A fetal electrocardio recognition method based on full channel features

Abstract

The application discloses a fetal electrocardiogram recognition method based on full-channel features, which can select a fetal electrocardiogram signal from multi-channel output of a blind source separation method and belongs to the technical field of fetal electrocardiogram recognition methods and comprises the following steps: signal preprocessing: normalizing multi-channel electrocardiogram data; fetal electrocardiogram recognition: recognizing fetal electrocardiogram by adopting a cross-domain consistency convolutional neural network; and outputting results: selecting fetal electrocardiogram in original signals according to the recognition results.

Description

Technical Field

[0001] This invention relates to a fetal heart rate identification method, and more particularly to a fetal heart rate identification method based on full-channel features. This method can select fetal heart rate signals from the multi-channel output of the blind source separation method, and belongs to the technical field of fetal heart rate identification methods. Background Technology

[0002] As a non-invasive monitoring method, fetal electrocardiography can not only be used to calculate fetal heart rate, but also provide information related to physiological status, and therefore has received increasing attention in the field of fetal heart rate monitoring.

[0003] During signal acquisition, placing electrodes on the pregnant woman's abdominal wall allows for non-invasive acquisition of mixed abdominal electrocardiogram signals containing fetal heart signals.

[0004] It is important to note that the fetal heart rate signal in the abdominal ECG is affected by maternal ECG and noise. Therefore, in order to obtain fetal health information, it is necessary to first extract the fetal heart rate from the abdominal ECG.

[0005] Blind source separation is a commonly used algorithm for fetal electrocardiogram (FCG) extraction. This algorithm takes a multi-channel abdominal ECG as input and decomposes it into maternal ECG, fetal ECG, and noise components, with each component occupying one channel of the multi-channel output. However, the channel position, amplitude, and positive / negative information of each component in the multi-channel output of the blind source separation method are uncertain, requiring manual selection of the fetal ECG components. For example, in the machine learning-based non-stimulation fetal heart rate monitoring classification method disclosed in the prior art (CN112971799A), only after selecting one of the output signals from the blind source separation method as the separated fetal ECG signal can the fetal heart rate monitoring type be classified and the fetal NST type be obtained.

[0006] Therefore, this invention designs a fetal heart rate identification method based on full-channel features to solve the above problems. Summary of the Invention

[0007] The main objective of this invention is to provide a fetal heart rate identification method based on full-channel features, which can select fetal heart rate signals from the multi-channel output of the blind source separation method.

[0008] The objective of this invention can be achieved by adopting the following technical solution:

[0009] A fetal cardiac electrocardiogram identification method based on full-channel features includes the following steps:

[0010] Step S1: Signal preprocessing normalizes the multi-channel ECG data;

[0011] Step S2: Fetal ECG identification uses a cross-domain consistency convolutional neural network to identify the fetal ECG.

[0012] Step S3: Output the fetal heart rate from the original signal based on the recognition results.

[0013] Preferably, in step S1, the multi-channel output signal of the blind source separation method is segmented into a-second intervals. When the sampling frequency is kHz, the sample length is m = a*k; when the number of channels is n, the sample size is m*n. The ECG data of each channel is normalized separately.

[0014] Preferably, in step S2, the target fetal ECG channel location is identified by classification, and this function is implemented by a cross-domain consistent convolutional neural network.

[0015] Preferably, the network consists of p convolutional layers and q fully connected layers, with each convolutional layer followed by a nonlinearization using, but not limited to, the LeakyReLU function.

[0016] Preferably, the input of the network is ECG data from n channels. The ECG data that best matches the characteristics of fetal heart rate needs to be selected. The output of the network is n bits, using one-hot encoding. Each bit corresponds to one channel. The output of the position where the fetal heart rate is located is 1, and the output of other bits is 0.

[0017] Preferably, the total loss function L of the network total It consists of two parts: the basic loss L bas And cross-domain consistency loss L cdc ;

[0018] Total loss function L total The expression is:

[0019] L total =L bas +α*L cdc

[0020] In the formula, α is the weighting coefficient of cross-domain consistency loss.

[0021] Preferred, basic loss L bas Choose a classification loss function, utilizing, but not limited to, the cross-entropy function:

[0022] L bas = f(Y, Y′);

[0023] In the formula, Y is the actual output label of the network, Y' is the target label, and f(.) is the selected classification loss function;

[0024] Basic loss L bas It is supervised loss, which measures the distance between the network's actual output label and the target label.

[0025] Preferably, the cross-domain consistency loss Lcdc Its expression is:

[0026] L cdc =||YY * ||1;

[0027] In the formula Y * The output of the network for the transformed samples is obtained by adding random noise to the original samples;

[0028] Cross-domain consistency loss L cdc It is an unsupervised loss that measures the difference between the original sample output and the transformed sample output of the network. The smaller the distance, the better the network performance.

[0029] Preferably, the ECG data features of the n channels output by the blind source separation method are simultaneously used as the input of the network.

[0030] Preferably, in step S3, based on the classification results, the ECG data of the corresponding channel in the original data is used as the target fetal ECG, thereby realizing the automatic identification of the fetal ECG.

[0031] Beneficial technical effects of the present invention:

[0032] This invention provides a fetal heart rate identification method based on full-channel features. This method can select fetal heart rate signals from the multi-channel output of the blind source separation method. The method takes the multi-channel output of the blind source separation method as input and automatically identifies the location of the fetal heart rate components in the channel through classification. The selected fetal heart rate signals provide support for subsequent tasks such as fetal heart rate monitoring type classification. Attached Figure Description

[0033] Figure 1 This is a flowchart of the steps of a fetal heart rate identification method based on full-channel features according to the present invention;

[0034] Figure 2 This is an example diagram of the multi-channel electrocardiogram data features to be identified in this invention;

[0035] Figure 3 This is a diagram of the network model structure used in this invention;

[0036] Figure 4 This is a schematic diagram of the loss function. Detailed Implementation

[0037] To enable those skilled in the art to understand the technical solution of the present invention more clearly, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0038] like Figures 1-4As shown in the figure, this embodiment provides a fetal heart rate identification method based on full-channel features, which includes the following steps:

[0039] Step S1: Signal preprocessing normalizes the multi-channel ECG data;

[0040] Step S2: Fetal ECG identification uses a cross-domain consistency convolutional neural network to identify the fetal ECG.

[0041] Step S3: Output the fetal heart rate from the original signal based on the recognition results.

[0042] In this embodiment, in step S1, the multi-channel output signal of the blind source separation method is segmented into a seconds. When the sampling frequency is k Hz, the sample length m = a*k. When the number of channels is n, the sample size is m*n.

[0043] The ECG data for each channel were normalized separately.

[0044] In this embodiment, in step S2, the target fetal electrocardiogram channel location is identified by classification. This function is implemented by a cross-domain consistent convolutional neural network.

[0045] In this embodiment, the network consists of p convolutional layers and q fully connected layers, with each convolutional layer being nonlinearized using, but not limited to, the LeakyReLU function.

[0046] In this embodiment, the input of the network is ECG data from n channels. The ECG data that best matches the characteristics of fetal heart rate needs to be selected. The output of the network is n bits, using one-hot encoding. Each bit corresponds to one channel. The output of the position where the fetal heart rate is located is 1, and the output of other bits is 0.

[0047] In this embodiment, the total loss function L of the network total It consists of two parts: the basic loss L bas And cross-domain consistency loss L cdc ;

[0048] The expression for the total loss function Ltotal is:

[0049] L total =L bas +α*L cdc

[0050] In the formula, α is the weighting coefficient of cross-domain consistency loss.

[0051] In this embodiment, the basic loss L bas Choose a classification loss function, utilizing, but not limited to, the cross-entropy function:

[0052] Lbas = f(Y, Y′);

[0053] In the formula, Y is the actual output label of the network, Y' is the target label, and f(.) is the selected classification loss function;

[0054] Basic loss L bas It is supervised loss, which measures the distance between the network's actual output label and the target label.

[0055] In this embodiment, the cross-domain consistency loss L cdc Its expression is:

[0056] L cdc =||YY * ||1;

[0057] In the formula Y * The output of the network for the transformed samples is obtained by adding random noise to the original samples;

[0058] Cross-domain consistency loss L cdc It is an unsupervised loss that measures the difference between the original sample output and the transformed sample output of the network. The smaller the distance, the better the network performance.

[0059] In this embodiment, the ECG data features of n channels output by the blind source separation method are simultaneously used as the input of the network.

[0060] In this embodiment, in step S3, based on the classification results, the ECG data of the corresponding channel in the original data is used as the target fetal ECG, thereby realizing the automatic selection of fetal ECG.

[0061] like Figure 2 As shown, the multi-channel output signal of the blind source separation method is segmented into 2-second intervals. When the sampling frequency is 250Hz, the sample length m = 2 * 250. When the number of channels is 4, the sample size is 500 * 4. The ECG data of each channel is normalized to [-1, 1] separately.

[0062] like Figure 3 As shown, the network consists of 5 convolutional layers and 2 fully connected layers, with each convolutional layer followed by a LeakyReLU function for nonlinearization.

[0063] Specifically, such as Figure 4 As shown, for 4-channel ECG data, the label is 4 bits, using one-hot encoding. Each bit corresponds to one channel, with the position of the fetal heart rate being output as 1 and other bits being output as 0.

[0064] Specifically, such as Figure 4 As shown, the total loss function L of this network totalIt consists of two parts: the basic loss L bas And cross-domain consistency loss L cdc Total loss function L total The expression is

[0065] L total =L bas +α*L cdc

[0066] In the formula, α is the weighting coefficient of cross-domain consistency loss, which is set to 0.001 in the example.

[0067] Specifically, based on the classification results of the network, the ECG data of the corresponding channel in the original data is used as the target fetal ECG, thereby realizing the automatic selection of fetal ECG.

[0068] The above are merely further embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A fetal cardiac electrocardiogram identification method based on full-channel features, characterized in that: Includes the following steps: Step S1: Signal preprocessing, normalizing the multi-channel ECG data; Step S2: Fetal ECG identification, using a cross-domain consistency convolutional neural network to identify the fetal ECG; Step S3: Output the results and select the fetal heart rate from the original signal based on the recognition results; In step S1, the multi-channel output signal of the blind source separation method is segmented into a seconds. When the sampling frequency is k Hz, the sample length m = a*k. When the number of channels is n, the sample size is m*n. Then, the ECG data of each channel is normalized separately. In step S2, the target fetal ECG channel location is identified by classification. This function is implemented by a cross-domain consistency convolutional neural network. The input to this network is ECG data from n channels. The network needs to select the ECG data that best matches the characteristics of fetal heart rate. The output of this network is n bits, using one-hot encoding. Each bit corresponds to one channel. The output of the position where the fetal heart rate is located is 1, and the output of other bits is 0. The total loss function Ltotal of the network consists of two parts: the basic loss Lbas and the cross-domain consistency loss Lcdc. The expression for the total loss function Ltotal is: L total = L bas + α*L cdc In the formula, α is the weighting coefficient of cross-domain consistency loss; The cross-domain consistency loss Lcdc is expressed as follows: L cdc = ||YY*||1; In the formula, Y* represents the network output of the transformed sample, which is obtained by adding random noise to the original sample; Cross-domain consistency loss (LCDC) is an unsupervised loss that measures the difference between the original sample output and the transformed sample output of the network. The smaller the distance, the better the network performance. The ECG data features from the n channels output by the blind source separation method are simultaneously used as the input to the network; In step S3, based on the classification results, the ECG data of the corresponding channel in the original data is used as the target fetal ECG, thereby realizing the automatic selection of fetal ECG.

2. The fetal cardiac electrocardiogram identification method based on full-channel features according to claim 1, characterized in that: The network consists of p convolutional layers and q fully connected layers, with each convolutional layer followed by a nonlinearization function, but not limited to the LeakyReLU function.

3. The fetal cardiac electrocardiogram identification method based on full-channel features according to claim 2, characterized in that: The basic loss is Lbas, which is a classification loss function, utilizing, but not limited to, the cross-entropy function. L bas = f(Y, Y′); In the formula, Y is the actual output label of the network, Y' is the target label, and f(.) is the selected classification loss function; The basic loss Lbas is a supervised loss that measures the distance between the network's actual output label and the target label.

Images