHI, the average number of apnea hypopnea per hour during sleep throughout the night) is not less than 5 times. Patients who meet the above three conditions can be diagnosed with OSAHS. Furthermore, clinically, OSAHS can be divided into three degrees: mild, moderate and severe, with reference to AHI and nighttime blood oxygen saturation SpO2. Among them, mild OSAHS is defined as 5 < AHI < 15, 85% < SpO2 < 90%; moderate OSAHS is defined as 5 < AHI < 30, 80% < SpO2 < 84%; severe OSAHS is defined as AHI > 30, SpO2 < 80%.
[0004] The diagnostic methods for OSAHS include polysomnography (PSG), fiberoptic nasal and laryngoscope examination, and impact examination. Among them, PSG is the "gold standard" for diagnosing OSAHS. PSG can simultaneously monitor nine physiological parameters, including blood oxygen saturation, oral and nasal airflow, snoring, respiratory movement, electroencephalogram, electrocardiogram, electromyogram, body position, and electrooculogram, providing patients with a detailed description of their physical condition and high-precision diagnostic results. However, PSG has high requirements for the use environment, complex operation, high cost, and a long monitoring process, which does not meet the needs of daily home monitoring. Studies have shown that the generation of snoring is closely related to the physiological structure of the respiratory tract, which can reflect the pathological conditions of the respiratory tract of OSAHS patients, and there are certain differences in the snoring of OSAHS patients with different degrees of illness. Summary of the invention
[0005] In order to solve the shortcomings of the prior art, the present invention uses computer means to realize automatic recognition of different snoring sounds and assist in the early screening of OSAHS patients at home. The present invention adopts the following technical solutions:
[0006] The automatic recognition system of hypopnea and apnea based on graph convolutional neural network provides a way to process time series. It uses visual graph method to convert snoring data into graph structure, and uses convolutional neural network to realize the classification of normal snoring, hypopnea-related snoring and apnea-related snoring. Finally, the number of hypopnea and apnea events during the whole 7h sleep is counted and the AHI index is calculated to assist in determining whether OSAHS is present or evaluating the condition of OSAHS patients.
[0007] Specifically include:
[0008] A hypopnea and apnea automatic recognition system based on a graph convolutional neural network, comprising a sleep breathing audio acquisition module, a beamforming noise reduction module, an effective snoring interception module, a snoring sequence conversion module, a graph convolutional neural network model, and a respiratory event-related snoring statistics module connected in sequence;
[0009] The sleep breathing audio collection module uses a group of signal receiving unit arrays to collect the subject's sleep breathing audio data throughout the night;
[0010] The beamforming noise reduction module synthesizes the collected multiple sleep breathing audios into mono audio data to improve the signal-to-noise ratio of the data;
[0011] The effective snoring sound segment interception module automatically extracts candidate snoring sound segments from the noise-reduced sleep breathing audio signal through wavelet transform, and professional sleep physicians mark normal snoring, snoring during hypopnea events, and snoring after apnea;
[0012] The snoring sequence conversion module converts the marked valid snoring sequence into a network structure;
[0013] The graph convolutional neural network model, through a labeled network structure, trains the graph convolutional neural network model to achieve classification of different snoring sounds, including a node sequence selection unit, a receptive field construction unit, a receptive field normalization unit and a convolutional neural network; the node sequence selection unit is used to select w center node sequences of the visual graph network; the receptive field construction unit constructs the receptive field of each center node; the receptive field normalization unit is used to construct a node attribute feature input channel and an edge attribute feature input channel; the convolutional neural network extracts and merges the node attribute features and the edge attribute features for classification;
[0014] The respiratory event-related snoring statistics module evaluates new subjects through a trained graph convolutional neural network model, and the statistical model outputs the number of hypopnea and apnea-related snoring.
[0015] Furthermore, the node sequence selection unit counts the number of connections of each node, that is, the degree of the node, for the visual graph network corresponding to each snoring segment, and arranges them in descending order to obtain a set of rearranged nodes y1, y2, ..., y n , select the top w nodes as the central nodes with a step size of s, and restore them to the original order to obtain a new node sequence, recorded as z1,z2,…,z w , if there are less than w central nodes, fill it with zero;
[0016] Furthermore, the receptive field construction unit, for each central node z i (i=1,2,…,w), use the generalized first search algorithm to find the k-1 neighbor nodes closest to it in the original network to form the receptive field R i .
[0017] Furthermore, the receptive field normalization unit, for each central node z i (i=1,2,…,w), its receptive field R i Each neighbor node in is arranged in the order of the original time series to form a one-dimensional vector V i =[v1,v2,…,v k ], where v j The attribute values corresponding to the corresponding moments of the original time series constitute the node attribute feature input channel; the edge attributes within the receptive field constitute a k*k two-dimensional vector, which constitutes the edge attribute feature input channel.
[0018] Furthermore, the convolutional neural network is used to extract network features, including a one-dimensional convolutional layer, a two-dimensional convolutional layer, a merging unit, and a classification unit. The one-dimensional convolutional layer extracts features of one-dimensional node attributes, and the two-dimensional convolutional layer extracts features of two-dimensional edge attributes. The features are then merged through the merging unit and then classified through the classification unit.
[0019] Furthermore, the beamforming noise reduction module obtains signals through M signal receiving units:
[0020]
[0021] Among them, x i (t) represents the signal received by the i-th signal receiving unit, and the variable a i (·) represents the direction vector of the transmitting source from the wave direction of arrival, θ e represents the direction of arrival of the desired signal, e(t) represents the desired signal in the far field space, and d j (t),j=1,2,…,J represents J interference signals, θ dj Indicates the direction of arrival of the interference signal, n i (t), t=1,2,…,M represents the additive noise on each signal receiving unit;
[0022] The matrix of the signal receiving unit is X(t)=[x1(t),x2(t),…,x M (t)], the output of the beamforming noise reduction module is the weighted sum of each received signal y(t) = W T (θ)X(t),W T represents the weight matrix, and beamforming is expressed as:
[0023]
[0024]
[0025] Where R = E[X(t)X T (t)] is the covariance matrix of the array output. The adaptive beam formed by MVDP is to find the weight of each signal receiving unit. Using Lagrange number multiplication, we get:
[0026] L(W)=W T RW+λ(W T a(θ e )-1)
[0027] Taking its derivative and setting it to 0, we get:
[0028] W=μR -1 a(θ e )
[0029] According to the constraint condition W T a(θ e )=1, the value of the proportional constant μ is calculated as:
[0030]
[0031] Therefore, the optimal weight obtained according to the MVDR criterion is:
[0032]
[0033] where λ is the Lagrange multiplier.
[0034] Furthermore, the effective snoring segment interception module performs windowing and frame processing on the snoring data M(t) output by the beamforming noise reduction module, and then uses the db4 small Bohm function to decompose each frame of data into k layers. The detail function is recorded as The average amplitude of the k-th layer wavelet coefficients is:
[0035]
[0036] Where i represents the i-th frame, L(k) represents the length of the k-th layer wavelet coefficient;
[0037] Divide the decomposed k layers into two segments and find the The maximum values of M1(i) and M2(i) are calculated:
[0038] MD i =M1(i)×M2(i)
[0039] If MD i >0, the data corresponding to the i-th frame is a candidate segment, and multiple consecutive windows that meet the conditions together constitute a candidate snoring segment.
[0040] Furthermore, the decomposition of each frame of data is to decompose each frame of data into 10 layers, the first 5 layers are the first part, the last 5 layers are the second part, and the detail function is recorded as Get each part Maximum value:
[0041]
[0042] Furthermore, the candidate snoring sound segment is a discrete time series in time. n =(x1,x2,…,x n ), the node corresponding to the visible graph network is the time node x i , the edges of the network are determined by the following rules:
[0043] where t a <t b <t c
[0044] Among them, a, b, c∈n, t represents the time point, if x a 、x c If the above relationship is satisfied, then x a With x c There are edges between them. Different from the original visibility graph method, the improved visibility graph method defines the weight of the edge as:
[0045]
[0046] Traverse the time series X n After all the time points of , the corresponding visual graph network is obtained.
[0047] Furthermore, the respiratory event-related snoring statistics module calculates the apnea-hypopnea index of the patient under test:
[0048]
[0049] The parameter T represents the sleep monitoring events of the subject throughout the night, in h; HI and AI represent the number of hypopnea and apnea events occurring within Th, respectively.
[0050] The advantages and beneficial effects of the present invention are:
[0051] The present invention converts the original snoring time series into graph data with a network structure, which provides an opportunity for the subsequent use of convolutional neural networks that are good at processing image information. The classification model that integrates the visual graph and the graph convolutional neural network can further improve the efficiency of automatic recognition of different snoring sounds, and the system can be transplanted and developed to a variety of mobile platforms such as mobile phones, computers and tablets, providing convenience for early diagnosis of OSAHS patients at home. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 It is a system block diagram of the present invention.
[0053] Figure 2 It is a system working process diagram of the present invention. DETAILED DESCRIPTION
[0054] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.
[0055] The embodiment of the present invention is based on a system for automatic detection and evaluation of OSAHS, and applies a visual graph and a graph convolutional neural network fusion model to snoring audio to achieve automatic recognition of a normal model, hypopnea-related snoring, and apnea-related snoring. The AHI of the subject is then calculated based on the statistical results to determine whether the subject suffers from OSAHS or to assess the severity of the disease.
[0056] like Figure 1 As shown, the automatic recognition system for hypopnea and apnea events based on visual graphs and graph convolutional neural networks includes: a sleep breathing audio acquisition module, a beamforming noise reduction module, an effective snoring segment capture module, a snoring sequence conversion module, a graph convolutional neural network model construction module and a respiratory event-related snoring statistics module.
[0057] The sleep breathing audio collection module uses multiple microphones or microphone arrays to collect the subject's sleep breathing audio data throughout the night. The specific implementation method is that multiple microphones or microphone arrays are placed within a range of about 1 meter near the subject's bedside to collect audio data. There is no restriction on the placement direction, and the arrangement form can be linear array, circular array, rectangular array and space array, etc.
[0058] The beamforming noise reduction module uses MVDR (Minimum Variance Distortion Response, adaptive beamforming) to synthesize multiple collected sleep breathing audios into mono audio data to improve the signal-to-noise ratio of the data. The specific implementation is as follows:
[0059] Assume that the desired signal in the far field is e(t) and its arrival direction is θ e In addition, there are J interference signals d j (t),j=1,2,,J, the arrival directions are θ dj Assume that there are M signal receiving units in the space, and the additive noise on each unit is n i (t), t=1,2,…,M, then the signal received by the mth unit is:
[0060]
[0061] The variable a i (·) represents the direction vector of the transmitting source from the wave direction. Then, the received signals of all receiving units can be expressed as a matrix X(t) = [x1(t), x2(t), …, x M The output of the beamformer is the weighted sum of the received signals, i.e., y(t) = W T (θ)X(t),W T represents the weight matrix. The beam travel problem can be expressed as:
[0062]
[0063]
[0064] Where R = E[X(t)X T (t)] is the covariance matrix of the array output.
[0065] The purpose of the MVDP beamformer is to find the weights of each receiving unit. Using Lagrange number multiplication, we get:
[0066] L(W)=W T RW+λ(W T a(θ e )-1)
[0067] Taking its derivative and setting it to 0, we get
[0068] W=μR -1 a(θ e )
[0069] According to the constraint condition W T a(θ e )=1, the value of the proportional constant μ is calculated as:
[0070]
[0071] Therefore, the optimal weight obtained according to the MVDR criterion is:
[0072]
[0073] The effective snoring sound segment extraction module automatically extracts candidate snoring sound segments from the noise-reduced sleep breathing audio signal through wavelet transform, and professional sleep physicians mark normal snoring, snoring during hypopnea events, and snoring after apnea. The specific implementation steps of this method are as follows:
[0074] a) After windowing and framing the snoring data M(t) output in step S20), the db4 small Bohm function is used to decompose each frame of data into 10 layers, where the first 5 layers are the first part and the last 5 layers are the second part. The detail function is recorded as The average amplitude of the k-th layer wavelet coefficients is
[0075]
[0076] Among them, i represents the i-th frame, and L(k) represents the length of the k-th layer wavelet coefficient.
[0077] b) Setting
[0078]
[0079]
[0080] c) Calculation
[0081] MD i =M1(i)×M2(i)
[0082] d) If MD i >0, the data corresponding to the i-th frame is a candidate segment, and multiple consecutive windows that meet the conditions together constitute a candidate snoring segment.
[0083] The snoring sequence conversion module converts the marked valid snoring sequence into a network structure using an improved visual graph method. The execution process is as follows: The duration of the extracted valid snoring segments is distributed between 1s and 3s. For a given discrete time sequence X n =(x1,x2,…,x n ), the node corresponding to the visible graph network is the time node x i , the edges of the network are determined by the following rules:
[0084] where t a <t b <t c
[0085] If x a ,x c If the above relationship is satisfied, then x a With x a There are edges between them. Different from the original visibility graph method, the improved visibility graph method defines the weight of the edge as:
[0086]
[0087] Traverse the time series X n After all the time points of , the corresponding visual graph network is obtained.
[0088] The graph convolutional neural network model construction module uses the labeled network structure obtained in the above steps to train the graph convolutional neural network model to classify three different snoring sounds. The specific implementation includes the following steps:
[0089] a) Select w central node sequences of the visual graph network. For each visual graph network corresponding to the snoring segment, count the degree of each node and sort them in descending order to obtain a set of rearranged nodes y1, y2, ..., y n Finally, the top w nodes are selected as the central nodes with a step size of s, and the new node sequence is obtained after restoring the original order, which is recorded as z1, z2, …, z wIf there are less than w central nodes, they will be supplemented with zero.
[0090] b) Construct the receptive field of each central node. For each central node z i (i=1,2,…,w), use the generalized first search algorithm to find the k-1 neighbor nodes closest to it in the original network to form the receptive field R i .
[0091] c) Receptive field normalization. For each central node z i (i=1,2,…,w), its receptive field R i Each neighbor node in the original time series is arranged in the order of the original time series to form a one-dimensional vector V i =[v1,v2,…,v k ], where v j The attribute values corresponding to the corresponding moments of the original time series constitute an input channel; in addition, the edge attributes within the receptive field constitute a k*k two-dimensional vector, constituting an input channel.
[0092] d) Convolutional neural network, extracting network features, is divided into two parts: the first part is the one-dimensional convolution layer to extract the features of one-dimensional node attributes, and the second part is the two-dimensional convolution layer to extract the two-dimensional edge attribute features. All features are then merged as input to achieve the purpose of classification.
[0093] The respiratory event-related snoring statistics module uses the constructed classification model to evaluate new subjects. The statistical model outputs the number of hypopnea and apnea-related snoring, calculates its AHI, and completes the evaluation of the subject's disease condition. Among them, the definition of AHI is as follows:
[0094]
[0095] The parameter T represents the sleep monitoring events of the subject throughout the night, in h; HI and AI represent the number of hypopnea and apnea events occurring within Th, respectively.
[0096] like Figure 2 As shown, the system execution process is as follows:
[0097] 1) Collect the subject's sleep breathing audio for 7 hours throughout the night through a microphone array {M i (t)}, i represents the i-th microphone;
[0098] 2) applying a minimum variance distortionless response (MVDR) beamforming method to the de-noised audio data M(t) obtained in 1);
[0099] 3) Using wavelet transform, we extract effective snore segments from the audio of the whole night’s sleep breathing. Then, we label the snores after the respiratory obstruction event as apnea snores, label the snores during the hypopnea event as hypopnea snores, and define the rest of the snores as normal snores.
[0100] 4) The snoring time series obtained in 3) is converted into a graph G(V,E) with time points as nodes and relationships between nodes as edges through an improved visibility graph (VG) method;
[0101] 5) Construct a Group Convolutional Neural Network (GCN) and train the model using the network data obtained in 4);
[0102] 6) Use the trained GCN model to classify unlabeled snoring segments, count the number of snoring related to hypopnea and apnea events, calculate the apnea-hypopnea index (AHI) of the tested patients, and assist in the diagnosis and evaluation of OSAHS patients.
[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A hypopnea and apnea automatic recognition system based on a graph convolutional neural network, comprising a sleep breathing audio acquisition module, a beamforming noise reduction module, an effective snoring segment interception module, a snoring sequence conversion module, a graph convolutional neural network model, and a respiratory event-related snoring statistics module, which are connected in sequence, and characterized by: The sleep breathing audio collection module uses a group of signal receiving unit arrays to collect the subject's sleep breathing audio data throughout the night; The beamforming noise reduction module synthesizes the collected multiple sleep breathing audios into mono audio data; The effective snoring sound segment interception module automatically extracts candidate snoring sound segments from the noise-reduced sleep breathing audio signal through wavelet transform, and marks normal snoring, snoring during hypopnea events, and snoring after apnea; The snoring sequence conversion module converts the time characteristics of the marked effective snoring sequence into a network structure; the candidate snoring segments are discrete time sequences in time, and the candidate snoring segments are divided into two groups according to the discrete time sequence X. n =(x1,x2,…,x n ), the node corresponding to the visible graph network is the time node x i , the edges of the network are determined by the following rules: Among them, a, b, c∈n, t represents the time point, if x a 、x c If the above rules are met, then x a With x c There is an edge between them, and the weight of the edge is: Traverse the time series X n After all the time points, the corresponding visual graph network is obtained; The graph convolutional neural network model, through a labeled network structure, trains the graph convolutional neural network model to achieve classification of different snoring sounds, including a node sequence selection unit, a receptive field construction unit, a receptive field normalization unit and a convolutional neural network; the node sequence selection unit is used to select w center node sequences of the visual graph network; the receptive field construction unit constructs the receptive field of each center node; the receptive field normalization unit is used to construct a node attribute feature input channel and an edge attribute feature input channel; Convolutional neural network, extracts and combines node attribute features and edge attribute features for classification; The node sequence selection unit counts the number of connections of each node, that is, the degree of the node, for each visual graph network corresponding to the snoring sound segment and arranges them in descending order to obtain a set of rearranged nodes y1, y2, ..., y n , select the top w nodes as the central nodes with a step size of s, and restore them to the original order to obtain a new node sequence, recorded as z1,z2,…,z w , if there are less than w central nodes, fill it with zero; The receptive field construction unit, for each central node z i (i=1,2,…,w), use the generalized first search algorithm to find the k-1 neighbor nodes closest to it in the original network to form the receptive field R i ; The receptive field normalization unit, for each central node z i (i=1,2,…,w), its receptive field R i Each neighbor node in is arranged in the order of the original time series to form a one-dimensional vector V i =[v1,v2,…,v k ], where v j The attribute values corresponding to the corresponding moments of the original time series constitute the node attribute feature input channel; The edge attributes within the receptive field form a k*k two-dimensional vector, which constitutes the edge attribute feature input channel; The convolutional neural network is used to extract network features, including a one-dimensional convolution layer, a two-dimensional convolution layer, a merging unit, and a classification unit. The one-dimensional convolution layer extracts features of one-dimensional node attributes, and the two-dimensional convolution layer extracts features of two-dimensional edge attributes. The merging unit then merges features, and the classification unit performs classification. The respiratory event-related snoring statistics module evaluates new subjects through a trained graph convolutional neural network model, and the statistical model outputs the number of hypopnea and apnea-related snoring.
2. The automatic recognition system for hypopnea and apnea based on graph convolutional neural network according to claim 1 is characterized in that The beamforming noise reduction module obtains signals through M signal receiving units: Among them, x i (t) represents the signal received by the i-th signal receiving unit, and the variable a i (·) represents the direction vector of the transmitting source from the wave direction of arrival, θ e represents the direction of arrival of the desired signal, e(t) represents the desired signal in the far field space, and d j (t), j = 1, 2, ..., J represents J interference signals, θ dj Indicates the direction of arrival of the interference signal, n i (t), t=1,2,…,M represents the additive noise on each signal receiving unit; The matrix of the signal receiving unit is X(t)=[x1(t),x2(t),…,x M (t)], the output of the beamforming noise reduction module is the weighted sum of each received signal y(t) = W T (θ)X(t),W T represents the weight matrix, and beamforming is expressed as: Where R = E[X(t)X T (t)] is the covariance matrix of the array output. The adaptive beam formed by MVDP is to find the weight of each signal receiving unit. Using Lagrange number multiplication, we get: L(W)=W T RW+λ(W T a(θ e )-1) Taking its derivative and setting it to 0, we get: W=μR -1 a(θ e ) According to the constraint condition W T a(θ e )=1, the value of the proportional constant μ is calculated as: Therefore, the optimal weight obtained according to the MVDR criterion is: where λ is the Lagrange multiplier.
3. The automatic recognition system for hypopnea and apnea based on graph convolutional neural network according to claim 1 is characterized in that The effective snoring segment interception module performs windowing and frame processing on the snoring data M(t) output by the beamforming noise reduction module, and then uses the db4 small Bohm function to decompose each frame of data into k layers. The detail function is recorded as The average amplitude of the k-th layer wavelet coefficients is: Where i represents the i-th frame, L(k) represents the length of the k-th layer wavelet coefficient; Divide the decomposed k layers into two segments and find their respective The maximum values of M1(i) and M2(i) are calculated: MD i =M1(i)×M2(i) If MD i >0, the data corresponding to the i-th frame is a candidate segment, and multiple consecutive windows that meet the conditions together constitute a candidate snoring segment.
4. The automatic recognition system for hypopnea and apnea based on graph convolutional neural network according to claim 3 is characterized in that The decomposition of each frame of data is to decompose each frame of data into 10 layers, the first 5 layers are the first part, and the last 5 layers are the second part. The detail function is recorded as Get each part Maximum value:
5. The automatic recognition system for hypopnea and apnea based on graph convolutional neural network according to claim 1 is characterized in that The respiratory event-related snoring statistics module calculates the apnea-hypopnea index of the patient under test: The parameter T represents the sleep monitoring events of the subject throughout the night, in h; HI and AI represent the number of hypopnea and apnea events occurring within Th, respectively.
