CN118787360A - An ECG signal classification method based on ECG spatiotemporal feature learning - Google Patents

Abstract

The present invention discloses an electrocardiogram signal classification method based on electrocardiogram spatiotemporal feature learning. The present invention comprises: collecting 12-lead electrocardiogram signals on the human body surface, dividing the electrocardiogram signals into a training set and a test set, reconstructing the signal phase space, calculating the Euclidean distance from the signal trajectory point to the spatial origin in the phase space as a nonlinear electrocardiogram dynamic parameter; extracting the deep nonlinear electrocardiogram dynamic characteristics of the Euclidean distance parameter in the signal phase space, and calculating the time complexity and space complexity of the nonlinear electrocardiogram dynamic characteristics; selecting a corresponding classifier based on the time complexity and space complexity, inputting the nonlinear electrocardiogram dynamic characteristics of the training set into the selected classifier for training, and inputting the nonlinear electrocardiogram dynamic characteristics of the test set into the trained network for recognition, and completing the classification of the electrocardiogram signal. The present invention extracts features from the body surface electrocardiogram signals, and the network automatically performs feature learning and recognition, without the need for complex image processing, and has simple operation and high recognition accuracy.

Description

An ECG signal classification method based on ECG spatiotemporal feature learning

Technical Field

The present invention belongs to the technical field of pattern recognition, and in particular relates to an electrocardiogram signal classification method based on electrocardiogram spatiotemporal feature learning.

Background Art

Electrocardiogram (ECG) detection is a key technology in medical diagnosis, which is used to monitor and analyze the ECG activity. Although a lot of progress has been made in intelligent ECG detection, most of these works focus on extracting static or statistical features of ECG signals, converting the classification problem of ECG patterns into a classification problem of static patterns. Since static features are limited, they are not enough to fully describe the temporal properties of ECG signals. If you want to fully describe the temporal properties of ECG signals, one possible approach is to extract the nonlinear dynamic features of ECG signals and establish an effective and dynamic ECG pattern framework. However, how to extract and represent the nonlinear dynamic features in the ECG process is still a challenging problem, and how to reuse the extracted nonlinear dynamic features for ECG classification and recognition is also a key issue.

Summary of the invention

The purpose of the present invention is to overcome the problems existing in the prior art and to provide a more concise and accurate ECG signal classification method based on nonlinear electrocardiographic dynamics indicators by extracting the nonlinear dynamic characteristics of ECG signals.

To solve the above problems, the specific technical solution of the present invention is implemented through the following steps:

Step 1: Collect 12-lead ECG signals from the human body surface, divide the ECG signals into training set and test set, reconstruct the signal phase space, and calculate the Euclidean distance from the signal trajectory point to the space origin in the phase space as the nonlinear ECG dynamic parameter.

Step 2: Extract the deep nonlinear electrocardiodynamic features of the Euclidean distance parameters in the signal phase space, and calculate the time complexity and space complexity of the nonlinear electrocardiodynamic features.

Step 3: Select the corresponding classifier based on time complexity and space complexity, input the nonlinear electrocardiodynamic features of the training set into the selected classifier for training, and input the nonlinear electrocardiodynamic features of the test set into the trained network for recognition to complete the classification of ECG signals.

Preferably, the step 1 specifically includes the following process:

(1) Remove noise and artifacts from ECG signals and divide them into training set and test set.

(2) Reconstruct the phase space of 12-lead ECG signals:

in, is the information of the jth coordinate point of the ith lead in the phase space, which is given by It is composed of several information points, τ is the time delay, and d represents the embedding dimension of the phase space.

(3) Calculate the Euclidean distance from the signal trajectory point to the origin of the space in the phase space as the nonlinear electrocardiographic dynamic parameter:

in, is the Euclidean distance from the jth coordinate vector of the ith lead in the phase space to the origin of the space.

Preferably, the step 2 mainly includes the following process:

(1) Establishing a radial basis function neural network of a nonlinear electrocardiographic dynamics model, inputting the nonlinear electrocardiographic dynamics parameters of the training set and the test set into the radial basis function neural network of the nonlinear electrocardiographic dynamics model, and extracting the nonlinear electrocardiographic dynamics features stored in a constant value neural network weight matrix;

(2) Calculate the time complexity and space complexity of each dimension of the nonlinear electrocardiodynamic characteristics to obtain the total time complexity and space complexity of the nonlinear electrocardiodynamics.

Preferably, the nonlinear electrocardiographic dynamics characteristic is specifically calculated as follows:

in, represents the nonlinear electrocardiographic dynamics model of the i-th lead, x represents the state variable of the input signal, and p represents the system parameter. It is a Gaussian radial basis function RBF neural network According to the input trajectory, the convergence weight mean is learned at different distances from the network neurons. A is a diagonal matrix, and the a _i inside the diagonal matrix is the gain parameter. It represents the weights automatically learned by the Gaussian radial basis function RBF neural network, which will eventually converge; S(x) represents the Gaussian radial basis function, which contains information such as the number of neurons and their position distribution; ε _i represents the learning error; It constitutes the nonlinear electrocardiographic dynamic characteristics obtained by continuously identifying the intrinsic nonlinear dynamics of the local nonlinear electrocardiographic dynamic parameter trajectory through the RBF neural network.

Preferably, the time complexity and space complexity of each dimension of the nonlinear electrocardiodynamic characteristics are specifically calculated as follows:

Among them, TC _k represents the time complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows:

Where l _k is the number of different substrings in the kth new sequence, O _k (n) represents the complexity count, and n represents the length of the new sequence; the calculation process is as follows:

(1) Take the extreme points of the nonlinear electrocardiodynamic feature x, sort them in ascending order, and transform them into a new sequence F(j′) of length l;

(2) Create a symbol set γ = {1,…,t}, transform the extreme points of the nonlinear electrocardiodynamic characteristics, and obtain a new symbol sequence y(i'):

(3) Decompose the symbol sequence y(i') into a series of substrings. The first substring is recorded as a = {y _k (1), y _k (2), ..., y _k (m'-1)}, and the second substring is recorded as b = {y _k (m')}. Then ab = {y _k (1), y _k (2), ..., y _k (m'-1), y _k (m')}. It means to remove the substring of the last element of ab, that is If b is If the substring is not a string, update b = {y _k (m'), y _k (m'+1)}, otherwise update a = {y _k (1), y _k (2), ..., y _k (m')}, update b = {y _k (m'+1)}, and increase the complexity count O _k (n) by 1;

S _k represents the spatial complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows:

Where l' _k is the number of different substrings in the kth new sequence, _Ok (m) represents the complexity count, and m represents the length of the new sequence; the calculation process is as follows:

(1) Calculate the directional derivative of each point of the nonlinear electrocardiodynamic characteristic x to obtain a spatial rate sequence r(i) reflecting each point of the nonlinear electrocardiodynamic characteristic x.

(2) Decompose r(i) into a series of substrings. The complexity calculation O _k (m) is consistent with the rule of time complexity counting O _k (n).

Preferably, in step 3, the method of selection is:

When the time complexity TC is much greater than the space complexity SC, that is, TC>>SC, the space-dominant classification model is selected for classification and recognition;

When the time complexity TC is much smaller than the space complexity SC, that is, TC＜＜SC, the time-dominant classification model is selected for classification and recognition;

When the time complexity TC and space complexity SC satisfy TC>SC and When , the time-space cascade model is selected for classification and recognition.

When the time complexity TC and space complexity SC satisfy SC>TC and When , the space-time cascade model is selected for classification and recognition.

Preferably, the spatially dominant classification model has the following specific structure:

The spatial dominance classification model includes an input processing layer, a network initialization layer, a dense connection layer, a global average pooling layer, and a fully connected layer, a total of five layers.

(1) The first layer is the input processing layer. For the 12-dimensional nonlinear electrocardiographic dynamics characteristics, the specific input method is as follows:

I ₁ =C[r(x ₁ ),r(x ₂ ),...,r(x ₁₂ )]

Where x ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r represents the dimension reconstruction operation of the features, C represents the stacking of the reconstructed feature matrix, and the stacked three-dimensional feature I ₁ is used as the input of the space-dominant classification model.

(2) The second layer is the network initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(3) The third layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(5) The fifth layer is the fully connected layer, which completes the classification of ECG signals.

Preferably, the time-dominant classification model has the following specific structure:

The time-dominant classification model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, and fully connected layer, a total of five layers.

(1) The first layer is the initial layer of the network, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64;

(4) The fourth layer is the feature fusion layer, and the specific fusion method is as follows:

I ₂ =A[y ₁ ,y ₂ ,...,y ₁₂ ]

Among them, y ₁ ,y ₂ ,...,y ₁₂ represent the features of the nonlinear ECG features of each dimension after the first three layers of the time-dominant classification model and the global average pooling layer outputs. A represents the element-by-element addition method. After feature fusion, the 12-dimensional time features become one-dimensional time features.

(5) The fifth layer is the fully connected layer, which completes the classification of ECG signals.

Preferably, the time-space cascade model has the following specific structure:

The time-space cascade model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, spatial initialization layer, dense connection layer, global average pooling layer, and fully connected layer, a total of eight layers.

(1) The first layer is the initial layer, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64;

(4) The fourth layer is the feature processing layer, which is mainly responsible for reconstructing and stacking the 12 one-dimensional time features output. The specific calculation method is as follows:

I ₃ =C[r(y ₁ ),r(y ₂ ),...,r(y ₁₂ )]

Among them, y ₁ ,y ₂ ,...,y ₁₂ are the one-dimensional time feature vectors extracted from the nonlinear electrocardiodynamic features through the first three layers, r represents the dimension reconstruction operation of the one-dimensional time feature vector, C represents the stacking of the depth of the reconstructed feature matrix, and the stacked three-dimensional feature I ₃ is used as the input of the subsequent layer;

(5) The fifth layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(6) The sixth layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random inactivation layer stacked in sequence. The dense layers are densely connected; the transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(8) The eighth layer is a fully connected layer, which completes the classification of ECG signals;

Preferably, the space-time cascade model has the following specific structure:

The spatial-temporal cascade model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, spatial initialization layer, dense connection layer, global average pooling layer, and fully connected layer, a total of eight layers.

(1) The first layer is the input processing layer. The specific processing method for the 12-dimensional nonlinear electrocardiographic dynamics characteristics is as follows:

I ₄ =C[r(x ₁ ),r(x ₂ ),...,r(x ₁₂ )]

Where k ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r represents the dimension reconstruction operation of the features, C represents the stacking of feature matrices, and the stacked three-dimensional features I ₄ are used as the input of the space-time cascade model;

(2) The second layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(3) The third layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(5) The fifth layer is the temporal initialization layer, which consists of a one-dimensional convolutional layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(6) The sixth layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 18;

(8) The eighth layer is the fully connected layer, which completes the classification of ECG signals.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The phase space of the 12-lead ECG signal on the body surface is reconstructed. The obtained nonlinear ECG dynamic parameter ED can better reflect the intrinsic information of the ECG signal. At the same time, nonlinear feature extraction is further performed. This nonlinear dynamic representation has a high discrimination ability for ECG recognition.

2. Different from the commonly used single-lead ECG classification, this method makes full use of the 12-lead ECG signals and extracts the spatiotemporal characteristics of each lead based on the combination of time complexity and space complexity to form a comprehensive ECG feature method with better recognition ability for ECG signals.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an electrocardiogram signal classification method based on nonlinear electrocardiogram dynamics proposed by the present invention.

FIG2 is a three-dimensional schematic diagram of reconstructing the phase space of the avR lead in the 12-lead signal in an embodiment.

FIG. 3 is a schematic diagram of a nonlinear electrocardiographic dynamic parameter ED of the avR lead in the 12-lead signal in an embodiment.

FIG. 4 is a schematic diagram of the structure of a space-dominant classification model in an embodiment.

FIG5 is a schematic diagram of the structure of a time-dominant classification model in an embodiment.

FIG6 is a schematic diagram of the structure of a time-space cascade model in an embodiment.

FIG. 7 is a schematic diagram of the structure of a space-time cascade model in an embodiment.

DETAILED DESCRIPTION

The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

Example

As shown in FIG1 , a method for classifying electrocardiogram signals based on nonlinear electrocardiogram dynamics includes the following steps:

Step 1: Collect 12-lead ECG signals from the human body surface, divide the ECG signals into training set and test set, reconstruct the signal phase space, and calculate the Euclidean distance from the signal trajectory point to the origin of the space in the phase space as the nonlinear ECG dynamic parameter. Specifically, the following steps are included:

(1) Remove noise and artifacts from ECG signals and divide them into training and test sets;

(2) Reconstruct the phase space of 12-lead ECG signals:

in, is the information of the jth coordinate point of the ith lead in the phase space, which is given by It is composed of several information points, τ is the time delay, and d is the embedding dimension of the phase space. In this embodiment, τ is set to 1, d is set to 3, and the visualization result of the phase space reconstruction of the ECG signal is shown in Figure 2.

(3) Calculate the Euclidean distance from the signal trajectory point to the origin of the space in the phase space as the nonlinear electrocardiographic dynamic parameter:

in, is the Euclidean distance between the jth coordinate vector of the ith lead in the phase space and the origin of the space. The visualization result of the ED sequence is shown in Figure 3.

Step 2: Extract the deep nonlinear electrocardiodynamic features of the Euclidean distance parameters in the signal phase space and calculate the time complexity and space complexity of the nonlinear electrocardiodynamic features. Specifically, the following steps are included:

(1) Establishing a radial basis function neural network of a nonlinear electrocardiographic dynamics model, inputting the nonlinear electrocardiographic dynamics parameters of the training set and the test set into the radial basis function neural network of the nonlinear electrocardiographic dynamics model, and extracting the nonlinear electrocardiographic dynamics features stored in a constant value neural network weight matrix;

Nonlinear electrocardiographic dynamics characteristics, the specific calculation method is as follows:

in, represents the nonlinear electrocardiographic dynamics model of the i-th lead, x represents the state variable of the input signal, and p represents the system parameter. It is a Gaussian radial basis function RBF neural network According to the input trajectory, the average of the convergence weights learned at different distances from the network neurons, A is a diagonal matrix, and a _i inside the diagonal matrix is a gain parameter, which is set to 0.9. represents the weights automatically learned by the Gaussian radial basis function RBF neural network, which will eventually converge; S(x) represents the Gaussian radial basis function, which contains information such as the number of neurons and the position distribution. In this embodiment, the RBF neurons are evenly distributed in the area [-2.5, 2.5] × [-2.5, 2.5] × [-2.5, 2.5], and the width is set to 0.07; ε _i represents the learning error; The nonlinear electrocardiodynamic characteristics are obtained by continuously identifying the intrinsic nonlinear dynamics of the local nonlinear electrocardiodynamic parameter trajectory through the RBF neural network, and finally the nonlinear electrocardiodynamic characteristics can be obtained.

in, represents the mean value of the network convergence weights learned by the radial basis function neural network for the ED sequence of 12 nonlinear ECG dynamic parameters, and It represents 12 nonlinear electrocardiodynamic features stored in a constant neural network weight matrix.

(2) Calculate the time complexity and space complexity of each dimension of the nonlinear electrocardiodynamic characteristics to obtain the total time complexity and space complexity of the nonlinear electrocardiodynamic characteristics. The specific calculation method for the time complexity and space complexity of each dimension of the nonlinear electrocardiodynamic characteristics is as follows:

Among them, TC _k represents the time complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows:

Where l _k is the number of different substrings in the Kth new sequence, O _k (n) represents the complexity count, and n represents the length of the new sequence; the calculation process is as follows:

(1) Take the extreme points of the nonlinear electrocardiodynamic feature x, sort them in ascending order, and transform them into a new sequence F(j′) of length l;

(2) Create a symbol set γ = {1,…,t}, transform the extreme points of the nonlinear electrocardiodynamic characteristics, and obtain a new symbol sequence y(i'):

(3) Decompose the symbol sequence y(i') into a series of substrings. The first substring is recorded as a = {y _k (1), y _k (2), ..., y _k (m'-10}, and the second substring is recorded as b = {y _k (m')}. Then ab = {y _k (1), y _k (2), ..., y _k (m'-1), y _k (m')}. It means to remove the substring of the last element of ab, that is If b is If the substring is not a string, update b = {y _k (m'), y _k (m'+1)}, otherwise update a = {y _k (1), y _k (2), ..., y _k (m')}, update b = {y _k (m'+1)}, and increase the complexity count O _k (n) by 1;

SC _k represents the spatial complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows:

Where l' _k is the number of different substrings in the Kth new sequence, _Ok (m) represents the complexity count, and m represents the length of the new sequence; the calculation process is as follows:

(1) Calculate the directional derivative of each point of the nonlinear electrocardiodynamic characteristic x to obtain a spatial rate sequence r(i) reflecting each point of the nonlinear electrocardiodynamic characteristic x.

(2) Decompose r(i) into a series of substrings. The complexity calculation O _k (m) is consistent with the rule of time complexity counting O _k (n).

Step 3: Based on the calculated time complexity and space complexity, the method for selecting the spatiotemporal network in this embodiment is:

When the time complexity TC is much greater than the space complexity SC, that is, TC>>SC, the space-dominant classification model is selected for classification and recognition. The space-dominant classification model is shown in Figure 4, and its specific structure is as follows:

The spatial dominance classification model includes an input processing layer, a network initialization layer, a dense connection layer, a global average pooling layer, and a fully connected layer, a total of five layers.

(1) The first layer is the input processing layer. For the 12-dimensional nonlinear electrocardiographic dynamics characteristics, the specific input method is as follows:

I ₁ =C[r(x ₁ ),r(x ₂ ),...,r(x ₁₂ )]

Wherein x ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r indicates that a dimension reconstruction operation is performed on the features. In this embodiment, the length of the nonlinear electrocardiodynamic feature is 2048, and the dimension reconstruction reconstructs each nonlinear electrocardiodynamic feature into a two-dimensional matrix feature map of dimension 32×64. C indicates stacking the reconstructed feature matrix, and the stacked three-dimensional feature I ₁ is used as the input of the spatially dominant classification model.

(2) The second layer is the network initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(3) The third layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(5) The fifth layer is the fully connected layer, which completes the classification of ECG signals.

When the time complexity TC is much smaller than the space complexity SC, that is, TC＜＜SC, the time-dominant classification model is selected for classification and recognition. The time-dominant classification model is shown in Figure 5, and its specific structure is as follows:

The time-dominant classification model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, and fully connected layer, a total of five layers.

(1) The first layer is the initial layer of the network, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64;

(4) The fourth layer is the feature fusion layer, and the specific fusion method is as follows:

I ₂ =A[y ₁ ,y ₂ ,...,y ₁₂ ]

Among them, y ₁ ,y ₂ ,...,y ₁₂ represent the features of the nonlinear ECG features of each dimension after the first three layers of the time-dominant classification model and the global average layer output, A represents the element-by-element addition method, and the 12-dimensional time features are transformed into one-dimensional time features after feature fusion;

(5) The fifth layer is the fully connected layer, which completes the classification of ECG signals.

When the time complexity TC and space complexity SC satisfy TC>SC and When , the time-space cascade model is selected for classification and recognition. The time-space cascade model is shown in Figure 6, and the specific structure is as follows:

The time-space cascade model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, spatial initialization layer, dense connection layer, global average pooling layer, and fully connected layer, a total of eight layers.

(1) The first layer is the initial layer, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64;

(4) The fourth layer is the feature processing layer, which is mainly responsible for reconstructing and stacking the 12 one-dimensional time features output. The specific calculation method is as follows:

I ₃ =C[r(y ₁ ),r(y ₂ ),...,r(y ₁₂ )]

Wherein, y ₁ ,y ₂ ,...,y ₁₂ are one-dimensional time feature vectors extracted by the nonlinear electrocardiodynamic features through the first three layers, r represents the dimension reconstruction operation performed on the one-dimensional time feature vector. In this embodiment, the length of the one-dimensional time feature vector output by the nonlinear electrocardiodynamic features is 2048, and the dimension reconstruction reconstructs each feature vector into a two-dimensional matrix feature map of dimension 32×64, C represents the stacking of the depth of the reconstructed feature matrix, and the stacked three-dimensional feature I ₃ is used as the input of the subsequent layer;

(5) The fifth layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(6) The sixth layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random inactivation layer stacked in sequence. The dense layers are densely connected; the transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(8) The eighth layer is a fully connected layer, which completes the classification of ECG signals;

When the time complexity TC and space complexity SC satisfy SC>TC and When , the space-time cascade model is selected for classification and recognition. The space-time cascade model is shown in Figure 7, and the specific structure is as follows:

The spatial-temporal cascade model includes an input processing layer, a spatial initialization layer, a dense connection layer, a global average pooling layer, a temporal initialization layer, a residual connection layer, a global average pooling layer, and a fully connected layer, a total of eight layers.

(1) The first layer is the input processing layer. The specific processing method for the 12-dimensional nonlinear electrocardiographic dynamics characteristics is as follows:

I ₄ =C[r(x ₁ ),r(x ₂ ),...,r(x ₁₂ )]

Wherein x ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r indicates that a dimension reconstruction operation is performed on the feature. In this embodiment, the length of the nonlinear electrocardiodynamic feature is 2048, and the dimension reconstruction reconstructs each nonlinear electrocardiodynamic feature into a two-dimensional matrix feature map with a dimension of 32×64. C indicates stacking the feature matrix, and the stacked three-dimensional feature I ₄ is used as the input of the space-time cascade model.

(2) The second layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels;

(3) The third layer is a dense connection layer, which is mainly composed of 3 dense modules and 2 transformation layers. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolution layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolution layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2.

(4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16;

(5) The fifth layer is the temporal initialization layer, which consists of a one-dimensional convolutional layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3;

(6) The sixth layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers.

(7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 18;

(8) The eighth layer is the fully connected layer, which completes the classification of ECG signals.

In the embodiment, the accuracy of ECG signal classification based on ECG spatiotemporal feature learning can reach up to 92%, and the accuracy of nonlinear feature extraction using the original signal without using the nonlinear ECG dynamic parameter ED can reach up to 86%.

In this embodiment, 12-lead ECG signals are subjected to ECG classification. Compared with the traditional ECG pattern classification method, this method fully considers the nonlinear electrocardiographic dynamics characteristics and the corresponding spatiotemporal characteristics, and its ECG recognition rate will be more accurate.

The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims (10)

1. A method for classifying electrocardiogram signals based on electrocardiogram spatiotemporal feature learning, characterized in that it comprises the following steps: Step 1: Collect 12-lead ECG signals from the human body surface, divide the ECG signals into training set and test set, reconstruct the signal phase space, and calculate the Euclidean distance from the signal trajectory point to the space origin in the phase space as the nonlinear ECG dynamic parameter; Step 2: Extract the deep nonlinear electrocardiodynamic features of the Euclidean distance parameters in the signal phase space, and calculate the time complexity and space complexity of the nonlinear electrocardiodynamic features; Step 3: Select the corresponding classifier based on time complexity and space complexity, input the nonlinear electrocardiodynamic features of the training set into the selected classifier for training, and input the nonlinear electrocardiodynamic features of the test set into the trained network for recognition to complete the classification of ECG signals. 2. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 1, characterized in that the step 1 specifically comprises the following steps: 1-1. Remove noise and artifacts from ECG signals and divide them into training set and test set; 1-2. Reconstruct the phase space of 12-lead ECG signals: in, is the information of the jth coordinate point of the ith lead in the phase space, which is given by It is composed of several information points, while τ is the time delay and d represents the embedding dimension of the phase space; 1-3. Calculate the Euclidean distance from the signal trajectory point to the origin of the space in the phase space as the nonlinear electrocardiographic dynamic parameter: in, is the Euclidean distance from the jth coordinate vector of the ith lead in the phase space to the origin of the space. 3. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 1 or 2, characterized in that the step 2 specifically comprises the following steps: 2-1. Establish a radial basis function neural network of a nonlinear electrocardiodynamic model, input the nonlinear electrocardiodynamic parameters of the training set and the test set into the radial basis function neural network of the nonlinear electrocardiodynamic model, and extract the nonlinear electrocardiodynamic features stored in a constant value neural network weight matrix; 2-2. Calculate the time complexity and space complexity of each dimension of the nonlinear electrocardiodynamic characteristics to obtain the total time complexity and space complexity of nonlinear electrocardiodynamics. 4. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 3 is characterized in that the nonlinear electrocardiographic dynamics feature described in step 2-1 is specifically calculated as follows: in, represents the nonlinear electrocardiographic dynamics model of the i-th lead, x represents the state variable of the input signal, and p represents the system parameter; It is a Gaussian radial basis function RBF neural network According to the input trajectory, the convergence weight mean is learned at different distances from the network neurons. A is a diagonal matrix, and the a _i inside the diagonal matrix is the gain parameter. It represents the weights automatically learned by the Gaussian radial basis function RBF neural network, which will eventually converge; S(x) represents the Gaussian radial basis function; ε _i represents the learning error; It constitutes the nonlinear electrocardiographic dynamic characteristics obtained by continuously identifying the intrinsic nonlinear dynamics of the local nonlinear electrocardiographic dynamic parameter trajectory through the RBF neural network. 5. According to claim 3, the ECG signal classification method based on ECG spatiotemporal feature learning is characterized in that the time complexity and space complexity of step 2-2 are specifically calculated as follows: Among them, TC _k represents the time complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows: Where l _k is the number of different substrings in the kth new sequence, O _k (n) represents the complexity count, and n represents the length of the new sequence; the calculation process is as follows: (1) Take the extreme points of the nonlinear electrocardiodynamic feature x, sort them in ascending order, and transform them into a new sequence 'F(j) with a length of l; (2) Create a symbol set γ = {1,…,t}, transform the extreme points of the nonlinear electrocardiodynamic characteristics, and obtain a new symbol sequence y(i'): (3) Decompose the symbol sequence y(i') into a series of substrings. The first substring is recorded as a = {y _k (1), y _k (2), ..., y _k (m'-1)}, and the second substring is recorded as b = {y _k (m')}. Then ab = {y _k (1), y _k (2), ..., y _k (m'-1), y _k (m')}. It means to remove the substring of the last element of ab, that is If b is If the substring is not a string, update b = {y _k (m'), y _k (m'+1)}, otherwise update a = {y _k (1), y _k (2), ..., y _k (m')}, update b = {y _k (m'+1)}, and increase the complexity count O _k (n) by 1; S _k represents the spatial complexity of the kth nonlinear electrocardiographic dynamic parameter ED sequence, and its calculation formula is as follows: Where l' _k is the number of different substrings in the kth new sequence, _Ok (m) represents the complexity count, and m represents the length of the new sequence; the calculation process is as follows: ① Calculate the directional derivative of each point of the nonlinear electrocardiodynamic characteristic x to obtain the spatial change rate sequence r(i) reflecting each point of the nonlinear electrocardiodynamic characteristic x; ② Decompose r(i) into a series of substrings. Its complexity calculation O _k (m) is consistent with the rule of time complexity counting O _k (n). 6. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 3, characterized in that in step 3, the method selected is: When the time complexity TC is much greater than the space complexity SC, that is, TC>SC, the space-dominant classification model is selected for classification and recognition; When the time complexity TC is much smaller than the space complexity SC, that is, TC>SC<SC, the time-dominant classification model is selected for classification and recognition; When the time complexity TC and space complexity SC satisfy TC>SC and When , the time-space cascade model is selected for classification and recognition; When the time complexity TC and space complexity SC satisfy SC>TC and When , the space-time cascade model is selected for classification and recognition. 7. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 6 is characterized in that the spatially dominant classification model has the following specific structure: The spatially dominant classification model includes an input processing layer, a network initialization layer, a dense connection layer, a global average pooling layer, and a fully connected layer, a total of five layers. (1) The first layer is the input processing layer. For the 12-dimensional nonlinear electrocardiographic dynamics characteristics, the specific input method is as follows: I ₁ =C[r(x ₁ ),r(x ₂ ),...,r(x ₁₂ )] Where x ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r represents the dimension reconstruction operation of the features, C represents the stacking of the reconstructed feature matrix, and the stacked three-dimensional feature I ₁ is used as the input of the spatially dominant classification model; (2) The second layer is the network initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels; (3) The third layer is a densely connected layer, which is mainly composed of 3 dense modules and 2 transformation layers alternating. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolutional layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolutional layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2. (4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16; (5) The fifth layer is the fully connected layer, which completes the classification of ECG signals. 8. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 6 is characterized in that the time-dominant classification model has the following specific structure: The time-dominant classification model includes a five-layer structure: the initial network layer, the residual connection layer, the global average pooling layer, the feature fusion layer, and the fully connected layer. (1) The first layer is the initial layer of the network, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3; (2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers. (3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64; (4) The fourth layer is the feature fusion layer, and the specific fusion method is as follows: I ₂ =A [y ₁ , y ₂ ,..., y ₁₂ ] Among them, y ₁ ,y ₂ ,...,y ₁₂ represent the features of the nonlinear ECG features of each dimension after the first three layers of the time-dominant classification model and the global average pooling layer outputs. A represents the element-by-element addition method. After feature fusion, the 12-dimensional time features become one-dimensional time features. (5) The fifth layer is the fully connected layer, which completes the classification of ECG signals. 9. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 6 is characterized in that the time-space cascade model has the following specific structure: The time-space cascade model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, spatial initialization layer, dense connection layer, global average pooling layer, and fully connected layer, a total of eight layers; (1) The first layer is the initial layer, which consists of a one-dimensional convolution layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3; (2) The second layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers. (3) The third layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 64; (4) The fourth layer is the feature processing layer, which is mainly responsible for reconstructing and stacking the 12 one-dimensional time features output. The specific calculation method is as follows: I ₃ =C[r(y ₁ ),r(y ₂ ),...,r(y ₁₂ )] Among them, y ₁ ,y ₂ ,...,y ₁₂ are the one-dimensional time feature vectors extracted from the nonlinear electrocardiodynamic features through the first three layers, r represents the dimension reconstruction operation of the one-dimensional time feature vector, C represents the stacking of the depth of the reconstructed feature matrix, and the stacked three-dimensional feature I ₃ is used as the input of the subsequent layer; (5) The fifth layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels; (6) The sixth layer is a densely connected layer, which is mainly composed of 3 dense modules and 2 transformation layers alternating. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolutional layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolutional layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2. (7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16; (8) The eighth layer is the fully connected layer, which completes the classification of ECG signals. 10. The ECG signal classification method based on ECG spatiotemporal feature learning according to claim 6, characterized in that the specific structure of the space-time cascade model is as follows: The spatial-temporal cascade model includes the network initial layer, residual connection layer, global average pooling layer, feature fusion layer, spatial initialization layer, dense connection layer, global average pooling layer, and fully connected layer, a total of eight layers; (1) The first layer is the input processing layer. The specific processing method for the 12-dimensional nonlinear electrocardiographic dynamics characteristics is as follows: I ₄ =C[r(x ₁ ),r(x ₂ ),..,r(x ₁₂ )] Where x ₁ , x ₂ , ..., x ₁₂ are nonlinear electrocardiodynamic features, r represents the dimension reconstruction operation of the features, C represents the stacking of the feature matrix, and the stacked three-dimensional features I ₄ are used as the input of the space-time cascade model; (2) The second layer is the spatial initialization layer, which includes a three-dimensional convolution layer composed of 64 3×3×3 convolution kernels; (3) The third layer is a densely connected layer, which is mainly composed of 3 dense modules and 2 transformation layers alternating. The 3 dense modules are composed of 3, 4, and 6 dense layers respectively. The dense layers are composed of a 3D convolutional layer with a kernel size of 3×3×3, a nonlinear activation layer, and a random dropout layer stacked in sequence. The dense layers are densely connected. The transformation layer is mainly composed of a 3D convolutional layer with a kernel size of 1×1×1 and a maximum pooling layer with a kernel size of 1×2×2. (4) The fourth layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 12×8×16; (5) The fifth layer is the temporal initialization layer, which consists of a one-dimensional convolutional layer composed of 64 convolution kernels with a kernel size of 7 and a one-dimensional maximum pooling layer with a kernel size of 3; (6) The sixth layer is the residual connection layer, which includes four residual modules. The four residual modules are composed of 3, 4, 6, and 3 residual layers respectively, totaling 16 residual layers. Each residual layer is mainly composed of a one-dimensional convolution layer, a normalization layer, and a nonlinear activation layer in sequence. Residual connections are used between layers. (7) The seventh layer is the global average pooling layer, which consists of an average pooling layer with a kernel size of 18; (8) The eighth layer is the fully connected layer, which completes the classification of ECG signals.