A method for classifying uterine myoelectric signals based on multivariate empirical mode decomposition

Abstract

The application discloses a kind of uterine myoelectricity signal classification method based on multivariate empirical mode decomposition, comprising: 1, using multichannel surface electrode acquisition a period of uterine contraction within uterine myoelectricity signal and carries out denoising pretreatment;2, using multivariate empirical mode decomposition to extend the uterine myoelectricity signal after denoising pretreatment to time-frequency domain, obtain the intrinsic mode component containing different time scale of uterine myoelectricity signal, according to variance contribution and correlation coefficient select effective component therein;3, to the intrinsic mode component after screening, sliding window is carried out, and six characteristic values are calculated;4, using chi-square test pre-screening and sequence bidirectional search combination feature selection algorithm, select optimal feature subset;5, based on support vector machine model, classify normal and abnormal activity in uterine myoelectricity signal.The application can utilize the local information of uterine myoelectricity signal in time-frequency domain, filter out irrelevant noise and redundant features, to complete the abnormal activity identification in uterine myoelectricity signal.

Description

Technical Field

[0001] This invention belongs to the field of electromyographic signal processing and analysis, specifically a method for classifying uterine electromyographic signals based on multivariate empirical mode decomposition. Background Technology

[0002] Eelctrohysterogram (EHG) signals, as electrophysiological signals accompanying uterine contractions, are a comprehensive representation of the individual electrical activity of countless uterine smooth muscle cells. Near the onset of labor, the increase in gap junctions among uterine smooth muscle cells provides a rapid pathway for the propagation of action potentials, reflecting enhanced cellular excitability. When action potentials propagate intermittently along the myometrium, they lead to uterine contractions. Regular and synchronized contractions serve as important indicators for determining whether a pregnant woman is in labor. With advancements in EHG acquisition technology, high-density surface electrodes can record the electrical activity generated during uterine contractions non-invasively from the pregnant woman's abdomen, acquiring EHG signals rich in spatiotemporal information. This objectively describes the duration, frequency, and intensity of uterine contractions, thereby identifying abnormal activity in the uterine EHG signal during contractions and providing early warning of labor.

[0003] Because the generation and propagation of EHG signals are caused by the complex electrical activity of countless uterine smooth muscle cells, they exhibit strong non-stationarity and nonlinearity. Therefore, analysis of the entire frequency range of EHG signals may not adequately describe the evolution of pregnancy and makes subsequent classification difficult. In recent years, many different time-frequency processing techniques have been used to improve feature representation capabilities by analyzing signals at different scales. Among these, compared to the fixed time and frequency resolution of the Short-Time Fourier Transform (SFT) globally, and the need for manually selecting basis functions and decomposition levels for wavelet transform, Empirical Mode Decomposition (EMD), as a fully data-driven method, can adaptively decompose non-stationary time series into intrinsic mode components at different time scales. However, for multi-channel signals, EMD can only be performed one-by-one on a single channel, leading to inconsistencies in the number and frequency distribution of intrinsic mode components across different channels, as well as spectral aliasing. This poses a significant challenge to subsequent signal analysis and the identification of abnormal uterine electromyography (EMG) activity. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a classification method for uterine electromyography (EHG) signals based on multivariate empirical mode decomposition. This method aims to adaptively decompose multi-channel EHG signals into intrinsic mode components at different time scales, while ensuring consistency in the number and spectrum of different channels and filtering out irrelevant noise and redundant features. This solves the problems of mode mismatch and mode aliasing, enabling the identification of abnormal activities in uterine EHG signals.

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

[0006] The present invention provides a method for classifying uterine electromyographic signals based on multivariate empirical mode decomposition, characterized by the following steps:

[0007] Step 1, Signal Acquisition:

[0008] Step 1.1: Using a uterine electromyography (EMG) measuring device and a Q-channel high-density surface electrode, collect surface EMG signal data of a subject at time t during uterine contraction, denoted as: x(t) = [x1(t), x2(t), ..., x...]. q (t), ..., x Q (t)] T , 1≤q≤Q, 1≤t≤T1, where x q (t) represents the uterine electromyography signal data at time t on the q-th channel, where Q is the total number of channels of the high-density electrodes and T1 is the total recording duration;

[0009] Step 1.2, at the sampling frequency f s The discrete sample data of the uterine electromyography signal x(t) are obtained below, denoted as: x(n) = [x1(n), x2(n), ..., x q (n), ..., x Q (n)] T , 1≤n≤N, where x q (n) represents the uterine electromyography (EMG) signal data of the nth sample point on the qth channel, and N is the total number of recorded sample points; thus, after measuring for a duration of T1, a uterine EMG time series x = [x1, x2, ..., xn] with a sample length of N is obtained. q , ..., x Q ] T , where x q Let x represent the uterine electromyography time series of the q-th channel, and x q ={x q Let the category label of the uterine electromyography time series x be y∈{1,0}, where y=1 represents normal and y=0 represents abnormal;

[0010] Step 2, Signal Preprocessing:

[0011] The uterine electromyography (EMG) time series x was denoised using a bandpass filter to remove interfering physiological signals and irrelevant noise, resulting in the denoised Q-channel uterine EMG time series x. * =[x * 1,x * 2..., x * q , ..., x *Q ] T , where x * q This represents the electromyographic time sequence of the q-th channel after noise reduction.

[0012] Step 3, Signal Time-Frequency Spreading:

[0013] Step 3.1: Use the multivariate empirical mode decomposition method shown in equation (1) to decompose the uterine electromyography signal time series x * The joint decomposition on the Q channel is divided into I intrinsic mode components and residual terms:

[0014]

[0015] In equation (1), Let r represent the i-th eigenmode component of the q-th channel. q This represents the residual term of the q-th channel;

[0016] Step 3.2: Calculate according to equations (2) and (3) respectively. Varc, the variance contribution of the i-th intrinsic mode component i Correlation coefficient i :

[0017]

[0018]

[0019] In equations (2) and (3), and var(x) * q ) respectively represent and x * q variance express and x * q cross variance;

[0020] According to {Varc i |1≤i≤I} and {Cor i Let |1≤i≤I} be the number of intrinsic modal components that contribute to and are relevant to uterine contraction, denoted as . in, This represents the i0th intrinsic mode component of the q-th channel;

[0021] Step 4, Feature Extraction:

[0022] Set the window length to N. w The overlap with the window is N L For each N-th length After performing sliding window frame segmentation, a total of ((NN) were obtained. w ) / N L +1) sliding window time series, and denote one of the sliding window time series as and in, Indicates a length of N w The n0th sample point data in the time series;

[0023] Six eigenvalues ​​are calculated for each sliding window time series, including: root mean square amplitude and autocorrelation zero-crossing number in the time domain, peak frequency and median frequency in the frequency domain, and sample entropy and fuzzy entropy in the nonlinear domain. The six eigenvalues ​​obtained from all sliding window time series are then averaged to obtain the three-dimensional feature matrix f = [f(i0, j), f2(i0, j), ..., f...]. q (i0, j), ..., f Q (i0, j)] T , 1≤i0≤I0, 1≤j≤6, where, f q (i0, j) represents the j-th eigenvalue of the i0-th intrinsic mode component in the q-th channel;

[0024] Step 5, Feature Selection:

[0025] Step 5.1: Convert the three-dimensional feature matrix f into a one-dimensional feature set F = {F(k) | k = 1, 2, ..., Q × 10 × 6}, and calculate the chi-square score of the k-th feature F(k) using the chi-square test. After sorting all chi-square scores in descending order, select K1 features that are greater than the preset threshold th to obtain the initial feature set F. * ={F * (k′)|k″=1,2,…,K1};where, F * (k′) represents the k′th feature after the initial screening, and K1 represents the number of features in the initial screening feature set;

[0026] Step 5.2: Based on the bidirectional sequence search algorithm, process the initially screened feature set F * Feature optimization is performed to obtain the optimal feature subset F. ** ={F ** (k″)|k″=1,2,…,K2},where, F ** (k″) represents the k″-th feature selected by the bidirectional search algorithm, and K2 represents the number of features in the optimal feature subset;

[0027] Step 6: Following the process of steps 1-5, obtain the optimal feature subsets of M objects to be tested and form a dataset R = {F}. ** 1(k″), F** 2(k″), ..., F ** M (k″)|k″=1,2,…,K2},where, F ** M (k″) represents the optimal feature subset of the Mth object to be tested;

[0028] Divide the dataset R into M a Training data with known labels and their corresponding category labels M b Test data with unknown labels Use support vector machines on the training data Establish category labels The mapping relationship between them is used to find the hyperplane γ that can classify different categories, and thus the test data is analyzed based on the hyperplane γ of the support vector machine model. Perform classification and obtain the predicted classification results.

[0029] The uterine electromyography signal classification method based on multivariate empirical mode decomposition described in this invention is also characterized in that step 5.2 includes:

[0030] Step 5.2.1: Set the total number of algorithm executions to T, from the feature set F * A feature subset F is randomly generated from the data. * (k0);

[0031] Step 5.2.2: From the feature set F * Select one feature F in sequence * (k + Add feature subset F * In (k0), determine F * If the F1 score of (k0) is optimal, proceed to step 5.2.3; otherwise, continue to step 5.2.2.

[0032] Step 5.2.3: Sequentially start from feature subset F * Deleting a feature F(k) from (k0) - And determine the feature subset F after deletion. * Does the F1 score of (k0) decrease? If so, then in the feature subset F * The feature F(k) is retained in (k0). - Otherwise, continue with step 5.2.3; thus obtaining the optimal feature subset F. ** ={F ** (k″)|k″=1,2,…,K2},where, F **(k″) represents the k″-th feature selected by the bidirectional search algorithm, and K2 represents the number of features in the optimal feature subset.

[0033] The present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the uterine electromyography signal classification method, and the processor is configured to execute the program stored in the memory.

[0034] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the uterine electromyography signal classification method.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. In step three of this invention, the uterine electromyography signal is adaptively decomposed into time and frequency components to obtain intrinsic mode components distributed in different frequency ranges. In step four, the intensity, frequency and complexity features of the uterine electromyography signal are extracted to characterize the propagation trend of electrical activity during uterine contraction. In step five, the recognition performance of abnormal activities in the uterine electromyography signal is improved by filtering out feature components and redundant noise that are not related to uterine contraction.

[0037] 2. In step three of this invention, multivariate empirical mode decomposition improves the time-frequency resolution of uterine electromyography signals at different scales and ensures the consistency of the number and spectrum of intrinsic mode components obtained between different channels, thus solving the problems of mode mismatch and mode aliasing.

[0038] 3. This invention obtains time-domain, frequency-domain, and nonlinear-domain features of each component, and uses a feature selection algorithm combining chi-square test and bidirectional sequence search in step five to filter out irrelevant feature components, eliminate redundant information and irrelevant noise, and better locate information reflecting uterine electromyographic contraction activity. Attached Figure Description

[0039] Figure 1 This is a flowchart of a uterine electromyography signal classification method based on multivariate empirical mode decomposition according to the present invention.

[0040] Figure 2 The diagrams show the intrinsic mode component waveforms and corresponding power spectral density distributions of the multi-channel EHG signal after MEMD decomposition according to the present invention.

[0041] Figure 3 Box plots showing the variance contributions of each level of IMF to the original EHG signal in this invention.

[0042] Figure 4 Box plots showing the correlation coefficients of each level of IMF in this invention with the original EHG signal;

[0043] Figure 5 This is a trend diagram showing the influence of the preset threshold th and the number of iterations T on the F1 score in the feature selection algorithm of this invention;

[0044] Figure 6 This is a bar chart comparing the performance of different preterm birth prediction methods of the present invention. Detailed Implementation

[0045] In this embodiment, a method for classifying uterine electromyographic signals based on multivariate empirical mode decomposition is described, such as... Figure 1 As shown, it includes the following steps:

[0046] Step 1, Signal Acquisition:

[0047] Step 1.1: Using a uterine electromyography (EMG) measuring device and a Q-channel high-density surface electrode, collect surface EMG signal data of a subject at time t during uterine contraction, denoted as: x(t) = [x1(t), x2(t), ..., x...]. q (t), ..., x Q (t)] T , 1≤q≤Q, 1≤t≤T1, where x q (t) represents the uterine electromyography signal data at time t on the q-th channel, where Q is the total number of channels of the high-density electrodes and T1 is the total recording duration;

[0048] Step 1.2, at the sampling frequency f s The discrete sample data of the uterine electromyography signal x(t) are obtained below, denoted as: x(n) = [x1(n), x2(n), ..., x q (n), ..., x Q (n)] T , 1≤n≤N, where x q (n) represents the uterine electromyography (EMG) signal data of the nth sample point on the qth channel, and N is the total number of recorded sample points; thus, after measuring for a duration of T1, a uterine EMG time series x = [x1, x2, ..., xn] with a sample length of N is obtained. q , ..., x Q ] T , where x q Let x represent the uterine electromyography time series of the q-th channel, and x q ={x q Let the category label of the uterine electromyography time series x be y∈{1,0}, where y=1 represents normal and y=0 represents abnormal;

[0049] The specific implementation includes:

[0050] (1) Recruit subjects with an M-position and a gestational age of 30 weeks or more. Instruct each subject to lie flat in a comfortable position. Four electrodes are symmetrically attached to the abdominal skin surface below the subject's navel. The reference electrode and ground electrode are placed on either side of the groin. The device operates at a sampling frequency f. s Electromyographic signals of the uterine smooth muscle surface were collected. To reduce noise interference from external equipment and body movement, four electrodes were paired differentially to form Q bipolar channels. The center-to-center distance between adjacent electrodes in the electrode array was d. For example, M = 300, f... s =20Hz, Q=6, d=7cm;

[0051] (2) Collect continuous time-series signals during uterine contractions of each subject, manually remove obvious noise interference and abnormal segments to obtain M segments of uterine electromyography (EMG) signals with a duration of T1. After sampling discretization, obtain discrete sample point data corresponding to the M groups of uterine EMG signals, and add a category label y∈{1,0} to the uterine EMG data of each subject, thereby storing it in the uterine EMG dataset. For example, T1 can be set as: T1 = 30min, category label y = 1 represents normal, y = 0 represents abnormal;

[0052] Step 2, Signal Preprocessing:

[0053] The uterine electromyography (EMG) time series x was denoised using a bandpass filter to remove interfering physiological signals and irrelevant noise, resulting in the denoised Q-channel uterine EMG time series x. * =[x * 1,x * 2..., x * q , ..., x * Q ] T , where x * q This represents the electromyographic time sequence of the q-th channel after noise reduction.

[0054] In this embodiment, a four-stage Butterworth causal IIR filter with a bandpass cutoff frequency of 0.3–3 Hz is used to filter out physiological signals often mixed in with the EHG signal, such as maternal ECG, fetal ECG, and maternal respiration, as well as noise such as 50 Hz power frequency interference and baseline drift that are easily introduced during the acquisition process. Furthermore, to avoid transient effects caused by filtering, the T0-minute segments before and after the entire T1 signal are removed, and the uterine electromyography signal with a duration of T1–2T0 is used for subsequent analysis. The total number of sample points for the corresponding discrete uterine electromyography data is N. For example, T0 = 5 min, N = 24000 can be set.

[0055] Step 3: Time-frequency expansion is performed on the denoised uterine signal to decompose it into individual components in different frequency ranges for analysis, which can better locate useful information. Multivariate empirical mode decomposition (EMD), as a fully data-driven time-frequency decomposition method, can adaptively decompose complex non-stationary signals into intrinsic mode components containing different time scales, while ensuring the consistency of energy distribution across channels. This invention obtains I intrinsic mode components after decomposition by performing steps 3.1 to 3.2, and selects I0 of them that are related to the uterine contraction frequency for subsequent analysis, eliminating interference from irrelevant frequency components.

[0056] Step 3.1: Use the multivariate empirical mode decomposition method shown in equation (1) to decompose the uterine electromyography signal time series x * The joint decomposition on the Q channel is divided into I intrinsic mode components and residual terms:

[0057]

[0058] In equation (1), Let r represent the i-th eigenmode component of the q-th channel. q This represents the residual term of the q-th channel;

[0059] The specific implementation process of multivariate empirical mode decomposition is as follows: For the discrete input uterine electromyography signal x of the Q channel * =[x * 1,x * 2..., x * q , ..., x * Q ] T Along the angle on the Q-1 dimensional hypersphere Projecting onto the given direction yields a set of k-dimensional projection vectors, defined as follows: Next, cubic spline interpolation is used to connect the extreme points of the projection vector in each direction, forming k envelopes. The average of these envelopes is then used to obtain the local mean of the multivariate signal, from the multivariate input signal x. * The process involves iteratively subtracting local means until a region no longer contains a sufficient number of extrema, and if both conditions for an IMF (Internal Maximum Factor) are met, it can be defined as an IMF. This process is repeated until the residual is less than a preset threshold, at which point the IMF selection process ends. It should be noted that the two conditions for defining an IMF are: the number of intersections between local extrema and zeros must be equal or differ by at most one; and the average of the local maximum (upper envelope) and local minimum (lower envelope) must be zero.

[0060] In this embodiment, as Figure 2As shown, after MEMD decomposition, a segment of uterine electromyography (EMG) signal yielded the same number of I intrinsic mode components (IMFs) across six channels, which were then aligned one by one on the frequency scale. The fluctuation trends of each IMF component decreased from dramatic to gradual according to the decomposition order, corresponding to a change from high frequency to low frequency. To observe the local details of the intrinsic mode components, a five-minute segment of the entire signal was extracted for demonstration. Typically, the value of I ranges from 10 to 15; for example, I = 12 can be set.

[0061] Step 3.2: Calculate according to equations (2) and (3) respectively. Varc, the variance contribution of the i-th intrinsic mode component i Correlation coefficient i :

[0062]

[0063]

[0064] In equations (2) and (3), and var(x) * q ) respectively represent and x * q variance express and x * q cross variance;

[0065] According to {Varc i |1≤i≤I} and {Cor i Let |1≤i≤I} be the number of intrinsic modal components that contribute to and are relevant to uterine contraction, denoted as . in, This represents the i0th intrinsic mode component of the q-th channel;

[0066] In this embodiment, the variance contribution and correlation coefficient of the 12 IMFs obtained in step 3.1 are calculated at each decomposition level, and box plots are drawn as follows. Figure 3 and Figure 4 As shown, the results indicate that the top 5 levels of all IMFs reflect high Varc and Cor values, suggesting a significant contribution and strong correlation to the EHG signal. We hypothesize that the lower values ​​of these two indices in the remaining IMFs (levels 6–12) indicate the inclusion of frequency components unrelated to preterm birth classification. Therefore, the top five IMFs are selected for subsequent feature extraction.

[0067] Step 4, Feature Extraction:

[0068] Set the window length to Nw and the window overlap to N. L For each N-th length After performing sliding window frame segmentation, a total of ((NN) were obtained. w ) / N L +1) sliding window time series, and denote one of the sliding window time series as and in, Indicates a length of N w The n0th sample point data in the time series;

[0069] Six eigenvalues ​​are calculated for each sliding window time series, including: root mean square amplitude and autocorrelation zero-crossing number in the time domain, peak frequency and median frequency in the frequency domain, and sample entropy and fuzzy entropy in the nonlinear domain. The six eigenvalues ​​obtained from all sliding window time series are then averaged to obtain the three-dimensional feature matrix f = [f(i0, j), f2(i0, j), ..., f...]. q (i0, j), ..., f Q (i0, j)] T , 1≤i0≤I0, 1≤j≤6, where, f q (i0, j) represents the j-th eigenvalue of the i0-th intrinsic mode component in the q-th channel;

[0070] In this embodiment, a window length of N is used. w The overlap with the window is N L Sliding window analysis is used to extract feature information from the detailed parts of the intrinsic mode components. The choice of window length and overlap size is a comprehensive consideration between computational complexity and feature representation continuity. For example, N can be set as follows: w =1200, N L =300. For each channel's specific intrinsic mode component, six features are extracted from each window, and the average of all windows is used as the final feature result. Therefore, for each subject, a matrix f is obtained covering three dimensions: channel, intrinsic mode component, and features. Wherein, the given input signal is... The specific calculation process for the six features is as follows:

[0071] 1) Root Mean Square Amplitude (RMS): This reflects the intensity of uterine contractions by calculating the power of the EHG signal. It is defined as the square root of the square of the value at all sample points divided by the number of sample points.

[0072]

[0073] 2) Autocorrelation zero-crossing point (AuZC): The autocorrelation function R(τ) measures the similarity of a signal at different times and can be used to distinguish between random and periodic activity in EHG signals. It is defined as:

[0074]

[0075] Therefore, AuZC is defined from the signal The first zero-crossing point from which the peak of the autocorrelation function R(τ) begins:

[0076] AuZc=min{τ,τ∈τ xx}, R(τ xx )=0 (6)

[0077] 3) Peak Frequency (PF): The power spectral density P(n0) of the input signal is obtained through Fast Fourier Transform. PF is defined as the frequency value corresponding to the maximum amplitude of P(n0).

[0078]

[0079] 4) Median Frequency (MDF): Defined as the frequency at which the sum of the upper and lower portions of the signal power spectrum P(n) is the same.

[0080]

[0081] 5) Sample Entropy (SampEn): Describes the irregularity of a finite-length time series. The more complex the time series, the higher the entropy.

[0082] The higher the value, the better. It can be defined as the negative logarithm of the conditional probability that two sequences are similar within a tolerance value r:

[0083]

[0084] Among them, A m+1 (r) and B m (r) represents the probability of two sequences matching m+1 and m points, respectively. For example, the input dimension m = 3 and the tolerance value r = 0.15SD (SD is the standard deviation of the time series).

[0085] 6) FuzzyEn: Measures the probability of generating new patterns when the dimensions of a time series change, using an exponential function. Fuzzyen is used to fuzzify the similarity measurement, making the entropy value change continuously and smoothly with parameter variations. This is suitable for situations where there are blurred edges between classes. FuzzyEn is calculated as:

[0086]

[0087] Where, φ m+1(r) and ψ m (r) represents the probability that two sequences match m+1 and m points respectively under weighted similarity s. For example, the embedding dimension m = 2, the tolerance value r = 0.2SD, and the weighted similarity s = 3 can be set.

[0088] Step 5, Feature Selection:

[0089] Step 5.1: Convert the three-dimensional feature matrix f into a one-dimensional feature set F = {F(k) | k = 1, 2, ..., Q × I0 × 6}, and calculate the chi-square score of the k-th feature F(k) using the chi-square test. After sorting all chi-square scores in descending order, select K1 features that are greater than the preset threshold th to obtain the initial feature set F. * ={F * (k′)|k′=1,2,…,K1};where, F * (k′) represents the k′th feature after the initial screening, and K1 represents the number of features in the initial screening feature set;

[0090] Step 5.2: Feature optimization based on the bidirectional sequential search algorithm, and set the total number of algorithm executions to T:

[0091] Step 5.2.1: From the feature set F * A feature subset F is randomly generated from the data. * (k0);

[0092] Step 5.2.2: From the feature set F * Select one feature F in sequence * (k + Add feature subset F * In (k0), determine F * If the F1 score of (k0) is optimal, proceed to step 5.2.3; otherwise, continue to step 5.2.2.

[0093] Step 5.2.3: Sequentially start from feature subset F * Deleting a feature F(k) from (k0) - And determine the feature subset F after deletion. * Does the F1 score of (k0) decrease? If so, then in the feature subset F * The feature F(k) is retained in (k0). - Otherwise, continue with step 5.2.3; thus obtaining the optimal feature subset F. ** ={F ** (k″)|k″=1,2,…,K2},where, F ** (k″) represents the k″-th feature selected by the bidirectional search algorithm, and K2 represents the number of features in the optimal feature subset;

[0094] In this embodiment, a feature selection algorithm is applied to the high-dimensional feature set f to remove redundant terms and irrelevant components, thereby improving feature quality. A two-step feature selection algorithm is designed: 5.1 Feature pre-screening and 5.2 Feature optimization. In step 5.1, it is assumed that features and categories are independent of each other; the larger the chi-square value, the stronger the correlation between features and categories. In step 5.2, a wrapper method is used to customize the optimal feature subset based on machine learning for the classifier. Based on the heuristic search idea, a sequence bidirectional search feature selection algorithm (SBS) is designed, and feature optimization is performed through performance feedback from the F1 score. The F1 score is a weighted average of the classifier model's precision and recall, taking into account both precision and recall, and is calculated as follows:

[0095]

[0096] In the above feature selection process, the chi-square test threshold *th* and the number of iterations *T* in the bidirectional sequential search algorithm are controllable. This means that the globally optimal solution can be selected by maximizing the F1 score of the validation data. Setting *th* to a range of 0 to 0.4 with an increment of 0.1, and *T* to a range of 1 to 5 with an increment of 1, the resulting F1 score variation curves are shown below. Figure 5 As shown, for example, th = 0.2 and T = 4 can be set, at which point the F1 score reaches optimal performance.

[0097] Step 6: Following the process of steps 1-5, obtain the optimal feature subsets of M objects to be tested and form a dataset R = {F}. ** 1(k″), F ** 2(k″), ..., F ** M (k″)|k″=1,2,…,K2},where, F ** M (k″) represents the optimal feature subset of the Mth object to be tested.

[0098] Divide the dataset R into M a Training data with known labels and their corresponding category labels M b Test data with unknown labels Use support vector machines on the training data Establish category labels The mapping relationship between them is used to find the hyperplane γ that can classify different categories, and thus the hyperplane γ based on the support vector machine model is used to analyze the test data. Perform classification and obtain the predicted classification results.

[0099] In this embodiment, the dataset R, consisting of the optimal feature set obtained from M test objects, is divided into a training set and a test set in an 8:2 ratio, ensuring that the proportion of different categories in the two sets is consistent, thus obtaining M a Training data with known labels and M b Test data with unknown labels. Since the feature parameters of the EHG signal are linearly inseparable, the Support Vector Machine model selects the RBF kernel to map the input features to a high-dimensional space, finding a hyperplane that can separate different categories, and then performing classification prediction on the test data. Furthermore, the settings of hyperparameters such as the penalty coefficient C and the kernel scale gamma have a significant impact on the classifier's performance and can be optimized through grid parameter search. For example, M can be set... a =240, M b =60, C={2 -3 ,2 -2 ,2 -1 ,2 1 ,2 2}, gamma={5 -4 5 -3 5 -2 5 -1}

[0100] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor in executing the above-described uterine electromyography signal classification method, and the processor is configured to execute the program stored in the memory.

[0101] In this embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the above-described uterine electromyography signal classification method.

[0102] To quantify the classification performance of this invention, common evaluation metrics including sensitivity, specificity, F1 score, accuracy, and area under the curve (AUC) were used to evaluate its preterm birth prediction performance. Experimental data from 300 different subjects were used to compare the method of this invention with traditional uterine electromyography signal classification methods. All data processing was performed using MATLAB 2020 software. a And run on.

[0103] The four comparison methods are as follows:

[0104] The classification method for uterine electromyography (EHG) signals based on short-time Fourier transform (SFT) proceeds as follows: The acquired surface EHG signals undergo SFT to expand to different frequency ranges. To meet the requirements for high frequency and time resolution, the window length is set to n = 200, and the step size is set to s = 100. When the sampling frequency of the EHG signal is 20Hz, the time and frequency resolutions after the SFT are 5s and 0.1Hz, respectively. Next, the approximate entropy and sample entropy features of different frequency components are calculated, and principal component analysis is applied to select the top 20 principal components as the optimized feature subset. Finally, a suitable machine learning model is selected to complete the classification of the uterine EHG signals.

[0105] The classification method for uterine electromyography (EMG) signals based on genetic algorithms proceeds as follows: Time-domain, frequency-domain, and nonlinear-domain features in four frequency bands (0.1–4 Hz, 0.2–34 Hz, 0.34–4 Hz, 0.34–1 Hz) are calculated for the collected surface uterine EMG signals, along with clinical information, resulting in a total of 222 features. Next, a genetic algorithm is used to optimize the feature subset and reduce the feature dimensionality. This is a population-based heuristic search method that simulates the natural evolutionary process through genetic function, completing iterative population optimization. Finally, the uterine EMG signals are classified based on a machine learning model.

[0106] The classification method for uterine electromyography (EMG) signals based on EMD and wavelet packet decomposition is as follows: The acquired surface uterine EMG signals are subjected to EMD decomposition to obtain 11 intrinsic mode functions (IMFs). Then, wavelet packet decomposition is used to obtain 12 low-frequency and high-frequency components on each IMF. Further, ten features, including fractal dimension, average energy, sample entropy, and standard deviation, are extracted. Next, particle swarm optimization (PSO) is used for feature selection, and the selected important features are ranked using the Patacharian distance. Finally, a classifier is used to process all ranked features to complete the classification of uterine EMG signals.

[0107] The classification method for uterine electromyography (EMG) signals based on Hilbert-Huang transform is as follows: The collected surface uterine EMG signals are subjected to Hilbert-Huang transform, which includes two steps: EMD and Hilbert transform. First, a set of IMFs is obtained through EMD. The entropy values ​​of the instantaneous amplitude and instantaneous frequency of the first ten IMFs are calculated, and the ratio of the two is used as the final feature. Finally, a model is established based on a machine learning algorithm, and the input samples are trained to complete the classification of uterine EMG signals.

[0108] The bar chart showing the mean and standard deviation of the performance of the method provided by this invention and the comparative method in the uterine electromyography signal classification task is shown below. Figure 6 As shown in the figure. Experimental results demonstrate that, under all five indicators, the uterine electromyography signal classification method based on multivariate empirical mode decomposition can provide high accuracy and stability in predicting abnormal uterine activity.

[0109] In summary, this embodiment presents a method for classifying uterine electromyography (EHG) signals based on multivariate empirical mode decomposition (EMD). This method includes: acquiring surface EHG signals during uterine contractions; preprocessing and denoising the signals according to the frequency range of uterine contractions; expanding the original EHG time-series signals to frequency components at different scales through EHG, ensuring spectral consistency across channels, and better observing the details of uterine electrical activity during contractions; extracting six indicators reflecting the intensity and complexity of EHG signals from the time domain, frequency domain, time-frequency domain, and nonlinear domain to form a macroscopic description of uterine muscle activity; integrating a chi-square test pre-screening and a bidirectional sequence search feature selection algorithm to obtain an optimized feature subset, reducing feature dimensionality, avoiding model overfitting, and removing redundant terms and irrelevant noise components; and finding the boundary plane separating normal and abnormal EHG signal categories based on machine learning algorithms, uncovering the potential relationship between EHG feature descriptions and uterine electrophysiological activity, and achieving objective and effective assessment of abnormal uterine activity. This invention is of great significance for clinical uterine contraction monitoring, preterm birth screening, and postpartum hemorrhage assessment.

Claims

1. A method for classifying uterine electromyographic signals based on multivariate empirical mode decomposition, characterized in that, The procedure is as follows: Step 1, Signal Acquisition: Step 1.1: Using uterine electromyography equipment and The high-density surface electrode of the channel collects data on a test object. The surface electromyographic signal data of the uterus during uterine contractions are denoted as: , Indicates the first The first one on the passage uterine electromyography signal data at different times, This represents the total number of channels in the high-density electrode. The total duration recorded; Step 1.2, at the sampling frequency Under these conditions, uterine electromyography signals were obtained. Discrete sample point data, denoted as: , ,in, Indicates the first The first one on the passage Uterine electromyography signal data from a sample point The total number of sample points recorded; thus, after... After measuring the duration, the sample length was obtained as follows: Uterine electromyography time series ,in, Indicates the first The time series of uterine electromyography in the channel, and | uterine electromyography time sequence The category label is ,in, =1 indicates normal. =0 indicates an error; Step 2, Signal Preprocessing: Using a bandpass filter to analyze the uterine electromyography time series After noise reduction to filter out interfering physiological signals and irrelevant noise, the uterine electromyography time sequence of the Q channel was obtained after noise reduction. ,in, Indicates the number after noise reduction processing The time series of uterine electromyography in the channel; Step 3, Signal Time-Frequency Spreading: Step 3.1: Use the multivariate empirical mode decomposition method shown in equation (1) to decompose the time series of uterine electromyography signals. Jointly decomposed on the Q channel into Each intrinsic modal component and residual term: ； In equation (1), Indicates the first The first channel Each intrinsic mode component This represents the residual term of the q-th channel; Step 3.2: Calculate according to equations (2) and (3) respectively. In the Variance contribution of each intrinsic modal component and correlation coefficient : ； ； In equations (2) and (3), and They represent and variance express and cross variance; according to | and | Select The intrinsic modal components that contribute to and are relevant to uterine contractions are denoted as... ;in, Indicates the first The first channel One intrinsic mode component; Step 4, Feature Extraction: Set the window length to Overlap with window For each length of of After performing sliding window frame division processing, a total of There are several sliding window time series, and one of the sliding window time series is denoted as . ,and ,in, Indicates length is The first time series Data for each sample point; Six eigenvalues ​​are calculated for each sliding window time series, including: root mean square amplitude and autocorrelation zero-crossing number in the time domain, peak frequency and median frequency in the frequency domain, and sample entropy and fuzzy entropy in the nonlinear domain. The six eigenvalues ​​obtained from all sliding window time series are then averaged to obtain the three-dimensional feature matrix. ,in, Indicates the first The first channel The eigenmode component of the eigenmode component One eigenvalue; Step 5, Feature Selection: Step 5.1: Convert the three-dimensional feature matrix Transform into a one-dimensional feature set And calculate the first by chi-square test Features The chi-square scores are calculated, and after sorting all chi-square scores in descending order, those greater than a preset threshold are selected. of Each feature is used to obtain the feature set after initial screening. ;in, Indicates the first [number] after the initial screening. One characteristic, This indicates the number of features in the initial feature set. Step 5.2: Analyze the initially selected feature set using a bidirectional sequence search algorithm. Feature optimization is performed to obtain the optimal feature subset. ,in, This indicates that the bidirectional search algorithm selects the first... One characteristic, This represents the number of features in the optimal feature subset; Step 5.2.1: Set the total number of algorithm executions to T, from the feature set... A feature subset is randomly generated from the data. ; Step 5.2.2: From the feature set Select one feature in sequence Add feature subset In, and judge If the F1 score is optimal, proceed to step 5.2.3; otherwise, continue to step 5.2.

2. Step 5.2.3: Sequentially from the feature subsets Deleting a feature And determine the feature subset after deletion. Does the F1 score decrease? If so, then in the feature subset... Preserved features Otherwise, continue with step 5.2.3; thus obtaining the optimal feature subset. ,in, This indicates that the bidirectional search algorithm selects the first... One characteristic, This represents the number of features in the optimal feature subset; Step 6: Obtain the results following the steps 1-5. The optimal feature subset of each object to be tested is used to form a dataset. ,in, This represents the optimal feature subset of the Mth object to be tested; Dataset Divided into Training data with known labels and their corresponding category tags , Test data with unknown labels Using support vector machines on training data Establish category labels The mapping relationship between them is used to find the hyperplanes that can classify different categories. Thus, the hyperplane based on the support vector machine model For test data Perform classification and obtain the predicted classification results. .

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the uterine electromyography signal classification method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when run by a processor, executes the steps of the uterine electromyography signal classification method of claim 1.

Images