Myocardial infarction global lead reconstruction method applied to real-time electrocardiogram monitoring device

Abstract

The application proposes a myocardial infarction global lead reconstruction method applied to real-time electrocardiogram monitoring equipment, comprising the following steps: step S1: data preprocessing is performed on myocardial infarction electrocardiogram signal data set and healthy control electrocardiogram signal data set; step S2: principal component analysis of each lead is performed, a variance contribution rate diagram is drawn, a highest contribution degree lead decision is realized, and a reconstructed lead group is determined; step S3: redundant information of the original data of the lead group is filtered out, nonlinear convolution pre-classification processing is performed on the data, and the lead reconstruction algorithm effect is improved; step S4: the conversion relationship between the leads is fitted by the lead reconstruction algorithm; step S5: a convolutional neural network is used as a lead reconstruction model framework, network structure optimization and parameter setting are performed; and step S6: based on the optimized network, global lead reconstruction models are respectively constructed, and state differences before and after lead reconstruction are verified; the application can realize portable electrocardiogram monitoring and clinical diagnosis lead intercommunication, and further improve the practicability of a portable electrocardiogram monitor.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electrocardio feature classification and machine learning, and in particular to a myocardial infarction global lead reconstruction method applied to real-time electrocardio monitoring equipment. BACKGROUND

[0002] Cardiovascular diseases (CVD) have become a dark cloud over people's heads with the aggravation of social population aging. Myocardial infarction (MI) is the most common factor causing sudden death. A large part of the patients with myocardial infarction do not have obvious symptoms in daily life, but if they are in a high-pressure fatigue state for a long time, it may induce massive myocardial cell necrosis. These patients with no obvious symptoms often do not actively use the examination method with radiation damage, and are more likely to cause continuous deterioration of the disease due to their own and the doctor's negligence. Electrocardiogram (ECG) diagnosis, as a non-invasive and low-cost examination method, has the ability to discover potential myocardial infarction lesions and is widely used in clinical practice. Electrocardio monitoring in clinical practice is generally divided into static standard twelve-lead monitoring and dynamic electrocardio monitoring.

[0003] Due to the limitation of current medical resources, it is difficult to achieve real-time monitoring of each patient for 365 days a year. However, acute myocardial infarction is sudden, and real-time monitoring can help discover and treat diseases in time. In addition, portable electrocardio monitors are generally limited in the number of lead channels due to considerations of portability, endurance, and other aspects, while hospitals generally use standard twelve-lead electrocardiogram as the basis for diagnostic interpretation. This creates a contradiction between portable design and insufficient clinical significance. In today's medical and engineering combination background, how to use the powerful data processing capability of computers and related data processing methods in the field of artificial intelligence to explore the method of lead interconnection between portable electrocardio monitoring and clinical diagnosis, and thus improve the practicality of portable electrocardio monitors is a problem that needs to be solved. SUMMARY

[0004] The present application proposes a myocardial infarction global lead reconstruction method applied to real-time electrocardio monitoring equipment, which can realize lead interconnection between portable electrocardio monitoring and clinical diagnosis, and thus improve the practicality of portable electrocardio monitors.

[0005] The present application adopts the following technical solutions.

[0006] The myocardial infarction global lead reconstruction method applied to real-time electrocardio monitoring equipment includes the following steps:

[0007] Step S1: data preprocessing is performed on the myocardial infarction electrocardio signal data set and the healthy control electrocardio signal data set;

[0008] Step S2: Principal component analysis is performed on the processed electrocardio data set P1 in step S1, a variance contribution rate graph is drawn, the highest contribution lead decision is made, and the reconstructed lead group is determined in combination with the electrode requirements of the dynamic monitoring device;

[0009] Step S3: Redundant information of the original data of the lead group is filtered out, and the myocardial infarction and healthy control electrocardio data are pre-classified by nonlinear convolution to improve the effect of the lead reconstruction algorithm;

[0010] Step S4: The lead reconstruction algorithm is designed based on physical reality, and the conversion relationship between the leads is fitted;

[0011] Step S5: The convolutional neural network is used as the lead reconstruction model framework, and the network structure optimization and parameter setting are performed;

[0012] Step S6: Based on the optimized network, the global lead reconstruction model for myocardial infarction and healthy control is constructed respectively, and the state difference before and after lead reconstruction is verified by root mean square error (RMSE) and correlation coefficient (CC) and the like.

[0013] Step S1 is specifically:

[0014] Step S11: The Mallat algorithm is used to perform discrete wavelet decomposition on the original signal, and the original signal is filtered through a low-pass filter and a high-pass filter to obtain low-frequency and high-frequency signals; Specifically: first, the low-frequency signal is down-sampled to obtain a new signal with half the length, which represents the low-frequency part of the original signal; second, the high-frequency signal is down-sampled to obtain a new signal with half the length, which represents the high-frequency part of the original signal, and the obtained new signal is used as the input signal; finally, repeat the above steps until the preset number of layers is reached; Through discrete wavelet decomposition, a group of low-frequency and high-frequency signal wavelet coefficients is obtained, and the decomposition layer number is set to eight; db6 wavelet is selected, and the scale function φ(x) and the low-pass filter coefficients h0, h1, h2, h3, h4, h5 are expressed as follows:

[0015]

[0016]

[0017] Based on the scale function and the low-pass filter coefficients, the wavelet basis function of the Daubechies wavelet is calculated;

[0018] The low-pass wavelet basis function ψ low (x) and the high-pass wavelet basis function ψ high(x) The calculation formula is as follows:

[0019]

[0020]

[0021] In wavelet analysis, the decomposition process of each layer is regarded as the convolution operation of the target signal and the low-pass wavelet base function and the high-pass wavelet base function, and then the result is down-sampled to obtain the signal low-frequency part coefficient a j-1 (n) and the signal high-frequency part coefficient d j-1 (n), the calculation formula is as follows:

[0022]

[0023]

[0024] Wherein, x j (k) represents the signal after the jth layer decomposition. The high-frequency noise in the electrocardiogram signal adopted in the embodiment mainly concentrates on the D1 scale, the low-frequency baseline drift noise concentrates on the A8 scale, and the D5 scale and the D6 scale contain most of the electrocardiogram signal.

[0025] Step S12: The second, third, fourth, fifth, sixth and seventh layer detail coefficients are used for wavelet reconstruction, and the first and eighth layer detail coefficients containing noise are all set to zero, so as to obtain the electrocardiogram signal filtered from noise;

[0026] Step S13: The polyphase filter is used to realize electrocardiogram signal down-sampling, and corresponding filtering processing, high-frequency component removal and fixed interval extraction operation are performed on the input signal in different frequency ranges, so as to obtain the down-sampled signal.

[0027] Step S2 is specifically:

[0028] Step S21: The principal component analysis is performed on the twelve leads, the lead with the highest effective degree in the twelve leads is determined, the original data in the higher dimensional space is projected and mapped to the low-dimensional space through the linear transformation method, the redundancy is reduced, and the most effective information of the original data is extracted in the data dimension reduction process; each lead is regarded as a data dimension; m pieces of n-dimensional data are combined into a matrix X m×n ;

[0029] Step S22: The matrix X m×n is zero-mean, that is, each row is subtracted by the mean value of the row, to obtain a matrix Y;

[0030] Step S23: The covariance matrix corresponding to Y is calculated:

[0031] C=(1 / m)*YY T

[0032] Step S24: calculate the eigenvalue a of C and the eigenvector A;

[0033] Step S25: sort the eigenvector matrix A in descending order according to the eigenvalue of the matrix corresponding to the eigenvector matrix A to form a new matrix, and take the first k (k≤n) rows to form a matrix P;

[0034] Step S26: draw a variance contribution diagram, and select I, II, and V1 as the known leads according to the result.

[0035] Step S26 uses a portable mobile monitoring device with five electrodes, and the known leads are I, II, and a chest lead.

[0036] Step S3 is specifically:

[0037] Step S31: before the data is reconstructed, the data is pre-classified, and the preprocessed electrocardiogram signal is input into the first network, and the input matrix channel number is 3 and the length is 1000;

[0038] Step S32: the convolution layer extracts the input vector features, and the activation layer performs nonlinear transformation on the network. After two convolution and normalization operations, the output feature vector enters the max pooling layer to realize the downsampling operation, and the above operation is repeated x times. Smaller convolution kernels are used in deeper networks to extract more subtle information, and the number of convolution kernels is increased to increase the channel number. After each convolution operation, the data is put into the normalization layer for processing, so that the data is quickly distributed in a reasonable interval range in the short term, and the network iteration speed is accelerated.

[0039] Step S33: input the feature vector into the fully connected layer, and finally output the myocardial infarction signal and normal electrocardiogram signal classification results.

[0040] Step S4 is specifically:

[0041] Step S41: mathematically analyze the relationship between the lead projection axis and the electrocardiogram vector ring, and obtain the real-time voltage values of I, II, and V1 leads by calculating the dot product of the two, specifically:

[0042] The pre-classified myocardial infarction electrocardiogram data and healthy control electrocardiogram data are reconstructed based on the physical actual design of the lead; the conversion relationship between the leads is fitted, and the III, aVL, aVF, aVR, V2, V3, V4, V5, V6 lead electrocardiogram data is reconstructed through the I, II, V1 lead electrocardiogram data, to form a complete twelve-lead electrocardiogram; the electrocardiogram vector ring moves in the order of P ring, QRS ring and T ring, representing the comprehensive electrocardiogram vector at different times. Calculate the dot product of the electrocardiogram vector at three times and the lead vector to obtain the current I lead voltage value GI , II lead voltage value G II , V1 lead voltage value Calculate the dot product of the cardiac vector ring and the projection axis , to obtain the I lead voltage value G I ; Calculate the dot product of the cardiac vector ring and the projection axis , to obtain the II lead voltage value G II ; Calculate the dot product of the cardiac vector ring and the projection axis , to obtain the V1 lead voltage value

[0043] Step S42: Simplify the lead reconstruction problem into a mathematical relationship between known and to-be-solved lead axes; Specifically: draw a twelve-lead projection axis spherical model to obtain an approximately orthogonal relationship between I, II, and V1 lead projection axes, calculate III, aVL, aVF, and aVR lead electrocardiogram data from I and II lead electrocardiogram data; use a convolutional neural network to fit the mathematical relationship between the known I, II, and V1 lead projection axes and the to-be-solved V2, V3, V4, V5, and V6 lead projection axes, establish a global nonlinear model, and adjust the weight parameters of the known I, II, and V1 leads according to individual differences,

[0044] The convolutional neural network is used to fit the conversion relationship between the lead axes, a global nonlinear model is established, the weight parameters are adjusted according to individual differences, and the method proposed in step S41 is used to build a lead reconstruction algorithm formula to calculate V2, V3, V4, V5, and V6 lead electrocardiogram data.

[0045]

[0046] where y(t n ) is the voltage value of the to-be-solved lead at time t n , x I (t n ), x II (t n ), and are the voltage values of the known leads at time t n ; f cnn is the convolutional neural network structure used.

[0047] Step S5 specifically includes:

[0048] Step S51: Preliminarily construct a convolutional neural network model framework for lead reconstruction, apply a mean square error loss function, a ReLU activation function, and a Dropout strategy, divide the data set P1 into a training set S1 and a test set S2, and the ratio of the training set S1 and the test set S2 is about 7:3; During the training process, 20% of the model training set S1 is used as an independent validation set S3 for cross-validation.

[0049] Step S52: The training set S1 is sent to the initial model, the number of network iterations is controlled using the early stopping method, and parameter optimization is performed based on the grid search method, with the optimization range including the optimizer, learning rate, Batch size, and Dropout ratio.

[0050] In step S51, a global lead reconstruction model is designed based on a convolutional neural network as a basic framework, a convolutional neural network structure is designed for a V2 lead reconstruction model model1 of a myocardial infarction patient, an activation function is used to make the network have a nonlinear characteristic, to strengthen the learning ability of the network, and to be able to simulate complex functions, and a ReLU function is used as an activation function, which is placed after the convolutional layer and the pooling layer; the formula of the ReLU activation function is as follows:

[0051] f(x) = max(0, x)

[0052] The lead reconstruction model is based on a one-dimensional convolutional neural network, a nonlinear regression network is constructed, a fully connected layer is placed between the input layer and the convolutional layer, the number of neurons of the fully connected layer is set to 16-18, and a ReLU activation function is added to provide nonlinear changes. The data is subjected to 2-3 times of convolution and pooling operations to realize feature extraction, the size of the convolution kernel is reduced before each convolution operation, and the number of convolution kernels is increased to increase the number of feature channels; a Dropout layer is added to prevent model overfitting; a global pooling layer is used in the output layer to perform one-dimensional operation on multi-channel information, to reduce the parameter quantity while ensuring the performance of the network model; and the output unit represents the real-time prediction value of the V2 lead voltage;

[0053] In the training process of the model, 20% of the model training set S1 is used as an independent validation set S3 for cross-validation to evaluate the training effect of the model. A mean square error loss function is applied, and its expression is as follows:

[0054]

[0055] where n is the dimension of the output sample, is the value of the i-th dimension of the predicted value, and yi is the value of the i-th dimension of the true value. The mean square error loss function evaluates the difference between the predicted value and the true value, and when the difference between the predicted value and the true value is large, the loss function will be large, and a larger gradient is obtained to speed up the model training;

[0056] In step S52, part of the training set S1 is set as the verification set S3, and testing is performed on the verification set S3 after each training iteration, the error of the verification set S3 is recorded each time, and when the error of the verification set S3 no longer changes, the weight of the network model is saved, thereby effectively controlling the number of network iterations, preventing overfitting, and saving the best model; the grid search method is used to optimize part of the hyperparameters in the network training process, including the optimizer, the learning rate, the Batch Size, and the Dropout proportion; after grid search, the Adam optimizer with strong robustness, the learning rate of 0.01, the Batch Size of 512 pieces of training data per batch, and the Dropout random inactivation proportion of 0.5 are finally selected, on this basis, the network model obtains the best performance, and the network structure is optimized.

[0057] Step S6 is specifically:

[0058] Step S61: Based on a one-dimensional convolutional neural network, five lead reconstruction models are constructed for myocardial infarction and healthy control respectively, wherein the models model1, model2, model3, model4, and model5 correspond to the V2, V3, V4, V5, and V6 leads of the myocardial infarction electrocardiogram, and the models model6, model7, model8, model9, and model10 correspond to the V2, V3, V4, V5, and V6 leads of the healthy control electrocardiogram.

[0059] Step S62: The root mean square error and the correlation coefficient are used as lead reconstruction effect indicators to perform a convolution layer comparison experiment to determine the number of network convolution layers; the morphological performance of the final lead signal depends on the reconstruction ability of the original lead group by the 10 models model1, model2, model3, model4, model5, model6, model7, model8, model9, and model10.

[0060] In step S61, a lead reconstruction model is constructed for myocardial infarction and healthy control groups, each lead reconstruction model reconstructing V2, V3, V4, V5, and V6 leads, a total of 10 lead reconstruction models being designed; wherein model 1, model 2, model 3, model 4, and model 5 correspond to V2, V3, V4, V5, and V6 leads of myocardial infarction electrocardiograms, and model 6, model 7, model 8, model 9, and model 10 correspond to V2, V3, V4, V5, and V6 leads of healthy control electrocardiograms; the lead reconstruction model network structures corresponding to model 2, model 3, model 4, model 5, model 6, model 7, model 8, model 9, and model 10 are the same as that of model 1, and the parameter optimization mode is the same as that of model 1;

[0061] In step S62, the root mean square error and the correlation coefficient are used to evaluate the difference between the waveforms before and after lead reconstruction; the root mean square error is used to show the difference between the corresponding sampling points of the two groups of signals, and its formula is as follows:

[0062]

[0063] wherein X i and Y i represent the voltage amplitude of the original electrocardiogram signal at the i-th point and the voltage amplitude of the reconstructed electrocardiogram signal at the i-th point, respectively, and n is the number of sampling points of each signal;

[0064] The correlation coefficient is used to show the difference in shape between the two waveforms, and when the change trends of the two waveforms are consistent, the correlation coefficient tends to 1, and the value range of the correlation coefficient is (0, 1), and its formula is:

[0065]

[0066] The pre-classified verification set data is respectively sent into the corresponding category lead reconstruction network, four levels of convolution layer numbers are set, the root mean square error and the correlation coefficient are used as the lead reconstruction effect index, the convolution layer number comparison experiment is carried out, and the convolution layer numbers suitable for the myocardial infarction lead reconstruction model and the healthy control lead reconstruction model are respectively determined.

[0067] This invention relates to a method for global lead reconstruction in myocardial infarction applied to real-time electrocardiogram (ECG) monitoring devices, achieving the effect of constructing the morphology of other unknown leads based on known lead groups, thereby improving the accuracy of real-time monitoring of myocardial infarction. The main techniques include: (1) preprocessing the ECG signal datasets of myocardial infarction and healthy controls, performing principal component analysis on each lead of the preprocessed ECG datasets to determine the reconstructed lead groups; (2) filtering out redundant information from the original data to obtain the ECG data of the reconstructed lead groups, performing pre-classification processing, and designing the lead reconstruction algorithm based on physical realities; (3) using a convolutional neural network as the lead reconstruction model framework, optimizing the network structure and setting parameters, constructing global lead reconstruction models for myocardial infarction and healthy controls based on the optimized network, and evaluating the lead reconstruction effect; compared with existing technologies, it has the following advantages:

[0068] 1. Compared with existing technologies, this study conducts research on global lead reconstruction models. Guided by the mathematical relationships between lead transitions, a convolutional neural network is constructed. The clinical electrocardiogram characteristics of myocardial infarction are analyzed, and performance indicators for evaluating lead reconstruction effects are proposed. This demonstrates the feasibility of a lead reconstruction model with a two-level convolutional neural network, thus overcoming the limitation of the limited number of leads in portable dynamic electrocardiogram monitors.

[0069] 2. Most existing technologies use lead reconstruction algorithms with the same group of people for both training and testing sets, which are only applicable to cases where patients have already had standard twelve-lead data collected. This patent independently established and optimized a model, and based on the selected reconstruction lead group, developed a global lead reconstruction model for myocardial infarction and healthy controls that can be mounted on a portable mobile dynamic electrocardiogram monitor. Furthermore, based on model training, the consistency between the reconstructed lead signal and the real signal is also guaranteed. Attached Figure Description

[0070] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0071] Appendix Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0072] Appendix Figure 2 This is a schematic diagram of the global lead reconstruction model process according to an embodiment of the present invention. Detailed Implementation

[0073] The specific implementation details of the method of this patent will now be described in more detail with reference to the accompanying drawings. Exemplary embodiments of this disclosure are shown in the drawings; however, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. These embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0074] As shown in the figure, the myocardial infarction global lead reconstruction method applied to real-time electrocardiogram monitoring equipment comprises the following steps:

[0075] Step S1: data preprocessing is performed on the myocardial infarction electrocardiogram signal dataset and the healthy control electrocardiogram signal dataset;

[0076] Step S2: principal component analysis is performed on the processed electrocardiogram dataset P1 in step S1, a variance contribution rate graph is drawn, a highest contribution degree lead decision is made, and a reconstruction lead group is determined in combination with the electrode requirements of the dynamic monitoring equipment;

[0077] Step S3: redundant information of the original data of the lead group is filtered out, and nonlinear convolution pre-classification processing is performed on the myocardial infarction and healthy control electrocardiogram data, so as to improve the effect of the lead reconstruction algorithm;

[0078] Step S4: the lead reconstruction algorithm is designed based on physical actuality, and the conversion relationship between the leads is fitted;

[0079] Step S5: a convolutional neural network is used as a lead reconstruction model framework, and network structure optimization and parameter setting are performed;

[0080] Step S6: based on the optimized network, global lead reconstruction models for myocardial infarction and healthy control are respectively constructed, and the state difference before and after lead reconstruction is verified through root mean square error (RMSE) and correlation coefficient (CC) and the like.

[0081] Step S1 specifically comprises:

[0082] Step S11: Mallat algorithm is used to perform discrete wavelet decomposition on the original signal, and the original signal is filtered through a low-pass filter and a high-pass filter to obtain low-frequency and high-frequency signals; specifically, first, the low-frequency signal is down-sampled to obtain a new signal with half the length, which represents the low-frequency part of the original signal; second, the high-frequency signal is down-sampled to obtain a new signal with half the length, which represents the high-frequency part of the original signal, and the obtained new signal is used as an input signal; finally, the above steps are repeated until a preset number of layers is reached; through the discrete wavelet decomposition method, a group of low-frequency and high-frequency signal wavelet coefficients is obtained, and the decomposition layer number is set to eight; db6 wavelet is selected, the scale function φ(x) and the low-pass filter coefficients h0, h1, h2, h3, h4, h5 are expressed as follows:

[0083]

[0084]

[0085] Based on the scale function and low-pass filter coefficient, the wavelet base function of Daubechies wavelet is calculated; the low-pass wavelet base function ψ low (x) and the high-pass wavelet base function ψ high (x) is calculated as follows:

[0086]

[0087]

[0088] In the wavelet analysis, the decomposition process of each layer is regarded as the convolution operation of the target signal and the low-pass wavelet base function and the high-pass wavelet base function, and then the result is down-sampled to obtain the signal low-frequency part coefficient a j-1 (n) and the signal high-frequency part coefficient d j-1 (n) is calculated as follows:

[0089]

[0090]

[0091] Wherein, x j (k) represents the signal after the jth layer decomposition. The high-frequency noise in the electrocardiogram signal adopted in the embodiment mainly concentrates in D1 scale, the low-frequency baseline drift noise concentrates in A8 scale, and D5 scale and D6 scale contain most of the electrocardiogram signal;

[0092] Step S12: using the second, third, fourth, fifth, sixth and seventh layer detail coefficients to perform wavelet reconstruction, setting all the first and eighth layer detail coefficients containing noise to zero to obtain the electrocardiogram signal filtered from noise;

[0093] Step S13: using a polyphase filter to realize electrocardiogram signal down-sampling, performing corresponding filtering processing, high-frequency component removal and fixed interval extraction operation on input signals of different frequency ranges to obtain a down-sampled signal.

[0094] Step S2 is specifically:

[0095] Step S21: performing principal component analysis on the twelve leads to determine the lead with the highest effective degree in the twelve leads, projecting and mapping the original data in a higher dimensional space to a low dimensional space through linear transformation to reduce redundancy, and extracting the most effective information of the original data in the data dimension reduction process; each lead is regarded as a data dimension; m pieces of n-dimensional data are combined into a matrix X m×n ;

[0096] Step S22: performing zero-meaning on the matrix X m×n , that is, subtracting the mean value of each row to obtain a matrix Y;

[0097] Step S23: calculate the covariance matrix corresponding to Y:

[0098] C = (1 / m) * YY T

[0099] Step S24: calculate the eigenvalue a of C and the eigenvector A;

[0100] Step S25: sort the eigenvector matrix A in descending order according to the size of the eigenvalue corresponding to the matrix, and express it as a new matrix, and take the first k (k≤n) rows to form a matrix P.

[0101] Step S26: draw a variance contribution chart, and select I, II, V1 as the known leads according to the result.

[0102] Step S26 uses a portable mobile monitoring device with five electrodes, and the known leads are I, II and a chest lead.

[0103] Step S3 is specifically:

[0104] Step S31: Before the data is reconstructed, the data is pre-classified, and the preprocessed electrocardio signal is input into the first network, the input matrix channel number is 3, and the length is 1000 sequence;

[0105] Step S32: the convolution layer extracts the input vector features, the activation layer performs nonlinear transformation on the network, and after two convolution and normalization operations, the output feature vector enters the maximum pooling layer to realize the downsampling operation, and the above operation is repeated x times; use smaller convolution kernel in deeper network to extract more subtle information, and increase the number of convolution kernel to improve the channel number; after each convolution operation, the data is put into the normalization layer for processing, so that the data is quickly distributed in a reasonable interval range in a short period of time, and the network iteration speed is accelerated;

[0106] Step S33: input the feature vector into the full connection layer, and finally output two types of classification results of myocardial infarction signal and normal electrocardio signal.

[0107] Step S4 is specifically:

[0108] Step S41: mathematically analyze the relationship between the lead projection axis and the electrocardio vector ring, and obtain the real-time voltage values of I, II and V1 leads by calculating the dot product of the two; specifically:

[0109] The lead reconstruction algorithm is designed based on the physical actuality for the pre-classified myocardial infarction electrocardiogram data and the healthy control electrocardiogram data; the conversion relationship between the leads is fitted, the electrocardiogram data of III, aVL, aVF, aVR, V2, V3, V4, V5, V6 leads are reconstructed from the electrocardiogram data of I, II, V1 leads, and a complete twelve-lead electrocardiogram is formed; the electrocardiogram vector loop moves in the order of P loop, QRS loop and T loop, and represents the integrated electrocardiogram vector at different times. The dot product of the electrocardiogram vector at three times and the lead vector is calculated to obtain the current time I lead voltage value G I , the II lead voltage value G II , and the V1 lead voltage value The dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the I lead voltage value G I ; the dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the II lead voltage value G II ; and the dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the V1 lead voltage value

[0110] Step S42: Simplify the lead reconstruction problem into a mathematical relationship between the known and to-be-solved lead axes; specifically: draw a twelve-lead projection axis spherical model to obtain an approximately orthogonal relationship between the I, II, V1 lead projection axes, calculate the III, aVL, aVF, aVR lead electrocardiogram data from the I, II lead electrocardiogram data; use a convolutional neural network to fit the mathematical relationship between the known I, II, V1 lead projection axes and the to-be-solved V2, V3, V4, V5, V6 lead projection axes, establish a global nonlinear model, and adjust the weight parameters of the known I, II, V1 leads according to the body differences,

[0111] The conversion relationship between the lead axes is fitted using a convolutional neural network to establish a global nonlinear model, the weight parameters are adjusted according to the individual differences, the method based on step S41 is proposed, the lead reconstruction algorithm formula is built, and the V2, V3, V4, V5, V6 lead electrocardiogram data is calculated;

[0112]

[0113] where y(t n ) is the voltage value of the to-be-solved lead at time t n , x I (t n ), x II (t n ) and are the voltage values of the known leads at time t n ; f cnn is the convolutional neural network structure used.

[0114] Step S5 is specifically:

[0115] Step S51: A convolutional neural network model framework for realizing lead reconstruction is initially constructed, a mean square error loss function, a ReLU activation function and a Dropout strategy are applied, and the data set P1 is divided into a training set S1 and a test set S2, wherein the proportion of the training set S1 and the test set S2 is about 7:3; in the training process, 20% of the model training set S1 is taken as an independent validation set S3 for cross-validation;

[0116] Step S52: The training set S1 is sent into the initial model, the early stopping method is used to control the number of network iterations, and the parameter optimization is performed based on the grid search method, and the optimization range includes the optimizer, the learning rate, the Batch size and the Dropout proportion.

[0117] In step S51, a global lead reconstruction model is designed based on a convolutional neural network as a basic framework; a V2 lead reconstruction model model1 of a myocardial infarction patient is designed, a convolutional neural network structure is designed; an activation function is used to make the network have a nonlinear characteristic, to strengthen the network learning ability, and to be able to simulate a complex function, a ReLU function is used as an activation function, and the ReLU function is placed after a convolutional layer and a pooling layer; the formula of the ReLU activation function is as follows:

[0118] f(x) = max(0, x)

[0119] The lead reconstruction model is based on a one-dimensional convolutional neural network, a nonlinear regression network is constructed, a fully connected layer is placed between an input layer and a convolutional layer, the number of neurons of the fully connected layer is set to 16-18, and a ReLU activation function is added to provide nonlinear changes. The data is subjected to 2-3 times of convolutional pooling operations to realize feature extraction, the convolutional kernel size is reduced before each convolutional operation, and the number of convolutional kernels is increased to increase the number of feature channels; a Dropout layer is added to prevent model overfitting; a global pooling layer is used in an output layer to perform one-dimensional operation on multi-channel information, to reduce the parameter amount while ensuring the network model performance; and an output unit represents a V2 lead voltage real-time prediction value;

[0120] In the training process of the model, 20% of the model training set S1 is taken as an independent validation set S3 for cross-validation, and the model training effect is evaluated. A mean square error loss function is applied, and its expression is as follows:

[0121]

[0122] wherein n is the dimension of an output sample, is the value of the i-th dimension of a prediction value, y iis the value of the i-th dimension of the true value. The mean square error loss function evaluates the difference between the predicted value and the true value, and when the difference between the predicted value and the true value is large, the loss function will be large, and a larger gradient is obtained to speed up the model training;

[0123] In step S52, part of the training set S1 is set as the verification set S3, and testing is performed on the verification set S3 after each training iteration, and the error of the verification set S3 is recorded each time, and when the error of the verification set S3 no longer changes, the weight of the network model is saved, thereby effectively controlling the number of network iterations, preventing overfitting while saving the best model; the grid search method is used to optimize part of the hyperparameters in the network training process, including the optimizer, the learning rate, the Batch Size, and the Dropout ratio; after grid search, the Adam optimizer with strong robustness, the learning rate of 0.01, the Batch Size of 512 training data per batch, and the Dropout random inactivation ratio of 0.5 are finally selected, on this basis, the network model obtains the best performance, and the network structure optimization is realized.

[0124] Step S6 is specifically:

[0125] Step S61: Based on the one-dimensional convolutional neural network, five lead reconstruction models are constructed for myocardial infarction and healthy control respectively, wherein the models model1, model2, model3, model4 and model5 correspond to the V2, V3, V4, V5 and V6 leads of the myocardial infarction electrocardiogram, and the models model6, model7, model8, model9 and model10 correspond to the V2, V3, V4, V5 and V6 leads of the healthy control electrocardiogram;

[0126] Step S62: The root mean square error and the correlation coefficient are used as lead reconstruction effect indicators, and a convolution layer comparison experiment is performed to determine the number of network convolution layers; the morphological performance of the final lead signal depends on the reconstruction ability of the original lead group by the 10 models model1, model2, model3, model4, model5, model6, model7, model8, model9 and model10.

[0127] In step S61, a lead reconstruction model is constructed for myocardial infarction and healthy control population, each lead reconstruction model is respectively reconstructed for V2, V3, V4, V5 and V6 leads, a total of 10 lead reconstruction models are designed; wherein model 1, model 2, model 3, model 4, model 5 correspond to V2, V3, V4, V5 and V6 leads of myocardial infarction electrocardiogram, model 6, model 7, model 8, model 9 and model 10 correspond to V2, V3, V4, V5 and V6 leads of healthy control electrocardiogram; the network structure setting and parameter optimization mode of model 2, model 3, model 4, model 5, model 6, model 7, model 8, model 9 and model 10 corresponding lead reconstruction model are consistent with model 1;

[0128] In step S62, the difference degree of the waveform before and after lead reconstruction is evaluated using root mean square error and correlation coefficient; the root mean square error is used to show the difference between the corresponding sampling points of the two groups of signals, and its formula is as follows:

[0129]

[0130] Wherein X i and Y i respectively represent the voltage amplitude of the original electrocardiogram signal at the i point and the voltage amplitude of the reconstructed electrocardiogram signal at the i point, and n is the number of sampling points of each signal;

[0131] The correlation coefficient is used to show the difference in shape between two waveforms, and when the change trend of two waveforms is consistent, the correlation coefficient tends to 1, and the value range of the correlation coefficient is (0, 1), and its formula is:

[0132]

[0133] The pre-classified verification set data is respectively sent into the corresponding category lead reconstruction network, four levels of convolution layer number are set, the root mean square error and the correlation coefficient are used as the lead reconstruction effect index, the convolution layer number comparison experiment is carried out, and the convolution layer number suitable for the myocardial infarction lead reconstruction model and the healthy control lead reconstruction model is determined.

[0134] Embodiment:

[0135] Referring to FIG. Figure 1 The example provides a myocardial infarction global lead reconstruction method applied to real-time electrocardiogram monitoring equipment, which comprises data preprocessing, principal component analysis and reconstruction lead group selection, convolution pre-classification processing, lead reconstruction algorithm design, global lead reconstruction model construction and parameter optimization, and finally lead reconstruction effect evaluation.

[0136] The data preprocessing is specifically:

[0137] The electrocardiogram data is obtained from the PTB Diagnostic ECG Database, from which 80 myocardial infarction electrocardiogram records and 80 healthy control electrocardiogram records are selected, totaling 160 records, to form a database P1, wherein each record contains 16 input channels, the sampling rate is 1000 Hz, and the recording duration is 1-2 minutes. Each record is cut into 10-second segments. The database P1 is divided into a training set S1 and a test set S2, the training set S1 contains 64 myocardial infarction electrocardiogram data and 64 healthy control electrocardiogram data; the test set S2 contains 16 myocardial infarction electrocardiogram data and 16 healthy control electrocardiogram data. The patient data between the training set S1 and the test set S2 is independent and does not cross.

[0138] The original electrocardiogram signal is decomposed by the Mallat algorithm, and the original signal is filtered by a low-pass filter and a high-pass filter to obtain the low-frequency part and the high-frequency part of the signal. First, the low-frequency part signal is down-sampled to obtain a new signal with half the length, which represents the low-frequency part of the original signal; second, the high-frequency part signal is down-sampled to obtain a new signal with half the length, which represents the high-frequency part of the original signal, and the obtained new signal is used as the input signal; finally, the above steps are repeated until the preset number of layers is reached. A set of wavelet coefficients of the low-frequency part and the high-frequency part of the signal is obtained by discrete wavelet decomposition. The db6 wavelet function is selected, and the decomposition layer is set to eight. The scale function φ(x) and the low-pass filter coefficients h0, h1, h2, h3, h4, h5 are expressed as follows:

[0139]

[0140]

[0141] The wavelet basis function is calculated based on the scale function and the low-pass filter coefficients, and the low-pass wavelet basis function ψ low (x) and the high-pass wavelet basis function ψ high (x) are calculated as follows:

[0142]

[0143]

[0144] In wavelet analysis, the decomposition process of each layer can be regarded as the convolution operation of the target signal with the low-pass wavelet basis function and the high-pass wavelet basis function, and then the result is down-sampled to obtain the signal low-frequency part coefficient a j-1 (n) and the signal high-frequency part coefficient d j-1 (n) are calculated as follows:

[0145]

[0146]

[0147] wherein x j (k) represents the signal after the decomposition of the jth layer. The high-frequency noise in the electrocardiogram signal in the embodiment is mainly concentrated in D1 scale, the low-frequency baseline drift noise is concentrated in A8 scale, and D5 scale and D6 scale contain most of the electrocardiogram signal. Therefore, the second, third, fourth, fifth, sixth and seventh layer detail coefficients are used for wavelet reconstruction in the embodiment, and the first and eighth layer detail coefficients containing noise are all set to zero, so as to obtain the electrocardiogram signal filtered from noise.

[0148] The polyphase filter is used to realize the down-sampling of the electrocardiogram signal, and the input signal in different frequency ranges is subjected to corresponding filtering processing, high-frequency component removal and fixed interval extraction operation, so as to obtain the down-sampled signal.

[0149] The lead group selection of the principal component analysis and reconstruction is specifically:

[0150] The principal component analysis (PCA) method is used to determine the lead with the highest effective degree in the twelve leads. The original data in a higher dimensional space is projected and mapped onto a low-dimensional space by a linear transformation method, so as to reduce the redundancy and extract the most effective information of the original data in the process of data dimension reduction. The specific steps of the principal component analysis method include: regarding each lead as a data dimension, combining m pieces of n-dimensional data into a matrix X m×n ; the matrix X m×n is subjected to zero mean, that is, the mean value of each row is subtracted to obtain a matrix Y; the covariance matrix C corresponding to Y is calculated, and the expression is as follows:

[0151] C=(1 / m)*YY T

[0152] After obtaining the covariance matrix C, the eigenvalue a and the eigenvector A of C are calculated; the eigenvector matrix A is sequentially sorted from top to bottom according to the size of the eigenvalue of the matrix corresponding to the eigenvector to form a new matrix, and the first l (l ≤ n) rows are taken to form a matrix P; a variance contribution rate graph is drawn, when the number of leads k is 4, the reduced data can contribute nearly 99% of the information, and the first three principal components have contributed 98% of the information. Since the four leads III, aVL, aVF and aVR can be calculated from the I and II leads data, and most portable mobile monitoring devices are generally composed of five electrodes, therefore, it is most appropriate to select the I, II and V1 leads as known leads for lead reconstruction.

[0153] The convolution pre-classification process is specifically:

[0154] In order to improve the effect of the lead reconstruction algorithm, based on a one-dimensional convolutional neural network, the pre-processed electrocardio signal is pre-classified before the lead reconstruction of the electrocardio signal data. The input matrix of the network is a sequence with a channel number of 3 and a length of 1000, the convolution kernel size used is gradually reduced, and the signal is gradually convolved, and smaller convolution kernels are used in deeper network layers to extract more subtle information. After each convolution operation, the data is put into the normalization layer for processing, so that the data is rapidly distributed in a reasonable interval range in the short term, and the network iteration speed is accelerated.

[0155] After two convolution and normalization operations, the output feature vector enters the max-pooling layer to realize the downsampling operation, and the above operation is repeated x times. With the network deepening, the number of convolution kernels is increased to increase the number of channels, and at the same time, the data of each channel is continuously reduced, and more attention is paid to subtle information. Finally, the multi-channel feature signals extracted in the network are mapped to the output layer through the full connection layer, and two classification results of myocardial infarction and healthy control are obtained.

[0156] The cross-validation method is used, the training set S1 is used to train the pre-classification model based on the one-dimensional convolutional neural network, the data of the test set S2 is put into the trained pre-classification model, and the model performance is evaluated according to the accuracy, sensitivity and specificity three indexes, and the optimal pre-classification model is selected.

[0157] The lead reconstruction algorithm design is specifically:

[0158] The lead reconstruction algorithm is designed based on the physical actuality for the pre-classified myocardial infarction electrocardiogram data and the healthy control electrocardiogram data. The conversion relationship between the leads is fitted, and the electrocardiogram data of III, aVL, aVF, aVR, V2, V3, V4, V5 and V6 leads are reconstructed from the electrocardiogram data of I, II and V1 leads to form a complete twelve-lead electrocardiogram. The electrocardiogram vector loop moves in the order of P loop, QRS loop and T loop, representing the integrated electrocardiogram vector at different times. The dot product of the electrocardiogram vector and the lead vector at three times is calculated to obtain the current time I lead voltage value G I , the II lead voltage value G II and the V1 lead voltage value The dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the I lead voltage value G I ; the dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the II lead voltage value G II ; and the dot product of the electrocardiogram vector loop and the projection axis is calculated to obtain the V1 lead voltage value

[0159] A twelve-lead projection axis spherical model is drawn to obtain an approximately orthogonal relationship among the I, II and V1 lead projection axes, and the electrocardiogram data of III, aVL, aVF and aVR leads are calculated from the electrocardiogram data of I and II leads. A convolutional neural network is used to fit the mathematical relationship between the known I, II and V1 lead projection axes and the to-be-solved V2, V3, V4, V5 and V6 lead projection axes, to establish a global nonlinear model, adjust the weight parameters of the known I, II and V1 leads according to the human body differences, and build the lead reconstruction algorithm formula as shown below to calculate the electrocardiogram data of V2, V3, V4, V5 and V6 leads:

[0160]

[0161] where y(t n ) is the voltage value of the to-be-solved lead at time t n , x I (t n ), x II (t n ) and are the voltage values of the known leads at time t n , and f cnn is the convolutional neural network structure used.

[0162] The global lead reconstruction model construction and parameter optimization are specifically as follows:

[0163] Take the convolutional neural network as the basic framework, design the global lead reconstruction model. Take the V2 lead reconstruction model model1 of the myocardial infarction patient as an example, design the convolutional neural network structure. The activation function makes the network have nonlinear characteristics, enhances the network learning ability, and can simulate complex functions. The formula of ReLU activation function is as follows:

[0164] f(x)=max(0,x)

[0165] ReLU activation function will not be affected by the extreme value of the input, that is, the gradient will not disappear completely, which effectively solves the problem of gradient disappearance; At the same time, ReLU activation function has the advantages of simple form and fixed gradient, which is conducive to the convergence of network and the improvement of convergence speed. In this embodiment, ReLU function is used as the activation function, and ReLU function is placed after the convolutional layer and the pooling layer.

[0166] The dropout method changes the number of neurons by setting a part of intermediate nodes to zero in each model training process to build a new neural network, so as to reduce the degree of network model forced fitting input data due to unreasonable parameter setting, thereby improving the overall performance of the network. In this embodiment, the dropout strategy is used to randomly inactivate the intermediate layer nodes in the lead reconstruction model to prevent overfitting.

[0167] In this embodiment, a one-dimensional convolutional neural network is used to build a nonlinear regression network. A fully connected layer is placed between the input layer and the convolutional layer, the number of neurons of the fully connected layer is set to 16-18, and ReLU activation function is added to provide nonlinear change. 2-3 times of convolution and pooling operation is performed on the data to realize feature extraction, the size of the convolution kernel is reduced before each convolution operation, and the number of convolution kernels is increased to increase the number of feature channels. A dropout layer is added to prevent model overfitting. A global pooling layer is used in the output layer to one-dimensionalize the multi-channel information, reduce the parameter quantity, and ensure the performance of the network model. The output unit represents the real-time prediction value of V2 lead voltage.

[0168] During the training of the model, 20% of the model training set S1 is used as an independent validation set S3 for cross-validation to evaluate the training effect of the model. The mean square error loss function is used, and its expression is as follows:

[0169]

[0170] Where n is the dimension of the output sample, is the value of the i-th dimension of the predicted value, y i is the value of the i-th dimension of the true value. The mean square error loss function evaluates the difference between the predicted value and the true value. When the difference between the predicted value and the true value is large, the loss function will be large, and a larger gradient will be obtained to speed up the model training.

[0171] Using early stopping method as the strategy of model training, part of the training set S1 is set as the validation set S3, test on the validation set S3 after each training iteration, record the error of the validation set S3 each iteration, save the weight of the network model when the error of the validation set S3 no longer changes, thereby effectively controlling the number of network iterations, preventing overfitting while saving the best model.

[0172] The grid search method refers to the method of obtaining the optimal parameter value by comparing the results obtained by enumerating the parameter values in a certain range. In this embodiment, the grid search method is used to optimize some hyperparameters in the network training process, including the optimizer, learning rate, Batch Size, and Dropout ratio. After grid search, the Adam optimizer with strong robustness, learning rate of 0.01, Batch Size of 512 training data per batch, and Dropout random inactivation ratio of 0.5 are finally selected. On this basis, the network model obtains the best performance, and the network structure optimization is realized.

[0173] The lead reconstruction effect evaluation is specifically:

[0174] To improve the accuracy of lead reconstruction, this embodiment is based on the optimized one-dimensional convolutional neural network, and lead reconstruction models are constructed for myocardial infarction and healthy control groups respectively. Each lead reconstruction model reconstructs V2, V3, V4, V5, and V6 leads, a total of 10 lead reconstruction models are designed. Among them, model1, model2, model3, model4, and model5 correspond to V2, V3, V4, V5, and V6 leads of myocardial infarction electrocardiogram, and model6, model7, model8, model9, and model10 correspond to V2, V3, V4, V5, and V6 leads of healthy control electrocardiogram. The network structure setting and parameter optimization method of model2, model3, model4, model5, model6, model7, model8, model9, and model10 corresponding lead reconstruction model are consistent with model1.

[0175] The root mean square error and the correlation coefficient are used to evaluate the difference between the waveforms before and after lead reconstruction. The root mean square error can show the difference between the corresponding sampling points of the two groups of signals, and its formula is as follows:

[0176]

[0177] Where X i and Y irespectively, n is the number of sampling points of each signal.

[0178] The correlation coefficient can show the difference in shape between two waveforms. When the change trend of two waveforms is consistent, the correlation coefficient tends to 1. The value range of the correlation coefficient is (0, 1), and the formula is:

[0179]

[0180] The electrocardiogram sequence used in this embodiment is a relatively simple time series, so too many convolutional neural network layers may not be suitable. The pre-classified validation set data is respectively sent to the corresponding lead reconstruction network, four levels of convolutional layer number are set, the root mean square error and the correlation coefficient are used as the lead reconstruction effect index, the convolutional layer number comparison experiment is carried out, and the convolutional layer number suitable for the myocardial infarction lead reconstruction model and the healthy control lead reconstruction model is determined respectively.

[0181] The root mean square error of the myocardial infarction lead reconstruction model with different convolutional layer numbers is compared. The root mean square error of the myocardial infarction six-layer convolutional model reconstruction result is 0.224, which is the highest value in all models. The root mean square error of the myocardial infarction two-layer convolutional model reconstruction result is 0.101. After comparison, the lead reconstruction scheme with more layers is abandoned. According to the overall experimental results, the two-layer convolutional model has a whole root mean square error value of 0.078, and the model effect is optimal.

[0182] The correlation coefficients of the myocardial infarction lead reconstruction model with different convolutional layer numbers are compared. According to the overall experimental results, the two-layer convolutional model has a whole correlation coefficient of 0.937, so that the lead reconstruction model has the largest correlation coefficient value. Through the evaluation of the root mean square error and the correlation coefficient, it is determined that the best convolutional layer number of the myocardial infarction lead reconstruction model is two layers.

[0183] The root mean square error and the correlation coefficient of the healthy control lead reconstruction model with different convolutional layer numbers are compared according to the above method. According to the overall experimental results, the two-layer convolutional model has a whole root mean square error value of 0.078 and a whole correlation coefficient of 0.951, and the model effect is optimal.

[0184] After determining the number of network convolutional layers of the myocardial infarction lead reconstruction model and the healthy control lead reconstruction model, the overall framework of the lead reconstruction model is completed. The electrocardiosignal with I, II and V1 lead channels is sent into the corresponding category lead reconstruction model to obtain the to-be-solved V2, V3, V4, V5 and V6 lead electrocardiosignal. The shape performance of the final lead signal depends on the reconstruction ability of the 10 models, i.e., model1, model2, model3, model4, model5, model6, model7, model8, model9 and model10, on the original lead group, the reconstructed twelve-lead electrocardiosignal provides great convenience for the portable dynamic electrocardiosignal monitor with limited lead number, helps to realize all-weather real-time electrocardiosignal monitoring, and improves the early prediction ability of myocardial infarction disease. Those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations fall within the scope of the claims of the present application and equivalent technologies thereof, the present application also intends to include these modifications and variations.

Claims

1. A method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring devices, characterized by: Comprise the following steps: Step S1: data preprocessing of myocardial infarction electrocardio signal data set and healthy control electrocardio signal data set; Step S2: principal component analysis of each lead of the processed electrocardio data set P1 in step S1, draw the variance contribution rate graph, realize the highest contribution lead decision, combine the electrode requirement of dynamic monitoring equipment, determine the reconstruction lead group; Step S3: filter out the redundant information of the original data of the lead group, and perform nonlinear convolution pre-classification processing on the myocardial infarction and healthy control electrocardio data to improve the effect of lead reconstruction algorithm; Step S4: design the lead reconstruction algorithm based on physical reality, and fit the conversion relationship between the leads; Step S5: take the convolutional neural network as the lead reconstruction model framework, and perform network structure optimization and parameter setting; Step S6: based on the optimized network, global lead reconstruction models for myocardial infarction and healthy control are respectively constructed, and the state difference before and after lead reconstruction is evaluated and verified through root mean square error RMSE and correlation coefficient CC; Step S4 specifically comprises: Step S41: mathematically analyze the relationship between lead projection axis and electrocardio vector ring, and respectively obtain the real-time voltage values of leads I, II and V1 by calculating the dot product of the two; Specifically: Based on the physical reality, the lead reconstruction algorithm is designed for the pre-classified myocardial infarction electrocardio data and healthy control electrocardio data; The conversion relationship between the leads is fitted, and the electrocardio data of leads III, aVL, aVF, aVR, V2, V3, V4, V5 and V6 is reconstructed from the electrocardio data of leads I, II and V1 to form a complete twelve-lead electrocardiogram; The electrocardio vector ring moves in the order of P ring, QRS ring and T ring, representing the comprehensive electrocardio vector at different times; The dot product of the three time instants electrocardio vector and lead vector is calculated to obtain the current time instants I lead voltage value G I , II lead voltage value G II , V1 lead voltage value The dot product of the electrocardio vector ring and the projection axis is calculated to obtain the I lead voltage value G I The dot product of the electrocardio vector ring and the projection axis is calculated to obtain the II lead voltage value G II The dot product of the electrocardio vector ring and the projection axis is calculated to obtain the V1 lead voltage value Step S42: simplify the lead reconstruction problem into the mathematical relationship between the known and to-be-solved lead axes; Specifically: Draw the twelve-lead projection axis spherical model to obtain the approximately orthogonal relationship between the projection axes of leads I, II and V1, calculate the electrocardio data of leads III, aVL, aVF and aVR from the electrocardio data of leads I and II; Use convolutional neural network to fit the mathematical relationship between the known projection axes of leads I, II and V1 and the to-be-solved projection axes of leads V2, V3, V4, V5 and V6, establish a global nonlinear model, and adjust the weight parameters of the known leads I, II and V1 according to the individual differences; The conversion relationship between the lead axes is fitted by using convolutional neural network to establish a global nonlinear model, adjust the weight parameters according to the individual differences, and build the lead reconstruction algorithm formula based on the method proposed in step S41 to calculate the electrocardio data of leads V2, V3, V4, V5 and V6; Where y(t) n ) is the lead to be determined at t n Voltage value at time x I (t n ), x II (t n )and Given leads at t n Voltage value at time f cnn This refers to the convolutional neural network structure used.

2. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 1, characterized in that: Step S1 specifically comprises: Step S11: use Mallat algorithm to perform discrete wavelet decomposition on the original signal, filter the original signal through low-pass filter and high-pass filter to obtain the low-frequency part and high-frequency part of the signal; Specifically: First, the low-frequency part signal is down-sampled to obtain a new signal of half length, which represents the low-frequency part of the original signal; second, the high-frequency part signal is down-sampled to obtain a new signal of half length, which represents the high-frequency part of the original signal, and the obtained new signal is taken as an input signal; finally, the above steps are repeated until a preset number of layers is obtained; a group of wavelet coefficients of the low-frequency part and the high-frequency part signals are obtained through a discrete wavelet decomposition mode; The decomposition layer number is set to eight; the db6 wavelet is selected, and the scale function φ(x) and the low-pass filter coefficients h0, h1, h2, h3, h4, h5 are expressed as follows: Based on the scale function and the low-pass filter coefficients, the wavelet basis function of the Daubechies wavelet is calculated; Low-pass wavelet basis function ψ low (x) and high-pass wavelet basis function ψ high (x) is calculated as follows: In wavelet analysis, the decomposition process of each layer is regarded as the convolution operation of the target signal with the low-pass wavelet base function and the high-pass wavelet base function, and then the result is down-sampled to obtain the signal low-frequency part coefficient a j-1 (n) and the signal high-frequency part coefficient d j-1 (n), whose calculation formula is as follows: wherein x j (k) represents the signal after the jth layer decomposition; the high-frequency noise in the adopted electrocardiogram signal mainly appears at the D1 scale, the low-frequency baseline drift noise is concentrated at the A8 scale, and the D5 scale and the D6 scale contain most of the electrocardiogram signal; Step S12: wavelet reconstruction is performed by using the second, third, fourth, fifth, sixth and seventh layer detail coefficients, and the first and eighth layer detail coefficients containing noise are all set to zero, so as to obtain the electrocardio signal filtered from noise; Step S13: the multi-phase filter is used to realize the down-sampling of the electrocardio signal, the input signals in different frequency ranges are subjected to corresponding filter processing, high-frequency component removal and fixed interval extraction operations, and the down-sampled signal is obtained.

3. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 1, characterized in that: Step S2 is specifically: Step S21: principal component analysis is performed on the twelve leads, the lead with the highest effective degree in the twelve leads is determined, the original data in the higher dimensional space is projected and mapped onto the low dimensional space through the linear transformation method, the redundancy is reduced, and the most effective information of the original data is extracted in the data dimension reduction process; each lead serves as a data dimension; m pieces of n-dimensional data are combined into a matrix X m×n ; Step S22: Subtracting the mean value of each row from the matrix X m×n Zero-meaning is performed, i.e. each row is subtracted by the mean value of this row, resulting in a matrix Y; Step S23: the covariance matrix corresponding to Y is calculated: C = (1 / m) * YY T Step S24: the eigenvalue a and the eigenvector A of C are calculated; Step S25: the eigenvector matrix A is sequentially sorted from top to bottom according to the size of the eigenvalues of the matrix corresponding to the eigenvector matrix A to form a new matrix, and the first k (k≤n) rows are taken to form a matrix P; Step S26: a variance contribution graph is drawn, and the results are selected according to the results to select three groups of leads I, II and V1 as known leads.

4. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 3, characterized in that: Step S26 uses a portable mobile monitoring device with five electrodes, and the known leads are lead I, lead II and a chest lead.

5. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 3, characterized in that: Step S3 is specifically: Step S31: before the data is subjected to lead reconstruction, the data is pre-classified, the preprocessed electrocardio signal is input into the first level network, the input matrix channel number is 3, and the sequence length is 1000; Step S32: the convolution layer extracts the input vector features, the activation layer performs nonlinear transformation on the network, after two convolution and normalization operations, the output feature vector enters the maximum pooling layer to realize the down-sampling operation, and the above operation is repeated x times; smaller convolution kernels are used in deeper networks to extract more subtle information, and the number of convolution kernels is increased to increase the channel number; after each convolution operation, the data is put into the normalization layer for processing, so that the data is rapidly distributed in a reasonable interval range in a short period of time, and the network iteration speed is accelerated; Step S33: the feature vector is input into the full connection layer, and finally the network outputs two types of classification results of myocardial infarction signals and normal electrocardio signals.

6. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 1, characterized in that: Step S5 is specifically: Step S51: initially construct a convolutional neural network model framework for realizing lead reconstruction, apply a mean square error loss function, a ReLU activation function and a Dropout strategy, divide the data set P1 into a training set S1 and a test set S2, wherein the proportion of the training set S1 and the test set S2 is 7:3; in the training process, 20% of the model training set S1 is taken as an independent validation set S3 for cross-validation; Step S52: send the training set S1 into the initial model, use the early stopping method to control the number of network iterations, and use the grid search method to optimize the parameters, and the optimization range includes the optimizer, the learning rate, the Batch size and the Dropout proportion.

7. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 6, characterized in that: In step S51, a global lead reconstruction model is designed based on a convolutional neural network as a basic framework; a convolutional neural network structure is designed for a V2 lead reconstruction model model1 of a myocardial infarction patient; an activation function is used to make the network have a nonlinear characteristic, to strengthen the network learning ability, and to be able to simulate a complex function, and a ReLU function is used as an activation function, which is placed after a convolutional layer and a pooling layer; the formula of the ReLU activation function is as follows: f(x) = max(0, x) The lead reconstruction model is based on a one-dimensional convolutional neural network, a nonlinear regression network is constructed, a fully connected layer is placed between an input layer and a convolutional layer, the number of neurons of the fully connected layer is set to 16-18, and a ReLU activation function is added to provide nonlinear changes; 2-3 convolutional pooling operations are performed on the data to realize feature extraction, the size of the convolutional kernel is reduced before each convolutional operation, and the number of convolutional kernels is increased to increase the number of feature channels; a Dropout layer is added to prevent model overfitting; a global pooling layer is used in the output layer to perform one-dimensional operation on multi-channel information, to reduce the parameter amount while ensuring the performance of the network model; and an output unit represents a real-time prediction value of the V2 lead voltage; In the training process of the model, 20% of the model training set S1 is taken as an independent validation set S3 for cross-validation to evaluate the training effect of the model; a mean square error loss function is applied, and its expression is as follows: where n is the dimension of the output sample, yi is the value of the i-th dimension of the predicted value, i yi is the value of the i-th dimension of the true value; the mean square error loss function evaluates the difference between the predicted value and the true value, and when the difference between the predicted value and the true value is large, the loss function will be large, and a larger gradient is obtained to speed up the model training; In step S52, part of the training set S1 is set as the validation set S3, testing is performed on the validation set S3 after each training iteration, the error of the validation set S3 is recorded each time, the weights of the network model are saved when the error of the validation set S3 no longer changes, thereby effectively controlling the number of network iterations, preventing overfitting while saving the best model; the grid search method is used to optimize some hyperparameters in the network training process, including the optimizer, the learning rate, the Batch size and the Dropout proportion; After grid search, the Adam optimizer with strong robustness, the learning rate of 0.01, the Batch size of 512 for each batch of training data, and the Dropout random inactivation proportion of 0.5 are finally selected, on this basis, the network model obtains the best performance, and the network structure is optimized.

8. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 1, characterized in that: Step S6 is specifically: Step S61: Based on a one-dimensional convolutional neural network, five lead reconstruction models are constructed for myocardial infarction and healthy controls, respectively, wherein model1, model2, model3, model4, and model5 correspond to V2, V3, V4, V5, and V6 leads of myocardial infarction electrocardiogram, and model6, model7, model8, model9, and model10 correspond to V2, V3, V4, V5, and V6 leads of healthy control electrocardiogram; Step S62: Using the root mean square error and the correlation coefficient as the lead reconstruction effect index, a convolution layer comparison experiment is performed to determine the number of network convolution layers; the morphological performance of the final lead signal depends on the reconstruction ability of the original lead group by the 10 models model1, model2, model3, model4, model5, model6, model7, model8, model9, and model10.

9. The method for myocardial infarction global lead reconstruction applied to real-time electrocardiogram monitoring device according to claim 8, characterized in that: In step S61, lead reconstruction models are constructed for myocardial infarction and healthy control populations, and each lead reconstruction model reconstructs V2, V3, V4, V5, and V6 leads, a total of 10 lead reconstruction models are designed; wherein model1, model2, model3, model4, and model5 correspond to V2, V3, V4, V5, and V6 leads of myocardial infarction electrocardiogram, and model6, model7, model8, model9, and model10 correspond to V2, V3, V4, V5, and V6 leads of healthy control electrocardiogram; the network structure settings and parameter optimization methods of the lead reconstruction models corresponding to model2, model3, model4, model5, model6, model7, model8, model9, and model10 are consistent with model1; In step S62, the root mean square error and the correlation coefficient are used to evaluate the difference between the waveforms before and after lead reconstruction; the root mean square error is used to show the difference between the corresponding sampling points of the two groups of signals, and its formula is as follows: wherein X i and Y i respectively represent the voltage amplitude of the original ECG signal at the i-th point and the voltage amplitude of the reconstructed ECG signal at the i-th point, and n is the number of sampling points of each signal; The correlation coefficient is used to show the difference in morphology between the two waveforms, and when the change trends of the two waveforms are consistent, the correlation coefficient tends to 1, and the value range of the correlation coefficient is (0, 1), and its formula is: The pre-classified verification set data is sent into the corresponding lead reconstruction network, four levels of convolution layer number are set, the root mean square error and the correlation coefficient are used as the lead reconstruction effect index, a convolution layer comparison experiment is performed, and the number of convolution layers suitable for the myocardial infarction lead reconstruction model and the healthy control lead reconstruction model is determined, respectively.

Images