Abdominal wall electric signal denoising method based on single-channel singular value decomposition and correlation analysis of partial resampling

Abstract

The invention provides an abdominal wall electric signal denoising method based on partial resampling single-channel singular value decomposition and correlation analysis. The method comprises the steps of abdominal wall electric signal preprocessing, Pan-Tompkin method maternal R wave recognition, maternal part resampling, single-channel singular value decomposition, correlation denoising and finally obtaining of denoised fetal electrocardiogram. The method has the remarkable advantages that the problem that SVD algorithm residual noise in abdominal wall electric signals with large heart rate variability is large is solved, meanwhile, residual maternal electrocardio is almost all filtered out through a related denoising analysis method, accuracy is high, and the method can be applied to the conditions that the signal-to-noise ratio of collecting equipment is low, and the signal-to-noise ratio of the collecting equipment is high. And the fetal electrocardiogram with the high signal-to-noise ratio can be obtained under the condition that the number of channels is limited.

Description

Technical Field

[0001] The present invention relates to an abdominal wall electrical signal denoising method and a fetal electrocardiogram detection method. Background Art

[0002] Fetal monitoring is of great significance for good prenatal and postnatal care, especially for elderly and high-risk pregnant women. Compared with traditional Doppler ultrasound monitoring, the fetal electrocardiogram detection method based on abdominal wall electrical signals has the advantages of being absolutely non-invasive, portable, and suitable for long-term home monitoring.

[0003] With the development of electronic technology, the fetal monitor CTG based on Doppler ultrasound and uterine contraction pressure has been popularized clinically. On the one hand, the CTG device emits 2 MHz ultrasonic signals to the fetal heart, calculates the fetal heart rate using the Doppler frequency shift of the echo, and on the other hand, measures the uterine contraction signal using a pressure sensor placed on the pregnant woman's abdomen, and combines the two to give a judgment on the fetal condition. As an ultrasonic device, CTG is not an absolutely non-invasive method. In particular, it has an angle dependence, requires the ultrasonic probe to always be aligned with the fetal heart valve, and is prone to misjudgment. Therefore, fetal monitoring means based on fetal electrocardiogram have always received extensive attention. The market needs a non-invasive fetal heart monitor that supports long-term monitoring and has a high monitoring accuracy. Currently, the methods for processing abdominal wall electrical signals are as follows:

[0004] Separation algorithm based on independent component analysis (ICA): The separation algorithm based on ICA cannot uniquely determine which of the separated signals is the maternal electrocardiogram signal and which is the fetal electrocardiogram signal, and the separated signals will show amplitude distortion. Generally, due to the slow convergence speed of ICA, it cannot meet the real-time requirement of fetal heart monitoring.

[0005] Separation algorithm based on principal component analysis (PCA): By calculating the covariance matrix between signals of different channels, the principal components are extracted to maximize the explained variance. PCA subtracts by projecting the signal into the principal component space, thereby removing the correlated noise in the multi-channel signals. However, in the case of complex noise components, this method has poor effects and needs to be combined with other denoising methods to improve performance.

[0006] Separation algorithm based on single-channel singular value decomposition (SSVD): SSVD is a method that takes the periodicity of electrocardiogram as a priori knowledge. Assume that the signal x(k) = {x(1), x(2),...} is a periodic signal with a period of n. Thus, m consecutive periods of length n of x(k) can be represented as a matrix A. Performing singular value decomposition on matrix A can obtain A = U∑V T, where ∑ is called the singular value matrix, U is the left singular value matrix, and V is the right singular value matrix. If the signal is strictly periodic, then matrix A is a rank-1 matrix. However, for quasi-periodic signals such as electrocardiograms, A may be full rank. After resampling, that is, resampling different periodic electrocardiogram signal segments to the same period, the eigenvalue σ1 in the singular value matrix should be much larger than other components. By retaining the first few singular values and the corresponding left and right singular value matrices and setting the rest to zero, denoising can be achieved. However, the problem is that the denoising ability for signals with large heart rate variability is not very good.

[0007] Decomposition algorithm based on comb filter (RR): The premise of comb filter is also to assume that the signal to be extracted is a periodic signal. From Fourier transform, it can be known that all periodic signals can be composed of the fundamental wave and its higher harmonics. And the comb filter is essentially n groups of band-pass / band-stop filters. By calibrating the position of the fundamental wave, the fundamental wave and its higher harmonics are filtered out to obtain a separation effect. However, for a good separation effect, the signal must be strictly periodic. The electrocardiogram signal has heart rate variability and cannot meet the periodic requirement. Therefore, the method based on comb filter often requires resampling, which is the same as SSVD. But neither comb filter nor singular value decomposition can ensure that the maternal component in the abdominal wall electrical signal and the estimated maternal electrocardiogram are exactly the same in the time domain. This leads to the residual maternal electrocardiogram signal affecting the judgment of the fetal heart position when the fetal heart amplitude is relatively small.

[0008] Correlation, as the most commonly used statistical measure, uses the correlation coefficient to describe the degree of association between two variables, and its value range is [-1, +1]. A negative value indicates that as the value of one variable increases, the other decreases; a positive value indicates that as the value of one variable increases, the other also increases; 0 indicates that the increase or decrease of one variable has no effect on the value of the other. For example, the Pearson correlation coefficient can well describe the linear correlation degree between two signals, so as to reduce the error caused by phase linear distortion. The present invention intends to introduce correlation analysis to overcome the deficiencies of existing methods. Summary of the Invention

[0009] The purpose of the present invention is to overcome the deficiencies existing in the above-mentioned prior art and propose a method for denoising abdominal wall electrical signals based on partial resampling, single-channel singular value decomposition, and correlation analysis (hereinafter referred to as CPSSVD for short). This method is applied to extract fetal heart signals from abdominal wall electrical signals. The abdominal wall electrical signals are preprocessed, maternal R wave is identified, partially resampled, singular value decomposed, and correlation analyzed for denoising, and finally the denoised fetal electrocardiogram is obtained. Compared with common methods for denoising abdominal wall electrical signals, it has high real-time performance and high accuracy, and can separate fetal heart signals when the fetal heart is relatively weak but still visible and the channel is limited. The technical solutions adopted to achieve the above-mentioned invention of the mother are specifically as follows:

[0010] The abdominal wall electrical signal denoising method based on partial resampling, single-channel singular value decomposition and correlation analysis proposed by the present invention is characterized by including abdominal wall electrical signal preprocessing, maternal R wave identification, maternal partial resampling, single-channel singular value decomposition, correlation analysis, and finally obtaining the denoised fetal electrocardiogram; for the abdominal wall electrical signal preprocessing, according to the characteristics of the abdominal wall electrical signal and the type of interference signal, a method of high-pass, low-pass, and notch preprocessing is adopted to obtain an abdominal wall electrical signal with a relatively stable baseline; for the maternal R wave identification, according to the Pan-Tompkin algorithm, the position of the maternal electrocardiogram R wave in the abdominal wall electrical signal is obtained; for the maternal partial resampling, based on the identified position of the maternal electrocardiogram R wave, the position of the T wave is detected backward, and the relative positions of the part Pb-R between the starting position Pb of the electrocardiogram P wave and the R wave and the part R-Te between the electrocardiogram R wave and the T wave are calibrated, and the part Te-Pb from the end position of the electrocardiogram T wave to the starting position of the P wave is resampled; for the single-channel singular value decomposition, the Pb-Te segment of each maternal electrocardiogram and the Te-Pb segment after resampling are spliced into a row vector, M is the length of each row vector, and the number of row vectors is determined by the number N of electrocardiogram cycles, and a matrix of N×M is combined. The matrix is subjected to singular value decomposition. If the first component of the eigenvalue matrix is greater than or equal to L times the second component, the matrix is restored through the first component to obtain the estimated maternal electrocardiogram; for the correlation analysis, a sliding window with a length of K is used to window the abdominal wall electrical signal and the estimated maternal electrocardiogram, and the correlation analysis is performed on the time-domain signal segments within the window. The threshold of the correlation coefficient is set, and the part higher than the threshold is the residual signal of the maternal electrocardiogram, which is directly set to zero. For the part less than the correlation coefficient, the preprocessed abdominal wall electrical signal and the estimated maternal electrocardiogram are subtracted to obtain the denoised fetal electrocardiogram.

[0011] Further, the abdominal wall electrical signal preprocessing method includes using a high-pass filter to remove baseline drift within 1 Hz, a notch filter to remove power frequency interference at 50 Hz, and a low-pass filter to filter out high-frequency noise above 100 Hz.

[0012] Further, the maternal R wave identification adopts the Pan-Tompkin maternal R wave identification algorithm, including 1) setting a first signal threshold, and setting half of the first signal threshold as the second signal threshold; 2) finding local peaks; 3) comparing the peak amplitude with the first signal threshold. If it is greater than the first signal threshold, it is identified as an R wave, otherwise it is regarded as interference. If no R wave appears within 2 s, the peak amplitude is compared with the second signal threshold. If it is greater than the second signal threshold and the peak appears 360 ms after the previous R wave, the peak is identified as an R wave.

[0013] Further, for the resampling of the abdominal wall electrical signals, the maternal part calibrates the positions of the P wave and T wave in the original abdominal wall electrical signals according to the electrocardiogram characteristics of the subject to be tested, determines the relative positions of Pb-R-Te, counts the lengths of the Te-Pb signal segments in each cycle within the statistical window and takes the maximum value as MaxNum, upsamples all the Te-Pb signal segments to MaxNum and splices them with the Te-R-Pb segment.

[0014] Further, for the singular value decomposition, the value range of M for single-channel singular value decomposition is 40 - 60, the value of N depends on the sampling rate Fs, and generally the range of N is 0.5 - 0.8 times Fs, and the value range of L is above 3.

[0015] Further, for the correlation analysis, assuming that the length of one electrocardiogram cycle is L, the sliding window intercepts the time-domain segment with a length of K = L / 5, and performs Pearson correlation analysis on the data within the window. The threshold of the correlation coefficient set by the correlation analysis denoising method is 0.95.

[0016] The present invention proposes a method for denoising abdominal wall electrical signals. This method estimates the fetal heart rate for the noisy signal based on partial resampling singular value decomposition and correlation analysis denoising. Its significant advantages are as follows:

[0017] The method of partial resampling avoids the problem that the Pb-R-Te segment is distorted during the resampling stage due to prominent heart rate variability, resulting in poor periodicity of the resampled signal and failure of restoration.

[0018] Using the method of correlation analysis denoising greatly suppresses the problem that the residual component of the maternal electrocardiogram is large due to the direct subtraction of the abdominal wall electrical signal and the estimated maternal signal caused by phase and amplitude distortion, and avoids the disadvantages of slow convergence speed and complex calculation of adaptive filtering methods such as LMS (Least Mean Square).

[0019] After being verified by the data set and through strict index comparison with methods such as wavelet denoising + Fast Independent Component Analysis (FastICA), single-channel singular value decomposition (SSVD), and comb-shaped filtering (RR). Description of the Drawings

[0020] Figure 1 The implementation flowchart of the method of the present invention.

[0021] Figure 2 Waveform diagrams of the abdominal wall electrical signals before and after preprocessing (a) Original abdominal wall electrical signal; (b) Abdominal wall electrical signal after preprocessing.

[0022] Figure 3 Recognition of the R wave of the maternal electrocardiogram.

[0023] Figure 4 Diagram of the electrocardiogram signal cycle division.

[0024] Figure 5 Comparison diagram of three methods for removing maternal electrocardiogram ability (a) Original signal; (b) Signal after comb filter; (c) Signal after single-channel singular value decomposition; (d) Signal processed by this method.

[0025] Figure 6 Effect diagram of the ability of abdominal wall electrical signal to remove the mother body (a) Original signal; (b) Estimated maternal electrocardiogram signal; (c) Fetal electrocardiogram signal.

[0026] Figure 7 Mean square error diagram of three methods under different heart rate variabilities. Specific implementation method

[0028] To further describe the effects and advantages of the present invention in detail, the present invention will be described in detail below in combination with the simulation results of a specific example. The following results are all obtained through simulation using MATLAB R2021b software.

[0029] In order to solve the problem that the fetal electrocardiogram morphology is damaged or submerged by noise, and at the same time to overcome the drawbacks of traditional mainstream methods in specific usage scenarios, the present invention proposes a method for denoising abdominal wall electrical signals based on partial resampling, single-channel singular value decomposition and correlation analysis. The algorithm framework is as shown in the appendix Figure 1 shown, including abdominal wall electrical signal preprocessing, maternal R wave identification, maternal partial resampling, single-channel singular value decomposition, correlation analysis, and finally obtaining the denoised fetal electrocardiogram. The specific process is as follows:

[0030] Abdominal wall electrical signal dataset: The dataset uses the public databases set-A and set-B used in the PhysioNet / Computing in Cardiology Challenge 2013 competition. Each data contains 4-channel abdominal wall electrical signals, the sampling rate is 1KHz, and the length of each data is 60s.

[0031] According to the characteristics of the abdominal wall electrical signal and the type of interference signal, a filtering preprocessing method is adopted to obtain a relatively stable baseline waveform of the abdominal wall electrical signal. The preprocessing method includes using a Butterworth high-pass filter to filter out the baseline drift within 1Hz, a notch filter to remove the 50Hz power frequency interference, and a low-pass filter to filter out the high-frequency noise above 100Hz. The appendix Figure 2 shows the waveform diagrams of the abdominal wall electrical signal before and after filtering.

[0032] For the identification of the maternal R wave, the Pan-Tompkin identification algorithm is used, including: 1) setting the first signal threshold and setting half of the first signal threshold as the second signal threshold; 2) finding the local peak; 3) comparing the peak amplitude with the first signal threshold. If it is greater than the first signal threshold, it is identified as the R wave, otherwise it is regarded as interference. If no R wave appears within 2 s, then compare the peak amplitude with the second signal threshold. If it is greater than the second signal threshold and the peak appears 360 ms after the previous R wave, then identify the peak as the R wave. In this embodiment, the first signal threshold is 23.307 and the second signal threshold is 11.6535. Attached Figure 3 shows the accuracy of the maternal electrocardiogram R wave marking.

[0033] For the partial resampling of the mother body, based on the identified position of the maternal electrocardiogram R wave, as shown in the attached Figure 4 figure, detect the position of the T wave backward, calibrate the relative positions of the part Pb-R between the starting position Pb of the electrocardiogram P wave and the R wave and the part R-Te between the R wave and the T wave of the electrocardiogram, resample the part Te-Pb from the end position of the electrocardiogram T wave to the starting position of the P wave, and count the length of each cycle of the Te-Pb signal segment within the window and take the maximum value as MaxNum. Upsample all Te-Pb signal segments to MaxNum and splice them with the Te-R-Pb segment. In this embodiment, MaxNum is 362.

[0034] For single-channel singular value decomposition, splice the Pb-Te of each maternal electrocardiogram and the resampled Te-Pb segment into a row vector. M is the length of each row vector. Determine the number of row vectors according to the number N of electrocardiogram cycles, and combine them into an N×M matrix. Perform singular value decomposition on the matrix. If the first component of the eigenvalue matrix is greater than or equal to L times the second component, restore the matrix through the first component to obtain the estimated maternal electrocardiogram. In this embodiment, M is 500, N is 14, the size of the composed matrix is 14×500, and L is 3.

[0035] Perform singular value decomposition on the matrix. If the first component of the eigenvalue matrix is greater than or equal to L times the second component, restore the matrix through the first component to obtain the estimated maternal electrocardiogram. For example, in this embodiment, the first component of the matrix is 4.66×10 3 and the second component of the matrix is 1.03×10 3 . The size of the first component exceeds three times that of the second component. Retain the first component and set the remaining components to zero to obtain the estimated maternal electrocardiogram, and upsample the upsampled part to the length of the original signal.

[0036] For correlation analysis, a sliding window with a length of K is used to window the abdominal wall electrical signal and the estimated maternal electrocardiogram (ECG), and correlation analysis is performed on the time-domain signal segments within the window. A threshold for the correlation coefficient is set. The part above the threshold is the residual maternal ECG signal, which is directly set to zero. For the part less than the correlation coefficient, the preprocessed abdominal wall electrical signal is subtracted from the estimated maternal ECG to obtain the denoised fetal ECG. In this embodiment, the length L of one ECG cycle is 500, and the sliding window intercepts time-domain segments with a length of K = L / 5, i.e., 100. Pearson correlation analysis is performed on the data within the window, and the threshold for the correlation coefficient set by the correlation analysis denoising method is 0.95.

[0037] Index verification: To verify the performance of the algorithm, an index verification part is set in the process of the present invention. The maternal ECG with a certain heart rate variability is generated according to the ECG model proposed by Patrick E. McSharry et al., and the heart rate variability can be changed by changing the Te-Pb segment. The larger the heart rate variability coefficient r, the more obvious the heart rate variability (HRV) of the electrocardiogram. Only RR, SSVD, and CPSSVD are used to estimate the ECG components respectively, and the generated estimation errors are as shown in the appendix Figure 5 shown. The error error = ECF - eECG, where eECG represents the estimated ECG component and ECG represents the original ECG component. For the real abdominal wall electrical signal as shown in the appendix Figure 6 shown, the fetal heart component can be more obvious after correlation denoising when the fetal heart component in the preprocessed abdominal wall electrical signal is small. At the same time, during the process of the heart rate variability coefficient r changing from 0 to 0.9 as shown in the appendix Figure 7 shown, the RMS value of the method of the present invention hardly changes as r increases. At the moment of r = 0.4, the root mean square (RMS) of the estimation errors obtained by the three methods are 0.0059, 0.0012, and 0.000775, indicating that the method of the present invention is much stronger than the first two methods in the ability to remove the maternal ECG.

[0038] The above embodiments show that a method for denoising abdominal wall electrical signals based on partial resampling, single-channel singular value decomposition, and correlation analysis proposed by the present invention is effective.

[0039] It should be noted that the above embodiments are not used to limit the protection scope of the present invention. Equivalent transformations, substitutions, and several improvements such as expansion based on the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A method for denoising abdominal wall electrical signals based on partial resampling single-channel singular value decomposition and correlation analysis, characterized in that: It includes abdominal wall electrical signal preprocessing, maternal R wave identification, maternal part resampling, single channel singular value decomposition, correlation analysis, and finally obtains the denoised fetal ECG; for abdominal wall electrical signal preprocessing, high-low pass and notch preprocessing methods are adopted to obtain abdominal wall electrical signals with relatively stable baselines: for maternal R wave identification, the maternal ECG R wave position in the abdominal wall electrical signal is obtained according to the Pan-Tompkin algorithm; for maternal part resampling, based on the identified maternal ECG R wave position, the T wave position is detected backward, and the relative positions of the ECG P wave starting position Pb to the R wave part Pb-R and the ECG R wave to T wave part R-Te are calibrated, and the ECG T wave end position to the P wave starting position part Te-Pb segment is resampled; for single channel singular value decomposition, The Pb-Te segment of each maternal ECG and the resampled Te-Pb segment are concatenated into a row vector, M is the length of each row vector, and the number of row vectors is determined by the number of ECG cycles N. The matrix is ​​combined into an N×M matrix, and the matrix is ​​subjected to singular value decomposition. If the first component of the eigenvalue matrix is ​​greater than or equal to L times the second component, the matrix is ​​restored by the first component to obtain the maternal estimated ECG; for correlation analysis, a sliding window of length K is used to open a window for the abdominal wall electrical signal and the maternal estimated ECG, and correlation analysis is performed on the time domain signal segments within the window, and a threshold of the correlation coefficient is set. The part above the threshold is the maternal ECG residual signal, which is directly zeroed. For the part less than the correlation coefficient, the preprocessed abdominal wall electrical signal and the estimated maternal ECG are subtracted to obtain the denoised fetal ECG.

2. According to claim 1, a method for denoising abdominal wall electrical signals based on partial resampling single channel singular value decomposition and correlation analysis, characterized in that Abdominal wall electrical signal preprocessing includes using a high-pass filter to remove baseline drift within 1 Hz, a notch filter to remove power frequency interference at 50 Hz, and a low-pass filter to remove high-frequency noise above 100 Hz.

3. The method for denoising abdominal wall electrical signals based on partial resampling single-channel singular value decomposition and correlation analysis according to claim 1, characterized in that Maternal R wave identification uses the Pan-Tompkin maternal R wave identification algorithm, which includes 1) setting a first signal threshold, and setting half of the first signal threshold as the second signal threshold; 2) finding the local peak; 3) comparing the peak amplitude with the first signal threshold. If it is greater than the first signal threshold, it is identified as an R wave, otherwise it is regarded as interference. If no R wave appears within 2 seconds, the peak amplitude is compared with the second signal threshold. If it is greater than the second signal threshold and the peak appears 360ms after the previous R wave, the peak is identified as an R wave.

4. The method for denoising abdominal wall electrical signals based on partial resampling single-channel singular value decomposition and correlation analysis according to claim 1, characterized in that: The maternal part calibrates the P wave and T wave positions in the original abdominal wall electrical signal according to the ECG characteristics of the subject, determines the relative position of Pb-R-Te, counts the length of the Te-Pb signal segment of each cycle in the window and takes the maximum value as MaxNum, upsamples all Te-Pb signal segments to MaxNum and splices them with the Te-R-Pb segment.

5. The method for denoising abdominal wall electrical signals based on partial resampling single-channel singular value decomposition and correlation analysis according to claim 1, characterized in that: The M value range of single-channel singular value decomposition is 40-60, the N value depends on the sampling rate Fs. Generally, the N range is 0.5-0.8 times Fs, and the L value range is above 3.

6. The method for denoising abdominal wall electrical signals based on partial resampling single-channel singular value decomposition and correlation analysis according to claim 1, characterized in that: In the correlation analysis, it is assumed that the length of an ECG cycle is L and the sliding window cuts the time domain segment with a length of K=L / 5. Pearson correlation analysis is performed on the data in the window, and the threshold of the correlation coefficient set by the correlation analysis denoising method is 0.95.

Images