Contactless sleep stage classification system and classification method based on visual graph

Abstract

The present invention provides a non-contact sleep stage classification system and classification method based on a visual graph. The system includes a non-contact physiological information acquisition module, a visual graph network construction module, and a sleep stage multivariate classification module. The system is used to train the network features extracted by the visual graph network construction module using a machine learning method. First, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Thereafter, the sleep stages are further sub-classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, and finally the division of the five sleep stages is completed. The present invention applies the visual graph network to the respiratory and heartbeat signals extracted by the non-contact vital signs monitoring radar, realizes the construction of two branch networks of respiratory and heartbeat, further explores the deep information of respiratory and heartbeat signals, and helps to improve the accuracy of sleep stage classification.

Description

Technical Field

[0001] The present invention relates to the field of non-contact sleep stage classification, and in particular to research on a sleep stage classification system and classification method based on a visual graph. Background Art

[0002] Sleep is crucial to human health, and accurately monitoring sleep stages helps assess sleep quality. Currently, traditional sleep monitoring relies primarily on polysomnography (PSG), which uses electrodes to collect biosignals such as electroencephalograms (EEGs) and electrocardiograms (ECGs) and is considered the gold standard for sleep assessment. However, PSG equipment is expensive and complex to operate, and the electrodes attached to the patient can affect their comfort, particularly for those with sleep disorders, potentially disrupting their natural sleep state. Therefore, existing technologies lack comfort and convenience.

[0003] To address these issues, non-contact bio-radar technology is emerging as an alternative. Bio-radar analyzes radar echo signals reflected from the human body to extract vital signs such as breathing and heartbeat, enabling sleep stage inference without the need for electrodes. This significantly improves monitoring comfort and convenience. However, existing non-contact monitoring technologies still have room for improvement in data processing and signal analysis, particularly in signal feature extraction and classification accuracy. Summary of the Invention

[0004] The purpose of the present invention is to address the problems existing in the above-mentioned prior art and provide a non-contact sleep stage classification system and classification method based on visual graphs.

[0005] The technical solutions for achieving the purpose of the present invention are:

[0006] A contactless sleep stage classification system based on a visual graph, the system comprising:

[0007] The non-contact physiological information acquisition module is used to collect the patient's vital sign signals in real time during sleep using a non-contact vital sign monitoring radar, and filter the signals to obtain breathing and heartbeat signals;

[0008] A visual graph network building module is used to map the respiratory signal and heartbeat signal obtained by the non-contact physiological information acquisition module into a respiratory network and a heartbeat network through a weighted finite traversal visual graph, and extract network features of the respiratory network and the heartbeat network respectively to complete the network building;

[0009] The sleep stage multi-classification module is used to train the network features extracted by the visual graph network building module using machine learning methods. First, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, finally completing the division of five sleep stages.

[0010] The overall architecture includes a non-contact acquisition module for physiological signals, a visual graph network construction module for respiratory and heartbeat signals, and a multi-dimensional classification module for sleep stages;

[0011] The construction of a visual graph network of respiratory and heartbeat signals obtained by non-contact vital signs monitoring radar and the realization of a coupling network for respiratory and heartbeat signals;

[0012] The multi-faceted division of sleep stages first divides the sleep stages into wakefulness and sleep, and then further subdivides the sleep stages into rapid eye movement (REM), light sleep (stage I and stage II), and deep sleep (stage III).

[0013] A non-contact sleep stage classification method based on a visual graph, the classification method comprising:

[0014] Use non-contact vital sign monitoring radar to collect the patient's vital sign signals in real time during sleep, and filter the signals to obtain respiratory and heartbeat signals;

[0015] The respiratory signal and heartbeat signal are mapped into the respiratory network and heartbeat network through the weighted finite traversal visibility graph, and the network features of the respiratory network and heartbeat network are extracted respectively to complete the network construction;

[0016] The extracted network features are trained using machine learning methods: first, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement (REM), sleep stage I, sleep stage II, and sleep stage III, and finally the division into five sleep stages is completed.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] 1) Using a non-contact vital signs monitoring radar to monitor sleep stages in real time, improving patient comfort and avoiding psychological discomfort caused by contact with the device. It is particularly suitable for long-term monitoring, allowing patients to maintain a natural sleep state;

[0019] 2) By analyzing respiratory and heartbeat signals using weighted finite traversal visibility graphs and mapping them into a network, we can mine deep information from the signals, improve the accuracy of sleep stage prediction, and capture more hidden information patterns than traditional methods.

[0020] 3) Combine the respiratory and heartbeat networks to construct a coupled network, study the relationship between the two networks, further improve the accuracy and complexity of signal processing, and enhance the comprehensive analysis of sleep states; 4) Propose a two-step classification method, first divide it into the awake period and the non-awake period, and then further subdivide the non-awake period, which solves the classification problem of rapid eye movement and non-rapid eye movement periods, and greatly improves the accuracy of sleep stage classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a contactless sleep stage classification system based on a visual graph in one embodiment.

[0022] Figure 2 FIG. 1 is a complex network diagram of respiratory and heartbeat signals based on a visual graph in one embodiment.

[0023] Figure 3 FIG. 1 is a schematic diagram showing a comparison between polysomnography and contactless sleep stage classification results based on a visual graph in one embodiment. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0025] In one embodiment, combined Figure 1-Figure 3 , provides a sleep stage classification system based on a visual graph, the system comprising:

[0026] The non-contact physiological information acquisition module is used to collect the patient's vital sign signals in real time during sleep using a non-contact vital sign monitoring radar, and filter the signals to obtain breathing and heartbeat signals;

[0027] A visual graph network building module is used to map the respiratory signal and heartbeat signal obtained by the non-contact physiological information acquisition module into a respiratory network and a heartbeat network through a weighted finite traversal visual graph, and extract network features of the respiratory network and the heartbeat network respectively to complete the network building;

[0028] The sleep stage multi-classification module is used to train the network features extracted by the visual graph network building module using machine learning methods. First, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, finally completing the division of five sleep stages.

[0029] Furthermore, in one embodiment, the non-contact physiological information acquisition module includes:

[0030] Signal acquisition unit, used to set up the vital signs monitoring radar above the patient's abdomen and collect radar echo signals , including respiratory signals, heartbeat signals and clutter signals, the vital signs monitoring radar is a continuous wave radar;

[0031] Respiratory signal acquisition unit, introducing a bandpass filter , for radar echo signal Bandpass filtering is performed to remove clutter and noise. The frequency range is 0.15~0.4Hz, which is the basic respiratory frequency band. After filtering, the respiratory signal is obtained. ;

[0032] Heartbeat signal acquisition unit, introducing a bandpass filter , for radar echo signal Bandpass filtering is performed to remove clutter and noise. The frequency range is 1~1.4Hz, which is the basic heartbeat frequency band. The heartbeat signal is obtained after filtering. .

[0033] Furthermore, in one embodiment, the visual graph network building module includes:

[0034] (1) Weighted finite traversal visibility graph initialization module:

[0035] ① Connection principle setting: The time series of breathing and heartbeat signals obtained by the non-contact physiological information acquisition module and , N is the total length of the time series. Assuming a finite crossing distance L = 1, the principle that any two points in the sequence are visible and connected is:

[0036] sequence Two points separated by m data points and When there is an edge, there are k data points between the two points ,in ,satisfy:

[0037] ;

[0038] And the rest data points satisfy:

[0039] ;

[0040] At this point, there is an edge between the two points.

[0041] ② Edge direction setting: Considering the irreversibility of time series, assuming that the sequence The two points with an edge are nodes and nodes , the direction of the edge between points AB is defined as:

[0042] ;

[0043] ③ Edge weight setting: defining sequence Two points and The edge weight for:

[0044] ;

[0045] (2) Respiration, heartbeat network and coupling network construction module:

[0046] ① Construction of respiratory network and heartbeat network: According to the settings in the weighted finite traversal visibility graph initialization module, the respiratory signal time series and heartbeat signal time series Send it to the initialization module and map it to obtain a respiratory weighted directed network according to the set principles , Heartbeat Weighted Directed Network ;

[0047] ② Construction of the respiratory and heartbeat coupling network. The specific process includes:

[0048] A. Computing Breathing Weighted Directed Networks The degree of the i-th node , degree value Defined as a respiratory weighted directed network The number of all other nodes connected to this node;

[0049] B. Computing Heartbeat Weighted Directed Networks The degree of the i-th node , degree value Defined as a heartbeat weighted directed network The number of all other nodes connected to this node;

[0050] C. When When , it is considered that the respiratory weighted directed network The i-th node and the heartbeat weighted directed network There is an edge before the i-th node, and the direction of the edge is from node i in the respiratory network to node i in the heartbeat network. The set of all edges is ;in, The coupling threshold is preset and is determined adaptively based on the physiological characteristics and sampling frequency of the respiratory and heartbeat signals. Here it is set to ;

[0051] D. Obtaining a respiratory and heartbeat coupling network ,in For two single-layer networks and Collection of For a single-layer network and The set of all edges between .

[0052] (3) Network feature extraction module: extract network features from the obtained network and extract characteristic parameters that reflect the changes in time series, including:

[0053] ① Average weight of respiratory network : reflects the connection density of the respiratory network;

[0054] ;

[0055] Where, It is a breathing network The edge weight between node i and node j;

[0056] ② Average clustering coefficient of respiratory network : reflects the overall connectivity of the respiratory network;

[0057] ;

[0058] Where, Represents all respiratory networks The number of nodes connected to node i, Indicates the actual number of edges between all nodes connected to node i, is the local clustering coefficient of node i;

[0059] ③ Respiratory network clustering coefficient entropy : describes the distribution of clustering coefficients of each node in the respiratory network;

[0060] ;

[0061] Where, For the Breathing Network The local clustering coefficient of node i;

[0062] ④ Weighted clustering coefficient entropy of respiratory network : describes the distribution of weighted clustering coefficients of each node in the respiratory network;

[0063] ;

[0064] ;

[0065] Where, For the Breathing Network The weighted clustering coefficients of node i, node j and node k are The remaining nodes except the middle node i;

[0066] ⑤ Average weight of heartbeat network : reflects the connection density of the heartbeat network;

[0067] ;

[0068] Where, It is a heartbeat network The edge weight between node i and node j;

[0069] ⑥ Average clustering coefficient of heartbeat network : reflects the overall connectivity of the heartbeat network;

[0070] ;

[0071] Where, Indicates all heartbeat networks The number of nodes connected to node i, Indicates the actual number of edges between all nodes connected to node i, is the local clustering coefficient of node i;

[0072] ⑦ Heartbeat network clustering coefficient entropy : describes the distribution of clustering coefficients of each node in the heartbeat network;

[0073] ;

[0074] Where, Heartbeat Network The local clustering coefficient of node i;

[0075] ⑧ Heartbeat network weighted clustering coefficient entropy : describes the distribution of weighted clustering coefficients of each node in the heartbeat network;

[0076] ;

[0077] ;

[0078] Where, Heartbeat Network The weighted clustering coefficients of node i, node j and node k are The remaining nodes except the middle node i;

[0079] ⑨ Respiratory-heartbeat coupling network edge crossing index INT: measures the probability that a pair of nodes are connected by an edge in all single layers;

[0080] ;

[0081] ;

[0082] Where M is the number of networks in the composite network. , Indicates the Layer Network The corresponding adjacency matrix Elements in .

[0083] ⑩ Inter-layer correlation of respiratory and heartbeat coupling networks: used to characterize multi-layer networks Chinese Respiratory Network and Heartbeat Network The correlation between the frequency series is used to quantify the physiological synchronization of respiratory and heartbeat signals during sleep stages (for example, high correlation corresponds to a tight structure in deep sleep); the calculation is done using the Pearson correlation coefficient;

[0084] ;

[0085] in, is the correlation value (range [-1,1]); and are the degrees of the i-th node in the G1 layer and the G2 layer respectively; and are the means of the degree sequences of G1 layer and G2 layer respectively; n is the number of nodes;

[0086] ⑪ Network similarity of the respiratory-cardiocoupled network: This measures the degree to which one network replaces another in a multilayer network, characterizing the similarity of network structures (for example, low similarity may correspond to representations of the arousal stage). It also highlights the role of C in bridging two layers by integrating the coupling edge C (the presence of the edge enhances similarity, reflecting the strength of physiological coupling). This is calculated using the Jaccard similarity coefficient.

[0087] ;

[0088] in, is the similarity value (range [0,1]); E1 and E2 are the edge sets of G1 layer and G2 layer respectively (ignoring weight thresholding); C is the set of all connected edges; is the number of shared edges; is the total number of unique edges.

[0089] Furthermore, in one embodiment, the sleep stage multivariate classification module includes:

[0090] (1) Network feature preprocessing unit: used to normalize the network parameters obtained by the visual graph network building module, and perform zero-mean normalization on the feature vector using the mean and standard deviation of the features;

[0091] ;

[0092] Where, represents the set of features of the i-th category, represents the mean of the i-th category feature, represents the variance of the i-th category feature;

[0093] (2) Awake / sleep stage classification unit, used to process and classify the extracted features, including:

[0094] A. Obtain polysomnographic data collected synchronously with the device of the present invention, and have professionally trained technicians stage the polysomnographic data according to the American Academy of Sleep Medicine's interpretation criteria to obtain the sleep stages for each time segment, specifically: wakefulness, non-rapid eye movement (N1), N2, N3, and rapid eye movement (REM);

[0095] B. While retaining the sleep stage for each time segment obtained in step A, relabel the normalized network feature samples: assign a new label of "0" to feature samples from the wakefulness period, and assign a uniform label of "1" to feature samples from all other sleep periods. This facilitates subsequent wakefulness / sleep binary classification model training or performance evaluation.

[0096] C. Randomly divide into training set and test set according to proportion;

[0097] D. Use machine learning to train the training set feature samples;

[0098] E. Use the trained machine learning module to classify the awake and awake periods in the test set;

[0099] (3) Sleep stage sub-classification units, including:

[0100] A. Based on the test set classification results obtained by the awake / sleep stage classification unit, retain the awake samples and remove the label "1" added to the remaining non-awake samples, restoring the original labels;

[0101] B. Remove the awake period and the labels added later from the training data and send it back to the machine learning module for training;

[0102] C. Use the machine learning module to predict the five sleep stages in the non-awake period in the training set;

[0103] D. Adding samples from the wakefulness period, we finally complete the refined classification of the five sleep stages, including wakefulness, rapid eye movement, sleep stage I, sleep stage II, and sleep stage III.

[0104] The non-contact sleep stage classification system based on visual graphs proposed in the present invention can be applied to home sleep monitoring. By using a non-contact vital signs monitoring radar, the user's breathing, heartbeat and other physiological indicators can be monitored in real time, avoiding the discomfort caused by traditional contact equipment and greatly improving the convenience and comfort of monitoring. It is particularly suitable for long-term monitoring, such as for the elderly, patients with chronic diseases or users with sleep disorders. By accurately classifying sleep stages, it helps detect abnormal conditions such as sleep apnea. During the postoperative recovery period or anesthesia observation period in a hospital or rehabilitation center, the patient's physiological condition is continuously and non-interferencely monitored to reduce the workload of medical staff. In addition, through cloud data sharing, medical staff can remotely monitor the condition and adjust the treatment plan in a timely manner, providing higher monitoring accuracy and efficiency.

[0105] In one embodiment, an overall architecture is provided, including a non-contact physiological signal acquisition module, a visual graph network construction module for respiratory and heartbeat signals, and a sleep stage multivariate classification module:

[0106] The non-contact physiological information acquisition module is used to collect the patient's vital sign signals in real time during sleep using a non-contact vital sign monitoring radar, and filter the signals to obtain breathing and heartbeat signals;

[0107] A visual graph network building module is used to map the respiratory signal and heartbeat signal obtained by the non-contact physiological information acquisition module into a respiratory network and a heartbeat network through a weighted finite traversal visual graph, and extract network features of the respiratory network and the heartbeat network respectively to complete the network building;

[0108] The sleep stage multi-classification module is used to train the network features extracted by the visual graph network building module using machine learning methods. First, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, finally completing the division of five sleep stages.

[0109] For the specific definition of each step, please refer to the definition of the non-contact sleep stage classification system based on the visual graph above, which will not be repeated here.

[0110] In one embodiment, a method for constructing a visual graph network of respiratory and heartbeat signals acquired by a non-contact vital signs monitoring radar and implementing a coupling network for the respiratory and heartbeat signals is provided:

[0111] The visual graph network construction module is used to map the respiratory signals and heartbeat signals obtained by the non-contact physiological information acquisition module into respiratory networks and heartbeat networks through weighted finite traversal visual graphs, and extract network features of the respiratory network and heartbeat network respectively to complete the network construction.

[0112] For the specific definition of each step, please refer to the definition of the non-contact sleep stage classification system based on the visual graph above, which will not be repeated here.

[0113] In one embodiment, a multivariate classification method for sleep stages is provided:

[0114] The sleep stage multi-classification module is used to train the network features extracted by the visual graph network building module using machine learning methods. First, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, finally completing the division of five sleep stages.

[0115] For the specific definition of each step, please refer to the definition of the non-contact sleep stage classification system based on the visual graph above, which will not be repeated here.

[0116] The present invention also discloses a non-contact sleep stage classification method based on a visual graph, the classification method comprising:

[0117] Use non-contact vital sign monitoring radar to collect the patient's vital sign signals in real time during sleep, and filter the signals to obtain respiratory and heartbeat signals;

[0118] The respiratory signal and heartbeat signal are mapped into the respiratory network and heartbeat network through the weighted finite traversal visibility graph, and the network features of the respiratory network and heartbeat network are extracted respectively to complete the network construction;

[0119] The extracted network features are trained using machine learning methods: first, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further subclassified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, and finally the division of five sleep stages is completed.

[0120] In summary, the present invention applies a visual graph network to the respiratory and heartbeat signals extracted by the non-contact vital signs monitoring radar, realizes the construction of two branch networks of respiratory and heartbeat, further mines the deep-level information of respiratory and heartbeat signals, and helps to improve the accuracy of sleep stage classification.

[0121] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A contactless sleep stage classification system based on a visual graph, characterized by: include: The non-contact physiological information acquisition module is used to collect the patient's vital sign signals in real time during sleep using a non-contact vital sign monitoring radar, and filter the signals to obtain breathing and heartbeat signals; A visual graph network building module is used to map the respiratory signal and heartbeat signal obtained by the non-contact physiological information acquisition module into a respiratory network and a heartbeat network through a weighted finite traversal visual graph, and extract network features of the respiratory network and the heartbeat network respectively to complete the network building; The sleep stage multivariate classification module is used to train the network features extracted by the visual graph network building module using machine learning methods. The extracted network features are first normalized, and then the sleep stages are divided into wakefulness and sleep. The sleep stages are then further classified into rapid eye movement, sleep stage I, sleep stage II, and sleep stage III, ultimately completing the classification of five sleep stages. The visual graph network building module includes: (1) Weighted finite traversal visibility graph initialization module; (2) Respiration, heartbeat network and coupling network construction modules; (3) Network feature extraction module: extract network features from the obtained network and extract characteristic parameters that reflect the changes in time series; The weighted finite traversal visibility graph initialization module includes: ① Connection principle setting: The time series of breathing and heartbeat signals obtained by the non-contact physiological information acquisition module and , N is the total length of the time series, and the finite crossing distance L=1 is set. The principle that any two points in the sequence are visible and connected is: sequence Two points separated by m data points and When there is an edge, there are k data points between the two points ,in ,satisfy: ； And the rest data points satisfy: ； At this point, there is an edge between the two points; ② Edge direction setting: Considering the irreversibility of time series, assuming that the sequence The two points with an edge are nodes and nodes , the direction of the edge between points AB is defined as: ； ③Edge weight setting: defining sequence Two points and The edge weight for: ； The breathing, heartbeat network and coupling network building modules include: ① Construction of respiratory network and heartbeat network: According to the settings in the weighted finite traversal visibility graph initialization module, the respiratory signal time series and heartbeat signal time series Send it to the initialization module and map it to obtain a respiratory weighted directed network according to the set principles , Heartbeat Weighted Directed Network ; ②Build a respiratory and heartbeat coupling network. The specific process includes: A. Computing Breathing Weighted Directed Networks The degree of the i-th node , degree value Defined as a respiratory weighted directed network The number of all other nodes connected to this node; B. Computing Heartbeat Weighted Directed Networks The degree of the i-th node , degree value Defined as a heartbeat weighted directed network The number of all other nodes connected to this node; C. When When , it is considered that the respiratory weighted directed network The i-th node and the heartbeat weighted directed network There is an edge before the i-th node, and the direction of the edge is from node i in the respiratory network to node i in the heartbeat network. The set of all edges is ;in, The coupling threshold is preset and is determined adaptively based on the physiological characteristics and sampling frequency of the respiratory and heartbeat signals. Here it is set to ; D. Obtaining a respiratory and heartbeat coupling network ,in For two single-layer networks and Collection of For a single-layer network and The set of all edges between them; The network feature extraction module includes: ① Average weight of respiratory network : reflects the connection density of the respiratory network; ； Where, It is a breathing network The edge weight between node i and node j; ② Average clustering coefficient of respiratory network : reflects the overall connectivity of the respiratory network; ； Where, Represents all respiratory networks The number of nodes connected to node i, Indicates the actual number of edges between all nodes connected to node i, is the local clustering coefficient of node i; ③ Respiratory network clustering coefficient entropy : describes the distribution of clustering coefficients of each node in the respiratory network; ； Where, For the Breathing Network The local clustering coefficient of node i; ④ Weighted clustering coefficient entropy of respiratory network : describes the distribution of weighted clustering coefficients of each node in the respiratory network; ； ； Where, For the Breathing Network The weighted clustering coefficients of node i, node j and node k are The remaining nodes except the middle node i; ⑤ Average weight of heartbeat network : reflects the connection density of the heartbeat network; ； Where, It is a heartbeat network The edge weight between node i and node j; ⑥ Average clustering coefficient of heartbeat network : reflects the overall connectivity of the heartbeat network; ； Where, Indicates all heartbeat networks The number of nodes connected to node i, Indicates the actual number of edges between all nodes connected to node i, is the local clustering coefficient of node i; ⑦ Heartbeat network clustering coefficient entropy : describes the distribution of clustering coefficients of each node in the heartbeat network; ； Where, Heartbeat Network The local clustering coefficient of node i; ⑧Heartbeat network weighted clustering coefficient entropy : describes the distribution of weighted clustering coefficients of each node in the heartbeat network; ； ； Where, Heartbeat Network The weighted clustering coefficients of node i, node j and node k are The remaining nodes except the middle node i; ⑨ Respiratory-heartbeat coupling network edge crossing index INT: measures the probability that a pair of nodes are connected by an edge in all single layers; ； ； Where M is the number of networks in the composite network. , Indicates the Layer Network The corresponding adjacency matrix Elements in ⑩Inter-layer correlation of respiratory and heartbeat coupling networks: used to characterize multi-layer networks Chinese Respiratory Network and Heartbeat Network The correlation between the above value sequences is used to quantify the physiological synchronization between respiratory and heartbeat signals during sleep stages; the calculation is done using the Pearson correlation coefficient; ； in, is the correlation value (range [-1,1]); and are the degrees of the i-th node in the G1 layer and the G2 layer respectively; and are the means of the degree sequences of G1 layer and G2 layer respectively; n is the number of nodes; ⑪ Network similarity of respiratory and heartbeat coupling networks: used to measure the degree to which one network replaces another in a multi-layer network, characterize the similarity of network structures, and highlight the role of C in bridging the two layers by integrating the coupling edge C. The calculation uses the Jaccard similarity coefficient. ; in, is the similarity value (range [0,1]); E1 and E2 are the edge sets of G1 layer and G2 layer respectively (ignoring weight thresholding); C is the set of all connected edges; is the number of shared edges; is the total number of unique edges.

2. The non-contact sleep stage classification system based on visual graph according to claim 1, characterized in that: The non-contact physiological information acquisition module includes: Signal acquisition unit, used to set up the vital signs monitoring radar above the patient's abdomen and collect radar echo signals , including respiratory signals, heartbeat signals and clutter signals; Respiratory signal acquisition unit, introducing a bandpass filter , for radar echo signal Bandpass filtering is performed to remove clutter and noise. The frequency range is 0.15~0.4Hz, which is the basic respiratory frequency band. After filtering, the respiratory signal is obtained. ; Heartbeat signal acquisition unit, introducing a bandpass filter , for radar echo signal Bandpass filtering is performed to remove clutter and noise. The frequency range is 1~1.4Hz, which is the basic heartbeat frequency band. The heartbeat signal is obtained after filtering. .

3. The non-contact sleep stage classification system based on visual graph according to claim 2, characterized in that: The sleep stage multivariate classification module includes: (1) Network feature preprocessing unit: used to normalize the network parameters obtained by the visual graph network building module, and perform zero-mean normalization on the feature vector using the mean and standard deviation of the features; ； Where, represents the set of features of the i-th category, represents the mean of the i-th category feature, represents the variance of the i-th category feature; (2) Awake / sleep stage classification unit, used to process and classify the extracted features, including: A. Obtain polysomnographic data collected synchronously with the device of the present invention, and have professionally trained technicians stage the polysomnographic data according to the American Academy of Sleep Medicine's interpretation criteria to obtain the sleep stages for each time segment, specifically: wakefulness, non-rapid eye movement (N1), N2, N3, and rapid eye movement (REM); B. While retaining the sleep stage for each time segment obtained in step A, relabel the normalized network feature samples: assign a new label of "0" to feature samples from the wake phase, and assign a uniform label of "1" to feature samples from all other sleep phases. This facilitates subsequent wakefulness / sleep binary classification model training or performance evaluation. C. Randomly divide into training set and test set according to proportion; D. Use machine learning to train the training set feature samples; E. Use the trained machine learning module to classify the awake and awake periods in the test set; (3) Sleep stage sub-classification units, including: A. Based on the test set classification results obtained by the awake / sleep stage classification unit, retain the awake samples and remove the label "1" added to the remaining non-awake samples, restoring the original labels; B. Remove the awake period and the labels added later from the training data and send it back to the machine learning module for training; C. Use the machine learning module to predict the five sleep stages in the non-awake period in the training set; D. Adding samples from the wakefulness period, we finally complete the refined classification of the five sleep stages, including wakefulness, rapid eye movement, sleep stage I, sleep stage II, and sleep stage III.

4. The non-contact sleep stage classification system based on a visual graph according to claim 2, characterized in that: The vital signs monitoring radar is a continuous wave radar.

5. The classification method of the non-contact sleep stage classification system based on visual graph according to claim 1, characterized in that: This classification method includes: Use non-contact vital sign monitoring radar to collect the patient's vital sign signals in real time during sleep, and filter the signals to obtain respiratory and heartbeat signals; The respiratory signal and heartbeat signal are mapped into the respiratory network and heartbeat network through the weighted finite traversal visibility graph, and the network features of the respiratory network and heartbeat network are extracted respectively to complete the network construction; The extracted network features are trained using machine learning methods: first, the extracted network features are normalized, and then the sleep stages are divided into wakefulness and sleep. Then, the sleep stages are further classified into rapid eye movement (REM), sleep stage I, sleep stage II, and sleep stage II, finally completing the division into five sleep stages.

Images