A fetal electrocardiogram signal extraction system based on an adversarial deep learning network

Abstract

The application discloses a kind of fetal electrocardiosignal extraction systems based on adversarial deep learning network, method includes: to the original abdominal electrocardiosignal AECG of puerpera is preprocessed noise reduction, the signal after preprocessing is as the input of binary classification signal quality evaluation network, and high-quality abdominal electrocardiosignal is screened out, subsequently screened AECG and the ECG data of corresponding newborn of puerpera are as input, using cyclic generative adversarial network, AECG is mapped to ECG, ECG is mapped to AECG, network learns to judge after discriminator, output reconstructed FECG, subsequently based on generative adversarial network is post-processed noise reduction, finally output pure, accurate FECG waveform.The application is accurately recovered fetal electrocardiosignal from maternal abdominal electrocardiosignal and denoised by using multi-scale convolution kernel, learning the abdominal electrocardiosignal of puerpera and fetal electrocardiosignal.The integrated fusion model proposed in the application solves the problem of low accuracy of reconstructed FECG signal, and can be widely applied to the field of biomedical signal processing and computer detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of biomedical signal processing, and particularly relates to a fetal electrocardiogram signal extraction system based on an adversarial deep learning network. BACKGROUND

[0002] Non-invasive fetal electrocardiogram (NI-FECG) is a very promising diagnostic method, which can detect various fetal heart conditions in early pregnancy and effectively identify arrhythmia, atrial septal defect, aortic valve stenosis and various genetic diseases. It achieves this purpose by analyzing basic electrocardiogram parameters such as fetal heart rate (FHR), heart rate variability (HRV) and electrocardiogram morphology information (including PR, ST and QT interval). Through these measurements, NI-FECG becomes an important tool for diagnosing and monitoring fetal health with minimal invasiveness, etc.

[0003] NI-FECG records AECG through surface electrodes placed on the abdomen of a pregnant woman, from which FECG is extracted, however, AECG usually contains various noises such as baseline drift, motion artifacts, power line noise, uterine and muscle contractions, loose electrode connection and white noise. Therefore, it is very difficult to extract FECG even if there is only slight distortion in AECG. The two types of signal extraction methods based on models and learning in the past mostly focus on FQRS waves, while ignoring FECG morphology, resulting in clinical application effects comparable to traditional fetal heart monitoring CTG, which limits the application of artificial intelligence in clinical electronic fetal heart monitoring. SUMMARY

[0004] The main purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a fetal electrocardiogram signal extraction system based on an adversarial deep learning network, which solves the problem of ignoring FECG signal morphology extraction due to excessive noise interference. The system is suitable for fetal electrocardiogram signal extraction process containing complex noise, and realizes automatic extraction of non-invasive fetal electrocardiogram signal based on advanced artificial intelligence algorithm integration on this basis.

[0005] In order to achieve the above purpose, the following technical solutions are adopted in the present application:

[0006] A fetal electrocardiogram signal extraction system based on an adversarial deep learning network, comprising the following steps:

[0007] S1, collecting one-dimensional abdominal electrical signals of hospital parturient and delivery women, and neonatal electrocardiogram signals;

[0008] S2, pre-processing and denoising the original one-dimensional abdominal electrical signals, and taking the pre-processed signals as the input of a binary classification signal quality evaluation network to screen out high-quality abdominal electrical signals;

[0009] S3, using the screened high-quality abdominal electrical signals as input, using the adversarial deep learning network, establishing the FECG reconstruction model;

[0010] S4, using the reconstructed FECG signal as input, using the adversarial deep learning network, constructing the FECG post-processing denoising model.

[0011] The system of the present application adopts the following technical scheme to realize it: fetal electrocardiogram signal extraction based on adversarial deep learning network, comprising:

[0012] The data input unit 21 is used for acquiring transabdominal electrical signal data of a parturient woman, a delivery woman, and electrocardiogram signal data of a newborn baby born by the delivery woman, 37 weeks+ is the full-term parturition screening standard, and full dilatation of the cervix is the delivery screening standard, and the transabdominal electrical signal of the woman is used as an input variable of a data quality evaluation model;

[0013] The data quality evaluation model unit 22 is used for constructing a transabdominal electrical signal quality evaluation model based on a deep learning method; the input is a single-channel transabdominal electrical signal, and the output is a signal quality binary classification label; a model containing a CNN and a fully connected network is used, the CNN extracts spatial features of the signal, and the diversity of the spatial characteristics is improved, and the fully connected network classifies and identifies the extracted features, so as to realize the binary classification quality evaluation of the abdominal electrical signal;

[0014] The FECG reconstruction model unit 23 is used for constructing an FECG reconstruction model based on a deep learning method; the input is four-channel abdominal electrical AECG data of a pregnant woman, and the output is an FECG signal morphological feature; based on a generative adversarial network GAN, two generators are used to learn two mapping pairs, two adversarial discriminators calculate the similarity between the basic true value and the mapping generated output, and the discriminator is connected to the associated generator to complete the cycle;

[0015] The FECG post-processing denoising model unit 24 is used for constructing an FECG post-processing denoising model based on a deep learning method; the input is a reconstructed FECG signal containing noise, and the output is a clean FECG signal with high signal-to-noise ratio SNR; the model includes three networks of an encoder, a decoder and a discriminator, the encoder network compresses the input noise fetal electrocardiogram signal into a low-dimensional feature vector, retaining the useful features of the signal, the decoder network takes the low-dimensional feature vector as input, learns to generate a high-dimensional vector as close as possible to the pure electrocardiogram signal, and the discriminator distinguishes between the generated signal and the pure signal, helping the decoder to generate the pure signal.

[0016] The extraction result output unit 25 is used for guiding the model input into the system for calculation to obtain the final pure fetal electrocardiogram FECG signal morphological graph.

[0017] Compared with the prior art method, the present application has the following advantages:

[0018] The present application collects the abdominal electrical signals of full-term and delivery pregnant women and the ECG signals of newborns in hospitals, and uses a deep learning method to successively carry out quality evaluation, reconstruction, post-processing noise reduction, comprehensively improves the extraction accuracy of the ECG signals of newborns and the morphological waveform, facilitates medical staff to early and timely find fetal hypoxia, acidosis, congenital heart genetic diseases and the like, and has certain significance and value for assisting clinical treatment decision and timely and early finding of fetal dangerous conditions. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a method flowchart of the present application;

[0020] Figure 2 is a deep learning method schematic diagram;

[0021] Figure 3 is an input and output flowchart of the system of the present application. DETAILED DESCRIPTION

[0022] The present application will be further described in detail below in combination with embodiments and drawings, but the embodiments of the present application are not limited thereto.

[0023] EMBODIMENT

[0024] As shown in the drawings, Figure 1 The present embodiment is based on a fetal ECG extraction method of abdominal electrical signals of pregnant women, a database of abdominal electrical signals of full-term pregnant women and delivery women and corresponding fetal ECG of newborns is established, all pregnant women in the hospital are included, fetal heart monitoring is carried out for at least more than 20 minutes, and the ECG of newborns is collected for 10 minutes. The fetal ECG extraction method based on abdominal electrical signals of pregnant women and ECG signals of newborns mainly includes the following steps:

[0025] S1, collecting one-dimensional abdominal electrical signals of labor and delivery women in the hospital and ECG signals of newborns;

[0026] S2, pre-processing and denoising the original one-dimensional abdominal electrical signals, and taking the pre-processed signals as the input of a binary classification signal quality evaluation network to screen out high-quality abdominal electrical signals;

[0027] S3, taking the screened high-quality abdominal electrical signals as the input, using an adversarial deep learning network to establish an FECG reconstruction model;

[0028] S4, taking the reconstructed FECG signals as the input, using an adversarial deep learning network to construct an FECG post-processing noise reduction model.

[0029] In the embodiment, the data collected in step S1 includes: abdominal electrical signal data of pregnant women, neonatal electrocardiogram data, electronic medical record system and delivery outcome data; wherein the abdominal electrical signal data of pregnant women is in the form of four-channel one-dimensional body surface signal, and is labeled by experts as two categories of signal quality; the neonatal electrocardiogram data is in the form of single-channel one-dimensional body surface signal, and the R-peak position is labeled by experts; the electronic medical record system and the delivery outcome data specifically include the demographic characteristics, medical history and physical examination results, laboratory examination results, fetal umbilical cord blood information, delivery outcome and other available patient characteristics of pregnant women, so as to guide the subsequent quality evaluation and signal screening of abdominal electrical signal data.

[0030] In the embodiment, the specific process of collecting data in step S1 is: collecting original electrical signal data and clinical information in the hospital according to the inclusion and exclusion criteria of pregnant women; wherein the inclusion criteria include: including full-term pregnant women and women with full cervical dilatation in the hospital, and continuously monitoring the abdominal electrical signal for 20 minutes or more during the delivery process; the exclusion criteria include: discontinuous collection time, original data missing rate greater than 50%, or less than 20 minutes of continuous monitoring time for emergency admission;

[0031] In the embodiment, after the data collection is completed, the data contains a large amount of noise and is difficult to directly input into the neural network algorithm. Such data must be preprocessed, and in order to make the neural network training effect outstanding, high-quality signals need to be selected as the input of the model.

[0032] In the example, the pre-processing process in step S2 includes using a 3-90Hz 5th order Butterworth band-pass filter to eliminate baseline drift in AECG, and using a 5th order 50Hz Butterworth notch filter to filter out power frequency interference, and using Z-score method to standardize the filtered AECG signal.

[0033] In the example, the signal quality evaluation process in step S2 includes using a sliding window method to segment the signal to adapt to the training and real-time processing needs of the quality evaluation network model, labeling the signal quality poor sequence as 0 and the quality good sequence as 1 according to expert labeling, inputting these sequences into a quality evaluation model composed of CNN and fully connected network, the model input single-channel AECG data, and the output AECG signal quality label, the target loss function is cross entropy, the training set and the test set are divided according to 7:3, this CNN-based signal quality evaluation model can filter out high-quality signals and discard low-quality signals, while saving unnecessary signal processing procedures, and the algorithm has fewer parameters and can be deployed on an embedded platform to improve hardware performance. As a general signal quality evaluation method, it increases the application range of the model.

[0034] In this example, the deep learning algorithm applied in step S3 reconstruction and step S4 post-processing noise reduction is as shown in Figure 2

[0035] The algorithm includes an FECG extraction model mapped with the abdominal electrical signal of the parturient and the neonatal ECG signal, and a post-processing noise reduction model mapped with the reconstructed FECG signal containing noise and the corresponding neonatal ECG signal. The integrated algorithm model is based on a generative adversarial GAN network.

[0036] In this example, the FECG extraction model in step S3 uses a recurrent generative adversarial network, uses two generators, one maps the maternal abdominal electrical signal to the fetal ECG signal, and the other maps the fetal ECG signal to the maternal abdominal electrical signal, and each generator corresponds to a discriminator, and the discrimination results of the two discriminators are incorporated into a loss function. In order to make the reconstructed FECG signal close to the actual true FECG signal in morphology, it is necessary to consider keeping the morphology, position and spectral components of the QRS complex of the FECG signal. Therefore, three loss components are added to the discriminator: spectral loss, time loss and power loss, so that the morphological features are preserved. Model gradient degradation problem will cause the accuracy of the generated FECG signal to decrease, therefore, a Resnet residual network is added in the middle link of encoding and decoding to alleviate the gradient degradation problem. Because the model is complex, using a typical discriminator will cause the calculation time to become longer, therefore, a PATCHGAN discriminator is used, the original signal is divided into 3*3 blocks, the threads are calculated, and each block is subjected to convolution operation and then discriminates the output, to speed up the operation efficiency.

[0037] In this example, the FECG signal reconstructed in step S3 often contains noise, and the FECG waveform after step S4 post-processing noise reduction is clear and the morphological features are obvious. The model is based on a deep learning method and trains a neural network in a supervised manner, the network uses paired data of noisy and clean fetal ECG signals to denoise the reconstructed FECG. The model is based on a generative adversarial network and includes a generator and a discriminator, the generator includes 8 dilated convolution layers for encoding and 9 shrinkage convolution layers for decoding, and the discriminator is composed of 4 rectified linear units to distinguish and identify the fetal ECG signal generated by the generator and the true ECG signal. Among them, due to the gradient disappearance problem occurring in the actual training process, the generated pure FECG signal is not obviously different from the input signal, the model connects the convolution layers in the encoder every other one with the transpose convolution layers in the decoder, that is, uses a residual network structure to solve the gradient disappearance problem.

[0038] As shown in Figure 3 The input and output flowchart of the FECG extraction based on the adversarial deep learning method proposed by the application is as shown in

[0039] ​The data input unit 21 is used to obtain the abdominal electrical signal data of the pregnant woman in labor and the newborn ECG signal data of the newborn born by the pregnant woman in labor, and take the abdominal electrical signal of the pregnant woman as the input variable of the data quality evaluation model;

[0040] The data quality evaluation model unit 22 is used to construct the abdominal electrical signal quality evaluation model based on the deep learning method; wherein the input is the single-channel abdominal electrical signal, and the output is the signal quality binary classification label;

[0041] The FECG reconstruction model unit 23 is used to construct the FECG reconstruction model based on the deep learning method; the input is the four-channel abdominal electrical AECG data of the pregnant woman, and the output is the FECG signal morphological feature;

[0042] The FECG post-processing noise reduction model unit 24 is used to construct the FECG post-processing noise reduction model based on the deep learning method; the input is the reconstructed FECG signal containing noise, and the output is the clean FECG signal with high SNR;

[0043] The extraction result output unit 25 is used to input the model into the system for calculation to obtain the final pure fetal ECG FECG signal morphological graph.

[0044] The above description of the disclosed examples is for a person skilled in the art to use the present application. Many modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein. Those skilled in the art can also make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope defined by the claims herein.

Claims

1. A method for fetal electrocardiogram extraction based on an adversarial deep learning network, comprising the following steps: S1, collect one-dimensional abdominal electrical signals of hospital puerpera and neonatal electrocardiogram signals; S2, pre-process and denoise the original one-dimensional abdominal electrical signals, and use the pre-processed signals as the input of a binary classification signal quality evaluation network to screen high-quality abdominal electrical signals; S3, use the screened high-quality abdominal electrical signals as the input, and use an adversarial deep learning network to establish an FECG reconstruction model; S4, use the reconstructed FECG signals as the input, and use an adversarial deep learning network to construct an FECG post-processing denoising model, the signal reconstruction in step S3 and the signal post-processing denoising algorithm in step S4 are integrated, based on the generative adversarial network, the abdominal electrical signals of the puerpera and the neonatal electrocardiogram signals are mapped for FECG extraction, the reconstructed noisy FECG signals and the corresponding neonatal electrocardiogram signals are mapped for post-processing denoising, and the loss function is defined based on the discriminator result, three loss components, i.e. spectral loss, time loss and power loss, are added based on the discriminator, so that the morphological characteristics of the FECG are retained.

2. The fetal electrocardiogram signal extraction method based on the adversarial deep learning network according to claim 1, wherein in step S1, the abdominal electrical signals of all selected puerpera are collected, and the signal quality is labeled by an expert, and the neonatal electrocardiogram signals of the puerpera are collected, and the R-peak position is labeled by an expert, it is assumed that the human physiological signal is robust and does not change significantly in a short time, it is considered that the difference between the fetal electrocardiogram signals in the uterus and outside the uterus is small, and the neonatal electrocardiogram signals can be used as the gold standard of the fetal electrocardiogram signals in the abdominal electrical signals of the puerpera.

3. The fetal electrocardiogram signal extraction method based on the adversarial deep learning network according to claim 1, wherein in step S2, the signal is pre-processed and denoised, the AECG signals in the last 20 minutes before delivery are used, the signals are subjected to Butterworth filtering and Z-score standardization, and then a model containing a CNN and a fully connected network is used, the CNN extracts the spatial features of the signals, the fully connected network classifies and identifies the extracted features, and the abdominal electrical signal binary classification quality evaluation is realized, since the model has a small number of parameters, it can be deployed on an embedded platform to improve the hardware performance, and as a general signal quality evaluation method, the application range of the evaluation model is increased.

4. Extraction system according to any one of claims 1-3, characterized in that A data input unit is configured to acquire abdominal electrical signal data of a puerpera and electrocardiogram signal data of a newborn baby born by the puerpera, and use the abdominal electrical signal of the puerpera as an input variable of a data quality evaluation model; A data quality evaluation model unit is configured to construct an abdominal electrical signal quality evaluation model based on a deep learning method, wherein the input is a single-channel abdominal electrical signal, and the output is a binary classification label of signal quality.

5. The extraction system according to claim 4, further comprising an FECG reconstruction model unit for constructing an FECG reconstruction model based on a deep learning method; inputting high-quality maternal single-channel abdominal electrocardiogram (AECG) data, and outputting FECG signal morphological features; an FECG post-processing noise reduction model unit for constructing an FECG post-processing noise reduction model based on a deep learning method; inputting the reconstructed FECG signal containing noise, and outputting a clean FECG signal with high SNR; and an extraction result output unit for inputting the model into the system for calculation to obtain a final pure fetal electrocardiogram (FECG) signal morphological pattern.

Images