An electrocardiosignal denoising method based on deep learning and electrocardiosimulation technology

Abstract

The application discloses a kind of based on deep learning and electrocardio simulation technology electrocardio signal denoising method, belong to medical technical field, a kind of based on deep learning and electrocardio simulation technology electrocardio signal denoising method, using following steps: step one, setting parameter, generates electrocardio signal;Step two, randomly generate the electric conductivity at the field point in range, inner interface electric conductivity, outer interface electric conductivity;Step three, with the electrocardio signal obtained by simulation as input, with electric conductivity at the field point, inner interface electric conductivity, outer interface electric conductivity as output, train deep learning network;Step four, using deep learning network to the electrocardio signal obtained by real collection carries out denoising;The application can avoid the problem that the noise interference amplitude is big due to movement, or device lead conductivity instability and the like, effectively removes the noise of dynamic electrocardiogram this part, improves the accuracy of use.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a method for denoising electrocardiogram (ECG) signals based on deep learning and ECG simulation technology. Background Technology

[0002] Extracting the R-wave peak position of an electrocardiogram (ECG) signal is often the first step in automating dynamic ECG processing. Many existing heartbeat classification algorithms based on physical models rely on the accuracy of the R-wave peak position. Therefore, improving the accuracy of R-wave peak extraction is of great significance. The main reason for the inaccuracy of the R-wave peak position in existing algorithms is the interference in the dynamic ECG.

[0003] Existing denoising algorithms mainly focus on power frequency interference, respiratory interference, and high-frequency noise in electrocardiograms (ECGs). However, unlike static ECGs, dynamic ECGs have a wider variety of interference types, and the noise interference amplitude is large due to motion or unstable conductivity of device leads. Since this interference does not have a fixed pattern, how to remove this noise has become a challenge in dynamic ECG processing. To address this issue, this application proposes an ECG signal denoising method based on a combination of deep learning algorithms and ECG simulation technology. Summary of the Invention

[0004] The purpose of this invention is to solve the problems existing in the prior art by proposing a method for denoising electrocardiogram (ECG) signals based on deep learning and ECG simulation technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for denoising electrocardiogram (ECG) signals based on deep learning and ECG simulation technology, comprising the following steps:

[0007] Step 1: Set parameters to generate an electrocardiogram (ECG) signal;

[0008] Step 2: Randomly generate the conductivity σ at the field point in the range of [1.2-4.5] and the internal interface conductivity σ in the range of [1.3-3.5]. l - The interfacial conductivity, ranging from [2.3 to 5.5], is σ. l + Randomly generate n*n integers between [0,36] and assign them to ε(i,j,0), then generate different electrocardiogram signals according to step one;

[0009] Step 3: Using the simulated electrocardiogram signal as input, and taking the conductivity σ at the field point and the conductivity σ at the internal interface as input... l - The external interface conductivity is σ l +with ε(i,j,0) as the output, to train a deep learning network;

[0010] Step 4: denoise the actually collected electrocardiographic signals using the deep learning network.

[0011] Preferably, in Step 1, the specific steps for generating electrocardiographic signals are as follows:

[0012] a. Determine the n*n-dimensional transfer coefficient matrix M according to the following equation t , where n is the number of nodes of the epicardial surface model;

[0013] wherein,

[0014]

[0015] b. Randomly generate n*n integers in the range of [0, 36], assign them to ε(i,j,0), and calculate M0 according to the formula provided in step a;

[0016] c. Solve for Φ according to the following equation 0 ;

[0017]

[0018] d. Solve for the epicardial potential;

[0019]

[0020] Further, in the formula ,

[0021] and, when 0<x<36, s(x) is respectively {-80,-80,-80,-80,55.9,30.3,8.0,-6.8,-14.2,-19.2,-23.3,-28.2,-32.4,-39.0,-44.7,-49.7,-54.6,-60.4,-64.5,-68.7,-72.8,-74.4,-76.1,-77.8,-78.6,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80}, and when x takes other values, s(x)=0.

[0022] Still further, the method is characterized in that, in the formula

[0023]

[0024] , δ is the number of surrounding elements whose values are in the range of [1-35], δ t h=2.

[0025] Furthermore, in the formula middle,

[0026] n is the number of nodes in the epicardial surface model, D j R represents the normalized vector from the center point of the heart to the node position of the i-th epicardial surface model, and R represents the fixed field point vector. This represents partial derivative operations.

[0027] Furthermore, in the formula

[0028]

[0029] In this context, r′ is the field point on the epicardial surface; σ is the conductivity at that field point; Ω h Represents the entire heart region; J s It is the equivalent dipole source for myocardium; S l (1=2,3,*m) represents a closed interface with uniform conductivity within the human torso, where the conductivity of the inner interface is σ. l - The conductivity of the outer interface is σ. l + S l Representative surface.

[0030] Preferably, in step three, the deep learning network is constructed using a 7-layer network plus a classifier, i.e., layer 1-layer 7;

[0031] Using the simulated electrocardiogram signal as input, and the conductivity σ at the field point and the conductivity σ at the inner interface are used as input. l - The external interface conductivity is σ l + ε(i,j,0) is used as the output to train the deep learning network. The training algorithm can be: stochastic gradient descent, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm, Adamax algorithm, etc.

[0032] Furthermore, layers 1-6 each consist of a convolutional layer and a pooling layer, while layer 7 is a fully connected layer. Specifically:

[0033] The convolutional layer in layer 1 contains 5 kernels, each with a kernel size of 224. The stride and kernel size in the pooling layer in layer 1 are both 2.

[0034] The convolutional layer 2 contains 5 kernels, each with a kernel size of 112. The stride and kernel size of the pooling layer in layer 2 are both 2.

[0035] Layer 3 convolutional layer contains 10 kernels, all with a kernel size of 100. The stride and kernel size of the pooling layer in layer 3 are both 2.

[0036] Layer 4 convolutional layer contains 10 kernels, all with a kernel size of 50. The stride and kernel size of the pooling layers in layer 4 are both 2.

[0037] Layer 5 convolutional layer contains 10 kernels, all with a kernel size of 48. The stride and kernel size of the pooling layer in layer 5 are both 2.

[0038] Layer 6 convolutional layer contains 10 kernels, all with a kernel size of 24. The stride and kernel size of the pooling layer in layer 4 are both 2.

[0039] The number of input neurons in layer 7 is the same as the number of output features in layer 6. The number of output neurons is 100. The classifier has 100 input features and 39 output neurons.

[0040] Furthermore, in step three, the training algorithm adopts any one of the following algorithms: stochastic gradient descent, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm, and Adamax algorithm.

[0041] Furthermore, in step four, the specific steps for denoising the actually acquired electrocardiogram signals are as follows:

[0042] ①. Input the electrocardiogram signal into the neural network trained in step three, and calculate the conductivity σ at the field point and the conductivity σ at the inner interface. l - The external interface conductivity is σ l + , ε(i,j,0);

[0043] ②. Calculate the 39 values ​​obtained in step ① according to step one to obtain a new ECG signal, which is the denoised ECG signal.

[0044] The parts of this device not covered herein are the same as or can be implemented using existing technologies. In this invention, different parameters are first input to obtain different ECG simulation signals. The simulation signals are used as input, and the parameters are used as output to train a deep learning network. Finally, the deep learning network is used to denoise the ECG signals obtained from actual acquisition. This application can avoid the problem of large noise interference caused by motion or unstable conductivity of device leads, effectively remove noise from this part of the dynamic ECG, make it easier to use, and improve the accuracy of use. Detailed Implementation

[0045] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicating the orientation or positional relationship are only for the convenience of describing this invention and simplifying the description, and therefore should not be construed as limiting this invention.

[0046] Example 1:

[0047] A method for denoising electrocardiogram (ECG) signals based on deep learning and ECG simulation technology, comprising the following steps:

[0048] Step 1: Set parameters to generate an electrocardiogram (ECG) signal;

[0049] Step 2: Randomly generate the conductivity σ at the field point in the range of [1.2-4.5] and the internal interface conductivity σ in the range of [1.3-3.5]. l - The interfacial conductivity, ranging from [2.3 to 5.5], is σ. l + Randomly generate n*n integers between [0,36] and assign them to ε(i,j,0), then generate different electrocardiogram signals according to step one;

[0050] Step 3: Using the simulated electrocardiogram signal as input, and taking the conductivity σ at the field point and the conductivity σ at the internal interface as input... l - The external interface conductivity is σ l + ε(i,j,0) is used as the output to train a deep learning network;

[0051] Step four: Use a deep learning network to denoise the real-collected electrocardiogram signals.

[0052] In step one, the specific steps for generating the electrocardiogram signal are as follows:

[0053] a. Determine the n*n dimensional transfer coefficient matrix M according to the following equation. t , where n is the number of nodes in the epicardial surface model;

[0054] in,

[0055]

[0056] b. Randomly generate n*n integers between [0, 36] and assign them to ε(i, j, 0). Calculate M0 according to the formula provided in step a.

[0057] c. Calculate Φ according to the following equation. 0 ;

[0058]

[0059] d. Determine the epicardial potential;

[0060]

[0061] In step three, the deep learning network is constructed using a 7-layer network plus a classifier, namely layer 1-layer 7;

[0062] Layers 1-6 each consist of a convolutional layer and a pooling layer, while layer 7 is a fully connected layer. Specifically:

[0063] The convolutional layer in layer 1 contains 5 kernels, each with a kernel size of 224. The stride and kernel size in the pooling layer in layer 1 are both 2.

[0064] The convolutional layer 2 contains 5 kernels, each with a kernel size of 112. The stride and kernel size of the pooling layer in layer 2 are both 2.

[0065] Layer 3 convolutional layer contains 10 kernels, all with a kernel size of 100. The stride and kernel size of the pooling layer in layer 3 are both 2.

[0066] Layer 4 convolutional layer contains 10 kernels, all with a kernel size of 50. The stride and kernel size of the pooling layers in layer 4 are both 2.

[0067] Layer 5 convolutional layer contains 10 kernels, all with a kernel size of 48. The stride and kernel size of the pooling layer in layer 5 are both 2.

[0068] Layer 6 convolutional layer contains 10 kernels, all with a kernel size of 24. The stride and kernel size of the pooling layer in layer 4 are both 2.

[0069] The number of input neurons in layer 7 is the same as the number of output features in layer 6. The number of output neurons is 100. The classifier has 100 input features and 39 output neurons.

[0070] In this embodiment, different ECG simulation signals are first obtained by inputting different parameters. The simulation signals are used as inputs and the parameters are used as outputs to train a deep learning network. Finally, the deep learning network is used to denoise the ECG signals obtained from real acquisition, thereby effectively removing noise interference in the dynamic ECG and improving the performance.

[0071] Example 2:

[0072] A denoising method for electrocardiogram signals based on deep learning and electrocardiogram simulation technology, operated according to the following steps:

[0073] Step 1: Set parameters to generate an electrocardiogram signal, and the specific steps are as follows:

[0074] a. Determine the n*n-dimensional transfer coefficient matrix M according to the following equation t , where n is the number of nodes of the epicardial surface model;

[0075]

[0076] where,

[0077] where, when 0<x<36, s(x) is respectively {-80,-80,-80,-80,55.9,30.3,8.0,-6.8,-14.2,-19.2,-23.3,-28.2,-32.4,-39.0,-44.7,-49.7,-54.6,-60.4,-64.5,-68.7,-72.8,-74.4,-76.1,-77.8,-78.6,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80}, and when x takes other values, s(x)=0.

[0078]

[0079] where, δ is the number of elements whose surrounding (adjacent, diagonally adjacent) values are in the range of [1-35], δ t h=2.

[0080] b. Randomly generate n*n integers in the range of [0,36], assign them to ε(i,j,0), and calculate M0 according to the formula provided in step a;

[0081] c. Solve for Φ according to the following equation 0 ;

[0082]

[0083] n is the number of nodes of the epicardial surface model, D j represents the normalized vector from the center point of the heart to the position of the i-th node of the epicardial surface model, R represents the fixed field point vector, represents the partial derivative operation;

[0084] d. Solve for the epicardial potential:

[0085]

[0086] Where, r′ is the field point on the epicardial surface; σ is the conductivity at that field point; Ω h Represents the entire heart region; J s It is the equivalent dipole source for myocardium; S l (1=2,3,*m) represents a closed interface with uniform conductivity within the human torso, where the conductivity of the inner interface is σ. l - The conductivity of the outer interface is σ. l + S l Representative surface.

[0087] Step 2: Randomly generate the conductivity σ at the field point in the range of [1.2-4.5], and the conductivity σ at the inner interface in the range of [1.3-3.5]. l - The interfacial conductivity, ranging from [2.3 to 5.5], is σ. l + Randomly generate n*n integers between [0,36] and assign them to ε(i,j,0), then generate different electrocardiogram signals according to step one;

[0088] Step 3: Using the simulated electrocardiogram signal as input, and the conductivity σ at the field point and the conductivity σ at the inner interface are used as inputs. l - The external interface conductivity is σ l + ε(i,j,0) is used as the output to train a deep learning network.

[0089] It should be noted that deep learning networks are constructed using a 7-layer network plus a classifier. The 7-layer deep learning network is called layer1-layer7.

[0090] Each of layers 1-6 consists of a convolutional layer and a pooling layer; specifically,

[0091] Layer 1 contains 5 convolutional layers with a kernel size of 224. The stride and kernel size in the pooling layers of Layer 1 are both 2. Layer 2 contains 5 convolutional layers with a kernel size of 112. The stride and kernel size in the pooling layers of Layer 2 are both 2. Layer 3 contains 10 convolutional layers with a kernel size of 100. The stride and kernel size in the pooling layers of Layer 3 are both 2. Layer 4 contains 10 convolutional layers with a kernel size of 50. The stride and kernel size in the pooling layers of Layer 4 are both 2. The stride and kernel size in layer 4 are both 2; layer 5 convolutional layer contains 10 kernels, all with a kernel size of 48, and the stride and kernel size in the pooling layers of layer 5 are both 2; layer 6 convolutional layer contains 10 kernels, all with a kernel size of 24, and the stride and kernel size in the pooling layers of layer 4 are both 2; the number of input neurons in the fully connected layer 7 is the same as the number of output features in layer 6, and the number of output neurons is 100. The classifier input features are 100, and the number of output neurons is 39.

[0092] Using the simulated electrocardiogram signal as input, and the conductivity σ at the field point and the conductivity σ at the inner interface are used as input. l - The external interface conductivity is σ l + ε(i,j,0) is used as the output to train a deep learning network.

[0093] The training algorithm can be any one of the following: Stochastic Gradient Descent (SGD), Adam, RMSProp, Adagrad, Adadelta, and Adamamax. SGD is one of the most basic and commonly used optimization algorithms. Because it uses only the gradient of a single sample for each update, it is computationally fast and efficient. SGD also has high memory utilization and low resource consumption, requiring only the storage of the gradient of a single sample. Adam is an adaptive learning rate optimization algorithm that adjusts the learning rate adaptively based on the gradient of each parameter. Specifically, it calculates a variable learning rate for each parameter, allowing for more precise control over the learning speed of each parameter during training. It combines the advantages of momentum and adaptive learning rates, enabling rapid convergence in the early stages of training and better adaptability to different parameters. RMSProp is another adaptive learning rate optimization algorithm that adjusts the learning rate by dividing by the moving average of the squared gradient. RMSProp can automatically adjust the learning rate, ensuring that the learning rate of different parameters changes during training. The learning speed can vary; RMSProp uses a moving average of the squared gradient to normalize the gradient magnitude, which helps alleviate the vanishing or exploding gradient problem; the Adagrad algorithm is an adaptive learning rate optimization algorithm that automatically adjusts the learning rate based on the historical gradient accumulation information of the parameters. This allows for different learning speeds for different parameters during training and provides some adaptability to sparse features, better handling those features that appear less frequently; the Adadelta algorithm is an improvement on the Adagrad algorithm, mainly addressing the problem of the monotonically decreasing learning rate in Adagrad. It offers more flexible learning rate adjustment compared to Adagrad's fixed learning rate and reduces dependence on hyperparameters such as the learning rate and initialization parameters; the Adam algorithm is a variant of the Adam algorithm, correcting the problem of excessively small learning rates. It can calculate an adaptive learning rate based on the gradient, improving training performance and exhibiting better robustness to large gradients compared to the Adam algorithm. The training algorithm used in this application can employ any one of the above-mentioned algorithms.

[0094] Step 4: Use a deep learning network to denoise the real-collected ECG signals; the specific steps are as follows:

[0095] ① Input the electrocardiogram signal into the neural network trained in step three, and calculate the conductivity σ at the field point and the conductivity σ at the inner interface. l - The external interface conductivity is σ l + , ε(i,j,0);

[0096] ② Calculate the 39 values ​​obtained in ① according to step one to obtain a new ECG signal, which is the denoised ECG signal.

[0097] In this embodiment, a method for denoising ECG signals based on a combination of deep learning algorithms and ECG simulation technology is proposed. First, different ECG simulation signals are obtained by inputting different parameters, specifically referring to the above equation operation. Then, the conductivity at the field point, the inner interface conductivity, and the outer interface conductivity within the range are randomly generated. n*n integers between [0, 36] are randomly generated and assigned to ε(i, j, 0). Then, different ECG signals are generated, using the simulated ECG signals as input, and the conductivity at the field point σ and the inner interface conductivity σ... l - σ of the external interface l + Using ε(i,j,0) as the output, a deep learning network is trained. This network is then used to denoise the real-world ECG signals. Specifically, the ECG signals are input into the trained neural network, and the conductivity σ at the field point and the conductivity σ at the inner interface are calculated. l - and the external interface conductivity σ l + ε(i,j,0), and then obtain the new ECG signal, which is the denoised ECG signal. This application can avoid the problem of large noise interference amplitude caused by movement or unstable conductivity of device leads, effectively remove the noise in this part of the dynamic ECG, make it easier to use, and improve the accuracy of use.

[0098] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for denoising electrocardiogram (ECG) signals based on deep learning and ECG simulation technology, characterized in that, Follow these steps: Step 1: Set parameters and generate an electrocardiogram (ECG) signal using an ECG physical model. The parameters involved in the ECG physical model include: conductivity σ at the field point and Ω in the heart region. h The equivalent dipole source Js of the myocardium, and the closed interface that serves as a uniformly distributed conductivity segment within the human torso. The conductivity of its inner interface is σl-, and the conductivity of its outer interface is σl+; and the specific steps for generating an electrocardiogram signal include: determining the n*n dimensional transmission coefficient matrix M according to the following equation. t , where n is the number of nodes in the epicardial surface model; ,in, ; ； It is the number of surrounding elements with values ​​between [1-35]. ; Step 2: Randomly generate conductivity values ​​at field points within the range of [1.2-4.5]. The interfacial conductivity is in the range of [1.3-3.5]. The interfacial conductivity, ranging from [2.3 to 5.5], is... Generate n*n random integers between [0, 36] and assign them to the epicardial activation sequence. Then, different electrocardiogram signals are generated according to step one, wherein the evolution of the epicardial activation sequence is based on adjacent and diagonally adjacent elements; Step 3: Using the simulated electrocardiogram signal as input, and the conductivity at the randomly generated field point in Step 2... Internal interface conductivity The external interface conductivity is Epicardial activation sequence As output, a deep learning network is trained; Step four involves using a deep learning network to denoise the real-world ECG signals, specifically including: ①. Input the actual acquired electrocardiogram (ECG) signal into the deep learning network trained in step three, and calculate a set of physical parameters, including the conductivity σ at the field point, the conductivity of the inner interface, the conductivity of the outer interface, and the epicardial activation sequence. ; ②. Calculate the set of physical parameters obtained in step ① in the same way as the ECG signal generated in step 1 to obtain a new ECG signal as the denoised ECG signal; Step one, the specific steps for generating an electrocardiogram signal, also include: b. Calculate according to the formula provided in step a. ; c. Solve according to the following equation ; ； n is the number of nodes in the epicardial surface model. R represents the normalized vector from the center point of the heart to the node position of the i-th epicardial surface model, and R represents the fixed field point vector. This represents partial derivative operations; d. Determine the epicardial potential; These are field points on the surface of the epicardium; It is the electrical conductivity at that point. Represents the entire heart region; It serves as an equivalent dipole source for the myocardium; A closed interface representing a uniformly distributed electrical conductivity within the human torso, whose internal interface conductivity is... The conductivity of the outer interface is M represents the surface area of ​​the body; In the formula middle, , and, when When x takes any value, s(x) is {-80,-80,-80,-80,55.9,30.3,8.0,-6.8,-14.2,-19.2,-23.3,-28.2,-32.4,-39.0,-44.7,-49.7,-54.6,-60.4,-64.5,-68.7,-72.8,-74.4,-76.1,-77.8,-78.6,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80,-80}. When x takes any other value, s(x) = 0. In step three, the deep learning network is constructed using a 7-layer network plus a classifier, i.e., layer 1-layer 7; the simulated electrocardiogram signal is used as input, and the conductivity at the field point is used as the input. Internal interface conductivity The external interface conductivity is , As output, the deep learning network is trained, and the training algorithm can be any one or a combination of the following: stochastic gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm, and Adamax algorithm. Layers 1-6 each consist of a convolutional layer and a pooling layer, while layer 7 is a fully connected layer. Specifically: The convolutional layer in layer 1 contains 5 kernels, each with a kernel size of 224. The stride and kernel size in the pooling layer in layer 1 are both 2. The convolutional layer 2 contains 5 kernels, each with a kernel size of 112. The stride and kernel size of the pooling layer in layer 2 are both 2. Layer 3 convolutional layer contains 10 kernels, all with a kernel size of 100. The stride and kernel size of the pooling layer in layer 3 are both 2. Layer 4 convolutional layer contains 10 kernels, all with a kernel size of 50. The stride and kernel size of the pooling layers in layer 4 are both 2. Layer 5 convolutional layer contains 10 kernels, all with a kernel size of 48. The stride and kernel size of the pooling layer in layer 5 are both 2. Layer 6 convolutional layer contains 10 kernels, all with a kernel size of 24. The stride and kernel size of the pooling layer in layer 4 are both 2. The number of input neurons in layer 7 is the same as the number of output features in layer 6. The number of output neurons is 100. The classifier has 100 input features and 39 output neurons.

Images