WO2021031155A1 - Method and device for multi-scale characteristic extraction based on ecg - Google Patents

Abstract

The present invention relates to a method and a device for multi-scale characteristic extraction based on ECG. The method and the device firstly obtain ECG signal identification units based on a lead ECG signal interception, and performs multi-scale decomposition of the ECG signal identification units to construct an ECG multi-scale space; and then subjects the ECG multi-scale space signals in the ECG multi-scale space to multi-scale characteristic extraction by a preset convolutional neural network. The ECG signal identification units comprise at least one cardiac cycle band which is more conducive for the convolutional neural network to conduct ECG variation characteristic and spatial characteristic learning of myocardial infarction, extract variation characteristics, which have a strong ability to discriminate diseases, from a deeper level from ECG signals through the preset convolutional neural network, and obtain spatial characteristics related to location of the disease according to the spatial learning ability of the convolutional neural network, which has important practical reference value for a doctor to predict the location of the myocardial infarction.

Description

An ECG-based multi-scale feature extraction method and device Technical field

The present invention relates to the field of medical information processing, in particular to an ECG-based multi-scale feature extraction method and device.

Background technique

Myocardial infarction is the most common cardiovascular disease. It is mainly caused by the hypoxia of the corresponding downstream myocardium caused by the blockage of the coronary arteries, which in turn leads to necrosis of the myocardium in this area. ECG is the main tool for measuring the electrical activity of the heart. The 12-lead ECG can correspond to the corresponding heart area and is widely used in the clinical diagnosis of myocardial infarction. Clinically, determining the location of myocardial infarction is of great significance for further treatment. Experienced clinicians can determine the location of the infarction based on the coupling relationship between multiple leads. For example, the ST segment is too high and pathological Q waves occur in leads V1 and V2, indicating that the location of myocardial infarction may occur in the anterior wall of the heart; if similar waveform changes occur in leads II, III, and aVF, then It indicates that the infarction occurred in the lower wall of the heart. However, this judgment often places high demands on the doctor's experience, and it is a time-consuming and laborious process. Therefore, it is necessary to develop an automatic detection system for the location of myocardial infarction. Current researches are mainly based on the combination of feature extraction and traditional machine learning. These methods first acquire features related to myocardial infarction through feature extraction algorithms, such as acquiring Q waves, R waves, S waves, and T waves through feature point detection algorithms, and then Based on these feature points, the corresponding waveform features are extracted, and finally, traditional machine learning algorithms (BP neural network, SVM, K-NN, etc.) are used to classify and recognize the features.

However, existing algorithms generally need to perform key point detection such as Q wave, R wave, S wave and T wave. The accuracy of feature extraction depends on the accuracy of key point detection, and the accuracy of key point detection is directly affected by noise. Therefore, the feature extraction method based on key point detection has the disadvantage of weak anti-interference ability; at the same time, the existing algorithm has poor model generalization ability and low accuracy.

Summary of the invention

The embodiment of the present invention provides an ECG-based multi-scale feature extraction method and device, so as to at least solve the existing technical problem that deeper variation features cannot be extracted when performing feature extraction on ECG signals.

According to an embodiment of the present invention, an ECG-based multi-scale feature extraction method is provided, which includes the following steps:

A number of ECG signal identification units are obtained based on the ECG signal interception of one lead, and the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

Multi-scale decomposition of several ECG signal recognition units to construct an ECG multi-scale space;

The ECG multi-scale spatial signal in the ECG multi-scale space is extracted through a preset convolutional neural network.

Further, obtaining several ECG signal identification units based on the ECG signal interception of one lead includes:

Perform key point detection on the ECG signal of a lead to obtain the R wave apex, and then perform the ECG signal recognition unit interception on each lead according to the detected R wave apex position. The interception expression is: ECG _cell = ECG[R(n +k)-R(n)]; where: R(n+k)-R(n) represents the ECG sequence between the nth R wave vertex and the n+kth R wave vertex, n, k=1, 2, 3, 4,...

Further, k takes the value 2.

Further, the multi-scale decomposition of several ECG signal recognition units to construct the ECG multi-scale space includes:

Take 12 leads separately

The horizontal arrangement is a two-dimensional matrix A1,

The horizontal arrangement is a two-dimensional matrix A2,

The horizontal arrangement is a two-dimensional matrix A3,...; then A1, A2, A3,... are superimposed into a multi-dimensional matrix;

…Is the wavelet frequency band decomposed by the ECG signal recognition unit according to the multi-scale space construction formula of the ECG signal;

The multi-scale space construction formula of ECG signal is as follows:

Where c and d represent the approximate and detailed wavelet coefficients of the lead signal, respectively, h and g are the corresponding low-pass filters and high-pass filters, where n, k = 1, 2, 3, 4,....

Further, the ECG signal identification unit is subjected to 3-scale decomposition according to the multi-scale space construction formula of the ECG signal.

Furthermore, the preset convolutional neural network structure is designed as: input layer→convolution layer 1→convolution layer 2→pooling layer 1→convolution layer 3→convolution layer 4→pooling layer 2→convolution layer 5→Pooling layer 3→Convolutional layer 6→Pooling layer 4→Convolutional layer 7→Fully connected layer 1→Fully connected layer 2→SoftMax classifier→output layer.

Further, in the preset convolutional neural network structure, the ReLu function is used as the activation function; the training phase of the preset convolutional neural network structure adds the Dropout operation to the fully connected layer 1 and the fully connected layer 2 respectively; In the proposed convolutional neural network structure, the L2 regular term is added to the objective function.

Further, before obtaining several ECG signal identification units based on the ECG signal interception of one lead, the method further includes:

Filter the collected lead ECG signals.

Further, performing filtering processing on the collected lead ECG signal includes:

The baseline drift of the lead ECG signal is removed by wavelet technology, and the power frequency interference of the ECG signal is removed by the combined denoising method of wavelet and Butterworth filter.

According to another embodiment of the present invention, there is provided an ECG-based multi-scale feature extraction device, including:

The identification unit intercepting unit is used to obtain several ECG signal identification units based on the ECG signal interception of one lead, and the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

The ECG multi-scale space construction unit is used for multi-scale decomposition of several ECG signal recognition units to construct the ECG multi-scale space;

The multi-scale feature extraction unit is configured to extract the ECG multi-scale spatial signals in the ECG multi-scale space through a preset convolutional neural network.

The ECG-based multi-scale feature extraction method and device in the embodiment of the present invention obtain several ECG signal identification units based on the interception of the ECG signal of one lead. The ECG signal identification unit is that the ECG signal of one lead includes at least one cardiac cycle. Band, it will be more conducive to the convolutional neural network to learn the ECG variability and spatial features of myocardial infarction, and the preset convolutional neural network can extract deeper variability features from the ECG signal. This variability feature has more Strong disease discrimination ability, and according to the spatial learning ability of the convolutional neural network to obtain the spatial characteristics related to the location of the disease, it has important practical reference value for doctors to predict the location of myocardial infarction.

Description of the drawings

Figure 1 is a flowchart of the ECG-based multi-scale feature extraction method of the present invention;

Figure 2 is a preferred flow chart of the ECG-based multi-scale feature extraction method of the present invention;

Fig. 3 is a denoising result diagram of the ECG-based multi-scale feature extraction method of the present invention;

4 is a schematic diagram of the ECG signal recognition unit intercepting the ECG-based multi-scale feature extraction method of the present invention;

FIG. 5 is a schematic diagram of the 3-dimensional multi-scale space structure of the ECG-based multi-scale feature extraction method of the present invention;

6 is a block diagram of the ECG-based multi-scale feature extraction device of the present invention;

Fig. 7 is a preferred module diagram of the ECG-based multi-scale feature extraction device of the present invention.

detailed description

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

Example one

According to an embodiment of the present invention, an ECG-based multi-scale feature extraction method is provided. Referring to FIG. 1, the method includes the following steps:

S101: Obtain several ECG signal identification units based on the ECG signal interception of one lead, where the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

S102: Perform multi-scale decomposition of several ECG signal recognition units to construct an ECG multi-scale space;

S103: Perform multi-scale feature extraction on the ECG multi-scale spatial signal in the ECG multi-scale space through a preset convolutional neural network.

In the ECG-based multi-scale feature extraction method in the embodiment of the present invention, a number of ECG signal identification units are obtained based on the interception of the ECG signal of one lead, and the ECG signal identification unit is a band that includes at least one cardiac cycle in the ECG signal of one lead. It will be more conducive to the convolutional neural network to learn the ECG variability and spatial characteristics of myocardial infarction, and the preset convolutional neural network can extract deeper variability features from the ECG signal. This variability feature has strong The ability to discriminate diseases, and obtain the spatial characteristics related to the location of the disease according to the spatial learning ability of the convolutional neural network, has important practical reference value for doctors to predict the location of myocardial infarction.

Preferably, obtaining several ECG signal identification units based on the ECG signal interception of one lead includes:

Perform key point detection on the ECG signal of a lead to obtain the R wave apex, and then perform the ECG signal recognition unit interception on each lead according to the detected R wave apex position. The interception expression is: ECG _cell = ECG[R(n +k)-R(n)]; where: R(n+k)-R(n) represents the ECG sequence between the nth R wave vertex and the n+kth R wave vertex, n, k=1, 2, 3, 4,... The advantage of this method is that because each ECG signal recognition unit contains a continuous and complete cardiac cycle, it will be more conducive to the convolutional neural network to learn the ECG variability and spatial characteristics of myocardial infarction.

Preferably, k takes the value 2. This method actually intercepts the ECG sequence of two cycles, including the second half cycle of the nth cycle (the R wave front is the first half cycle, and the R wave after the R wave is the second half cycle), the n+1th complete cardiac cycle And the first half of the n+2th cycle. The advantage of this method is that each ECG signal recognition unit contains a continuous and complete cardiac cycle.

Preferably, performing multi-scale decomposition of several ECG signal recognition units, and constructing an ECG multi-scale space includes:

Take 12 leads separately

The horizontal arrangement is a two-dimensional matrix A1,

The horizontal arrangement is a two-dimensional matrix A2,

The horizontal arrangement is a two-dimensional matrix A3,...; then A1, A2, A3,... are superimposed into a multi-dimensional matrix;

…Is the wavelet frequency band decomposed by the ECG signal recognition unit according to the multi-scale space construction formula of the ECG signal;

The multi-scale space construction formula of ECG signal is as follows:

Where c and d represent the approximate and detailed wavelet coefficients of the lead signal, respectively, h and g are the corresponding low-pass filters and high-pass filters, where n, k = 1, 2, 3, 4,.... Multi-scale signals are sufficient to express the ECG variation characteristics of deep-level myocardial infarction.

Preferably, the ECG signal identification unit performs 3-scale decomposition according to the multi-scale space construction formula of the ECG signal. In the previous experiment of the present invention, it was decomposed into 3 scales, and the experiment considered that the signal of 3 scales was enough to express the ECG variation characteristics of deep-level myocardial infarction.

Preferably, the preset convolutional neural network structure is designed as: input layer → convolution layer 1 → convolution layer 2 → pooling layer 1 → convolution layer 3 → convolution layer 4 → pooling layer 2 → convolution layer 5→Pooling layer 3→Convolutional layer 6→Pooling layer 4→Convolutional layer 7→Fully connected layer 1→Fully connected layer 2→SoftMax classifier→output layer.

Preferably, in the preset convolutional neural network structure, the ReLu function is used as the activation function; in the training phase of the preset convolutional neural network structure, the dropout operation is added to the fully connected layer 1 and the fully connected layer 2 respectively; In the proposed convolutional neural network structure, the L2 regular term is added to the objective function. Since the ReLu function has the characteristics of gradient unsaturation and fast calculation speed, which can quickly achieve convergence, the present invention uses the ReLu function as the activation function. To prevent over-fitting the training model, the present invention is in the training phase are connected in the whole layer 1 and the layer 2 to the fully connected Dro _p out operation. In order to further prevent the training model from overfitting, the present invention adds an L2 regular term to the objective function to obtain sparse model parameters and improve the generalization ability of the model.

Preferably, referring to Fig. 2, the method before obtaining several ECG signal identification units based on the ECG signal interception of one lead further includes:

S100: Perform filtering processing on the collected lead ECG signals to remove related noise interference.

Preferably, performing filtering processing on the collected lead ECG signal includes:

The baseline drift of the lead ECG signal is removed by wavelet technology, and the power frequency interference of the ECG signal is removed by the combined denoising method of wavelet and Butterworth filter.

In the following, specific embodiments are used to describe in detail the ECG-based multi-scale feature extraction method of the present invention:

The present invention proposes a multi-scale feature extraction method based on the combination of wavelet transform and convolutional neural network. The method can extract deeper variation features from ECG signals. This variation feature not only has a strong ability to discriminate diseases, but also The feature's anti-noise ability is improved, and the spatial features related to the onset location are obtained according to the spatial learning ability of the convolutional neural network, which has important practical reference value for doctors to predict the location of myocardial infarction.

The method of the present invention mainly analyzes the 12-lead ECG signal through wavelet transform and convolutional neural network technology to obtain the spatial characteristics related to the location of the onset, thereby facilitating the prediction of the location of myocardial infarction, and providing important information for the doctor to predict the location of the lesion in accordance with. The technology of the present invention is mainly used for auxiliary diagnosis of doctors in hospitals, and cannot directly diagnose diseases. The content is as follows:

First, the ECG signal is pre-processed, and this step is mainly to achieve signal denoising through filtering technology; the second step is to obtain the R-wave apex of the ECG waveform through the waveform detection algorithm for the segmentation of the ECG signal identification unit; third, the segmentation The ECG unit performs multi-scale decomposition and multi-scale spatial construction to obtain multi-scale ECG spatial signals; finally, the disease-related features and spatial features are extracted through the convolutional neural network, and the disease location is classified through the softmax classifier.

The detailed process of the method of the present invention is as follows:

(1). Filter the collected ECG signals, including removing baseline drift, power frequency interference and other noises. First, the wavelet technique is used to remove the baseline drift, and then the wavelet and Butterworth filter joint denoising method is used to remove other ECG interference noise. The denoising result is shown in Figure 3.

(2). Perform key point detection on the ECG signal to obtain the apex of the R wave. For the 12-lead ECG signal, since the R wave position difference of each lead is extremely small, only one of the leads needs to be R wave detection, and then According to the detected R wave apex position, the ECG signal recognition unit is intercepted for each lead. The interception expression is: ECG _cell =ECG[R(n+k)-R(n)]; where: R(n+k) )-R(n) represents the ECG sequence between the nth R wave apex and the n+kth R wave apex (where n, k=1, 2, 3, 4,...), in the previous experiment of the present invention Set k to 2, intuitively as shown in Figure 4. It can be seen that this method actually intercepts the ECG sequence of two cycles, including the second half cycle of the nth cycle (the R wave front is the first half cycle, and the R wave is the second half cycle), the n+1th cycle The complete cardiac cycle and the first half of the n+2 cycle. The advantage of this method is that because each ECG signal recognition unit contains a continuous and complete cardiac cycle, it will be more conducive to the convolutional neural network to learn the ECG variability and spatial characteristics of myocardial infarction, and then normalize the ECG signal recognition unit Converted to a length of 400 sampling points.

(3) ECG multi-scale space construction: Multi-scale decomposition of the intercepted ECG signal recognition unit is carried out through wavelet transform technology. In the preliminary experiment of the present invention, it is decomposed in 3 scales. The experiment thinks that the signal of 3 scales is sufficient Express the ECG variation characteristics of deep-level myocardial infarction, but not limited to 3-scale decomposition. Take the 3-scale wavelet decomposition used in previous experimental studies as an example, the multi-scale space construction formula of ECG signal is as follows:

Where c and d represent the approximate and detailed wavelet coefficients of the lead signal, respectively, h and g are the corresponding low-pass filters and high-pass filters, where n, k = 1, 2, 3, 4,.... According to the wavelet decomposition formula, the original ECG signal identification unit is decomposed into 3 wavelet frequency bands, which are respectively used

Said. The steps for constructing ECG multi-scale space are as follows: firstly take 12 leads

The horizontal arrangement is a two-dimensional matrix A1,

The horizontal arrangement is a two-dimensional matrix A2,

The horizontal arrangement is a two-dimensional matrix A3; then A1, A2, and A3 are superimposed into a three-dimensional matrix, which is represented as shown in Figure 5.

(4). The ECG multi-scale spatial signal constructed in step (3) is used for feature extraction and disease identification through a convolutional neural network. The convolutional neural network structure is designed as: input layer → convolution layer 1 → convolution layer 2 → pooling layer 1 → convolution layer 3 → convolution layer 4 → pooling layer 2 → convolution layer 5 → pooling layer 3→Convolutional layer 6→Pooling layer 4→Convolutional layer 7→Fully connected layer 1→Fully connected layer 2→SoftMax classifier→output layer. Since the ReLu function has the characteristics of gradient unsaturation and fast calculation speed, which can quickly achieve convergence, the present invention uses the ReLu function as the activation function. In order to prevent the training model from overfitting, the present invention adds dropout operations to the fully connected layer 1 and the fully connected layer 2 respectively during the training phase. In order to further prevent the training model from overfitting, the present invention adds an L2 regular term to the objective function to obtain sparse model parameters and improve the generalization ability of the model.

In order to more clearly define the convolutional neural network structure used in the present invention, the following description is made in symbolic language, and the symbols are defined as follows:

l: the first convolutional layer;

f ^l : the size of the first layer convolution kernel;

p ^l : the size of the first layer padding;

s ^l : the size of the step size of the first layer;

The number of layer 1 channels;

The number of convolution kernels in the first layer;

enter:

Indicates the height, width and channel number of layer 1-1 input;

Output:

Indicates the height, width and channel number of layer 1-1 output;

The size of the output image:

According to the above symbols, the convolutional neural network structure parameters are defined as follows:

Input layer size: 400*12*3,

Convolutional layer 1 hyperparameters: f ¹ ＝3*2*3, s ¹ =1, p ¹ =0,

(Number of convolution kernels);

Convolutional layer 1 output size: 398*11*32,

(Number of channels);

Convolutional layer 2 hyperparameters: f ² ＝3*2*32, s ² =1, p ² =0,

(Number of convolution kernels);

Convolutional layer 2 output size: 396*10*64,

Pooling layer 1 hyperparameters: average pooling filter size f ³ = 2*1, s ³ = 2*1, p ³ =0,

The output size of pooling layer 1: 198*10*64,

Convolutional layer 3 hyperparameters: f ⁴ ＝3*2*64, s ⁴ =1, p ⁴ =0,

(Number of convolution kernels);

Convolutional layer 3 output size: 196*9*128,

Convolutional layer 4 hyperparameters: f ⁵ ＝3*2*128, s ⁵ =1, p ⁵ =0,

(Number of convolution kernels);

Convolutional layer 4 output size: 194*8*128,

Pooling layer 2 hyperparameters: average pooling filter size f ⁶ =2*1, s ⁶ =2*1, p ⁶ =0,

Pooling layer 2 output size: 97*8*128,

Convolutional layer 5 hyperparameters: f ⁷ ＝3*3*128, s ⁷ =1, p ⁷ =0,

(Number of convolution kernels);

Convolutional layer 5 output size: 95*6*64,

Pooling layer 3 hyperparameters: average pooling filter size f ⁸ = 2*1, s ⁸ = 2*1, p ⁸ =0,

Output size of pooling layer 3: 47*6*64,

Convolutional layer 6 hyperparameters: f ⁹ ＝3*3*64, s ⁹ =1, p ⁹ =0,

(Number of convolution kernels);

Convolutional layer 6 output size: 45*3*64,

Pooling layer 4 hyperparameters: average pooling filter size f ¹⁰ ＝2*2, s ¹⁰ ＝2*1, p ¹⁰ =0,

Pooling layer 4 output size: 22*2*64,

Convolutional layer 7 hyperparameters: f ¹¹ ＝3*2*32, s ¹¹ =1, p ¹¹ =0,

(Number of convolution kernels);

Convolutional layer 7 output size: 20*1*32,

The fully connected layer 1 is expanded into a one-dimensional vector of 20*32=640, that is, the number of neuron nodes is 640. During the training process, this layer performs a dropout operation, and the retention probability of neurons is p=0.8;

The number of neuron nodes in the fully connected layer 2 is 128, and this layer performs a dropout operation during the training process, and the retention probability of neurons is p=0.8;

Output layer: This layer has 6 output nodes. The output results of this layer are judged by the SoftMax classifier, and the output results are divided into 6 categories (anterior wall myocardial infarction, inferior wall myocardial infarction, anterior wall myocardial infarction, anterior septal myocardial infarction , Inferior wall myocardial infarction and inferior posterior wall myocardial infarction);

(5). The preliminary experiments of the present invention are trained and tested on the PTB database.

Example two

According to another embodiment of the present invention, an ECG-based multi-scale feature extraction device is provided. See FIG. 6, including:

The identification unit intercepting unit 201 is configured to obtain several ECG signal identification units based on the ECG signal interception of one lead, and the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

The ECG multi-scale space construction unit 202 is used for multi-scale decomposition of several ECG signal recognition units to construct an ECG multi-scale space;

The multi-scale feature extraction unit 203 is configured to extract the ECG multi-scale spatial signals in the ECG multi-scale space through a preset convolutional neural network.

The ECG-based multi-scale feature extraction device in the embodiment of the present invention obtains several ECG signal identification units based on the interception of the ECG signal of one lead, and the ECG signal identification unit is a band of at least one cardiac cycle included in the ECG signal of one lead, It will be more conducive to the convolutional neural network to learn the ECG variability and spatial characteristics of myocardial infarction, and the preset convolutional neural network can extract deeper variability features from the ECG signal. This variability feature has strong The ability to discriminate diseases, and obtain the spatial characteristics related to the location of the disease according to the spatial learning ability of the convolutional neural network, has important practical reference value for doctors to predict the location of myocardial infarction.

Preferably, referring to Figure 7, the device further includes:

The filtering processing unit 200 is configured to perform filtering processing on the collected lead ECG signals. The filter processing unit 200 removes the baseline drift of the lead ECG signal through wavelet technology, and then removes the power frequency interference of the ECG signal through a combined wavelet and Butterworth filter denoising method.

In the following, specific embodiments are used to describe in detail the ECG-based multi-scale feature extraction device of the present invention:

The present invention proposes a multi-scale feature extraction device based on the combination of wavelet transform and convolutional neural network. The device can extract deeper variation features from ECG signals. This variation feature not only has a strong ability to discriminate diseases, but also The feature's anti-noise ability is improved, and the spatial features related to the onset location are obtained according to the spatial learning ability of the convolutional neural network, which has important practical reference value for doctors to predict the location of myocardial infarction.

The device of the present invention mainly analyzes the 12-lead ECG signal through wavelet transform and convolutional neural network technology to obtain the spatial characteristics related to the location of the onset, thereby facilitating the prediction of the location of myocardial infarction, and providing important information for the doctor to predict the location of the lesion in accordance with. The technology of the present invention is mainly used for auxiliary diagnosis of doctors in hospitals, and cannot directly diagnose diseases. The content is as follows:

First, the ECG signal is pre-processed, and this step is mainly to achieve signal denoising through filtering technology; the second step is to obtain the R-wave apex of the ECG waveform through the waveform detection algorithm for the segmentation of the ECG signal identification unit; third, the segmentation The ECG unit performs multi-scale decomposition and multi-scale spatial construction to obtain multi-scale ECG spatial signals; finally, the disease-related features and spatial features are extracted through the convolutional neural network, and the disease location is classified through the softmax classifier.

The detailed process of the device of the present invention is as follows:

Filter processing unit 200: filter the collected ECG signals, including removing baseline drift, power frequency interference and other noises. First, the wavelet technique is used to remove the baseline drift, and then the wavelet and Butterworth filter joint denoising method is used to remove other ECG interference noise. The denoising result is shown in Figure 3.

Recognition unit intercepting unit 201: Perform key point detection on the ECG signal to obtain the apex of the R wave. For the 12-lead ECG signal, since the R wave position difference of each lead is extremely small, only one of the leads needs to be R wave detection , And then perform ECG signal recognition unit interception for each lead according to the detected R wave apex position, the interception expression is: ECG _cell = ECG[R(n+k)-R(m)]; where: R(n +k)-R(n) represents the ECG sequence between the nth R wave apex and the n+kth R wave apex (where n, k=1, 2, 3, 4,...), in the early stage of the present invention In the experiment, k is set to 2, which is intuitively shown in Figure 4. It can be seen that this method actually intercepts the ECG sequence of two cycles, including the second half cycle of the nth cycle (the R wave front is the first half cycle, and the R wave is the second half cycle), the n+1th cycle The complete cardiac cycle and the first half of the n+2 cycle. The advantage of this method is that because each ECG signal recognition unit contains a continuous and complete cardiac cycle, it will be more conducive to the convolutional neural network to learn the ECG variability and spatial characteristics of myocardial infarction, and then normalize the ECG signal recognition unit Converted to a length of 400 sampling points.

ECG multi-scale space construction unit 202: Multi-scale decomposition is performed on the intercepted ECG signal recognition unit through wavelet transform technology. In the previous experiment of the present invention, it was decomposed into 3 scales. The experiment considered that the 3-scale signal is sufficient to express The ECG variability characteristics of deep-level myocardial infarction, but not limited to 3-scale decomposition. Take the 3-scale wavelet decomposition used in previous experimental studies as an example, the multi-scale space construction formula of ECG signal is as follows:

Where c and d represent the approximate and detailed wavelet coefficients of the lead signal, respectively, h and g are the corresponding low-pass filters and high-pass filters, where n, k = 1, 2, 3, 4,.... According to the wavelet decomposition formula, the original ECG signal identification unit is decomposed into 3 wavelet frequency bands, which are respectively used

Said. The steps for constructing ECG multi-scale space are as follows: firstly take 12 leads

The horizontal arrangement is a two-dimensional matrix A1,

The horizontal arrangement is a two-dimensional matrix A2,

The horizontal arrangement is a two-dimensional matrix A3; then A1, A2, and A3 are superimposed into a three-dimensional matrix, which is represented as shown in Figure 5.

The multi-scale feature extraction unit 203: the ECG multi-scale spatial signal constructed in the ECG multi-scale space construction unit 202 is used for feature extraction and disease identification through a convolutional neural network. The convolutional neural network structure is designed as: input layer → convolution layer 1 → convolution layer 2 → pooling layer 1 → convolution layer 3 → convolution layer 4 → pooling layer 2 → convolution layer 5 → pooling layer 3→Convolutional layer 6→Pooling layer 4→Convolutional layer 7→Fully connected layer 1→Fully connected layer 2→SoftMax classifier→output layer. Since the ReLu function has the characteristics of gradient unsaturation and fast calculation speed, which can quickly achieve convergence, the present invention uses the ReLu function as the activation function. In order to prevent the training model from overfitting, the present invention adds dropout operations to the fully connected layer 1 and the fully connected layer 2 respectively during the training phase. In order to further prevent the training model from overfitting, the present invention adds an L2 regular term to the objective function to obtain sparse model parameters and improve the generalization ability of the model.

In order to more clearly define the convolutional neural network structure used in the present invention, the following description is made in symbolic language, and the symbols are defined as follows:

l: the first convolutional layer;

f ^l : the size of the first layer convolution kernel;

p ^l : the size of the first layer padding;

s ^l : the size of the step size of the first layer;

The number of layer 1 channels;

The number of convolution kernels in the first layer;

enter:

Indicates the height, width and channel number of layer 1-1 input;

Output:

Indicates the height, width and channel number of layer 1-1 output;

The size of the output image:

According to the above symbols, the convolutional neural network structure parameters are defined as follows:

Input layer size: 400*12*3,

Convolutional layer 1 hyperparameters: f ¹ ＝3*2*3, s ¹ =1, p ¹ =0,

(Number of convolution kernels);

Convolutional layer 1 output size: 398*11*32,

(Number of channels);

Convolutional layer 2 hyperparameters: f ² ＝3*2*32, s ² =1, p ² =0,

(Number of convolution kernels);

Convolutional layer 2 output size: 396*10*64,

Pooling layer 1 hyperparameters: average pooling filter size f ³ = 2*1, s ³ = 2*1, p ³ =0,

The output size of pooling layer 1: 198*10*64,

Convolutional layer 3 hyperparameters: f ⁴ ＝3*2*64, s ⁴ =1, p ⁴ =0,

(Number of convolution kernels);

Convolutional layer 3 output size: 196*9*128,

Convolutional layer 4 hyperparameters: f ⁵ ＝3*2*128, s ⁵ =1, p ⁵ =0,

(Number of convolution kernels);

Convolutional layer 4 output size: 194*8*128,

Pooling layer 2 hyperparameters: average pooling filter size f ⁶ =2*1, s ⁶ =2=1, p ⁶ =0,

Pooling layer 2 output size: 97*8*128,

Convolutional layer 5 hyperparameters: f ⁷ ＝3*3*128, s ⁷ =1, p ⁷ =0,

(Number of convolution kernels);

Convolutional layer 5 output size: 95*6*64,

Pooling layer 3 hyperparameters: average pooling filter size f ⁸ = 2*1, s ⁸ = 2*1, p ⁸ =0,

Output size of pooling layer 3: 47*6*64,

Convolutional layer 6 hyperparameters: f ⁹ ＝3*3*64, s ⁹ =1, p ⁹ =0,

(Number of convolution kernels);

Convolutional layer 6 output size: 45*3*64,

Pooling layer 4 hyperparameters: average pooling filter size f ¹⁰ ＝2*2, s ¹⁰ ＝2*1, p ¹⁰ =0,

Pooling layer 4 output size: 22*2*64,

Convolutional layer 7 hyperparameters: f ¹¹ ＝3*2*32, s ¹¹ =1, p ¹¹ =0,

(Number of convolution kernels);

Convolutional layer 7 output size: 20*1*32,

The fully connected layer 1 is expanded into a one-dimensional vector of 20*32=640, that is, the number of neuron nodes is 640. During the training process, this layer performs a dropout operation, and the retention probability of neurons is p=0.8;

The number of neuron nodes in the fully connected layer 2 is 128, and this layer performs a dropout operation during the training process, and the retention probability of neurons is p=0.8;

Output layer: This layer has 6 output nodes. The output results of this layer are judged by the SoftMax classifier, and the output results are divided into 6 categories (anterior wall myocardial infarction, inferior wall myocardial infarction, anterior wall myocardial infarction, anterior septal myocardial infarction , Inferior wall myocardial infarction and inferior posterior wall myocardial infarction);

The preliminary experiments of the present invention are trained and tested on the Penn Tree Bank (PTB) database.

The innovations of the ECG-based multi-scale feature extraction method and device of the present invention are at least:

1. The segmentation method of the ECG signal recognition unit, that is, by intercepting the segment between the nth R wave and the n+kth R wave apex as the recognition unit;

2. Using the method of multi-scale decomposition of the ECG signal recognition unit to obtain the multi-scale ECG signal recognition unit to obtain deep-level myocardial infarction related features;

3. The present invention designs an ECG multi-scale space construction method suitable for convolutional neural network input;

4. The uniquely designed convolutional neural network structure of the present invention;

5. In order to prevent over-fitting, the present invention adds Dropout operations to fully connected layer 1 and fully connected layer 2 respectively in the training phase, and adds L2 regular term to the objective function.

The present invention has the advantages of high accuracy rate and strong anti-interference ability; the present invention can complete the positioning of the myocardial infarction area with at least two ECG cycles, saving golden time for patient rescue. The present invention has carried out model training on a professional PTB database, and has carried out test verification on an independent data set, and the model accuracy rate reaches 97%.

The sequence numbers of the foregoing embodiments of the present invention are only for description, and do not represent the superiority of the embodiments.

In the above-mentioned embodiments of the present invention, the description of each embodiment has its own focus. For parts that are not described in detail in an embodiment, reference may be made to related descriptions of other embodiments.

In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiment described above is only illustrative. For example, the division of units may be a logical function division, and there may be other divisions in actual implementation. For example, multiple units or components may be combined or integrated into Another system, or some features can be ignored, or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of units or modules, and may be in electrical or other forms.

The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , Including several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods in the various embodiments of the present invention. The aforementioned storage media include: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program code .

The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications are also It should be regarded as the protection scope of the present invention.

Claims (10)

1. An ECG-based multi-scale feature extraction method is characterized by including the following steps:

   A number of ECG signal identification units are obtained based on an ECG signal interception of one lead, where the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

   Performing multi-scale decomposition of several ECG signal recognition units to construct an ECG multi-scale space;

   The ECG multi-scale spatial signal in the ECG multi-scale space is subjected to multi-scale feature extraction through a preset convolutional neural network.
2. The ECG-based multi-scale feature extraction method according to claim 1, wherein the ECG signal interception based on one lead to obtain several ECG signal identification units comprises:

   Perform key point detection on the ECG signal of one lead to obtain the R wave apex, and then perform the ECG signal recognition unit interception on each lead according to the detected R wave apex position. The interception expression is: ECG _cell =ECG[R (n+k)-R(n)]; where: R(n+k)-R(n) represents the ECG sequence between the nth R wave vertex and the n+kth R wave vertex, n, k= 1, 2, 3, 4,..., 11.
3. The ECG-based multi-scale feature extraction method according to claim 2, wherein the value of k is 2.
4. The ECG-based multi-scale feature extraction method according to claim 1, wherein the multi-scale decomposition of the plurality of ECG signal recognition units to construct an ECG multi-scale space comprises:

   Take 12 leads separately

   

   The horizontal arrangement is a two-dimensional matrix A1,

   

   The horizontal arrangement is a two-dimensional matrix A2,

   

   The horizontal arrangement is a two-dimensional matrix A3,...; then A1, A2, A3,... are superimposed into a multi-dimensional matrix;

   

   Construct a wavelet frequency band decomposed by a formula for the ECG signal identification unit according to the multi-scale space of the ECG signal;

   The multi-scale space construction formula of the ECG signal is as follows:

   

   

   Where c and d represent the approximate and detailed wavelet coefficients of the lead signal, respectively, h and g are the corresponding low-pass filters and high-pass filters, where n, k = 1, 2, 3, 4,....
5. The ECG-based multi-scale feature extraction method according to claim 4, wherein the ECG signal identification unit is subjected to 3-scale decomposition according to the multi-scale space construction formula of the ECG signal.
6. The ECG-based multi-scale feature extraction method according to claim 1, wherein the preset convolutional neural network structure is designed as: input layer → convolution layer 1 → convolution layer 2 → pooling layer 1 → Convolutional layer 3 → Convolutional layer 4 → Pooling layer 2 → Convolutional layer 5 → Pooling layer 3 → Convolutional layer 6 → Pooling layer 4 → Convolutional layer 7 → Fully connected layer 1 → Fully connected layer 2 →SoftMax classifier→output layer.
7. The ECG-based multi-scale feature extraction method according to claim 6, wherein in the preset convolutional neural network structure, a ReLu function is used as an activation function; the preset convolutional neural network structure In the training phase of, the Dropout operation is added to the fully connected layer 1 and the fully connected layer 2 respectively; in the preset convolutional neural network structure, the L2 regular term is added to the objective function.
8. The ECG-based multi-scale feature extraction method according to claim 1, wherein the method further comprises: before obtaining several ECG signal identification units based on the ECG signal interception of one lead:

   Filter the collected lead ECG signals.
9. The ECG-based multi-scale feature extraction method according to claim 8, wherein the filtering processing of the collected lead ECG signals comprises:

   The baseline drift of the lead ECG signal is removed by wavelet technology, and the power frequency interference of the ECG signal is removed by the combined denoising method of wavelet and Butterworth filter.
10. An ECG-based multi-scale feature extraction device is characterized in that it includes:

    An identification unit intercepting unit, configured to obtain several ECG signal identification units based on the interception of the ECG signal of one lead, where the ECG signal identification unit is a band including at least one cardiac cycle in the ECG signal of one lead;

    The ECG multi-scale space construction unit is used for multi-scale decomposition of several ECG signal recognition units to construct an ECG multi-scale space;

    The multi-scale feature extraction unit is configured to extract the ECG multi-scale space signal in the ECG multi-scale space through a preset convolutional neural network.

Images