CN111990989A - Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network - Google Patents

Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network Download PDF

Info

Publication number
CN111990989A
CN111990989A CN202010894295.7A CN202010894295A CN111990989A CN 111990989 A CN111990989 A CN 111990989A CN 202010894295 A CN202010894295 A CN 202010894295A CN 111990989 A CN111990989 A CN 111990989A
Authority
CN
China
Prior art keywords
data
network
neural network
model
convolution
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010894295.7A
Other languages
Chinese (zh)
Inventor
刘娟
胡鹏
冯晶
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN202010894295.7A priority Critical patent/CN111990989A/en
Publication of CN111990989A publication Critical patent/CN111990989A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Signal Processing (AREA)
  • Fuzzy Systems (AREA)
  • Evolutionary Computation (AREA)
  • Pathology (AREA)
  • Mathematical Physics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention provides a single-lead electrocardio abnormal signal identification method based on a generation countermeasure network and a convolution cyclic neural network, which mainly solves the problem of unbalanced samples in a data set, enhances the data of the category with less data volume in the data set, and then identifies and classifies the electrocardio abnormal signal to assist in providing reference for doctors, reduce misdiagnosis and missed diagnosis rate and reduce the workload of the doctors; the application of the generated countermeasure network enables samples in the data set to reach relative balance, so that training of the convolution cyclic neural network is performed, and a better classification effect is achieved.

Description

Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network
Technical Field
The invention relates to the field of electrocardiosignal identification and classification, in particular to a single-lead electrocardio abnormal signal identification method based on a generation countermeasure network and a convolution cyclic neural network.
Background
Cardiovascular disease (CVD) refers to a series of diseases associated with the heart or blood vessels, also known as circulatory diseases. For the diagnosis of heart electrical diseases, Electrocardiogram (ECG or EKG) is a diagnostic technique for recording the electrophysiological activity of the heart in time units through the thorax, capturing its electrical signals by electrodes placed on the skin and plotting them into lines. As a non-invasive recording mode, the application of the electrocardiogram is the most extensive and authoritative.
In recent years, the level of technologies such as fuzzy recognition, artificial intelligence, and neural networks has been increasing. With the development of big data and artificial intelligence, the research on the automatic electrocardiogram diagnosis algorithm and system designed based on electrocardiogram signal data has been more in recent years, but most of the results still stay in the experimental stage, and a great distance is still needed to be left for the real commercial investment. Even if the part is put into commercial use, the precision is insufficient, the disease discrimination is not specific enough, and the like, so that the help of doctors is very limited.
The PDF of a 12-lead electrocardiogram is currently the most readily available data to hospitals or physicians. However, for some cardioelectric types of disease, the small sample size is a significant problem. The existing problems comprise the problem of unbalanced samples in a data set, and the prediction effect of a model on a certain type of electrocardio diseases is not ideal due to the unbalanced samples in the data set, so that the patent provides a single-lead electrocardio abnormal signal identification method based on generation of an antagonistic network and a convolutional recurrent neural network.
Disclosure of Invention
The invention aims to provide a single-lead electrocardio abnormal signal identification method based on a generation countermeasure network and a convolution cyclic neural network, which mainly solves the problem of unbalanced samples in a data set, performs data enhancement on the categories with less data volume in the data set, and then performs identification classification on electrocardio abnormal signals, assists in providing reference for doctors, reduces misdiagnosis and missed diagnosis rate, and reduces the workload of the doctors; the application of the generated countermeasure network enables samples in the data set to reach relative balance, so that training of the convolution cyclic neural network is performed, and a better classification effect is achieved.
Specifically, the invention provides a single-lead electrocardio abnormal signal identification method based on a generation countermeasure network and a convolution cyclic neural network, which comprises the following steps:
step 1, denoising electrocardiosignal data in an actual data set;
step 2, generating the electrocardiogram data with a small number of samples in the actual data set by using a generator of the anti-network DCGAN generated by deep convolution, and solving the problem of unbalanced samples in the actual data set, so as to perform data enhancement on the sample types with small data volume in the actual data set to obtain a final training set;
step 3, constructing a convolution cyclic neural network model, wherein the model comprises three parts: the first part is a convolutional neural network CNN used for capturing detail characteristics in the electrocardiogram data, and comprises 5 groups of convolutional layers, a BN layer, a LeakyReLU layer, a Dropout layer and a pooling layer; the second part is a bidirectional long and short term memory network (BilSTM) which is used for processing the time sequence relation in the characteristic sequence output by the convolutional neural network; the third part is an attention model, namely an attention model, and is used for dividing the weight of an output signal of the bidirectional long-short term memory network BilSTM;
step 4, circularly training a convolution cyclic neural network model by using a back propagation algorithm and adopting cross validation, wherein the total loss function of the convolution cyclic neural network model is defined as the sum of cross entropy loss functions of all the electrocardiographic data participating in training, and the labels and the prediction results of the single electrocardiographic data are respectively label and prediction results of labeliAnd predictioniThen the model total loss function is defined as follows:
Figure BDA0002657933790000031
where n represents the total number of samples on the training set and cross _ entry () represents the cross entropy loss function, predictioniThe model prediction value, label, of the ith electrocardiogram data in the training set is representediPresentation trainingConcentrating the label value of the ith electrocardiogram data;
and 5, classifying the test set by using the trained convolution cyclic neural network model.
Further, in step 2, the loss function of DCGAN is as follows,
Figure BDA0002657933790000032
wherein G represents a generator, D represents a discriminator, and z represents random noise, and an image is generated from the random noise and denoted as G (z); pdata(x) P (x | data) represents the probability of obtaining x from an actual data set, which is an electrocardiographic data set with a small number of samples, and PZP (Z) ═ P (Z | Z) represents the probability of obtaining Z from the generated dataset, which is generated using a dataset generated using DCGAN, and discriminator D is a neural network used to evaluate its authenticity, that is, the discriminator is used to discriminate whether each instance of data belongs to the real training dataset, where the discriminator input is an image x, and the output D (x) represents the probability that x is the real image, the image x being from the real dataset and the generated dataset.
Furthermore, in step 3, the calculation formula of the output value of the convolutional layer is,
Figure BDA0002657933790000041
wherein x isi，jAn ith row and a jth column element representing an image; wm，nRepresents the weight of the mth row and the nth column; wbA bias term representing filter; a isi，jAn ith row and a jth column element representing the feature map; f denotes an activation function, fhHigh, f, representing filterwIndicating the width of the filter.
Furthermore, in step 3, the updating calculation formula of the t-th step in the bidirectional LSTM is as follows,
it＝σ(Wixt+Uiht-1+bi)
ft＝σ(Wfxt+Ufht-1+bf)
ot＝σ(Woxt+Uoht-1+bo)
Figure BDA0002657933790000042
Figure BDA0002657933790000044
ht＝ot⊙Tanh(ct)
wherein t is the number of cycles, Wf、Wi、WC、WoTraining parameters which are all threshold recurrent neural networks, bf、bi、bC、boσ (x) is a first activation function, and tanh (x) is a second activation function, which are bias parameters of the recurrent neural network; i.e. itIs an input gate, ftIs a forgetting door otIs an output gate, ctIs an internal memory cell, xtFor the input signal, i.e. the extracted signature sequence from CNN, ht-1For the hidden layer output of the previous step, Wi、UiTwo matrices and vector biFor inputting the parameters of the gate, htIs the output of the hidden layer in step t.
Furthermore, in step 3, the attention model divides the weight of the output signal of the bidirectional long-short term memory network BiLSTM into the following specific implementation manner,
si＝f(si-1，yi-1，ci)
p(yi|y1，y2，......，yi-1)＝g(yi-1，si，ci)
Figure BDA0002657933790000043
Figure BDA0002657933790000051
eij＝a(si-1，hj)
wherein h isjIs a hidden state corresponding to the input state, xjI.e. the corresponding characteristic sequence, y, extracted from the CNN modeliIs a state in the output sequence, siIs a corresponding hidden state, ciIs the complete hidden state h of the input sequence1，h2，......，hTT represents the number of all hidden states of the input sequence, aijIs the attention weight parameter, eijIs xiAnd yiAligned value, p denotes given y1，y2，......，yi-1In the case of (1), the value yiG denotes a non-linear transformation and a refers to a neural network.
Further, in step 4, BatchNormalization is performed on each convolutional layer to accelerate the training of the circularly trained convolutional neural network model, and a LeakyRelu activation function is used to alleviate the gradient vanishing problem.
Further, in the step 1, denoising processing is performed on the electrocardiosignal data in the actual data set by adopting a denoising method combining an integrated empirical mode decomposition algorithm and a wavelet soft threshold.
Compared with the prior art, the invention has the following advantages:
1. in practical application, the sample data volume of certain electrocardio-type diseases is less, so that the problem of unbalanced sample in data set can occur. The invention proposes to use GAN (generation countermeasure network) for data enhancement, thereby solving the problem of unbalanced sample data in the data set.
2. For faster model prediction, the convolution operation is performed using one-dimensional CNN, which is easier to train, with only a few tens of Back Propagation (BP) periods.
ECG signals are a time series of data, so we consider the extracted features to also have time series characteristics. The invention refers to a semantic recognition method, and puts the characteristics captured by the CNN into a bidirectional cyclic neural network.
4. The invention improves the accuracy and provides reliable assistance and reference for medical personnel. Through repeated training of a large amount of data and continuous optimization of the algorithm, the accuracy of the abnormal electrocardio identification and classification is improved to a certain extent, reliable assistance and reference are provided for medical personnel, and misdiagnosis and missed diagnosis rates are reduced.
Drawings
FIG. 1 is a schematic flow chart of the present invention;
FIG. 2 is a comparison graph of the original data of the electrocardiograph signal and the de-noised data;
fig. 3 is a graph comparing the raw electrocardiographic signal data with electrocardiographic data generated by DCGAN, where (a) is the raw data and (b) is the generated data.
FIG. 4 is a schematic diagram of a GAN network structure;
FIG. 5 is a schematic diagram of a DCGAN network generator;
fig. 6 is a diagram of a convolutional recurrent neural network structure according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments, features and aspects of the present invention will be described in detail below with reference to the accompanying drawings. In the drawings, like reference numbers can indicate functionally identical or similar elements. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.
Specifically, the invention provides a method for identifying a single-lead electrocardiogram abnormal signal based on a generated countermeasure network and a convolution cyclic neural network, taking a kaggle data set as an example, as shown in fig. 1, and the method comprises the following steps:
the method comprises the following steps: denoising the data; before and after denoising, as shown in fig. 2, the electrocardiosignal is a bioelectric signal collected from the body surface of a human body, and has the commonality of the bioelectric signal: weak amplitude, low frequency, large impedance, randomness and the like, most energy of electrocardiosignals is concentrated at 0.05-100 Hz, QRS complex energy is concentrated at 5-45 Hz, and P, T wave frequency is generally below 10 Hz. Three kinds of interference mainly exist in the electrocardio data, namely 50Hz power frequency interference; the base line drifts, the frequency range is usually between 0.15 Hz and 0.3Hz, and sometimes reaches 1 Hz; myoelectric interference and wide frequency range. The method based on empirical mode decomposition is the best selection for baseline wander correction, and the best processing mode of power frequency interference is to use an equiripple notch filter; for eliminating electromyographic interference and motion artifacts, wavelet transformation is generally utilized, and a threshold function is used for denoising. Therefore, the invention adopts a denoising method combining an integrated empirical mode decomposition algorithm and a wavelet soft threshold, on one hand, the integrated empirical mode decomposition algorithm avoids the occurrence of an aliasing phenomenon of the empirical mode decomposition mode, and on the other hand, the wavelet soft threshold reduces the loss of useful information caused in the coefficient threshold processing process.
Step two: building a GAN model: the generator of DCGAN (deep convolution to generate countermeasure network) is used to generate some cardiac data with few samples, and the loss function of DCGAN is as follows:
Figure BDA0002657933790000071
where G denotes a generator, D denotes a discriminator, and z denotes random noise, and an image is generated from this noise and denoted as G (z). Pdata(x) P (x | data) represents the probability of obtaining x from the actual dataset (in the present embodiment, the fused beat type data in the kaggle dataset), and PZP (Z) ═ P (Z | Z) represents the probability of obtaining Z from the generated dataset (the dataset of the fused beat class generated using DCGAN). The discriminator D is a neural network for evaluating its authenticity, that is to say it is used to discriminate whether each data instance belongs to a real training data set. Where the discriminator input is an image x (derived from the fused beat class in the actual kaggle dataset and the fused beat class generated using DCGAN), and the output d (x) represents the probability that x is a true picture. The raw data and the electrocardiographic data generated by the DCGAN are shown in fig. 3, the network structure of the GAN is shown in fig. 4, and the network structure of the DCGAN is shown in fig. 5.
Step three: building a CNN model: the network mainly comprises a convolution Layer (Conv Layer), a BN Layer (Batch Normalization), a LeakyReLU, a Dropout Layer and a Pooling Layer (Pooling Layer), and a specific network flow chart is shown in figure 6, and five convolution parts with the same structure are connected in sequence. The method is used for capturing detail characteristics in the electrocardio data and ensuring the accuracy of the model. The calculation formula of the convolution layer output value is as follows:
Figure BDA0002657933790000072
wherein x isi，jThe ith row and the jth column of elements of the image in the training set (the training set divided from the data set after data enhancement); wm，nRepresents the weight of the mth row and the nth column; wbA bias term representing the filter in the convolutional layer; a isi，jAn ith row and a jth column element representing a Feature Map; f denotes an activation function, fhIndicating the height of the filter in the convolutional layer, fwIndicating the width of the filter in the convolutional layer.
Step four: constructing a bidirectional LSTM model: we use bi-directional LSTM (i.e. BiLSTM) to process the timing relationships in the signature sequence output by step three, and see fig. 6 for a specific network structure.
The LSTM is used for solving the problem of long-term dependence which can not be processed by the standard RNN, and is in the form of an RNN after other neural networks (gates) are added on the basis of the standard RNN, and the updating calculation formula in the t step is as follows:
it＝σ(Wixt+Uiht-1+bi)
ft＝σ(Wfxt+Ufht-1+bf)
ot＝σ(Woxt+Uoht-1+bo)
Figure BDA0002657933790000081
Figure BDA0002657933790000082
ht＝ot⊙Tanh(ct)
wherein t is the number of cycles, Wf、Wi、WC、WoTraining parameters which are all threshold recurrent neural networks, bf、bi、bC、boσ (x) is a first activation function, typically sigmoid, and tanh (x) is a second activation function, which is a bias parameter for the recurrent neural network. i.e. itAre input into people, ftIs a forgetting door otIs an output gate, ctIs an internal memory cell, which is a memory cell,
Figure BDA0002657933790000083
as candidate memory cells, xtFor the input signal (signature sequence extracted in CNN model), ht-1For the hidden layer output of the previous step, Wi、UiTwo matrices and vector biFor inputting the parameters of the gate, htIs the output of the hidden layer in this step.
Compared with the LSTM, the BilSTM not only can combine a plurality of input information before the current moment, but also can effectively combine a plurality of input information after the current moment, which is very important for the classification and prediction of the electrocardiosignals, and the classification result can be more accurately obtained by combining the information of the whole signal.
Step five: and (3) establishing an Attention model: attention model, i.e. the output signal of the processed signal (BilSTM, i.e. hidden state sequence h)1，h2，，…，hT) The time is divided by a certain weight, and the formula of the algorithm is as follows:
si＝f(si-1，yi-1，ci)
p(yi|y1，y2，......，yi-1)＝g(yi-1，si，ci)
Figure BDA0002657933790000091
Figure BDA0002657933790000092
eij＝a(si-1，hj)
wherein h isjIs an input state xj(i.e., the corresponding feature sequence extracted from the CNN model) corresponding hidden state, yiIs a state in the output sequence, siIs a corresponding hidden state, ciIs the complete hidden state h of the input sequence1，h2，......，hTT represents the number of all hidden states of the input sequence, aijIs the attention weight parameter, eijIs xiAnd yiThe value of the alignment. p represents given y1，y2，......，yi-1In the case of (1), the value yiG denotes a non-linear transformation, usually referred to as a multi-layer neural network, and a refers to a neural network. Step six: optimizing a convolution cyclic neural network model; the construction of the model is mainly divided into three parts: the first part is a convolution part and mainly comprises a convolution Layer (Conv Layer), a BN Layer (Batch Normalization), a LeakyReLU, a Dropout Layer and a Pooling Layer (Pooling Layer), a specific network flow chart is shown in figure 6, and five convolution parts with the same structure are connected in sequence. The second part is bidirectional LSTM, electrocardiosignals belong to time sequence data, and information in the context can be better combined by using a recurrent neural network. The third part is an attention model, namely an attention model, which can focus more on the detail part and some important information in the electrocardiogram data. Optimizing the established CNN and RNN models, performing BatchNormalization on each convolution layer to accelerate neural network training, reducing sensitivity to network initialization, and using LeakyRelu activation function to reduce gradient disappearance problem.
Further, the LeakyRelu activation function is
Figure BDA0002657933790000101
Wherein x isiThe value of the input activation function is shown, where a is a small positive constant, thus achieving both unilateral suppression and retaining part of the negative gradient information so that it is not completely lost. y isiThe value after activation of the activation function.
Furthermore, the initialization mode of the whole weight W of the convolution cyclic neural network is normal distribution initialization, an optimizer selects SGD in the network training process, the accuracy of selecting the gradient direction every time is guaranteed while the generalization performance of the model is guaranteed, and the learning effect is better.
Step seven: training a convolution cyclic neural network model; the training algorithm of the CNN model is a back propagation algorithm, cross validation is adopted for cyclic training, ECG data are randomly divided into 10 equal parts, nine parts are selected for training, the rest are used for testing, the cyclic repetition is performed for ten times, and a classification model is trained based on a random gradient descent algorithm.
Further, in the loss function, all the labels of the electrocardiographic data adopt a one-hot form, the total loss function of the convolution cyclic neural network model is defined as the sum of all the cross entropy loss functions of the electrocardiographic data participating in training, and the labels and the prediction results of the single electrocardiographic data are respectively labeliAnd predictioniThen the model total loss function is defined as follows:
Figure BDA0002657933790000102
where n represents the total number of samples on the training set and cross _ entropy () represents the cross entropy loss function. predictioniThe model prediction value, label, of the ith electrocardiogram data in the training set is representediAnd the label value represents the ith electrocardiogram data in the training set.
Further, the random gradient descent algorithm-based training classification model is characterized in that an SGD optimizer is adopted, the learning rate is exponentially attenuated, preferably, the initial learning rate is 0.001, the EPOCH value is 100, the batch _ size value is 64, the learning rate attenuation step size is 8000, and the single learning rate attenuation rate is 0.96.
The effect of the single-lead electrocardio abnormal signal identification method based on the generation countermeasure network and the convolution cyclic neural network is as follows:
in order to verify the effectiveness and feasibility of the method, experiments are carried out on a kaggle open-source arrhythmia data set, the data set is mainly based on an MIT-BIH arrhythmia data set, original 30-second electrocardio records are divided into lengths of about 1 second by RR intervals, and 125Hz is used as a sampling frequency. The number of classes classified is 5, and the corresponding class name is: normal electrocardiographic data (Normal), supraventricular ectopic beats (SVEB), Ventricular Ectopic Beats (VEB), fusion beats (F), and unknown beats (Q).
The raw data distribution is shown in the following table:
Category TrainingSet TestingSet
N 72471 18118
S 2223 556
V 5788 1448
F 641 162
Q 6431 1608
the length of the daily record was 187, and the sampling frequency was 125 Hz. As can be seen from this, the fused beat (F) category is too small in data amount compared with the other four types of electrocardiographic data, and has a problem of unbalanced samples, and therefore, data enhancement is required. In terms of generating the network, the input noise dimension of the generator is set to 20 and the size of the batch size is set to 32. In the aspect of network discrimination, original 187-dimensional electrocardiogram data is firstly filled to 256, then reshape is formed to [16, 16] and used as the input of a discriminator, Adam is selected as an optimizer of the network, the learning rate of Adam is set to 0.0002, beta 1 is set to 0.5, and beta 2 is set to 0.999. And (3) performing data enhancement on the fused pulsation sample by adopting a DCGAN model, wherein the enhanced electrocardio data distribution is shown as the following graph:
Figure BDA0002657933790000111
Figure BDA0002657933790000121
and finally, performing performance test by using the trained model to obtain the performance index of the model. We evaluated the performance of the method on the test set.
To prove that the methods proposed in this patent are feasible, therefore, each method is subjected to a comparative experiment in a separate experiment, where CNN represents the convolutional neural network model used in this patent, DCGAN represents the deep convolution generation countermeasure network used in this patent for performing data enhancement operation on the electrocardiographic data, BiLSTM represents the bidirectional long-short term memory network, and Attention represents the Attention mechanism used in this patent, and the results of the comparative experiment are shown in the following table:
Approach Average Accuracy
CNN 93.70％
CNN+DCGAN 94.20％
CNN+Attention+DCGAN 94.80％
CNN+BiLSTM+DCGAN 95.80％
CNN+BiLSTM+Attention+DCGAN 96.10％
as can be seen from the results of the comparative experiments in the above table, the accuracy of the method using the combination of the mechanisms CNN, DCGAN, BilSTM and Attention proposed in the present patent is significantly improved compared with the prior method using the mechanisms CNN, DCGAN, BilSTM and Attention alone, which indicates that the method proposed in the present patent is feasible.
Finally, the accuracy, recall, and F1-score of the test set using the DCGAN for data enhancement and the convolutional neural network, the bi-directional LSTM, and the Attention mechanism are shown in the table below.
Precision Recall F1-score support
N 0.99 1 0.99 18118
S 0.91 0.75 0.82 556
V 0.96 0.93 0.95 1448
F 0.75 0.72 0.73 162
Q 0.97 0.98 0.98 1608
avg/total 0.96 0.96 0.96 21892
As can be seen from the above table, the average Accuracy value of the method of the present invention on the test set is 96.10%, the average Precision value is 96.00%, the average Recall value is 96.00%, and the average F1-score value is 96.00%. We can see that the DCGAN is used for data enhancement, the convolutional neural network is used for extracting features, and the bidirectional cyclic neural network is used for extracting the time sequence relation among the features, so that the model has an obvious improvement effect on single-lead short-time data, the stability of the model is improved, in addition, the 1-D convolutional (namely one-dimensional convolutional) neural network is used for improving the overall efficiency of the network, and the robustness of the model is ensured. The experimental result shows that the method provided by the invention is effective, and can greatly improve the detection rate, the identification precision and the identification efficiency of abnormal data in the short-time electrocardiogram data, so that the method has great social and practical values.
Finally, it should be noted that: the above-mentioned embodiments are only used for illustrating the technical solution of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (7)

1. An electrocardiosignal identification method based on a generation countermeasure and convolution cycle network is characterized by comprising the following steps:
step 1, denoising electrocardiosignal data in an actual data set;
step 2, generating the electrocardiogram data with a small number of samples in the actual data set by using a generator of the anti-network DCGAN generated by deep convolution, and solving the problem of unbalanced samples in the actual data set, so as to perform data enhancement on the sample types with small data volume in the actual data set to obtain a final training set;
step 3, constructing a convolution cyclic neural network model, wherein the model comprises three parts: the first part is a convolutional neural network CNN used for capturing detail characteristics in the electrocardiogram data, and comprises 5 groups of convolutional layers, a BN layer, a LeakyReLU layer, a Dropout layer and a pooling layer; the second part is a bidirectional long and short term memory network (BilSTM) which is used for processing the time sequence relation in the characteristic sequence output by the convolutional neural network; the third part is an attention model, namely an attention model, and is used for dividing the weight of an output signal of the bidirectional long-short term memory network BilSTM;
step 4, circularly training a convolution cyclic neural network model by using a back propagation algorithm and adopting cross validation, wherein the total loss function of the convolution cyclic neural network model is defined as the sum of cross entropy loss functions of all the electrocardiographic data participating in training, and the labels and the prediction results of the single electrocardiographic data are respectively label and prediction results of labeliAnd predictioniThen the model total loss function is defined as follows:
Figure FDA0002657933780000011
where n represents the total number of samples on the training set and cross _ entropy () represents the cross entropy lossLoss function, predictioniThe model prediction value, label, of the ith electrocardiogram data in the training set is representediA label value representing the ith electrocardiogram data in the training set;
and 5, classifying the test set by using the trained convolution cyclic neural network model.
2. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: in step 2, the loss function of DCGAN is as follows,
Figure FDA0002657933780000012
wherein G represents a generator, D represents a discriminator, and z represents random noise, and an image is generated from the random noise and denoted as G (z); pdata(x) P (x | data) represents the probability of obtaining x from an actual data set, which is an electrocardiographic data set with a small number of samples, and PZP (Z) ═ P (Z | Z) represents the probability of obtaining Z from the generated dataset, which is generated using a dataset generated using DCGAN, and discriminator D is a neural network used to evaluate its authenticity, that is, the discriminator is used to discriminate whether each instance of data belongs to the real training dataset, where the discriminator input is an image x, and the output D (x) represents the probability that x is the real image, the image x being from the real dataset and the generated dataset.
3. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: in step 3, the calculation formula of the output value of the convolution layer is as follows,
Figure FDA0002657933780000021
wherein x isi，jAn ith row and a jth column element representing an image; wm，nRepresents the m-th rowThe nth column weight; wbA bias term representing filter; a isi，jAn ith row and a jth column element representing the feature map; f denotes an activation function, fhHigh, f, representing filterwIndicating the width of the filter.
4. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: in step 3, the updating calculation formula of the t step in the bidirectional LSTM is as follows,
it＝σ(Wixt+Uiht-1+bi)
ft＝σ(Wfxt+Ufht-1+bf)
ot＝σ(Woxt+Uoht-1+bo)
Figure FDA0002657933780000022
Figure FDA0002657933780000023
ht＝ot⊙Tanh(ct)
wherein t is the number of cycles, Wf、Wi、WC、WoTraining parameters which are all threshold recurrent neural networks, bf、bi、bC、boσ (x) is a first activation function, and tanh (x) is a second activation function, which are bias parameters of the recurrent neural network; i.e. itIs an input gate, ftIs a forgetting door otIs an output gate, ctIs an internal memory cell, xtFor the input signal, i.e. the extracted signature sequence from CNN, ht-1For the hidden layer output of the previous step, Wi、UiTwo matrices and vector biFor inputting the parameters of the gate, htIs the output of the hidden layer in step t.
5. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: in step 3, the attention model divides the weight of the output signal of the bidirectional long-short term memory network BilSTM in a concrete way as follows,
si＝f(si-1，yi-1，ci)
p(yi|y1，y2，......，yi-1)＝g(yi-1，si，ci)
Figure FDA0002657933780000024
Figure FDA0002657933780000025
eij＝a(si-1，hj)
wherein h isjIs a hidden state corresponding to the input state, xjI.e. the corresponding characteristic sequence, y, extracted from the CNN modeliIs a state in the output sequence, siIs a corresponding hidden state, ciIs the complete hidden state h of the input sequence1，h2，……，hTT represents the number of all hidden states of the input sequence, aijIs the attention weight parameter, eijIs xiAnd yiAligned value, p denotes given y1，y2，......，yi-1In the case of (1), the value yiG denotes a non-linear transformation and a refers to a neural network.
6. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: and in step 4, performing BatchNormalization on each convolutional layer to accelerate the training of the circularly trained convolutional neural network model, and using a LeakyRelu activation function to alleviate the gradient disappearance problem.
7. The method for recognizing the electrocardiosignals based on the generative confrontation and convolution cyclic network as claimed in claim 1, wherein: in the step 1, denoising processing is carried out on the electrocardiosignal data in the actual data set by adopting a denoising method combining an integrated empirical mode decomposition algorithm and a wavelet soft threshold.
CN202010894295.7A 2020-08-31 2020-08-31 Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network Pending CN111990989A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010894295.7A CN111990989A (en) 2020-08-31 2020-08-31 Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010894295.7A CN111990989A (en) 2020-08-31 2020-08-31 Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network

Publications (1)

Publication Number Publication Date
CN111990989A true CN111990989A (en) 2020-11-27

Family

ID=73465638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010894295.7A Pending CN111990989A (en) 2020-08-31 2020-08-31 Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network

Country Status (1)

Country Link
CN (1) CN111990989A (en)

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112508110A (en) * 2020-12-11 2021-03-16 哈尔滨理工大学 Deep learning-based electrocardiosignal graph classification method
CN112668507A (en) * 2020-12-31 2021-04-16 南京信息工程大学 Sea clutter prediction method and system based on hybrid neural network and attention mechanism
CN112687391A (en) * 2021-03-15 2021-04-20 四川大学 Construction method of intelligent identification and omission-prevention system for breast ultrasonic image focus
CN112729825A (en) * 2020-12-18 2021-04-30 中车永济电机有限公司 Method for constructing bearing fault diagnosis model based on convolution cyclic neural network
CN112766182A (en) * 2021-01-22 2021-05-07 浙江工业大学 Method for identifying wear state of broaching tool
CN112861604A (en) * 2020-12-25 2021-05-28 中国科学技术大学 Myoelectric action recognition and control method irrelevant to user
CN112932433A (en) * 2021-02-05 2021-06-11 青岛大学 Arrhythmia automatic classification method of electrocardiogram
CN113080992A (en) * 2021-03-30 2021-07-09 北京芯动卫士科技有限公司 Atrial fibrillation identification method based on DCGAN and deep neural network method
CN113080990A (en) * 2021-03-25 2021-07-09 南京蝶谷健康科技有限公司 Heart beat anomaly detection method based on CycleGAN and BilSTM neural network method
CN113180684A (en) * 2021-06-03 2021-07-30 哈尔滨理工大学 II-lead electrocardiosignal classification method
CN113180679A (en) * 2021-05-12 2021-07-30 武汉大学 Single-lead electrocardio abnormal signal identification method
CN113343924A (en) * 2021-07-01 2021-09-03 齐鲁工业大学 Modulation signal identification method based on multi-scale cyclic spectrum feature and self-attention generation countermeasure network
CN113436089A (en) * 2021-06-15 2021-09-24 山东省人工智能研究院 ECG noise reduction method based on combination of BilSTM and generation countermeasure network
CN113712542A (en) * 2021-08-27 2021-11-30 中山大学孙逸仙纪念医院 Tinnitus electroencephalogram convolutional neural network model construction method based on deep learning
CN113990499A (en) * 2021-10-22 2022-01-28 安徽心之声医疗科技有限公司 Heart age assessment method and device, electronic equipment and storage medium
CN114041801A (en) * 2021-11-15 2022-02-15 哈尔滨工业大学 Electrocardiosignal reconstruction method and system based on PSA-EWT and DCGAN
CN114159069A (en) * 2021-12-14 2022-03-11 南京邮电大学 System and method for generating confrontation network heartbeat signal based on space attention
CN116383649A (en) * 2023-04-03 2023-07-04 山东省人工智能研究院 Electrocardiosignal enhancement method based on novel generation countermeasure network
CN116958644A (en) * 2023-06-07 2023-10-27 中国人民解放军国防科技大学 Image classification countermeasure sample defense method based on feature consistency constraint
CN117743719A (en) * 2023-12-22 2024-03-22 北京京航计算通讯研究所 Page element identification method
CN116958644B (en) * 2023-06-07 2024-06-11 中国人民解放军国防科技大学 Image classification countermeasure sample defense method based on feature consistency constraint

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109036573A (en) * 2018-07-06 2018-12-18 苏州大学 Electrocardiogram based on production confrontation network technology generates and classification method
CN109998500A (en) * 2019-04-30 2019-07-12 陕西师范大学 A kind of pulse signal generation method and system based on production confrontation network
CN110584649A (en) * 2019-09-09 2019-12-20 武汉大学 Method and system for identifying abnormal electrocardiosignals in dynamic single-lead electrocardiogram
CN110619322A (en) * 2019-10-18 2019-12-27 武汉大学 Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network
CN111134662A (en) * 2020-02-17 2020-05-12 武汉大学 Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection
US20200204546A1 (en) * 2017-09-21 2020-06-25 Nokia Technologies Oy Apparatus, method and computer program product for biometric recognition
CN111449644A (en) * 2020-03-19 2020-07-28 复旦大学 Bioelectricity signal classification method based on time-frequency transformation and data enhancement technology

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200204546A1 (en) * 2017-09-21 2020-06-25 Nokia Technologies Oy Apparatus, method and computer program product for biometric recognition
CN109036573A (en) * 2018-07-06 2018-12-18 苏州大学 Electrocardiogram based on production confrontation network technology generates and classification method
CN109998500A (en) * 2019-04-30 2019-07-12 陕西师范大学 A kind of pulse signal generation method and system based on production confrontation network
CN110584649A (en) * 2019-09-09 2019-12-20 武汉大学 Method and system for identifying abnormal electrocardiosignals in dynamic single-lead electrocardiogram
CN110619322A (en) * 2019-10-18 2019-12-27 武汉大学 Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network
CN111134662A (en) * 2020-02-17 2020-05-12 武汉大学 Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection
CN111449644A (en) * 2020-03-19 2020-07-28 复旦大学 Bioelectricity signal classification method based on time-frequency transformation and data enhancement technology

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112508110A (en) * 2020-12-11 2021-03-16 哈尔滨理工大学 Deep learning-based electrocardiosignal graph classification method
CN112729825A (en) * 2020-12-18 2021-04-30 中车永济电机有限公司 Method for constructing bearing fault diagnosis model based on convolution cyclic neural network
CN112861604B (en) * 2020-12-25 2022-09-06 中国科学技术大学 Myoelectric action recognition and control method irrelevant to user
CN112861604A (en) * 2020-12-25 2021-05-28 中国科学技术大学 Myoelectric action recognition and control method irrelevant to user
CN112668507A (en) * 2020-12-31 2021-04-16 南京信息工程大学 Sea clutter prediction method and system based on hybrid neural network and attention mechanism
CN112766182A (en) * 2021-01-22 2021-05-07 浙江工业大学 Method for identifying wear state of broaching tool
CN112766182B (en) * 2021-01-22 2024-02-02 浙江工业大学 Broaching tool wear state identification method
CN112932433A (en) * 2021-02-05 2021-06-11 青岛大学 Arrhythmia automatic classification method of electrocardiogram
CN112687391A (en) * 2021-03-15 2021-04-20 四川大学 Construction method of intelligent identification and omission-prevention system for breast ultrasonic image focus
CN113080990B (en) * 2021-03-25 2023-01-10 南京蝶谷健康科技有限公司 Heart beat anomaly detection method based on CycleGAN and BilSTM neural network method
CN113080990A (en) * 2021-03-25 2021-07-09 南京蝶谷健康科技有限公司 Heart beat anomaly detection method based on CycleGAN and BilSTM neural network method
CN113080992A (en) * 2021-03-30 2021-07-09 北京芯动卫士科技有限公司 Atrial fibrillation identification method based on DCGAN and deep neural network method
CN113180679A (en) * 2021-05-12 2021-07-30 武汉大学 Single-lead electrocardio abnormal signal identification method
CN113180684B (en) * 2021-06-03 2023-10-24 哈尔滨理工大学 II-lead electrocardiosignal classification method
CN113180684A (en) * 2021-06-03 2021-07-30 哈尔滨理工大学 II-lead electrocardiosignal classification method
CN113436089A (en) * 2021-06-15 2021-09-24 山东省人工智能研究院 ECG noise reduction method based on combination of BilSTM and generation countermeasure network
CN113343924A (en) * 2021-07-01 2021-09-03 齐鲁工业大学 Modulation signal identification method based on multi-scale cyclic spectrum feature and self-attention generation countermeasure network
CN113712542A (en) * 2021-08-27 2021-11-30 中山大学孙逸仙纪念医院 Tinnitus electroencephalogram convolutional neural network model construction method based on deep learning
CN113990499A (en) * 2021-10-22 2022-01-28 安徽心之声医疗科技有限公司 Heart age assessment method and device, electronic equipment and storage medium
CN114041801A (en) * 2021-11-15 2022-02-15 哈尔滨工业大学 Electrocardiosignal reconstruction method and system based on PSA-EWT and DCGAN
CN114041801B (en) * 2021-11-15 2024-03-08 哈尔滨工业大学 Electrocardiosignal reconstruction method and system based on PSA-EWT and DCGAN
CN114159069A (en) * 2021-12-14 2022-03-11 南京邮电大学 System and method for generating confrontation network heartbeat signal based on space attention
CN116383649A (en) * 2023-04-03 2023-07-04 山东省人工智能研究院 Electrocardiosignal enhancement method based on novel generation countermeasure network
CN116383649B (en) * 2023-04-03 2024-01-23 山东省人工智能研究院 Electrocardiosignal enhancement method based on novel generation countermeasure network
CN116958644A (en) * 2023-06-07 2023-10-27 中国人民解放军国防科技大学 Image classification countermeasure sample defense method based on feature consistency constraint
CN116958644B (en) * 2023-06-07 2024-06-11 中国人民解放军国防科技大学 Image classification countermeasure sample defense method based on feature consistency constraint
CN117743719A (en) * 2023-12-22 2024-03-22 北京京航计算通讯研究所 Page element identification method

Similar Documents

Publication Publication Date Title
CN111990989A (en) Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network
Liu et al. Arrhythmia classification of LSTM autoencoder based on time series anomaly detection
Li et al. Deep convolutional neural network based ECG classification system using information fusion and one‐hot encoding techniques
CN110619322A (en) Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network
CN111134662B (en) Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection
CN110974214A (en) Automatic electrocardiogram classification method, system and equipment based on deep learning
CN111297349A (en) Machine learning-based heart rhythm classification system
CN110638430B (en) Method for building cascade neural network ECG signal arrhythmia classification model
CN111329469A (en) Arrhythmia prediction method
CN112906748A (en) 12-lead ECG arrhythmia detection classification model construction method based on residual error network
Gawande et al. Heart diseases classification using convolutional neural network
JP7487965B2 (en) Prediction method of electrocardiogram heart rate multi-type based on graph convolution
CN113180679A (en) Single-lead electrocardio abnormal signal identification method
Singh et al. Short and noisy electrocardiogram classification based on deep learning
Dang et al. A deep biometric recognition and diagnosis network with residual learning for arrhythmia screening using electrocardiogram recordings
Shi et al. Automated heartbeat classification based on convolutional neural network with multiple kernel sizes
Lu et al. Identification of arrhythmia by using a decision tree and gated network fusion model
Wang et al. Pay attention and watch temporal correlation: a novel 1-D convolutional neural network for ECG record classification
Benchaib et al. Cardiac arrhythmias classification using artificial metaplasticity algorithm
CN114091530A (en) Electrocardiogram classification method and domain-adaptive-based electrocardiogram classification model training method
Luo et al. Exploring adaptive graph topologies and temporal graph networks for eeg-based depression detection
Rajesh et al. A Silent Cardiac Atrial Fibrillation Detection and Classification using Deep Learning Approach
Nanarkar et al. A Survey on Classification and identification of Arrhythmia using Machine Learning techniques
Nasim et al. An Evolutionary-Neural Mechanism for Arrhythmia Classification With Optimum Features Using Single-Lead Electrocardiogram
Abo-Zahhad et al. Classification of ECG Signals for Detecting Coronary Heart Diseases Using Deep Transfer Learning Techniques

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination