A noise removal method for electrocardiosignal

Abstract

The present application relates to a kind of noise removal methods of electrocardiosignal, specifically comprising the following steps: D1, feature extraction: model construction is carried out to each lead in 12 leads, relationship extraction is carried out in the process of simulating signal through other 11 leads;D2, network training: the first 10000 sampling points of 12 lead signals are input into network, the nonlinear association between leads is extracted and stored by implicit layer, and the current electrocardio information is retained in the network after training;D3, network test: x1~x11 lead as input enters network, 1~2 second noise is randomly added in any position of 11 lead signals to simulate real environment, at this time, network uses the relationship between leads extracted, deduces y single lead signal from 11 leads, and the remaining sampling points of the whole data are tested and verified;D4, according to additive noise model, average all output data, and average value is used as the final output data.

Description

[0001] Field of study

[0002] This invention relates to a method for processing electrocardiogram (ECG) signals, specifically a method for noise removal from ECG signals. Background Technology

[0003] Early attempts at noise removal algorithms for electrocardiogram (ECG) signals were based on classic linear filters, such as low-pass filters, adaptive filters, and filter banks. To address various types of noise, researchers began combining different methods to achieve better denoising results. EEMD decomposes the noisy ECG signal into a series of intrinsic mode functions (IMFs), then applies local mean filtering (LM) to the decomposed signal. Finally, the denoised ECG signal is recovered by reconstructing the filtered IMFs or detail and approximate components.

[0004] To avoid excessive influence of external noise on the original signal, some researchers decompose the signal into different frequencies for processing. The signal's inherent characteristics are more pronounced at different frequencies, making external noise easier to detect and remove. Combining wavelet decomposition and variational mode decomposition (VMD) allows for the differentiation of noise and valid signals using correlation coefficients, enabling targeted processing of signals at different frequencies. Alternatively, CEEMDAN can be used to decompose the signal first, followed by wavelet and mathematical morphology processing for high-frequency and low-frequency signals respectively. Mode decomposition breaks down a complete signal according to frequency, allowing for the selection of different methods for different frequency ranges to achieve better noise reduction.

[0005] To acquire and utilize more electrocardiogram (ECG) information, some have considered using historical ECG signal data to predict the signal at the next moment. The Kalman filter can use historical ECG signal data for prediction, dividing the ECG signal into different bands and processing the high-frequency and low-frequency signals separately using two Kalman filters. The Kalman filter itself aims to predict the signal to come based on the ECG activity that has already occurred.

[0006] To compensate for the inadequacy of single-lead information, researchers have attempted to utilize the correlation relationships among 12 leads for denoising, employing a sparse autoencoder (SAE) to fit the model. This model requires fitting to each lead other than the noisy lead separately, providing a unique fitting model for each patient. Each lead forms its own independent channel, the sparse autoencoder mines the signal features of a single lead, and finally, correlation coefficients are used to weight the other leads except the target lead.

[0007] Current technology, when dealing with sudden illnesses, sometimes mixes pathological features with noise signals and removes them because the models do not analyze individual patients enough. Summary of the Invention

[0008] The purpose of this invention is to provide a method for noise removal from electrocardiogram (ECG) signals, in order to solve the problem that pathological features in ECG signals are removed because they cannot be distinguished from noise signals.

[0009] The objective of this invention is achieved as follows:

[0010] A method for noise removal from electrocardiogram (ECG) signals, comprising the following steps:

[0011] S1. Feature extraction: A model is constructed for each of the 12 leads, and the correlation is extracted as the simulated signal passes through the other 11 leads besides the target denoising lead.

[0012] S2. Network Training: The first 10,000 sampling points of the 12-lead signal are input into the GRNN neural network. The hidden layer in the GRNN neural network extracts and stores the nonlinear correlation between the 12 leads using the historical data obtained from these 10,000 sampling points. After training, the hidden layer in the GRNN neural network retains the information features of the patient's previous historical data, including the specificity of each lead and the correlation between the 12 leads.

[0013] S3. Network Testing: The other 11 leads besides the target denoising lead are used as inputs into the GRNN neural network to obtain a clean ECG signal from the multi-lead network. Noise for 1 to 2 seconds is randomly added at any position of the 11-lead signal to simulate real-world noise. The GRNN neural network uses the extracted correlation between leads to deduce the single-lead signal from the 11 leads. The remaining sampling points of the entire data are then tested and verified.

[0014] The S4 and 11 leads are fitted using a GRNN neural network to obtain the guidance signal. This guidance signal is then used to filter the noisy ECG signal. The clean ECG signal q is obtained based on the additive noise model. i :q i =p i -N i ; where p i For the input signal, N i It is noise.

[0015] Average all the output data and use the average value as the final output data:

[0016]

[0017] Among them, a k and b k All are constant coefficients, ω k It's a window, I i It is a guiding signal;

[0018] In this way, the noisy ECG signal can be filtered to obtain the final result.

[0019] Furthermore, step D1 of the present invention includes the following sub-steps:

[0020] The GRNN neural network described in S1-1 consists of four basic layers: an input layer, a pattern layer, a summation layer, and an output layer.

[0021] p(h|d)=p(d|h)p(h) / p(d)

[0022] Where h represents a model with a specific network architecture, and d represents the target dataset.

[0023] The S1-2 input layer is used to receive input information and has the same number of units as the number of subordinate (input) parameters.

[0024] After passing through the input layer of the GRNN neural network, S1-3 outputs:

[0025] I(k) = x(k), k = 1, 2, ..., n

[0026] Where n is the number of samples, and the feature vector of each sample is x(i).

[0027] The S1-4 pattern layer is used to perform statistical analysis on the sample feature vectors, and its output is the mean and standard deviation of each feature. For the j-th feature, the output T(j) of the pattern layer is:

[0028] T(j)=[mean(x(1,j)),std(x(1,j)),mean(x(2,j)),std(x(2,j)),...,mean(x(n,j)),std(x(n,j))].

[0029] S1-5 The layer that calculates the radial basis function based on the output of the template layer and the input of the input layer is the radial base layer, and the output of the radial base layer is the value of the radial basis function. Let the radial basis function be:

[0030] G(x)=exp(–||x-t||^2 / 2σ^2)

[0031] Where x is the input vector, t is the template vector, and σ is the bandwidth parameter of the radial basis function, the output of the radial basis function is:

[0032] R(j)=G(||I-T(j)||),j=1,2,...,m

[0033] Where m is the dimension of the eigenvector, and ||·|| represents the Euclidean norm of the vector.

[0034] The output O of the S1-6 GRNN neural network is:

[0035]

[0036] Here, w(j) is the weight of the output layer, which can be learned using the least squares method or regularization method.

[0037] S1-7 explores the interrelationships among the 11 leads and stores their characteristics in the constructed model.

[0038] This invention utilizes the multi-lead correlation of electrocardiogram (ECG) signals for denoising. It employs a generalized regression neural network and guided filtering to denoise ECG signals, focusing on experimental analysis of three types of noise in ECG signals. Comparison experiments and analyses were conducted with the PTB database and a 12-lead arrhythmia database, respectively, against NLMT and SAE_GF. The algorithm performs well in most cases, effectively removing noise while preserving signal features. The algorithm's model retains the nonlinear features between 11 leads through multiple hidden layer extractions, exhibiting a fast response speed and rapid denoising of the signal. In subsequent guided filtering, edge-preserving characteristics retain the derived guided signal features, as reflected in the mean squared error. Attached Figure Description

[0039] Figure 1 This is the overall flowchart of the present invention.

[0040] Figure 2 This is a diagram of the GRNN neural network structure for extracting the correlation between leads. Detailed Implementation

[0041] The invention will now be described in further detail with reference to the accompanying drawings.

[0042] like Figure 1 As shown, the noise removal method for electrocardiogram signals of the present invention includes the following steps:

[0043] I. Feature Extraction: A model was constructed for each of the 12 leads. As the simulated signal passed through the other 11 leads (excluding the target denoising lead), correlations were extracted to effectively preserve the characteristics and trajectory of the ECG signal. The posterior probabilities of multiple models were compared.

[0044] Step one includes the following sub-steps:

[0045] 1-1. A GRNN neural network consists of four basic layers: an input layer, a pattern layer, a summation layer, and an output layer.

[0046] p(h|d)=p(d|h)p(h) / p(d)

[0047] Where h represents a model with a specific network architecture, and d represents the target dataset.

[0048] 1-2. The input layer is used to receive input information and has the same number of units as the number of subordinate (input) parameters.

[0049] 1-3. After passing through the input layer of the GRNN neural network, the output is:

[0050] I(k) = x(k), k = 1, 2, ..., n

[0051] Where n is the number of samples, and the feature vector of each sample is x(i).

[0052] 1-4. The pattern layer is used to perform statistical analysis on the sample feature vectors, and its output is the mean and standard deviation of each feature. For the j-th feature, the output T(j) of the pattern layer is:

[0053] T(j)=[mean(x(1,j)),std(x(1,j)),mean(x(2,j)),std(x(2,j)),...,mean(x(n,j)),std(x(n,j))].

[0054] 1-5. The radial base layer is a layer that calculates the radial basis function based on the output of the template layer and the input of the input layer. Its output is the value of the radial basis function. Assume the radial basis function is:

[0055] G(x)=exp(–||x-t||^2 / 2σ^2)

[0056] Where x is the input vector, t is the template vector, and σ is the bandwidth parameter of the radial basis function, then the output of the radial basis function is:

[0057] R(j)=G(||I-T(j)||),j=1,2,...,m

[0058] Where m is the dimension of the eigenvector, and ||·|| represents the Euclidean norm of the vector.

[0059] 1-6. The output O of the GRNN neural network is:

[0060]

[0061] Here, w(j) is the weight of the output layer, which can be learned by least squares or regularization.

[0062] 1-7. Discover the correlations within the 11 leads and store their characteristics in the constructed model.

[0063] II. Network Training: The first 10,000 sampling points of the 12-lead signal are input into the GRNN neural network. The hidden layer in the GRNN neural network extracts and stores the nonlinear correlation between the 12 leads using the historical data obtained from these 10,000 sampling points. After training, the hidden layer in the GRNN neural network retains the information features of the patient's previous historical data, including the specificity of each lead and the correlation between the 12 leads.

[0064] III. Network Testing: To obtain a clean ECG signal from the multi-lead dataset, leads x1-x11 were input into the GRNN neural network. Noise signals of 1-2 seconds were randomly added at any point in the 11-lead signal to simulate real-world noise. The GRNN then used the extracted inter-lead relationships to derive the single-lead signal y from the 11 leads. In the experimental verification section, the remaining sampling points of the entire dataset were tested and verified.

[0065] IV. Baseline drift and electrode interference in ECG signals are additive noise. Additive noise generally refers to thermal noise, shot noise, etc. Their relationship with the ECG signal is additive; noise exists regardless of whether an ECG signal is present. After fitting the 11-lead signal using a GRNN neural network to obtain the guiding signal, a clean ECG signal q is obtained according to the additive noise model. i :q i =p i -N i ; where p i For the input signal, N i It is noise.

[0066] If each local window does not overlap, then the output signal will definitely exhibit blocking artifacts at the boundaries of the local windows. Averaging all the output data and using the average as the final output data yields:

[0067]

[0068] In the local window wk, a k and b k All are constant coefficients, ω k It's a window, I i It is a guiding signal.

[0069] In this way, the noisy ECG signal is filtered to obtain the final result.

[0070] This invention primarily focuses on the experimental analysis of three types of noise in electrocardiogram (ECG) signals. Comparison experiments and analyses were conducted with the NLMT and SAE_GF databases using the PTB database and a 12-lead arrhythmia database, respectively. The algorithm performed well in most cases, effectively removing noise while preserving signal characteristics. The algorithm's model retains the nonlinear characteristics between 11 leads through multiple hidden layer mining operations, and the model has a fast response speed, enabling rapid denoising of the signals to be processed. In subsequent guided filtering, edge-preserving characteristics retain the obtained guided signal features, which can be reflected in the mean squared error.

Claims

1. A method of noise removal of an electrocardiosignal, characterized by, The method comprises the following steps: S1, feature extraction: model construction is performed on each of the 12 leads, and correlation extraction is performed on the signals in the process of passing through the other 11 leads except the target de-noising lead; S2, network training: the first 10000 sampling points of the 12-lead signals are input into the GRNN neural network, the history data obtained by the hidden layer of the GRNN neural network through the 10000 sampling points are used to extract and store the nonlinear correlation between the 12 leads, and after training, the hidden layer of the GRNN neural network retains the information features of the patient's past history data, including the specificity of each lead and the correlation between the 12 leads; S3, network testing: the other 11 leads except the target de-noising lead are input into the GRNN neural network as input to obtain a clean electrocardio signal in the multi-lead; noise with a time length of 1-2 seconds is randomly added at any position of the 11-lead signals to simulate the noise in the real environment, and the GRNN neural network uses the extracted correlation between the leads to deduce the y single-lead signal from the 11 leads, and test and verify the remaining sampling points of the whole data; S4, 11 lead signal is fitted through GRNN neural network to obtain a guide signal, and the ECG signal containing noise is guided to filter; and a clean ECG signal q is obtained according to an additive noise model i : q i = p i -N i ; where p i is the input signal, N i is the noise; The average of all output data is taken as the final output data: wherein a k and b k are constant coefficients, ω k is the window, I i is the pilot signal; The noisy electrocardio signal is filtered through the guide filter to obtain the final result.

2. The noise removing method of electrocardiosignal according to claim 1, characterized in that, Step D1 comprises the following sub-steps: S1-1 the GRNN neural network comprises four basic layers: an input layer, a pattern layer, a sum layer and an output layer: p(h|d) = p(d|h) p(h) / p(d) Wherein, h represents a model with a specific network architecture, and d represents a target data set; S1-2 the input layer is used for receiving input information and has the same number of units as the number of dependent (input) parameters; D1-3 after passing through the input layer of the GRNN neural network, the output is: I(k) = x(k), k = 1, 2,..., n Wherein, n is the number of samples, and the feature vector of each sample is x(i); S1-4 the pattern layer is used for statistical analysis of the sample feature vector, and the output of the pattern layer is the mean and standard deviation of each feature; for the jth feature, the output T(j) of the pattern layer is: T(j) = [mean(x(1,j)), std(x(1,j)), mean(x(2,j)), std(x(2,j)),..., mean(x(n,j)), std(x(n,j))]; S1-5 the layer for calculating the radial basis function according to the output of the template layer and the input of the input layer is the radial basis layer, and the output of the radial basis layer is the value of the radial basis function; let the radial basis function be: G(x) = exp(-||x-t||^2 / 2σ^2) Wherein, x is the input vector, t is the template vector, and σ is the bandwidth parameter of the radial basis function, and the output of the radial basis layer is: R(j) = G(||I-T(j)||), j = 1, 2,..., m Wherein, m is the dimension of the feature vector, and ||·|| represents the Euclidean norm of the vector; S1-6 the output O of the GRNN neural network is: Wherein, w(j) is the weight of the output layer, which is learned by least square method or regularization method; S1-7 excavate the correlation between 11 leads, and the characteristics are stored in the constructed model.

Images