Uterine electromyographic signal classification method based on deep learning

Abstract

The invention discloses a uterine electromyographic signal classification method based on deep learning. The method comprises the following steps: 1, acquiring uterine electromyographic signals by using a multi-channel electrode, and carrying out filtering and noise reduction on the uterine electromyographic signals; 2, segmenting the signal by adopting a sliding window; 3, performing continuous wavelet transform on each window signal to obtain a two-dimensional time-frequency graph, and extracting the normalized peak amplitude of a specific frequency band of the whole section of signal as an auxiliary feature; 4, constructing a deep learning classification model combining cost-sensitive learning and Focal Loss, and training the deep learning classification model; and 5, applying the trained model to unknown user signals, summarizing window-level prediction results, and obtaining a final classification result through user-level post-processing. According to the method, the multi-scale time-frequency characteristics and auxiliary discrimination information of the uterine electromyographic signals can be fully excavated, so that the classification accuracy of the uterine electromyographic signals is improved.

Description

Technical Field

[0001] This invention belongs to the field of electromyography signal processing and analysis, specifically a deep learning-based method for classifying uterine electromyography signals. Background Technology

[0002] Electromyography (EHG) is a type of electrophysiological signal obtained by placing multichannel electrodes on the surface of the pregnant woman's abdomen to record the electrical activity generated in the uterine smooth muscle layer. This signal contains rich information on uterine dynamics, characterizing the onset time, intensity, and direction of electrical activity propagation within the uterine tissue. However, due to the significant non-stationarity, low-frequency characteristics, and susceptibility to various physiological and external noise interferences inherent in EHG signals, their processing and analysis still face considerable challenges.

[0003] The core task of uterine electromyography (EHG) signal classification is primarily focused on distinguishing between preterm and full-term delivery, i.e., preterm birth prediction. Existing preterm birth prediction methods largely rely on manually designed global features, such as time-domain, frequency-domain, and nonlinear indices, and use traditional machine learning classifiers like support vector machines, linear discriminant analysis, or random forests to establish a mapping relationship between features and delivery outcomes to achieve prediction. Essentially, preterm birth prediction is a classification problem; its goal is to learn the discriminative boundaries between different categories (preterm and full-term) in the feature space through the analysis of EHG data. However, these methods have significant limitations. First, the feature extraction process is highly dependent on human experience, making it difficult to guarantee the comprehensiveness and robustness of the features. Second, traditional classifiers have limited performance when dealing with high-dimensional and complex features, making it difficult to effectively capture the non-stationarity and complex features of EHG signals. More importantly, class imbalance is prevalent in clinically collected EHG data, resulting in insufficient model recognition of the minority class (preterm samples). These problems collectively limit further improvements in the accuracy and practicality of existing methods for uterine electromyography signal classification.

[0004] In recent years, the development of deep learning technology has provided new solutions for the intelligent analysis of complex physiological signals. However, this method still does not fully mine the deep discriminative information of EHG and has performance bottlenecks when dealing with data imbalance, which limits the prediction accuracy and clinical applicability. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies by proposing a deep learning-based method for classifying uterine electromyography (EMG) signals. This method aims to fully exploit the multi-scale time-frequency features and auxiliary discrimination information of EMG signals, thereby achieving accurate classification of EMG signals. This overcomes the problems of existing EMG signal classification methods, such as reliance on manual features, insufficient extraction of features from non-stationary signals, and decreased recognition performance under class imbalance conditions.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a deep learning-based method for classifying uterine electromyographic signals, characterized by the following steps: Step 1: Using a uterine electromyography (EMG) signal measurement device and an M-channel array electrode according to the sampling rate collection After collecting uterine electromyography (EMG) signals from several users within time T, a bandpass filter is used to denoise the EMG signals, resulting in a denoised uterine EMG signal dataset. ,in, This represents the denoised uterine electromyography signal of the j-th user, and This represents the uterine electromyography (EMG) signal sequence of the j-th user after noise reduction in the i-th channel, where N represents the total number of users. The length of the denoised uterine electromyography signal sequence; Data set of uterine electromyography signals The tag set is ,in, The true category is ,when 0 represents This is a normal signal, when ,express This is an abnormal signal; Step 2: According to the window length W and step size H, Perform sliding window segmentation to obtain the window sequence of the j-th user in the i-th channel. ,in, This represents the uterine electromyography (EMG) signal window segment of the j-th user in the i-th channel under the k-th window. Let represent the total number of windows created by the j-th user, and , This is a rounding operation; ; Step 3: [Regarding...] Two-dimensional time-frequency extraction is performed to obtain the window-level time-frequency plot sequence of the j-th user in the i-th channel. At the same time, for Global feature extraction is performed to obtain Peak amplitude characteristics ; Step 4: Construct a deep learning model for window classification, including: a multi-branch convolutional module, a feature mapping module, and a classification module, and then... and Process the data and output the predicted category probability value for the j-th user in the k-th window. and predictive classification ; Step 5: Based on X and its label set Y, and according to... and Construction cost sensitive loss It is used to train deep learning models to obtain a trained uterine electromyography signal classification model; Step 6: Input the uterine electromyography (EMG) signal of the unknown category user u into the trained uterine EMG signal classification model for processing to obtain window-level prediction results. After calculating statistical features, the results are input into the classifier for category determination to obtain the final predicted category of the user u.

[0007] The characteristic of the deep learning-based uterine electromyography signal classification method described in this invention is that step 3 is performed according to the following steps: Step 3.1: For Perform continuous wavelet transform to obtain the two-dimensional time-frequency representation of the i-th channel for the j-th user in the k-th window. Thus, the window time-frequency plot sequence of the j-th user in the i-th channel is obtained. ; Step 3.2: Use equation (1) to obtain the normalized peak amplitude characteristics of the j-th user. : (1) In equation (1), The denoised uterine electromyography signal sequence of the c-th channel selected for the j-th user Power spectrum at frequency q For the selected frequency band, and These represent the lower and upper frequency limits of the selected frequency band, respectively. The normalized peak amplitude is the reference frequency band.

[0008] Furthermore, step 4 is performed as follows: Step 4.1: The multi-branch convolutional module consists of M branches; each branch includes a residual block and a pooling layer; The inputs are processed by M branches, and each branch is processed by residual blocks and pooling layers. The corresponding output is the convolutional feature set of the j-th user in the k-th window. ,in, This represents the convolutional feature vector of the j-th user in the i-th channel within the k-th window; Step 4.2: The feature mapping module sequentially processes the data through a fully connected layer and a Sigmoid activation function layer. The process is performed to obtain the global embedding feature vector of the j-th user. ; Step 4.3: The classification module will and After fusion, and further processing by a fully connected layer and a sigmoid activation function, the predicted category probability value for the j-th user in the k-th window is obtained. The category corresponding to the larger predicted probability value will be used as the final predicted category. .

[0009] Furthermore, step 5 is performed as follows: Step 5.1: Construct a cost-sensitive loss function using equation (2) : (2) In equation (2), It is the modulation factor. for Cost-sensitive weights; Step 5.2: Based on set X and its label set Y, train the deep learning model using the Adam optimizer and calculate... To update the model parameters until the maximum number of iterations is reached or The process continues until convergence, thus obtaining a well-trained uterine electromyography signal classification model.

[0010] Furthermore, step 6 is performed as follows: Step 6.1: Input the uterine electromyography (EMG) signal of the user u with an unknown category into the trained uterine EMG signal classification model, and output the predicted category probability of the user u in the k-th window. and prediction categories ; Step 6.2: Based on Calculate the statistical characteristics of the user u to be tested, including: the ratio of the number of windows predicted as an anomaly category to the total number of windows. and the average predicted probability of the window ;in, This represents the total number of windows divided for the user u under test; Step 6.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and The input is fed into the classifier to obtain the final predicted category of the user u to be tested. The classifier uses a shallow decision tree with linear threshold judgment as the classification rule.

[0011] The present invention provides an electronic device, including 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, and the processor is configured to execute the program stored in the memory.

[0012] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the uterine electromyography signal classification method.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes an end-to-end deep learning method that can automatically extract discrimination patterns directly from raw uterine electromyography signals, replacing the traditional method that relies on manual feature engineering. This method can not only better capture non-stationary, multi-scale dynamic features, but also improve the ability to identify a few abnormal samples, thereby avoiding the problems of one-sidedness and insufficient robustness caused by manually designed features.

[0014] 2. This invention utilizes continuous wavelet transform to obtain multi-scale time-frequency representations of uterine electromyography signals, and combines peak amplitudes of specific frequency bands as auxiliary features, significantly enhancing feature expression capabilities. This enables the model to more comprehensively characterize dynamic features related to delivery outcomes, thereby improving the overall accuracy and robustness of classification.

[0015] 3. This invention employs a cost-sensitive learning strategy based on focal loss, which effectively mitigates the adverse effects of a severe imbalance between preterm and full-term samples on model training and significantly improves the ability to identify a few abnormal classes.

[0016] 4. Through a dual-level prediction strategy at the window level and the user level, this invention demonstrates superior discrimination ability and stability compared to existing methods in experimental verification. It can accurately distinguish between normal and abnormal uterine electromyographic signals and has promising clinical application prospects. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the deep learning model of the present invention; Figure 3 The graph shows the classification results and training loss function trends of the proposed aspects at the window level when the focus factor γ takes different values. Figure 4 This is a bar chart comparing the performance of different preterm birth prediction methods of the present invention at the user level. Detailed Implementation

[0018] In this embodiment, a deep learning-based method for classifying uterine electromyography signals is described, such as... Figure 1 As shown, the procedure is as follows: Step 1: Using a uterine electromyography (EMG) signal measurement device and an M-channel array electrode according to the sampling rate collection After collecting uterine electromyography (EMG) signals from several users within time T, a bandpass filter is used to denoise the EMG signals, resulting in a denoised uterine EMG signal dataset. ,in, This represents the denoised uterine electromyography signal of the j-th user, and This represents the uterine electromyography (EMG) signal sequence of the j-th user after noise reduction in the i-th channel, where N represents the total number of users. The length of the denoised uterine electromyography signal sequence; Data set of uterine electromyography signals The tag set is ,in The true category is ,when 0 indicates that the uterine electromyography signal of the j-th user is normal. , indicating that the uterine electromyography signal of the j-th user is abnormal; In this example, 164 pregnant women were recruited (N), and data were collected during gestational ages greater than 31 weeks. Among them, 132 samples showed normal uterine electromyography (EMG) signals, and 32 samples showed abnormal EMG signals. Before data collection, each pregnant woman lay comfortably on her back. The abdomen was cleaned with alcohol wipes. A 4-channel electrode (M=4 channels) was used, with E1, E2, E3, and E4 symmetrically placed around the navel, with a 7cm spacing between the electrodes. The sampling rate was set to fs=20Hz, and the acquisition time T was 30 minutes. To avoid external noise interference, the original 4 channels were differentially divided pairwise to obtain three bipolar channels: S1 = E2 − E1, S2 = E2 − E3, and S3 = E4 − E3. S2 and S3 were selected as the final EMG signals. A fourth-order Butterworth filter with a cutoff frequency of 0.08-5Hz was used to filter out noise such as maternal respiration, cardiac activity, motion artifacts, and power line interference present in the original EMG signals. To eliminate the edge effect caused by the transient response of the filter, 150s of data were truncated from both ends of the signal, resulting in a final length of L of 30000.

[0019] Step 2: According to the window length W and step size H, Perform sliding window segmentation to obtain the window sequence of the j-th user in the i-th channel. ,in, This represents the uterine electromyography signal segment of the j-th user in the i-th channel under the k-th window. Let represent the total number of windows created by the j-th user, and , This is a rounding operation; ; In this example, the sliding window size W = 512, and the step size H = 256. The size is 116.

[0020] Step 3: [Regarding...] Two-dimensional time-frequency extraction is performed to obtain the window-level time-frequency plot sequence of the j-th user in the i-th channel. At the same time, for Global feature extraction is performed to obtain Peak amplitude characteristics ; Step 3.1: For Perform continuous wavelet transform to obtain the two-dimensional time-frequency representation of the i-th channel for the j-th user in the k-th window. Thus, the window time-frequency plot sequence of the j-th user in the i-th channel is obtained. ; In this example, the continuous wavelet transform is performed according to equation (1a): (1a) In equation (1a), x[m] represents the original uterine electromyography signal within the window. It is the complex conjugate of the mother wavelet. The relationship between the scaling parameter s and the frequency f is given by equation (1b): (1b) The Morlet wavelet is chosen as the mother wavelet in wavelet transform because it is more suitable for analyzing low-frequency and non-stationary electrophysiological signals such as uterine electromyography. The center frequency of the Morlet mother wavelet... According to the sampling frequency The scale range is calculated to be 3–204 based on the filter bandwidth of 0.08–5 Hz.

[0021] Step 3.2: Use equation (1) to obtain the normalized peak amplitude characteristics of the j-th user. : (1) In equation (1), The denoised uterine electromyography signal sequence of the c-th channel selected for the j-th user Power spectrum at frequency q For the selected frequency band, and These represent the lower and upper frequency limits of the selected frequency band, respectively. Normalized peak amplitude for the reference frequency band; In this example, a grid search of the frequency band range is used, and a two-sample t-test is employed to screen for the most statistically significant frequency band intervals, thereby determining the lower and upper frequency limits. In the grid search of the frequency band range, the step size for the lower and upper cutoff frequencies is set to 0.025 Hz, the reference frequency band is selected as 0.08-1 Hz, and the frequency band range with the smallest p-value in the two-sample t-test is chosen. In this dataset, the lower frequency limit is... and upper limit frequency At that time, the two types of samples showed the greatest significant difference.

[0022] Step 4: Construct a deep learning model for window classification, including: a multi-branch convolutional module, a feature mapping module, and a classification module, and then... and Process the data and output the predicted category probability value for the j-th user in the k-th window. and predictive classification ; Step 4.1: Construct a multi-branch convolutional module consisting of M branches; each branch includes a residual block and a pooling layer; The inputs are processed by M branches, and each branch is processed by residual blocks and pooling layers. The corresponding output is the convolutional feature set of the j-th user in the k-th window. ,in, This represents the convolutional feature vector of the j-th user in the i-th channel within the k-th window; In this example, each branch of the multi-branch convolutional module includes three residual blocks and one max-pooling layer. The size of the two-dimensional convolutional kernel in the residual block is... The number of channels in residual blocks with 2 or 3 elements are 32, 64, and 128 respectively. The overall structure of the model is as follows: Figure 2 As shown.

[0023] Step 4.2: The feature mapping module sequentially processes the data through a fully connected layer and a Sigmoid activation function layer. The process is performed to obtain the global embedding feature vector of the j-th user. ; Step 4.3: The classification module will... and After fusion, and further processing by a fully connected layer and a sigmoid activation function, the predicted category probability value for the j-th user in the k-th window is obtained. The category corresponding to the larger predicted probability value will be used as the final predicted category. .

[0024] Step 5: Based on the uterine electromyography signal dataset X and its label set Y, and according to... and Construction cost sensitive loss It is used to train deep learning models to obtain a trained uterine electromyography signal classification model; Step 5.1: Construct a cost-sensitive loss function using equation (2) : (2) In equation (2), It is the modulation factor. for Cost-sensitive weights; in this example, the focus factor Focus factor is used to adjust the weights of hard-to-separate samples. The value was determined experimentally. The value is set to the ratio of the number of normal to the number of abnormal samples for abnormal samples, for example, 4.125, and 1 for normal samples.

[0025] Step 5.2: Based on the uterine electromyography signal dataset X and its label set The deep learning model is trained using the Adam optimizer, and computation is performed. To update the model parameters until the maximum number of iterations is reached or The process continues until convergence, thus obtaining a well-trained uterine electromyography signal classification model.

[0026] Step 6: Input the uterine electromyography (EMG) signal of the unknown category user u into the trained uterine EMG signal classification model for processing, obtain window-level prediction results, and use them to calculate statistical features before inputting them into the classifier for category determination. Step 6.1: Input the uterine electromyography (EMG) signal of the user u with an unknown category into the trained uterine EMG signal classification model, and output the predicted category probability of the user u in the k-th window. and prediction categories ; Step 6.2: Based on Calculate the statistical characteristics of the user u to be tested, including: the ratio of the number of windows predicted as an anomaly category to the total number of windows. and the average predicted probability of the window ;in, This represents the total number of windows divided for the user u under test. Step 6.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and The input is fed into the classifier to obtain the final predicted category of the user u to be tested. The classifier uses a shallow decision tree with linear threshold judgment as the classification rule.

[0027] In this example, the decision tree classifier is designed to perform rule-based linear integration based on window-level discrimination results to output the final user-level classification result. This classifier does not participate in additional model training and is only used for automated category discrimination. Its classification rule is: when both statistical features R^{(u)} and P^{(u)} are greater than a set threshold, the sample to be tested is determined to be an anomalous category; for example, the thresholds are both set to 0.5. Simultaneously, the signal data of the sample to be tested is included in the subsequent dataset.

[0028] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0029] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0030] To quantify the classification performance of this invention, commonly used evaluation metrics, including accuracy, F1 score, sensitivity, specificity, and area under the ROC curve (AUC), were used to evaluate the classification performance of uterine electromyography (EMG) signals. The method of this invention was compared with three other uterine EMG signal classification methods using 164 different subjects. All data processing and experimental procedures were performed on the same device. Figure 3 Verified different Under the given values, the performance of the proposed method in window-level classification (left figure) and the trend of training loss (right figure) are shown. The results indicate that as γ increases, both AUC and F1-score show a trend of first increasing and then decreasing. When γ=2, the model performance reaches its optimal level, with AUC and F1-score of 78.11%±5.10% and 67.32%±1.86%, respectively.

[0031] Three comparison methods were set up as follows: (1) CNN + STFT + PA: CWT was replaced with Short-Time Fourier Transform (STFT) to extract the EHG time-frequency map. The STFT parameters (window length, overlap rate, frequency distribution) were set in the same way as the Continuous Wavelet Transform (CWT) scheme to ensure the comparability of time-frequency resolution and frequency band range; (2) CNN + CWT: In this experiment, the PA features of the S3 channel were removed and only the time-frequency map generated by CWT was retained as input; (3) TCNN + Time Windows (baseline model): A one-dimensional convolutional neural network (Traditional Convolutional Neural Network, TCNN) was used to directly input the preprocessed EHG original time domain signal into the model after sliding window segmentation, omitting time-frequency transformation and PA feature extraction, as the baseline of the traditional method.

[0032] The mean and standard deviation histograms of the proposed method and the comparative method in the uterine electromyography signal classification task are shown below. Figure 4 As shown in the figure. The results show that, under all five indicators, the performance of the method proposed in this invention is significantly better than the comparative method, verifying the effectiveness and accuracy of this method in the uterine electromyography signal classification task.

Claims

1. A deep learning-based method for classifying uterine electromyographic signals, characterized in that, The procedure is as follows: Step 1: Using a uterine electromyography (EMG) signal measurement device and an M-channel array electrode according to the sampling rate collection After collecting uterine electromyography (EMG) signals from several users within time T, a bandpass filter is used to denoise the EMG signals, resulting in a denoised uterine EMG signal dataset. ,in, This represents the denoised uterine electromyography signal of the j-th user, and This represents the uterine electromyography (EMG) signal sequence of the j-th user after noise reduction in the i-th channel, where N represents the total number of users. The length of the denoised uterine electromyography signal sequence; Data set of uterine electromyography signals The tag set is ,in, The true category is ,when 0 represents This is a normal signal, when ,express This is an abnormal signal; Step 2: According to the window length W and step size H, Perform sliding window segmentation to obtain the window sequence of the j-th user in the i-th channel. ,in, This represents the uterine electromyography (EMG) signal window segment of the j-th user in the i-th channel under the k-th window. Let represent the total number of windows created by the j-th user, and , This is a rounding operation; ; Step 3: [Regarding...] Two-dimensional time-frequency extraction is performed to obtain the window-level time-frequency plot sequence of the j-th user in the i-th channel. At the same time, for Global feature extraction is performed to obtain Peak amplitude characteristics ; Step 4: Construct a deep learning model for window classification, including: a multi-branch convolutional module, a feature mapping module, and a classification module, and then... and Process the data and output the predicted category probability value for the j-th user in the k-th window. and predictive classification ; Step 5: Based on X and its label set Y, and according to... and Construction cost sensitive loss It is used to train deep learning models to obtain a trained uterine electromyography signal classification model; Step 6: Input the uterine electromyography (EMG) signal of the unknown category user u into the trained uterine EMG signal classification model for processing to obtain window-level prediction results. After calculating statistical features, the results are input into the classifier for category determination to obtain the final predicted category of the user u.

2. The method for classifying uterine electromyographic signals based on deep learning according to claim 1, characterized in that, Step 3 is performed as follows: Step 3.1: For Perform continuous wavelet transform to obtain the two-dimensional time-frequency representation of the i-th channel for the j-th user in the k-th window. Thus, the window time-frequency plot sequence of the j-th user in the i-th channel is obtained. ; Step 3.2: Use equation (1) to obtain the normalized peak amplitude characteristics of the j-th user. : (1) In equation (1), The denoised uterine electromyography signal sequence of the c-th channel selected for the j-th user Power spectrum at frequency q For the selected frequency band, and These represent the lower and upper frequency limits of the selected frequency band, respectively. The normalized peak amplitude is the reference frequency band.

3. The method for classifying uterine electromyographic signals based on deep learning according to claim 2, characterized in that, Step 4 is performed as follows: Step 4.1: The multi-branch convolutional module consists of M branches; each branch includes a residual block and a pooling layer; The inputs are processed by M branches, and each branch is processed by residual blocks and pooling layers. The corresponding output is the convolutional feature set of the j-th user in the k-th window. ,in, This represents the convolutional feature vector of the j-th user in the i-th channel within the k-th window; Step 4.2: The feature mapping module sequentially processes the data through a fully connected layer and a Sigmoid activation function layer. The process is performed to obtain the global embedding feature vector of the j-th user. ; Step 4.3: The classification module will and After fusion, and further processing by a fully connected layer and a sigmoid activation function, the predicted category probability value for the j-th user in the k-th window is obtained. The category corresponding to the larger predicted probability value will be used as the final predicted category. .

4. The method for classifying uterine electromyographic signals based on deep learning according to claim 1, characterized in that, Step 5 is performed as follows: Step 5.1: Construct a cost-sensitive loss function using equation (2) : (2) In equation (2), It is the modulation factor. for Cost-sensitive weights; Step 5.2: Based on set X and its label set Y, train the deep learning model using the Adam optimizer and calculate... To update the model parameters until the maximum number of iterations is reached or The process continues until convergence, thus obtaining a well-trained uterine electromyography signal classification model.

5. The method for classifying uterine electromyographic signals based on deep learning according to claim 1, characterized in that, Step 6 is performed as follows: Step 6.1: Input the uterine electromyography (EMG) signal of the user u with an unknown category into the trained uterine EMG signal classification model, and output the predicted category probability of the user u in the k-th window. and prediction categories ; Step 6.2: Based on Calculate the statistical characteristics of the user u to be tested, including: the ratio of the number of windows predicted as an anomaly category to the total number of windows. and the average predicted probability of the window ;in, This represents the total number of windows divided for the user u under test; Step 6.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and The input is fed into the classifier to obtain the final predicted category of the user u to be tested. The classifier uses a shallow decision tree with linear threshold judgment as the classification rule.

6. 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 according to any one of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the uterine electromyography signal classification method according to any one of claims 1-5.

Images