阅读说明：本技术 双通道混合网络模型的心血管疾病识别方法、装置及介质 (Cardiovascular disease identification method, device and medium of two-channel hybrid network model ) 是由 司玉娟 杨维熠 张弓 冯文轩 范伟 孙美琪 于 2021-07-07 设计创作，主要内容包括：本发明涉及一种双通道混合网络模型的心血管疾病识别方法、装置及介质的技术方案,包括：对双导联心电图的心电信号进行波段对齐分割处理,得到双导联心拍；通过第一混合卷积网络提取双导联心拍的融合特征；通过第二混合卷积网络提取双导联心拍的两个单导联特异性特征；通过线性支持向量机处理融合特征及单导联特异性特征以得到对应的三组决策值；将三组决策值映射为三组决策概率；使用D-S模型融合三组决策概率得到双导联心电图的分类结果。本发明的有益效果为：解决现有技术中在患者间数据集、不平衡数据集和含噪数据集中分类效果较差的问题,具有较好的分类结果。(The invention relates to a cardiovascular disease identification method, a device and a medium of a dual-channel hybrid network model, which comprises the following steps: performing wave band alignment segmentation processing on the electrocardiosignals of the double-lead electrocardiogram to obtain a double-lead heart beat; extracting fusion characteristics of the dual-lead heart beat through a first hybrid convolution network; extracting two single-lead specific characteristics of the double-lead heart beat through a second hybrid convolution network; processing the fusion characteristics and the single-lead specific characteristics through a linear support vector machine to obtain three corresponding groups of decision values; mapping the three sets of decision values into three sets of decision probabilities; and fusing three groups of decision probabilities by using a D-S model to obtain a classification result of the double-lead electrocardiogram. The invention has the beneficial effects that: the problem of among the prior art between the patient data set, unbalanced data set and contain the relatively poor classification effect in the data set that makes an uproar is solved, better classification result has.)

1. A cardiovascular disease identification method of a two-channel hybrid network model is characterized by comprising the following steps:

performing wave band alignment segmentation processing on the electrocardiosignals of the double-lead electrocardiogram to obtain a double-lead heart beat;

extracting fusion characteristics of the dual-lead heart beat through a first hybrid convolution network;

extracting two single-lead specific characteristics of the dual-lead heart beat through a second hybrid convolution network;

processing the fusion characteristics and the single-lead specific characteristics through a linear support vector machine to obtain three corresponding groups of decision values;

mapping the three sets of decision values into three sets of decision probabilities;

and fusing the three groups of decision probabilities by using a D-S model to obtain a classification result of the double-lead electrocardiogram.

2. The cardiovascular disease identification method of the two-channel hybrid network model according to claim 1, wherein the performing of the band-aligned segmentation process on the electrocardiographic signals of the dual-lead electrocardiogram comprises:

taking the R peak of the electrocardiosignal of the double-lead electrocardiogram as a reference, and taking the former R peak and the latter R peak as an R1 peak and an R2 peak respectively;

samples of 0.10s after the R1 peak and before the R2 peak were represented as points a1 and a2, respectively, and samples of 0.06 seconds before and after the R peak were represented as points B1 and B2, respectively;

respectively calling a sampling point between A1 and B1, a sampling point between B1 and B2 and a sampling point between B2 and A2 as a waveband X, a waveband Y and a waveband Z, and resampling each waveband to obtain a plurality of sampling points of the corresponding waveband;

connecting all the re-sampled wave bands X, Y and Z to obtain the heart beats of the sum of the sampling points of the wave bands X, Y and Z;

the amplitude of each heart beat was normalized to the interval [0,1] using a dispersion normalization method.

3. The cardiovascular disease identification method based on the two-channel hybrid network model according to claim 1, wherein the first and second dual-channel hybrid convolutional networks are the CCA-PCA convolutional network and the ICA-PCA convolutional network, respectively, wherein CCA is a canonical correlation analysis algorithm, PCA is a principal component analysis algorithm, and ICA independent component analysis algorithm.

4. The cardiovascular disease identification method of the dual-channel hybrid network model according to claim 1, wherein the extracting the fusion features of the dual-lead heart beat through the first hybrid convolutional network comprises heart beat preprocessing, CCA-PCA convolutional network construction and CCA-PCA convolutional network prediction;

the heart beat preprocessing comprises the following steps: remodeling the size of the double-lead heart beat to obtain an electrocardiogram matrix; sampling the electrocardio matrix by a window with a set size to obtain a plurality of sampling blocks, and vectorizing the sampling blocks; removing the mean value of each vector to obtain centralized vectors, and combining the centralized vectors into an initial-order matrix to be processed;

the construction of the CCA-PCA convolutional network comprises the following steps: constructing a typical related convolution kernel of the CCA-PCA convolution network convolution layer through the first eigenvector and the second eigenvector; performing two-dimensional convolution processing on the typical correlation convolution kernel and the electrocardio matrix to obtain an initial-order feature block set;

the CCA-PCA convolutional network prediction comprises the following steps: extracting a principal component convolution kernel from the initial order feature block set according to the first feature vector and the second feature vector; calculating a corresponding secondary order characteristic matrix according to the principal component convolution kernel; carrying out binarization processing on the secondary order feature matrix to obtain a processing result, and mapping the processing result to be 0 or 1; calculating to obtain a corresponding decimal matrix according to the secondary order characteristic matrix and the mapped processing result; dividing each decimal matrix into a plurality of sample blocks according to a set size and a set overlapping rate, obtaining the numerical values of all the sample blocks by using a histogram statistical mode, and further converting the numerical values of the sample blocks into the characteristic vectors corresponding to the single electrocardiogram matrix.

5. The cardiovascular disease identification method of the two-channel hybrid network model according to claim 4, wherein the heartbeat preprocessing comprises:

reshaping each heart beat into an electrocardio matrix with the size of m multiplied by n;

the ECG matrices obtained from the two-lead ECG signals are respectively expressed asAnd

to be provided withIs centered on each element, and has a size k₁×k₂In the windowUp-extracting a series of sample blocks and reconstructing the sample blocks into vectors

Removing each vectorTo obtain a centralized vector

All derived from single lead electrocardiographyCombined into a preliminary-stage pending matrix

6. The method for cardiovascular disease identification of the two-channel hybrid network model of claim 5, wherein the CCA-PCA convolutional network construction comprises:

calculating the vector a according to equation (1)₁And b₁To construct a typical correlation convolution kernel that is,

wherein S₁₂Is X¹And X²Of the covariance matrix, S₁₁And S₂₂Are each X¹And X²The autocovariance matrix of (a);

according to the formula (2), the formula (1) is optimized by adopting a Lagrange multiplier method,

wherein a is₁And b₁By maximizing J (a)₁,b₁) Obtaining, λ and ν representing lagrange multipliers;

calculating J (a) according to equation (3)₁,b₁) The partial derivative of (a) of (b),

wherein a is₁And b₁Are respectively asAndthe feature vector of (2);

calculating the first vector set according to the formula (4), wherein the vector set is a₁，b₁；

At the acquisition of L₁After the vector set, the typical correlation convolution kernel of the convolution layer is obtained according to the formula (5),

whereinA is to_iAnd b_iRespectively converted into matrix W_l ¹And W_l ²To be X¹And X²The first convolution kernel of (1);

according to the formulaCalculate 2 xL₁A primary feature block, wherein a represents a two-dimensional convolution process;

for each of the c-th leadsComputing all centered sample blocksWhereinIs thatThe jth sample block of (a);

vectorizing all sample blocks and combining into

And

all the initial-order feature blocks are processed through the above process to obtain

7. The method of claim 6, wherein the CCA-PCA convolutional network prediction comprises:

l extraction by equation (6)₂A convolution kernel of a main component of the image,

whereinThe covariance matrix (Y)^c)(Y^c)^TIs/are as followsFeature vector mapping as principal component convolution kernelAccording to

Calculating a sub-order feature block

According to

Calculated to obtain 2 XL₁×L₂A sub-level feature block connected with the sub-level feature block derived from the bi-core electrical connection to obtain

Binarizing all the sub-order feature matrices O with a function H (·)_i,lWhere H (-) maps values greater than 0 and other values to 1 and 0, respectively;

according to the formulaTo obtain L₁A decimal matrix in which the numerical range of the elements is

Partitioning each decimal matrix into sizes u₁×u₂B sample blocks with the overlapping rate of R;

according toTo process all values in the sample block using histogram statistics;

according to

To obtain the eigenvectors f of a single ECG matrix_i,1。

8. The cardiovascular disease identification method of the two-channel hybrid network model according to claim 5, wherein the extracting two single-lead specific features of the two-lead heart beat through the second hybrid convolutional network comprises:

remodeling each heart beat into an electrocardio matrix with the size of n multiplied by m

Electrocardiogram matrix for N-source derived from c-th electrocardiogram lead

All the electrocardio matrixes are processed through the initial stage of the first convolution layer in the canonical correlation analysis-principal component analysis convolution network and the initial-stage matrix to be processed is obtained

Performing blind source separation on X according to the formula S-BVX;

B. the V and independent component convolution kernel is obtained according to the following steps:

processing X with principal component analysis algorithm^cTo obtain a whitening matrix

The whitening matrix includes a matrix corresponding to covariance (X)^c)(X^c)^TL of₁Feature vector of maximum feature value

Obtaining the matrix Z according to equation (7)^c，

Z^c＝V^cX^c (7)

Using a processing matrix Z having a Gaussian non-linear function^cTo obtain an orthogonal matrix B^c；

According to D^c＝B^cV^cTo obtain a bagDraw L₁A column vectorOf the de-confusion matrix

The independent component convolution kernel is calculated according to equation (8),

wherein the content of the first and second substances,will vector d₁Mapping to matrix W_l ^c；

According toComputing a first order feature matrix

Processing first order feature matricesIs consistent with the principal component analysis convolution layer in a typical correlation analysis-principal component analysis convolution network to obtain a principal component convolution kernel

According to

Computing a sub-order feature matrix

According to the formula

And

calculating the eigenvector f of the ith electrocardio matrix_iWherein the function H (-) and the function Bhist (-) with the overlapping rate R are consistent with the output layer in the typical correlation analysis-principal component analysis network, and two groups of characteristics, namely f, are obtained after the data of the two electrocardioleads are respectively processed through the operations_i,2I is 1,2, …, N and f_i,3,i＝1,2,…,N。

9. The method of claim 8, wherein the processing the fused features and the single-derivative-specific features by a linear support vector machine to obtain three corresponding sets of decision values, and the mapping the three sets of decision values into three sets of decision probabilities comprises:

for each pair of feature vectors f_i,qAnd a label h_iThe calculation was performed using a linear support vector machine with L2 regularized L2 loss function according to equation (9)

Wherein h is_iIs an original label of the ith electrocardio matrix, and C represents a penalty coefficient and is set to be 1;

collecting corresponding labels h_iAnd a feature vector f_i,qAll decision values of

According to equation (10), all decision values of the forward propagation part of the softmax function are setAndconversion to decision probabilityAnd

whereinThe representation corresponds to a decision valueThe decision probability of (c).

10. The method for cardiovascular disease recognition based on the two-channel hybrid network model of claim 9, wherein the fusing the three sets of decision probabilities using the D-S model to obtain the classification result of the dual-lead electrocardiogram comprises:

the D-S combination rule is performed according to equation (11) to obtain a final classification result,

wherein I_iIs the ith heartElectric matrix, get m at maximum_a(I_i) Label h of (a) is the final predicted label.

11. A cardiovascular disease identification apparatus of a two-channel hybrid network model, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method steps of any of claims 1 to 10 when executing the computer program.

12. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the method of any one of claims 1 to 10.

Technical Field

The invention relates to the field of computer machine learning and biomedical signal processing, in particular to a cardiovascular disease identification method, a cardiovascular disease identification device and a cardiovascular disease identification medium of a two-channel hybrid network model.

Background

The number of patients experiencing chest pain increases each year, one of the major factors being cardiovascular disease. In recent years, cardiovascular disease has become a fatal epidemic accounting for one third of the worldwide deaths and 48% of deaths based on non-infectious diseases. According to the world health organization report, cardiovascular disease will cause death of 2300 million people in 2030. Therefore, it is important to discover and treat potential cardiovascular disease patients in an early stage.

Currently, most doctors diagnose coronary heart disease and heart failure by observing the electrocardiogram waveform of patients. However, manual detection of these diseases is difficult, time consuming and laborious due to possible visual fatigue and minor changes in the long-term electrocardiogram. Therefore, to solve these problems, an intelligent identification system based on electrocardiogram will play an important role in the automatic diagnosis of coronary heart disease and heart failure.

Currently, most relevant studies show good performance in the classification of coronary heart disease and heart failure. For example, in 2019, Acharya et al propose a convolutional neural network model for classification and feature extraction of electrocardiosignals, and obtain 95.98% heart failure recognition accuracy; lih et al in 2020 propose a hybrid network model based on a convolutional neural network and a long-and-short-term memory neural network as an electrocardiosignal classification method, and obtain an overall accuracy of 98.5% in the process of identifying normality, coronary heart disease, heart failure and myocardial infarction.

However, these studies still have some unsolved problems. First, most researchers have performed experiments on data in patients, which may lead to overfitting of the proposed method. In particular, the in-patient experiment may enable training and testing data to be selected from the same person. Since the heart rate differences between the electrocardiograms of the same person are usually small, trained algorithms may yield significantly poorer performance in identifying electrocardiograms of new individuals if their heart rates differ significantly from those in the training data due to overfitting. Therefore, in order to ensure optimal performance of the proposed method, inter-patient experiments have to be performed. Second, most researchers have not evaluated the noise robustness of the proposed method. Generally, electrocardiograms collected from an actual scene have a certain degree of noise, thereby distorting the electrocardiographic waveform and making it difficult to distinguish. However, most of the noise in the electrocardiogram has been removed from databases such as the best of the best possible ways to be considered. Thus, the electrocardiograms in these databases were recorded with clear waveform characteristics and were used directly by most researchers, resulting in overlooking the ability of the proposed method to handle noisy data. Therefore, in order to effectively evaluate the robustness of noise, the proposed method should be tested based on multi-level noise data with different signal-to-noise ratios. These studies lack quantitative and standard assessments of the ability of the proposed method to process imbalance data. In these studies, researchers randomly selected a proportion of abnormal electrocardiograms as experimental data. However, the proportion of abnormal electrocardiograms collected from a real situation is variable and unpredictable, thereby making it difficult to ensure the performance of these proposed methods and the ability of the proposed methods to process skewed data. Furthermore, in a real environment, a normal electrocardiogram is usually much more than an abnormal electrocardiogram. Therefore, it is extremely important to perform experiments in a multi-level unbalanced data set with less abnormal electrocardiograms.

Disclosure of Invention

The invention aims to at least solve one of the technical problems in the prior art, provides a cardiovascular disease identification method, a cardiovascular disease identification device and a cardiovascular disease identification medium of a dual-channel hybrid network model, solves the problem of poor classification effect of inter-patient data sets, unbalanced data sets and noisy data sets in the prior art, and has a good classification result.

The technical scheme of the invention comprises a cardiovascular disease identification method of a two-channel hybrid network model, which is characterized by comprising the following steps: performing wave band alignment segmentation processing on the electrocardiosignals of the double-lead electrocardiogram to obtain a double-lead heart beat; extracting fusion characteristics of the dual-lead heart beat through a first hybrid convolution network; extracting two single-lead specific characteristics of the dual-lead heart beat through a second hybrid convolution network; processing the fusion characteristics and the single-lead specific characteristics through a linear support vector machine to obtain three corresponding groups of decision values; mapping the three sets of decision values into three sets of decision probabilities; and fusing the three groups of decision probabilities by using a D-S model to obtain a classification result of the double-lead electrocardiogram.

According to the cardiovascular disease identification method of the two-channel hybrid network model, the wave band alignment segmentation processing of the electrocardiosignals of the double-lead electrocardiogram comprises the following steps: taking the R peak of the electrocardiosignal of the double-lead electrocardiogram as a reference, and taking the former R peak and the latter R peak as an R1 peak and an R2 peak respectively; samples of 0.10s after the R1 peak and before the R2 peak were represented as points a1 and a2, respectively, and samples of 0.06 seconds before and after the R peak were represented as points B1 and B2, respectively; respectively calling a sampling point between A1 and B1, a sampling point between B1 and B2 and a sampling point between B2 and A2 as a waveband X, a waveband Y and a waveband Z, and resampling each waveband to obtain a plurality of sampling points of the corresponding waveband; connecting all the re-sampled wave bands X, Y and Z to obtain the heart beats of the sum of the sampling points of the wave bands X, Y and Z; the amplitude of each heart beat was normalized to the interval [0,1] using a dispersion normalization method.

According to the cardiovascular disease identification method of the two-channel hybrid network model, the first two-channel hybrid convolutional network and the second two-channel hybrid convolutional network are the CCA-PCA convolutional network and the ICA-PCA convolutional network respectively, wherein the CCA is a typical correlation analysis algorithm, the PCA is a principal component analysis algorithm, and the ICA is an independent component analysis algorithm.

According to the cardiovascular disease identification method of the dual-channel hybrid network model, extracting the fusion characteristics of the dual-lead heart beat through a first hybrid convolutional network comprises heart beat preprocessing, CCA-PCA convolutional network construction and CCA-PCA convolutional network prediction;

the heart beat preprocessing comprises the following steps: remodeling the size of the double-lead heart beat to obtain an electrocardiogram matrix; sampling the electrocardio matrix by a window with a set size to obtain a plurality of sampling blocks, and vectorizing the sampling blocks; removing the mean value of each vector to obtain centralized vectors, and combining the centralized vectors into an initial-order matrix to be processed;

the construction of the CCA-PCA convolutional network comprises the following steps: constructing a typical related convolution kernel of the CCA-PCA convolution network convolution layer through the first eigenvector and the second eigenvector; performing two-dimensional convolution processing on the typical correlation convolution kernel and all electrocardiograms to obtain an initial-order feature block set;

the CCA-PCA convolutional network prediction comprises the following steps: extracting a principal component convolution kernel from the initial order feature block set according to the first feature vector and the second feature vector; calculating a corresponding secondary order feature matrix according to the principal component convolution kernel; carrying out binarization processing on the secondary order feature matrix to obtain a processing result, and mapping the processing result to be 0 or 1; calculating to obtain a corresponding decimal matrix according to the secondary order characteristic matrix and the mapped processing result; dividing each decimal matrix into a plurality of sample blocks according to a set size and a set overlapping rate, obtaining the numerical values of all the sample blocks by using a histogram statistical mode, and further converting the numerical values of the sample blocks into the characteristic vectors corresponding to the single electrocardiogram matrix.

According to the cardiovascular disease identification method of the two-channel hybrid network model, the central beat preprocessing comprises the following steps:

reshaping each heart beat into an electrocardio matrix with the size of m multiplied by n;

the ECG matrices obtained from the two-lead ECG signals are respectively expressed asAnd

to be provided withIs centered on each element, and has a size k₁×k₂In the windowUp-extracting a series of sample blocks and reconstructing the sample blocks into vectors

Removing each vectorTo obtain a centralized vector

All derived from single lead electrocardiographyCombined into a preliminary-stage pending matrix

According to the cardiovascular disease identification method of the two-channel hybrid network model, the CCA-PCA convolutional network construction comprises the following steps:

calculating the vector a according to equation (1)₁And b₁To construct a typical correlation convolution kernel that is,

wherein S₁₂Is X¹And X²Of the covariance matrix, S₁₁And S₂₂Are each X¹And X²The autocovariance matrix of (a);

according to the formula (2), the formula (1) is optimized by adopting a Lagrange multiplier method,

wherein a is₁And b₁By maximizing J (a)₁,b₁) Obtaining, λ and ν representing lagrange multipliers;

calculating J (a) according to equation (3)₁,b₁) The partial derivative of (a) of (b),

wherein a is₁And b₁Are respectively asAndthe feature vector of (2);

calculating the first vector set according to the formula (4), wherein the vector set is a₁，b₁；

At the acquisition of L₁After the vector set, the typical correlation convolution kernel of the convolution layer is obtained according to the formula (5),

whereinA is to_iAnd b_iRespectively converted into matrix W_l ¹And W_l ²To be X¹And X²The first convolution kernel of (1);

according to the formulaCalculate 2 xL₁A primary feature block, wherein a represents a two-dimensional convolution process;

for each of the c-th leadsComputing all centered sample blocksWhereinIs thatThe jth sample block of (a);

vectorizing all sample blocks and combining into

And

all the initial-order feature blocks are processed through the above process to obtain

According to the cardiovascular disease identification method of the two-channel hybrid network model, CCA-PCA convolutional network prediction comprises the following steps:

l extraction by equation (6)₂A convolution kernel of a main component of the image,

whereinThe covariance matrix (Y)^c)(Y^c)^TIs/are as followsFeature vector mapping as principal component convolution kernelAccording to

Calculating a sub-order feature block

According to

Calculated to obtain 2 XL₁×L₂A sub-level feature block connected with the sub-level feature block derived from the bi-core electrical connection to obtain

Binarizing all the sub-order feature matrices O with a function H (·)_i,lWhere H (-) maps values greater than 0 and other values to 1 and 0, respectively;

according to the formulaTo obtain L₁A decimal matrix of elementsThe numerical range of

Partitioning each decimal matrix into sizes u₁×u₂B sample blocks with the overlapping rate of R;

according toTo process all values in the sample block using histogram statistics;

according to

To obtain the eigenvectors f of a single ECG matrix_i,1。

The cardiovascular disease identification method of the two-channel hybrid network model, wherein the extracting of the two single-lead specific features of the two-lead heart beat through the second hybrid convolutional network comprises the following steps:

remodeling each heart beat into an electrocardio matrix with the size of n multiplied by m

Electrocardiogram matrix for N-source derived from c-th electrocardiogram lead

All the electrocardio matrixes are processed through the initial stage of the first convolution layer in the canonical correlation analysis-principal component analysis convolution network and the initial-stage matrix to be processed is obtained

Performing blind source separation on X according to the formula S-BVX;

B. the V and independent component convolution kernel is obtained according to the following steps:

processing X with principal component analysis algorithm^cTo obtain a whitening matrix

The whitening matrix includes a matrix corresponding to covariance (X)^c)(X^c)^TL of₁Feature vector of maximum feature value

Obtaining the matrix Z according to equation (7)^c，

Z^c＝V^cX^c (7)

Using a processing matrix Z having a Gaussian non-linear function^cTo obtain an orthogonal matrix B^c；

According to D^c＝B^cV^cTo obtain a composition comprising L₁A column vectorOf the de-confusion matrix

The independent component convolution kernel is calculated according to equation (8),

wherein the content of the first and second substances,will vector d₁Mapping to matrix W_l ^c；

According toComputing a first order feature matrix

Processing first order feature matricesIs consistent with the principal component analysis convolution layer in a typical correlation analysis-principal component analysis convolution network to obtain a principal component convolution kernel

According to

Computing a sub-order feature matrix

According to the formula

And

calculating the eigenvector f of the ith electrocardio matrix_iWherein the function H (-) and the function Bhist (-) with the overlapping rate R are consistent with the output layer in the typical correlation analysis-principal component analysis network, and two groups of characteristics, namely f, are obtained after the data of the two electrocardioleads are respectively processed through the operations_i,2I is 1,2, …, N and f_i,3,i＝1,2,…,N。

The method for identifying cardiovascular diseases according to the two-channel hybrid network model, wherein the fusion features and the single-lead specificity features are processed by a linear support vector machine to obtain three corresponding sets of decision values, and mapping the three sets of decision values into three sets of decision probabilities comprises:

for each pair of feature vectors f_i,qAnd a label h_iThe calculation was performed using a linear support vector machine with L2 regularized L2 loss function according to equation (9)

Wherein h is_iIs an original label of the ith electrocardio matrix, and C represents a penalty coefficient and is set to be 1;

collecting corresponding labels h_iAnd a feature vector f_i,qAll decision values of

According to equation (10), all decision values of the forward propagation part of the softmax function are setAndconversion to decision probabilityAnd

whereinThe representation corresponds to a decision valueThe decision probability of (c).

The cardiovascular disease identification method based on the two-channel hybrid network model, wherein the step of fusing three groups of decision probabilities by using a D-S model to obtain the classification result of the double-lead electrocardiogram comprises the following steps:

the D-S combination rule is performed according to equation (11) to obtain a final classification result,

wherein I_iIs the ith ECG matrix to get the maximum m_a(I_i) Label h of (a) is the final predicted label.

The technical solution of the present invention further includes a cardiovascular disease recognition apparatus of a two-channel hybrid network model, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any one of the method steps when executing the computer program.

The solution of the invention also includes a computer-readable storage medium on which a computer program is stored, which computer program, when being executed by a processor, carries out any of the methods.

The invention has the beneficial effects that: the invention has the beneficial effects that: the problem of among the prior art between the patient data set, unbalanced data set and contain the relatively poor classification effect in the data set that makes an uproar is solved, better classification result has.

Drawings

The invention is further described below with reference to the accompanying drawings and examples;

FIG. 1 shows a general flow diagram according to an embodiment of the invention.

Fig. 2 is a block diagram showing the overall structure according to the embodiment of the present invention.

Fig. 3 is a flowchart illustrating a heart beat segmentation method according to an embodiment of the invention.

FIG. 4 is a flow chart of deep lead correlation feature extraction for the CCA-PCA convolutional network in the system of the present invention

FIG. 5 is a flow chart of deep lead specific feature extraction by the ICA-PCA convolution network in the system of the present invention.

Fig. 6 is a graph showing an analysis of experimental results according to an embodiment of the present invention.

FIG. 7 is a graph illustrating experimental results on noisy data sets according to an embodiment of the present invention.

FIG. 8 is a graph illustrating experimental results on an unbalanced data set according to an embodiment of the present invention.

Fig. 9 is a diagram of an apparatus according to an embodiment of the present invention.

Detailed Description

Reference will now be made in detail to the present preferred embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to like elements throughout.

In the description of the present invention, the meaning of a plurality of means is one or more, the meaning of a plurality of means is two or more, and larger, smaller, larger, etc. are understood as excluding the number, and larger, smaller, inner, etc. are understood as including the number.

In the description of the present invention, the consecutive reference numbers of the method steps are for convenience of examination and understanding, and the implementation order between the steps is adjusted without affecting the technical effect achieved by the technical solution of the present invention by combining the whole technical solution of the present invention and the logical relationship between the steps.

In the description of the present invention, unless otherwise explicitly defined, terms such as set, etc. should be broadly construed, and those skilled in the art can reasonably determine the specific meanings of the above terms in the present invention in combination with the detailed contents of the technical solutions.

FIG. 1 shows a general flow diagram of the system of the present invention. It includes: A. performing wave band alignment segmentation processing on the electrocardiosignal; B. extracting the fusion characteristics of the double-lead electrocardiosignals by adopting a CCA-PCA convolution network to be used as characteristics 1; C. extracting single-lead specific characteristics as characteristics 2 and 3 by adopting an ICA-PCA convolution network; D. processing the three groups of characteristics by adopting a linear support vector machine to obtain three groups of decision values; mapping the three groups of decision values into three groups of decision probabilities by adopting a softmax function; D. and fusing three groups of decision probabilities by using a D-S model to obtain a final classification result, thereby realizing the diagnosis of the coronary heart disease and the heart failure. Wherein steps A, B and C are further described in fig. 3, 4, and 5, respectively.

Fig. 2 is a simplified block diagram of the system of the present invention. It includes: the device comprises a segmentation module, a feature extraction module and a classification module. The segmentation module segments the long-time electrocardiosignal into a single-cycle heart beat for subsequent processing. The feature extraction module is used for mining depth lead correlation features and depth lead specificity features from the heartbeat by adopting a CCA-PCA convolutional network and an ICA-PCA convolutional network; and the classification module adopts a linear support vector machine, a softmax function and a D-S model to perform decision layer information fusion so as to realize the automatic identification of the normality of electrocardiosignals, coronary heart disease and heart failure.

FIG. 3 is a flow chart of a heart beat segmentation method in the system of the present invention. The method is used for further explaining the wave band alignment segmentation method and mainly comprises the following steps:

the former R peak and the latter R peak are respectively called as R1 peak and R2 peak based on the R peak of the electrocardiosignal.

Samples 0.10s after the R1 peak and before the R2 peak were denoted as point a1 and point a2, respectively, and samples 0.06 seconds before and after the R peak were denoted as point B1 and point B2, respectively.

The sample points between a1 and B1, between B1 and B2, and between B2 and a2 are referred to as bands X, Y and Z, respectively, and each band is resampled to 100 sample points.

All resampled bands X, Y and Z are connected into heartbeats with 300 sample points.

The amplitude of each heart beat was normalized to the interval [0,1] using a dispersion normalization method.

Fig. 4 is a flow chart of deep lead correlation feature extraction for the CCA-PCA convolution network in the system of the present invention. The method is used for further explaining the CCA-PCA convolutional network and mainly comprises the following steps:

each heart beat is reshaped into an electrocardio matrix with the size of n multiplied by m.

The ECG matrices obtained from the two-lead ECG signals are respectively expressed asAnd

to be provided withIs centered on each element, and has a size k₁×k₂In the windowUp-extracting a series of sample blocks and reconstructing the sample blocks into vectors

Removing each vectorTo obtain a centralized vector

All derived from single lead electrocardiographyCombined into a preliminary-stage pending matrix

Calculating the vector a according to equation (1)₁And b₁To construct a representative correlation convolution kernel.

Wherein S₁₂Is X¹And X²Of the covariance matrix, S₁₁And S₂₂Are each X¹And X²The autocovariance matrix of (2).

And (3) optimizing the formula (1) by adopting a Lagrange multiplier method according to the formula (2).

Wherein a is₁And b₁By maximizing J (a)₁,b₁) Obtained, λ and ν represent lagrange multipliers.

Calculating J (a) according to equation (3)₁,b₁) Partial derivatives of (a).

Wherein a is₁And b₁Are respectively asAndthe feature vector of (2).

The l-th vector (a) is calculated according to equation (4)_lAnd b_l) And (4) collecting.

At the acquisition of L₁After each vector set, the typical associated convolution kernel for the convolutional layer is obtained according to equation (5).

WhereinA is to_lAnd b_lRespectively converted into matricesAndas X¹And X²The ith convolution kernel of (1). I according to the formulaCalculate 2 xL₁And (4) a primary feature block, wherein the primary feature block represents a two-dimensional convolution process.

For each of the c-th leadsComputing all centered sample blocksWhereinIs thatThe jth sample block of (1).

Vectorizing all sample blocks and combining intoAnd

all the initial-order feature blocks are processed through the above process to obtain

First, root ofExtracting L according to equation (6)₂A principal component convolution kernel.

WhereinThe covariance matrix (Y)^c)(Y^c)^TIs/are as followsFeature vector mapping as principal component convolution kernelThen, according toCalculating a sub-order feature block

According toCalculated to obtain 2 XL₁×L₂A sub-level feature block.

Connecting the sub-order feature blocks derived from the diplocentric electrical interconnect to obtain

Binarizing all the sub-order feature matrices O with a function H (·)_i,lWhere H (-) maps values greater than 0 and other values to 1 and 0, respectively.

Then, according to the formulaTo obtain L₁A decimal matrix in which the numerical range of the elements is

Partitioning each decimal matrix into sizes u₁×u₂And B sample blocks with the overlapping rate of R.

According toTo process all values in the sample block using a histogram statistical process.

According toTo obtain the eigenvectors f of a single ECG matrix_i，1。

FIG. 5 is a flow chart of deep lead specific feature extraction by the ICA-PCA convolution network in the system of the present invention. The method is used for further explaining the ICA-PCA convolutional network and mainly comprises the following steps:

remodeling each heart beat into an electrocardio matrix with the size of n multiplied by m

Electrocardiogram matrix for N-source derived from c-th electrocardiogram leadSimilar to the initial stage of the first convolution layer in the typical correlation analysis-principal component analysis convolution network, the method is used for processing all the electrocardio matrixes and obtaining a first-stage matrix to be processed

The purpose of the independent component analysis is to perform blind source separation on X according to the formula S-BVX. Here, B, V and the independent component convolution kernel are obtained according to the following steps:

first, X is processed using a principal component analysis algorithm^cTo obtain a whitening matrixThe whitening matrix includes a matrix (X) corresponding to a covariance^c)(X^c)^TL of₁Feature vector of maximum feature value

Obtaining the matrix Z according to equation (7)^c。

Z^c＝V^cX^c (7)

Processing matrix Z with FastICA toolkit having Gaussian non-linear function^cTo obtain an orthogonal matrix B^c。

According to D^c＝B^cV^cTo obtain a composition comprising L₁A column vectorOf the de-confusion matrix

The independent component convolution kernel is calculated according to equation (8).

Wherein the content of the first and second substances,will vector d_lMapping to a matrix

According toComputing a first order feature matrix

Processing preliminary stage featuresSign blockIs consistent with the principal component analysis convolution layer in a typical correlation analysis-principal component analysis convolution network to obtain a principal component convolution kernel

According toComputing a sub-order feature matrix

According to the formulaAndcalculating the eigenvector f of the ith electrocardio matrix_iWherein the function H (-) and the function Bhist (-) with the overlapping rate R are consistent with the output layer in the typical correlation analysis-principal component analysis network, and after the data of the two electrocardio leads are respectively processed by the operations, two groups of characteristics are obtained, namely f_i,2I is 1,2, …, N and f_i,3,i＝1,2,…,N。

The main flow of the linear support vector machine is as follows:

for each pair of feature vectors f_i,qAnd a label h_iA linear support vector machine with L2 regularized L2 loss function is employed to solve the optimization problem according to equation (9).

Wherein h is_iIs the original label of the ith cardiac electric matrix and C represents a penalty coefficient, set to 1.

Collecting corresponding labels h_iAnd a feature vector f_i,qAll decision values of

The major flow of the Softmax function is as follows:

according to equation (10), all decision values of the forward propagation part of the softmax function are setAndconversion to decision probabilityAnd

whereinThe representation corresponds to a decision valueThe decision probability of (c).

The main flow of the D-S model is as follows:

the D-S combination rule is executed according to equation (11) to obtain the final classification result.

Wherein I_iIs the ith cardiac electric matrix. Finally, has a maximum of m_a(I_i) Is considered as the final predicted tag.

In a specific embodiment, the first and second electrodes are,using the International traffic Normal Sinus Rhythm Database (Nsrdb), St.Petersburg 12-lead arrhythmia Database (Nstdb)St PetersburgINCART 12-lead Arrhythmia DatabaseIncardb) and the congestive heart failure database (Incardb)BIDMC Congestive Heart Failure DatabaseChfdb) as a data source for normal, coronary heart disease and heart failure. The invention adopts MTLAB2018 software to carry out performance verification. The data sets used included 18 normal individuals, 7 coronary heart disease patients and 15 heart failure patients in the Nsrdb database, the Incartdb database, and the Chfdb database. Firstly, 6 experimental data sets are set, wherein, A comprises the heart beats of normal and heart failure in the patient, B comprises the heart beats of normal and coronary heart disease in the patient, C comprises the heart beats of normal, coronary heart disease and heart failure in the patient, D comprises the heart beats of normal and heart failure between the patients, E comprises the heart beats of normal and coronary heart disease between the patients, and F comprises the heart beats of normal, heart failure and coronary heart disease between the patients. Then processing the double-lead electrocardiosignals of the heart beats by adopting a heart beat alignment segmentation method to obtain a single-cycle heart beat, extracting depth lead correlation characteristics from the double-lead heart beat by using a CCA-PCA convolution network model to realize characteristic layer information fusion and extraction, extracting two groups of single-lead specific characteristics from the single-lead heart beat by adopting an ICA-PCA convolution network model, sending the three groups of characteristics into a linear support vector machine to obtain decision values, converting the three groups of decision values into three groups of decision probabilities by using a softmax function, and finally realizing decision layer information fusion on the three groups of decision probabilities by using a D-S model to obtain a final classification result.

FIG. 6 is a graph of the results of an inter-patient experiment with the system of the present invention. In distinguishing between normal, CHF and CAD heartbeats, the method achieved an overall accuracy of 95.54% with a 4.46% error rate. All indicators exceeded 91% in identifying CHF heartbeats. In identifying CAD heartbeats, the method achieved precision and specificity in excess of 96% and an F1 score of 87.02%. Taken together, this method can distinguish well between normal, CHF and CAD heartbeats in an inter-patient experiment.

FIG. 7 is a graph of experimental results of the system of the present invention on noisy data sets. In these heartbeats, all noise was added by MATLAB 2018, and the ∞ dB signal-to-noise ratio indicates that the corresponding heartbeats do not have noise mixed in. Overall, as the signal-to-noise ratio decreases, the overall accuracy obtained by identifying each class also decreases. However, at a signal-to-noise ratio of 18dB for both group D and group E, the overall accuracy obtained by this method was only 1.16% and 2.63% lower than that obtained based on the noise-free data experiment. In addition, although the experimental results of the F-set with the multi-level noise are significantly worse than those of the other F-sets, the method can still obtain an overall accuracy of about 92% when the signal-to-noise ratio is 18 dB. More importantly, at a signal-to-noise ratio of 6dB, the corresponding heartbeats in group D and group E were hardly distinguishable by the naked eye, but the overall accuracy of the method reached 92.2% and 86.76%, respectively. Thus, the above results indicate that the method has excellent noise robustness in inter-patient CAD and CHF identification experiments.

FIG. 8 is a graph of experimental results of the system of the present invention on an unbalanced data set. In this work, the value of N representing the level of imbalance varies between 2 and 25, i.e., the number of normal beats is N times the number of abnormal beats. Wherein all precision, recall, specificity and F1 scores exceeded 97.5%, these factors floated less than 2.44% when different levels of imbalance were used. The comprehensive consideration of recall ratio and precision ratio F1 score is taken as a key index. In group a, the highest and lowest F1 scores obtained for identifying CHF were 99.93% (N ═ 2) and 99.75% (N ═ 25), respectively. In group B, the highest and lowest F1 scores obtained for identifying the CAD were 99.97% (N ═ 6) and 98.86% (N ═ 10), respectively. In addition, in group C, the highest and lowest F1 scores obtained for identifying CHF were 99.85% (N-2, 15) and 99.75% (N-25 for classifying CHF), respectively, while the highest and lowest F1 scores obtained for identifying CAD were 99.97% (N-6) and 98.73% (N-25), respectively. More importantly, even though the number of CAD or CHF heartbeats is 1/25 that of normal heartbeats, each index obtained by identifying CAD and CHF heartbeats exceeds 97.5%. The above results show that the system is able to produce superior performance in processing multi-level unbalanced heartbeat data when discriminating CAD and CHF heartbeats from mixed heartbeats.

Fig. 9 is a diagram of an apparatus according to an embodiment of the present invention. The apparatus comprises a memory 100 and a processor 200, wherein the processor 200 stores a computer program for performing: performing wave band alignment segmentation processing on the electrocardiosignals of the double-lead electrocardiogram to obtain a double-lead heart beat; extracting fusion characteristics of the dual-lead heart beat through a first hybrid convolution network; extracting two single-lead specific characteristics of the double-lead heart beat through a second hybrid convolution network; processing the fusion characteristics and the single-lead specific characteristics through a linear support vector machine to obtain three corresponding groups of decision values; mapping the three sets of decision values into three sets of decision probabilities; and fusing three groups of decision probabilities by using a D-S model to obtain a classification result of the double-lead electrocardiogram. Wherein the memory 100 is used for storing data.

It should be recognized that the method steps in embodiments of the present invention may be embodied or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer readable memory. The method may use standard programming techniques. Each program may be implemented in a high level procedural or object oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language. Furthermore, the program can be run on a programmed application specific integrated circuit for this purpose.

Further, the operations of processes described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and/or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) collectively executed on one or more processors, by hardware, or combinations thereof. The computer program includes a plurality of instructions executable by one or more processors.

Further, the method may be implemented in any type of computing platform operatively connected to a suitable interface, including but not limited to a personal computer, mini computer, mainframe, workstation, networked or distributed computing environment, separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, and the like. Aspects of the invention may be embodied in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optically read and/or write storage medium, RAM, ROM, or the like, such that it may be read by a programmable computer, which when read by the storage medium or device, is operative to configure and operate the computer to perform the procedures described herein. Further, the machine-readable code, or portions thereof, may be transmitted over a wired or wireless network. The invention described herein includes these and other different types of non-transitory computer-readable storage media when such media include instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. The invention also includes the computer itself when programmed according to the methods and techniques described herein.

A computer program can be applied to input data to perform the functions described herein to transform the input data to generate output data that is stored to non-volatile memory. The output information may also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including particular visual depictions of physical and tangible objects produced on a display.

The embodiments of the present invention have been described in detail with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.