Multi-parameter real-time wireless monitoring integrated device and method

Abstract

The present application relates to the field of real-time wireless monitoring, in particular to a multi-parameter real-time wireless monitoring integrated device and method device, which comprises a multi-sensor data acquisition module, a multi-modal fusion processing module, an abnormality detection analysis module and a multi-level alarm module; the multi-sensor data acquisition module uses a generative adversarial network to eliminate motion artifacts and acquires multiple physiological signals; the multi-modal fusion processing module realizes spatio-temporal alignment and correlation modeling by means of linear interpolation and graph convolution network, and generates a joint feature vector; the abnormality detection analysis module presets causal relationships based on a causal diagram, optimizes a weight matrix model, and uses counterfactual reasoning to determine the root cause; the multi-level alarm module distinguishes alarm levels and processes according to a risk scoring model; the device and method solve the problems of low precision and poor cooperativity of traditional monitoring technology, realize multi-parameter accurate monitoring, intelligent analysis and efficient alarm, and can be widely applied to remote medical monitoring and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of multi-parameter real-time wireless monitoring integration, in particular to a multi-parameter real-time wireless monitoring integration device and method. BACKGROUND

[0002] In today's medical field, multi-parameter real-time wireless monitoring technology is crucial for disease prevention, diagnosis and patient health management. Traditional multi-parameter monitoring methods have many drawbacks and cannot meet the needs of modern medicine.

[0003] Traditional monitoring equipment often relies on a single sensor or simple combination, which cannot accurately capture complex physiological signal changes. Various physiological parameters are interrelated, but traditional devices lack effective fusion and analysis means. The data collected by different sensors are often processed independently, and the potential relationship between parameters cannot be tapped. For example, there is a close relationship between respiratory rate and blood oxygen saturation. When the respiration is abnormal, the blood oxygen saturation will usually change, but the traditional monitoring is difficult to use this association for comprehensive judgment, and important pathological information is easily missed.

[0004] Without accurate root cause analysis, traditional monitoring is mainly based on preset threshold to judge abnormalities, and cannot analyze the causes of abnormalities in depth. When the blood oxygen saturation is lower than the normal range, it is difficult to determine whether it is caused by respiratory problems, cardiovascular problems or other factors. In order to solve the above defects, the present application provides a technical solution. SUMMARY

[0005] In order to solve the technical problems in the background art, the present application provides a multi-parameter real-time wireless monitoring integration device and method.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] The first aspect of the present application provides a multi-parameter real-time wireless monitoring integration device, comprising a multi-sensor data acquisition module, a multi-modal fusion processing module, an abnormality detection and analysis module, a multi-level alarm module and a database.

[0008] The multi-sensor data acquisition module acquires the user state according to the multi-sensor, and establishes a generative adversarial network to eliminate motion artifacts. The specific process is as follows:

[0009] The multi-sensor includes a photoelectric sensor, a heart rate piezoelectric sensor, an impedance sensor, a temperature sensor and a MEMS acceleration; the photoelectric sensor is used to acquire blood oxygen saturation signal in real time, the heart rate piezoelectric sensor is used to acquire electrocardiogram signal in real time, the temperature sensor is used to acquire body temperature signal in real time, the impedance sensor is used to acquire respiratory frequency signal in real time, and the MEMS accelerometer acquires three-axis dynamic acceleration signal in real time under the user motion state.

[0010] The generative adversarial network architecture comprises a noise generator, a signal generator and a discriminator, the dynamic acceleration signal is input to the noise generator, the dynamic acceleration signal is subjected to time domain feature extraction through a feature extraction layer of a one-dimensional convolutional neural network, three convolutional layers are included, high frequency vibration and low frequency displacement features of the motion signal are extracted layer by layer, and the convolutional features are mapped into a motion feature matrix through a fully connected layer; the electrocardiogram signal, the respiratory frequency signal and the motion feature matrix are input to the signal generator, and a noise-free signal without artifacts is output through a residual convolutional network; the noise-free signal is input to the discriminator, global features of the signal are extracted, the global features include waveform periodicity, amplitude distribution and spectral energy, and a discrimination probability p is output through a Sigmoid activation function in the output layer, p is in the range of [0, 1], if p is approximately equal to 1, it indicates that the input is a real signal and the motion artifacts have been removed, if p is approximately equal to 0, it indicates that the input is a generated signal, and the above steps are repeated until a real signal is obtained;

[0011] The multi-modal fusion processing module realizes spatio-temporal alignment and correlation modeling according to linear interpolation and graph convolution network (GCN), generates joint features through a feedforward network, and improves the accuracy of anomaly detection, and the specific steps are as follows:

[0012] According to linear interpolation, the blood oxygen saturation signal, the electrocardiogram signal, the body temperature signal, the respiratory frequency signal and the dynamic acceleration signal are synchronized to a unified time axis;

[0013] After synchronization, the GCN is modeled, specifically: a node set V = {X1, X2,..., X5} is set, corresponding to five kinds of nodes, each node includes each feature dimension, X1 includes amplitude, waveform rising edge time and pulse transmission time in the blood oxygen saturation node, X2 includes RR interval coefficient of variation and QRS wave amplitude in the electrocardiogram node, X3 includes real-time temperature and temperature change rate in the body temperature signal, X4 includes tidal volume and respiratory period in the respiratory frequency signal, and X5 includes acceleration amplitude, acceleration and motion direction; the edge set δ is initialized as the causal relationship preset by the medical guide, and the initial adjacency matrix is formed by setting multiple directed edges as is represented as the number of directed edges;

[0014] The initial adjacency matrix is converted into a learnable weight matrix by learning the edge weight through three layers of GCN, and the spatial correlation features of the nodes are obtained, specifically: the adjacency matrix is normalized by a symmetric normalization method through the first layer of GCN, and the calculation logic is as follows: wherein D is a degree matrix, ensuring that the node features are normalized according to the number of neighbors, and the calculation logic is as follows: wherein H (0) is the input feature vector, W (0) is the weight matrix, and σ is the activation function, and H (1)Capture the feature interaction of the first-order neighbor; the second layer GCN repeats the adjacency matrix normalization step, introduces residual connection to avoid gradient disappearance, and its calculation logic is: Output H (2) Model the feature coupling of the second-order neighbor, such as heart rate variability → change in cardiac output → oxygen transport efficiency; the third layer GCN outputs an edge weight matrix according to linear transformation, representing the causal strength between nodes as: The target function optimizes the adjacency matrix to output the learnable weight matrix A θ , whose edge weight range is (0, 1);

[0015] The features of each node after processing are integrated through a feedforward network to output a joint feature vector with a dimension of 128, which includes time-domain and frequency-domain indicators of heart rate variability, and spatial correlation features such as respiratory-heart rate coupling phase difference and mapping relationship between conduction time and blood pressure.

[0016] The anomaly detection analysis module is based on the preset physiological parameter causal relationship of the causal graph, and the causal graph model is optimized by the weight matrix, and the root cause analysis algorithm is introduced to determine the root cause through counterfactual reasoning, outputting the abnormal type, probability, and trend chart. The specific process is as follows: define a directed acyclic graph The joint feature vector with a dimension of 128 output by the feedforward network is mapped to a node state vector X ∈ R T×5 , where T is the time step.

[0017] Input the node state vector into the target function , where y t represents the actual observed physiological parameter vector at time t, represents the parameter vector predicted by the causal graph model, which calculates the causal transmission between nodes through the graph neural network, and λ = 0.01 represents the L2 regularization coefficient.

[0018] Calculate the causal graph model output according to the learnable weight matrix, and its calculation logic is:

[0019] Calculate the prediction error MES through backpropagation, and its calculation logic is: Use the Adam optimizer to accelerate convergence, stop when the MSR is less than 10 -5 for 50 consecutive rounds, and get the optimized causal graph model.

[0020] The root cause analysis algorithm is introduced to determine the root cause by counterfactual reasoning, and the abnormal type and probability value are output. Specifically, when the blood oxygen saturation is detected to be less than 90% or the heart rate interval coefficient of variation is greater than 0.15 or the body temperature is greater than 37.5℃ for three consecutive time steps, the root cause analysis is triggered. Taking blood oxygen saturation as an example, the blood oxygen saturation, electrocardiogram, body temperature, respiratory rate and dynamic acceleration are marked as TY, TR, TE, TW and TQ respectively. The intervention variables of blood oxygen saturation are sorted. If the first sorting is respiratory rate, the normal value of user respiratory rate is calculated to obtain TW normal , the current respiratory rate is intervened and set to TW normal , the trained causal graph model is used, the intervened respiratory rate value is input into the model, and the blood oxygen saturation theoretical value is calculated through the forward propagation of the model , the difference between the expected value of blood oxygen saturation E[TW normal |TY] after intervention and the currently observed blood oxygen saturation value E[TW|TY] is calculated, and the obtained difference value is divided by the historical standard deviation of blood oxygen saturation TW σ , to obtain the standardized causal effect SCE, and the calculation logic is:

[0021]

[0022] The difference between the theoretical value of blood oxygen saturation after intervention and the actual observed value E[TW|TY] is obtained, and the standardization processing is performed, and the calculation logic is: The corresponding preset threshold value in the database is extracted to determine the strength of causal association: when ZS>2, it is determined that there is a strong causal association; when 1.5≤ZS<2, it is determined that there is a weak causal association;

[0023] The logical regression model is used to integrate multiple signal evidences to calculate the root cause probability P (r) of respiratory rate, and the calculation logic is:

[0024] Wherein, 2, 1.5 and 1 are fixed weight coefficients, TR 偏离度 is the deviation of current electrocardiogram from the normal range, TE 偏差 is the difference between the current body temperature and the normal body temperature range; in this way, the root cause probability of electrocardiogram, body temperature and dynamic acceleration is calculated, and the maximum root cause probability is selected as the abnormal type and sent to the multi-level alarm module.

[0025] The multi-level alarm module distinguishes and processes the alarm level through risk scoring according to the abnormal detection and root cause analysis results, combined with the physiological parameter abnormality degree and clinical risk level. The specific process is as follows:

[0026] The root cause probability value P (r)and parameter absolute value deviation P 偏差 , substitute into the risk score calculation model: SW=P (r) ·0.7+P 偏差 ·0.3 to obtain a risk score value, wherein 0.7 and 0.3 are fixed weight coefficients of the root cause probability value and the parameter absolute value deviation; the risk score value is divided into three levels, including a blue warning, a yellow warning and a red warning, and specifically: 30≤S<40 is a blue warning, health suggestions are pushed through a short message and parameter fluctuations are recorded; 40≤S<70 is a yellow warning, a flashing warning light + voice reminder is given to a user's mobile terminal, and a parameter trend report is generated; and S≥70 is a red warning, an emergency contact person is notified through a telephone, a sound and light alarm is triggered and positioning tracking is sent to a corresponding doctor's mobile terminal.

[0027] The second aspect of the application provides a multi-parameter real-time wireless monitoring integrated method, and the specific steps are as follows:

[0028] Step one, multi-sensor data acquisition: according to the multi-sensor, the user state is collected, and a generative adversarial network is established to eliminate motion artifacts;

[0029] Step two, multi-modal fusion processing: according to linear interpolation and GCN, time-space alignment and correlation modeling are realized, joint features are generated through a feedforward network, and the accuracy of abnormal detection analysis is improved;

[0030] Step three, abnormal detection analysis: based on the preset physiological parameter causal relationship of a causal graph, the causal graph model is optimized through a weight matrix, a root cause analysis algorithm is introduced to determine the root cause through counterfactual reasoning, and an abnormal type, a probability and a trend graph are output;

[0031] Step four, multi-level alarm: according to a risk score calculation model, the risk score function is calculated by comprehensively considering the root cause probability, the parameter deviation degree and the trend risk, and the alarm level is distinguished and processed through the risk score.

[0032] Compared with the prior art, the application has the following beneficial effects:

[0033] Enhance multi-parameter fusion analysis capability: linear interpolation and graph convolution network are used to realize time-space alignment and correlation modeling of multi-physiological signals, and the potential relationship between parameters is excavated. The joint feature vector generated through the feedforward network contains rich time sequence and space correlation information, such as respiratory-heart rate coupling phase difference, so that the abnormal detection accuracy is greatly improved.

[0034] Accurate anomaly detection and root cause analysis: based on the preset physiological parameter causal relationship of the causal diagram, and through the weight matrix optimization model, combined with counterfactual reasoning to determine the root cause. When detecting abnormal blood oxygen saturation, it can accurately judge whether it is caused by respiratory problems, cardiovascular problems or other factors, which helps doctors quickly develop targeted treatment plans, avoids misdiagnosis and missed diagnosis, and improves treatment effect. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. The following drawings are not drawn in proportion to the actual size, and the focus is on showing the main idea of the present application.

[0036] Figure 1 The module connection block diagram of the present application.

[0037] Figure 2 The method step diagram of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor also belong to the scope of protection of the present application.

[0039] Please refer to Figure 1 The first aspect of the present application provides a multi-parameter real-time wireless monitoring integrated device, which comprises a multi-sensor data acquisition module, a multi-modal fusion processing module, an anomaly detection analysis module, a multi-level alarm module and a database.

[0040] The multi-sensor data acquisition module acquires the user state according to the multi-sensor, establishes a generative adversarial network to eliminate motion artifacts, and the specific process is as follows:

[0041] The multi-sensor includes a photoelectric sensor, a heart rate piezoelectric sensor, an impedance sensor, a temperature sensor and a MEMS acceleration; the photoelectric sensor is used to acquire blood oxygen saturation signal in real time, the heart rate piezoelectric sensor is used to acquire ECG signal in real time, the temperature sensor is used to acquire body temperature signal in real time, the impedance sensor is used to acquire respiratory frequency signal in real time, and the MEMS accelerometer acquires three-axis dynamic acceleration signal in real time under the user motion state;

[0042] The generative adversarial network architecture comprises a noise generator, a signal generator and a discriminator, the dynamic acceleration signal is input to the noise generator, the dynamic acceleration signal is subjected to time domain feature extraction through a feature extraction layer of a one-dimensional convolutional neural network, three convolutional layers are included, high frequency vibration and low frequency displacement features of the motion signal are extracted layer by layer, and the convolutional features are mapped into a motion feature matrix through a fully connected layer; the electrocardiogram signal, the respiratory frequency signal and the motion feature matrix are input to the signal generator, and a de-noised artifact-free signal is output through a residual convolutional network; the artifact-free signal is input to the discriminator, global features of the signal are extracted, the global features include waveform periodicity, amplitude distribution and spectral energy, and a sigmoid activation function is used in the output layer to output a discrimination probability p∈[0, 1], if p≈1, it indicates that the input is a real signal and the motion artifact has been removed, and if p≈0, it indicates that the input is a generated signal, and the above steps are repeated until a real signal is obtained;

[0043] The multi-modal fusion processing module realizes spatio-temporal alignment and correlation modeling according to linear interpolation and graph convolution network (GCN), generates joint features through a feedforward network, and improves the accuracy of anomaly detection, and the specific steps are as follows:

[0044] The blood oxygen saturation signal, the electrocardiogram signal, the body temperature signal, the respiratory frequency signal and the dynamic acceleration signal are synchronized to a unified time axis according to linear interpolation;

[0045] After synchronization, the GCN is modeled, specifically: a node set V={X1, X2,..., X5} is set, corresponding to five kinds of nodes, each node includes feature dimensions, X1 includes amplitude, waveform rising edge time and pulse transmission time in the blood oxygen saturation node, X2 includes RR interval variation coefficient and QRS wave amplitude in the electrocardiogram node, X3 includes real-time temperature and temperature change rate in the body temperature signal, X4 includes tidal volume and respiratory period in the respiratory frequency signal, and X5 includes acceleration amplitude, acceleration and motion direction; the edge set δ is initialized as the causal relationship preset by the medical guide, such as respiratory abnormality→blood oxygen decrease or heart rate acceleration→stroke volume change, and an initial adjacency matrix is formed by setting multiple directed edges as is represented as the number of directed edges;

[0046] The initial adjacency matrix is converted into a learnable weight matrix through three layers of GCN to learn the edge weight, and the spatial correlation features of the nodes are obtained, specifically: the adjacency matrix is normalized through a symmetric normalization method in the first layer of GCN, and the calculation logic is as follows: wherein D is a degree matrix, ensuring that the node features are normalized according to the number of neighbors, and the feature change and activation are performed, and the calculation logic is as follows: wherein the input H (0) is a feature vector, W (0) is a weight matrix, and σ is an activation function, and the output H (1)Capture the feature interaction of the first-order neighbor; the second layer GCN repeats the adjacency matrix normalization step, introduces residual connection to avoid gradient disappearance, and its calculation logic is: Output H (2) Model the feature coupling of the second-order neighbor, such as heart rate variability → change in cardiac output → oxygen transport efficiency; the third layer GCN outputs an edge weight matrix according to linear transformation, representing the causal strength between nodes as: Optimize the adjacency matrix through the objective function to output the learnable weight matrix A θ , whose edge weight range is (0, 1); for example, the edge weight of heart rate variability → respiratory rate change is optimized from the initial 1 to 0.85, indicating a strong causal relationship, capturing the spatial correlation features of physiological coupling such as ECG-oxygen, respiration-heart rate, etc.; the features of each node after processing are integrated through a feedforward network to output a joint feature vector with a dimension of 128, which contains time-domain and frequency-domain indicators of heart rate variability, spatial correlation features of respiratory-heart rate coupling phase difference and conduction time, and mapping relationship with blood pressure;

[0047] The anomaly detection analysis module is based on the causal graph and the preset physiological parameter causal relationship, and the weight matrix optimizes the causal graph model, introduces the root cause analysis algorithm to determine the counterfactual reasoning, outputs the abnormal type, probability and trend chart, and the specific process is as follows: define a directed acyclic graph The joint feature vector with a dimension of 128 output by the feedforward network is mapped to a node state vector X ∈ R T×5 , where T is the time step;

[0048] Input the node state vector into the objective function , where y t represents the actual observed physiological parameter vector at time t, represents the parameter vector predicted by the causal graph model, which calculates the causal transmission between nodes through the graph neural network, and λ = 0.01 represents the L2 regularization coefficient to avoid overfitting;

[0049] Calculate the causal graph model output according to the learnable weight matrix, and its calculation logic is:

[0050] Calculate the prediction error MES through backpropagation, and its calculation logic is: Use the Adam optimizer to accelerate convergence, stop when the MSR is less than 10 -5 for 50 consecutive rounds, and finally get the optimized causal graph model;

[0051] The root cause analysis algorithm is introduced to determine the root cause by counterfactual reasoning, and the abnormal type, probability and trend chart are output. Specifically, when the blood oxygen saturation is less than 90% or the heart rate interval coefficient of variation is greater than 0.15 or the body temperature is greater than 37.5℃ for 3 consecutive time steps, the root cause analysis is triggered. Taking blood oxygen saturation as an example, the blood oxygen saturation, electrocardiogram, body temperature, respiratory rate and dynamic acceleration are marked as TY, TR, TE, TW and TQ respectively. The intervention variables of blood oxygen saturation are sorted. If the first sorting is respiratory rate, the normal value of user respiratory rate is calculated as TW normal . It should be noted that the average value of the respiratory rate data of the individual in the past week is calculated as the normal value; the current respiratory rate is intervened and set as TW normal . The trained causal graph model is used to input the intervened respiratory rate value into the model, and the theoretical value of blood oxygen saturation is calculated by forward propagation of the model . The difference between the expected value of blood oxygen saturation E[TW normal |TY] after intervention and the currently observed blood oxygen saturation value E[TW|TY] is calculated, and the obtained difference value is divided by the historical standard deviation of blood oxygen saturation TW σ . The standardized causal effect SCE is obtained, and the calculation logic is as follows:

[0052]

[0053] The signals of electrocardiogram and body temperature are detected, and whether the related characteristics of electrocardiogram change after the respiratory rate is intervened is observed. The ST segment depression amplitude of electrocardiogram is checked to see if it decreases. If the ST segment depression amplitude indeed decreases, it may indicate that the change of respiratory rate has an impact on the electrical activity of the heart, thereby supporting the inference that there is a causal relationship between respiratory rate and blood oxygen saturation. The correlation between body temperature and heat metabolism rate is checked to see if it returns to normal. If the correlation returns to normal, it also provides additional supporting evidence for respiratory rate as the root cause.

[0054] The difference between the theoretical value of blood oxygen saturation after intervention and the actual observed value E[TWTY] is obtained, and the standardized processing is performed, and the calculation logic is as follows: The corresponding preset threshold value in the database is extracted to determine the strength of causal association: when ZS>2, it is determined that there is a strong causal association; when 1.5≤ZS<2, it is determined that there is a weak causal association.

[0055] The logical regression model is used to integrate multiple signal evidences to calculate the root cause probability P (r) of respiratory rate, and the calculation logic is as follows: wherein 2, 1.5 and 1 are fixed weight coefficients, TR 偏离度 is the deviation degree of current electrocardiogram from the normal range, and TE 偏差The current temperature is the difference value of the normal temperature range; in this way, the root cause probability of the electrocardio, body temperature and dynamic acceleration is selected as the maximum value of the abnormal type and sent to the multi-level alarm module.

[0056] The multi-level alarm module distinguishes and processes the alarm level through risk scoring according to the abnormal detection and root cause analysis results, in combination with the physiological parameter abnormality degree and clinical risk level, and the specific process is as follows:

[0057] The root cause probability value P of the current abnormal type is obtained (r) and the parameter absolute value deviation P 偏差 , the risk score calculation model is substituted: SW=P (r) ·0.7+P 偏差 ·0.3, wherein 0.7 and 0.3 are fixed weight coefficients of the root cause probability value and the parameter absolute value deviation; the risk score value is divided into three levels, including blue pre-warning, yellow alarm and red alarm, and the specific process is as follows: 30≤S<40 is blue pre-warning, health suggestions are pushed through short message and parameter fluctuations are recorded; 40≤S<70 is yellow alarm, a flashing warning light+voice reminder is given to the user's mobile terminal, and a parameter trend report is generated; S≥70 is red alarm, emergency contacts are notified through telephone, sound and light alarms are triggered and positioning tracking is sent to the corresponding doctor's mobile terminal.

[0058] As shown in Figure 2 , the second aspect of the present application provides a multi-parameter real-time wireless monitoring integrated method, and the specific steps are as follows:

[0059] Step one, multi-sensor data acquisition: according to the multi-sensor, the user state is collected, and the generated adversarial network is used to eliminate the motion artifact elimination;

[0060] Step two, multi-modal fusion processing: according to linear interpolation and GCN, time-space alignment and correlation modeling are realized, joint features are generated through a feedforward network, and the accuracy of abnormal detection analysis is improved;

[0061] Step three, abnormal detection analysis: based on the preset physiological parameter causal relationship of the causal graph, the causal graph model is optimized through the weight matrix, the root cause analysis algorithm is introduced to determine the counterfactual reasoning root cause, and the abnormal type, probability and trend graph are output;

[0062] Step four, multi-level alarm: according to the risk score calculation model, the risk score function is calculated by comprehensively considering the root cause probability, parameter deviation degree and trend risk, and the alarm level is distinguished and processed through the risk score.

[0063] The foregoing is illustrative of the present application, and is not to be construed as limiting thereof. While a number of exemplary embodiments of the application have been described, those skilled in the art will readily appreciate that many modifications are possible in the exemplary embodiments without materially departing from the novel teachings and advantages of the application. Accordingly, all such modifications are intended to be included within the scope of the present application as defined in the claims. It is to be understood that the above description is intended to be illustrative, and not restrictive. Many other embodiments will be apparent to those of skill in the art upon reviewing the above description. The scope of the application should, therefore, be determined not with reference to the above description, but should instead be determined with reference to the appended claims, along with their full scope of equivalents.

Claims

1. A multi-parameter real-time wireless monitoring integrated device, comprising a multi-sensor data acquisition module, a multi-modal fusion processing module, an anomaly detection analysis module, a multi-level alarm module and a database, characterized in that: the multi-modal fusion processing module realizes spatio-temporal alignment and correlation modeling according to linear interpolation and GCN, and generates joint features through a feedforward network; the multi-level alarm module calculates a risk score function by comprehensively calculating root cause probability, parameter deviation degree and trend risk according to a risk score calculation model, and distinguishes and processes alarm levels through the risk score; the anomaly detection analysis module introduces a root cause analysis algorithm to determine the root cause through counterfactual reasoning, and outputs the abnormal type and probability value, specifically: The abnormality detection analysis module presets physiological parameter causal relationship based on a causal diagram, and optimizes the causal diagram model through a weight matrix. Specifically, a directed acyclic graph is defined The joint feature vector with a feedforward network output dimension of 128 is mapped to a node state vector through a full connection layer where T is a time step number, represents a node set, is an edge set, containing preset causal relationships, is a causal strength between nodes; the node state vector is input into a target function where, represents an actually observed physiological parameter vector at t, represents a parameter vector predicted by the causal diagram model, is a node state feature matrix at t, which is calculated through graph neural network and causal transmission between nodes, represents an L2 regularization coefficient; The causal graph model output is calculated according to a learnable weight matrix, and the calculation logic is as follows: Wherein is an optimized adjacency weight matrix of the causal graph; a prediction error MES is calculated by back propagation, and the calculation logic is as follows: An Adam optimizer is used to accelerate convergence, and when a moving average MSR of multiple rounds of MSE is less than for 50 consecutive rounds, the optimization of the causal graph model is stopped. When the blood oxygen saturation is detected to be less than 90% or the coefficient of variation of the electrocardio interval is greater than 0.15 or the body temperature is greater than 37.5℃ for three consecutive time steps, the root cause analysis is triggered, the blood oxygen saturation, electrocardio, body temperature, respiratory rate and dynamic acceleration are marked as TY, TR, TE, TW and TQ respectively, the intervention variables of the blood oxygen saturation are sorted, if the first sorting is the respiratory rate, the normal value of the user's respiratory rate is calculated to obtain , the current respiratory rate is intervened and set to , the trained causal graph model is used, the intervened respiratory rate value is input into the model, the forward propagation of the model is calculated to obtain the theoretical value of the blood oxygen saturation ; the expected value of the blood oxygen saturation after intervention is calculated , the difference between the expected value and the currently observed blood oxygen saturation value is calculated, and the obtained difference value is divided by the historical standard deviation of the blood oxygen saturation , to obtain the standardized causal effect SCE, and the calculation logic is: ; Obtain the theoretical value of blood oxygen saturation after intervention Compared with actual observed values The difference between them is then standardized, and the calculation logic is as follows: ; Extract the corresponding preset threshold from the database to determine the strength of the causal relationship: when If the value is greater than 2, a strong causal relationship is determined; if 1.5 ≤ If the causal relationship is less than 2, then a weak causal relationship is determined to exist; the multi-modal fusion processing module realizes spatio-temporal alignment and correlation modeling according to linear interpolation and graph convolution network (GCN), and the specific steps are as follows: The root cause probability of respiratory rate is calculated by integrating multi-signal evidence through a logistic regression model The calculation logic is as follows: wherein 2, 1.5 and 1 are fixed weight coefficients, is the deviation degree of the current electrocardiogram from the normal range, is the difference between the current body temperature and the normal body temperature range; similarly, the root cause probability of electrocardiogram, body temperature and dynamic acceleration is selected, the maximum root cause probability is selected as the abnormal type and sent to the multi-level alarm module.

2. The multi-parameter real-time wireless monitoring integrated device of claim 1, wherein, synchronize the blood oxygen saturation signal, the electrocardiogram signal, the body temperature signal, the respiratory frequency signal and the dynamic acceleration signal to a unified time axis according to linear interpolation; the multi-modal fusion processing module generates joint features through a feedforward network, specifically: After synchronization, the GCN is modeled as follows: a node set is set , corresponding to 5 kinds of nodes, each node includes each feature dimension, The blood oxygen saturation node includes amplitude, waveform rising edge time and pulse transmission time, The ECG node includes RR interval coefficient of variation and QRS amplitude, The body temperature signal includes real-time temperature and temperature change rate, The respiratory rate signal includes tidal volume and respiratory cycle, Acceleration amplitude, acceleration and motion direction; edge set Initialized as the causal relationship preset by the medical guide, set multiple directed edges to form the initial adjacency matrix as , Indicated as the number of directed edges; Edge weights are learned through a 3-layer GCN, transforming the initial adjacency matrix into a learnable weight matrix, and then the spatial association features of nodes are obtained. Specifically, the first layer of the GCN normalizes the adjacency matrix using a symmetric normalization method, the calculation logic of which is as follows: ,in Given a degree matrix, to ensure that node features are normalized according to the number of neighbors, feature transformations and activations are performed on it. The calculation logic is as follows: , where input Feature vector This is the weight matrix. For the activation function, the output is... Capture the characteristic interactions of first-order neighbors; The second layer GCN repeats the adjacency matrix normalization step, introduces residual connection to avoid gradient vanishing, and the calculation logic is: , output Modeling the feature coupling of the second-order neighbor, the third layer GCN outputs an edge weight matrix according to a linear transformation, representing the causal strength between nodes as: , the target function optimizes the adjacency matrix to output the updated learnable weight matrix , and the edge weight range is .

3. The multi-parameter real-time wireless monitoring integrated device of claim 2, wherein, integrate the features of each node after processing through a feedforward network to output a joint feature vector with a dimension of 128, and the joint feature vector includes time-domain indicators of heart rate variability and frequency-domain indicators, spatial correlation features of respiratory-heart rate coupling phase difference and mapping relationship between conduction time and blood pressure. the multi-level alarm module distinguishes and processes alarm levels through a risk score according to the results of anomaly detection and root cause analysis, combined with the physiological parameter abnormality degree and the clinical risk level, and the specific process is as follows:

4. The multi-parameter real-time wireless monitoring integrated device of claim 1, wherein, the multi-sensor data acquisition module acquires the user state according to multiple sensors, specifically: obtaining a root cause probability value of the current abnormal type and a parameter absolute value deviation , substituting into a risk score calculation model: obtaining a risk score value, wherein 0.7 and 0.3 are fixed weight coefficients of the root cause probability value and the parameter absolute value deviation; dividing the risk score value into three levels, including blue warning, yellow warning and red warning, specifically: 30≤ <40 is blue warning, health suggestions are pushed through short message and parameter fluctuations are recorded; 40≤ <70 is yellow warning, a parameter trend report is generated by flashing warning light + voice reminding the user's mobile terminal; ≥70 is red warning, an emergency contact is notified by telephone, a sound and light alarm is triggered and positioning tracking is sent to the corresponding doctor's mobile terminal.

5. The multi-parameter real-time wireless monitoring integrated device of claim 1, wherein, the multiple sensors include a photoelectric sensor, a heart rate piezoelectric sensor, an impedance sensor, a temperature sensor and a MEMS acceleration; the photoelectric sensor is used to acquire the blood oxygen saturation signal in real time, the heart rate piezoelectric sensor is used to acquire the electrocardiogram signal in real time, the temperature sensor is used to acquire the body temperature signal in real time, the impedance sensor is used to acquire the respiratory frequency signal in real time, and the MEMS accelerometer acquires the three-axis dynamic acceleration signal in real time under the user motion state. the multi-sensor data acquisition module establishes a generative adversarial network to eliminate motion artifacts, and the specific process is as follows:

6. The multi-parameter real-time wireless monitoring integrated device of claim 5, wherein, the generative adversarial network architecture includes a noise generator, a signal generator and a discriminator, the dynamic acceleration signal is input into the noise generator, the time-domain features of the dynamic acceleration signal are extracted through a one-dimensional convolutional neural network feature extraction layer, which includes three convolutional layers, the high-frequency vibration and low-frequency displacement features of the motion signal are extracted layer by layer, and then the convolutional features are mapped into a motion feature matrix through a fully connected layer; the electrocardiogram signal, the respiratory frequency signal and the motion feature matrix are input into the signal generator, and a denoised artifact-free signal is output through a residual convolutional network; the specific steps are as follows: The artifact-free signal is input to the discriminator, global features of the signal are extracted, the global features include waveform periodicity, amplitude distribution and spectral energy, and a discrimination probability is output by a Sigmoid activation function of an output layer , if , it indicates that the input is a real signal and the motion artifact has been removed, if , it indicates that the input is a generated signal, and the above steps are repeated until it is a real signal.

7. A multi-parameter real-time wireless monitoring integrated method for implementing the multi-parameter real-time wireless monitoring integrated device of any one of claims 1-6, characterized in that, Step 1: Multi-sensor data acquisition: acquire the user state according to multiple sensors, and establish a generative adversarial network to eliminate motion artifacts; ​ Step two, multi-modal fusion processing: according to linear interpolation and GCN, spatio-temporal alignment and correlation modeling are realized, joint features are generated through feedforward network, and the accuracy of anomaly detection analysis is improved; Step three, anomaly detection analysis: based on the preset causal relationship of physiological parameters in the causal graph, the causal graph model is optimized through the weight matrix, the root cause analysis algorithm is introduced to judge the counterfactual reasoning, and the output includes the type, probability and trend of the anomaly; Step four, multi-level alarm: according to the risk score calculation model, the risk score function is calculated by comprehensively considering the root cause probability, parameter deviation degree and trend risk, and the alarm level is distinguished and processed through the risk score.

Images