Fetal electrocardiosignal extraction system based on adversarial deep learning network

Abstract

The invention discloses a fetal electrocardiosignal extraction system based on an adversarial deep learning network, and the method comprises the steps: carrying out the preprocessing and noise reduction of an original abdominal electrical signal AECG of a puerpera, taking the preprocessed signal as the input of a binary classification signal quality evaluation network, screening out a high-quality abdominal electrical signal, and carrying out the extraction of the high-quality abdominal electrical signal; then, the screened AECG and newborn ECG data corresponding to the puerpera serve as input, a cyclic generative adversarial network is used, mapping of the AECG to the ECG, mapping of the ECG to the AECG and a discriminator after network learning mapping are used for judgment, reconstructed FECG is output, post-processing noise reduction is conducted based on the generative adversarial network, and finally pure and accurate FECG waveforms are output. By using the multi-scale convolution kernel, the maternal abdominal electric signal and the fetal electrocardiosignal are subjected to adversarial learning, the fetal electrocardiosignal is accurately recovered from the maternal abdominal electric signal, and noise is reduced. The integrated fusion model provided by the invention solves the problem of low accuracy of FECG signal reconstruction at present, and can be widely applied to the fields of biomedical signal processing and computer detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of biomedical signal processing, and in particular relates to a fetal electrocardiogram signal extraction system based on an adversarial deep learning network. Background Art

[0002] Non-invasive fetal electrocardiography (NI-FECG) is a very promising diagnostic method that can detect various fetal heart conditions early in pregnancy and effectively identify arrhythmias, atrial septal defects, aortic valve stenosis, and multiple genetic diseases. It does this by analyzing basic ECG parameters such as fetal heart rate (FHR), heart rate variability (HRV), and ECG morphology information (including PR, ST, and QT intervals). With these measurements, NI-FECG becomes an important tool for diagnosing and monitoring fetal health while being minimally invasive, etc.

[0003] NI-FECG records AECG through surface electrodes placed on the abdomen of pregnant women and extracts FECG from it. However, AECG usually contains noise from various sources, such as baseline drift, motion artifacts, power line noise, uterine and muscle contractions, loose electrode connections, and white noise. Therefore, it is very difficult to extract FECG even if there is only slight distortion in AECG. In the past, the two types of signal extraction methods, model-based and learning-based, mostly focused on the FQRS wave and ignored the FECG morphology, resulting in clinical application effects comparable to those of traditional fetal heart monitoring CTG, which limits the application of artificial intelligence in clinical electronic fetal heart monitoring. Summary of the invention

[0004] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and propose a fetal ECG signal extraction system based on an adversarial deep learning network to solve the problem of ignoring the FECG signal morphology extraction due to excessive noise interference. The system is suitable for the fetal ECG signal extraction process in the abdomen of pregnant women with complex noise, and on this basis, realizes non-invasive automatic extraction of fetal ECG signals based on the integration of advanced artificial intelligence algorithms.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

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

[0007] S1, collect one-dimensional abdominal electrical signals of laboring and delivering women in the hospital, as well as electrocardiogram signals of newborns;

[0008] S2, preprocessing and denoising the original one-dimensional abdominal electrical signals, and using the preprocessed signals as inputs of a binary classification signal quality assessment network to screen out high-quality abdominal electrical signals;

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

[0010] S4. Take the reconstructed FECG signal as input and use the adversarial deep learning network to build a FECG post-processing denoising model.

[0011] The system of the present invention adopts the following technical solutions to achieve: fetal electrocardiogram signal extraction based on adversarial deep learning network, including:

[0012] The data input unit 21 is used to obtain the transabdominal electrical signal data of the laboring women and the delivery women, as well as the electrocardiogram signal data of the newborns born by the delivery women. 37 weeks+ is the screening standard for full-term labor, and the full cervical dilation is the screening standard for delivery. The transabdominal electrical signal of the parturient is used as the input variable of the data quality assessment model;

[0013] The data quality assessment model unit 22 is used to construct a transabdominal electrical signal quality assessment 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 including a CNN and a fully connected network is used, the CNN extracts the spatial features of the signal to improve the diversity of the spatial characteristics, and the fully connected network classifies and identifies the extracted features, thereby realizing the binary classification quality assessment of the abdominal electrical signal;

[0014] The FECG reconstruction model unit 23 is used to construct a FECG reconstruction model based on a deep learning method; the input is the four-channel AECG data of the pregnant woman's abdomen, and the output is the morphological features of the FECG signal; based on the generative adversarial network GAN, two generators are used to learn two mapping pairs, and two adversarial discriminators calculate the similarity between the ground truth value and the mapping generation 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 to construct a FECG post-processing denoising model based on a deep learning method; the input is a reconstructed noisy FECG signal, and the output is a clean FECG signal with a high signal-to-noise ratio (SNR); the model includes three networks: an encoder, a decoder, and a discriminator. The encoder network compresses the input noisy fetal ECG signal into a low-dimensional feature vector to retain the useful features of the signal. The decoder network takes the low-dimensional feature vector as input and learns to generate a high-dimensional vector that is as close to the pure ECG signal as possible. The discriminator helps the decoder generate a pure signal by distinguishing the generated signal from the pure signal.

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

[0017] Compared with the prior art methods, the present invention has the following advantages:

[0018] The present invention applies deep learning methods to the abdominal electrical signals of full-term and postpartum pregnant women and the ECG signals of newborns collected in the hospital, and successively carries out quality assessment, reconstruction, and post-processing noise reduction, thereby comprehensively improving the accuracy of neonatal ECG signal extraction and morphological waveforms, making it easier for medical staff to promptly detect fetal hypoxia, acidosis, congenital heart genetic diseases, etc., and has certain significance and value in assisting clinical treatment decisions and promptly detecting dangerous fetal conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

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

[0021] Figure 3 It is an input and output flow chart of the system of the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0023] Example

[0024] like Figure 1 As shown, this embodiment is based on the fetal electrocardiogram extraction method of the pregnant woman's abdominal electrical signal. By establishing a database of full-term pregnant women and parturient women's abdominal electrical signals and corresponding newborn fetal electrocardiograms, all pregnant women in the hospital are included, and fetal heart monitoring is performed for at least 20 minutes, and the neonatal electrocardiogram is collected for 10 minutes. The fetal electrocardiogram extraction method based on the pregnant woman's abdominal electrical signal and the neonatal electrocardiogram signal mainly includes the following steps:

[0025] S1, collect one-dimensional abdominal electrical signals of laboring and delivering women in the hospital, as well as electrocardiogram signals of newborns;

[0026] S2, preprocessing and denoising the original one-dimensional abdominal electrical signals, and using the preprocessed signals as inputs of a binary classification signal quality assessment network to screen out high-quality abdominal electrical signals;

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

[0028] S4. Take the reconstructed FECG signal as input and use the adversarial deep learning network to build a FECG post-processing denoising model.

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

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

[0031] In this embodiment, after the data is collected, the data contains a lot 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, it is necessary to screen out high-quality signals as the input of the model.

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

[0033] In this example, the signal quality assessment process in step S2 includes segmenting the signal using a sliding window method to meet the needs of training and real-time processing of the quality assessment network model. According to expert annotation, sequences with poor signal quality are labeled as 0, and sequences with good quality are labeled as 1. These sequences are input into a quality assessment model composed of CNN and a fully connected network. The model inputs single-channel AECG data and outputs AECG signal quality labels. The target loss function is cross entropy, and the training set and test set are divided into 7:3. This CNN-based signal quality assessment model can screen out signals with good quality and discard signals with relatively low quality, while eliminating useless signal processing processes. The algorithm has few parameters and can be deployed on embedded platforms to improve hardware performance. As a general signal quality assessment method, it increases the scope of model application.

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

[0035] The algorithm includes a FECG extraction model that maps the maternal transabdominal electrical signal and the neonatal ECG signal, and a post-processing denoising model that maps the reconstructed noisy FECG signal and the corresponding neonatal ECG signal. The integrated algorithm model is based on the generative adversarial GAN ​​network.

[0036] In this example, the FECG extraction model in step S3 uses a cyclic generative adversarial network, using two generators, one mapping the maternal abdominal electrical signal to the fetal ECG signal, and the other mapping the fetal ECG signal to the maternal abdominal electrical signal. Each generator corresponds to a discriminator, and the identification results of the two discriminators are incorporated into the composition loss function. In order to make the reconstructed FECG signal morphologically close to the actual FECG signal, it is necessary to consider maintaining the morphology, position and spectral components of the QRS complex wave of the FECG signal. Therefore, three loss components are added on the basis of the discriminator: spectrum loss, time loss, and power loss, so that the morphological features are retained. The model gradient degradation problem will cause the accuracy of the generated FECG signal to decrease, so the Resnet residual network is added in the middle link of encoding and decoding to alleviate the gradient degradation problem. Due to the complexity of the model, using a typical discriminator will lead to a longer calculation time, so the PATCHGAN discriminator is used to divide the original signal into 3*3 blocks, which are calculated by threads, and each block is convolved and then discriminated and output to speed up the calculation efficiency.

[0037] In this example, the FECG signal reconstructed in step S3 often contains noise. After post-processing and denoising in step S4, the FECG waveform is clear and has obvious morphological features. The model is based on a deep learning method and trains a neural network in a supervised manner. The network uses noisy and clean fetal ECG signal pairing data 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 convolutional layers for encoding and 9 contracted convolutional layers for decoding. The discriminator consists of 4 rectified linear units to distinguish the fetal ECG signal generated by the generator from the real ECG signal. Among them, due to the gradient vanishing problem in the actual training process, the generated pure FECG signal is not obviously different from the input signal. The model connects every other convolutional layer in the encoder to the transposed convolutional layer in the decoder, that is, using a residual network structure to solve this gradient vanishing problem.

[0038] like Figure 3 As shown, the FECG extraction input and output flow chart based on the adversarial deep learning method proposed in the present invention is:

[0039] The data input unit 21 is used to obtain the transabdominal electrical signal data of the laboring women and the delivery women, as well as the electrocardiogram signal data of the newborns born by the delivery women, and use the transabdominal electrical signal of the delivery women as the input variable of the data quality assessment model;

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

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

[0042] The FECG post-processing denoising model unit 24 is used to construct a 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 a high signal-to-noise ratio 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 electrocardiogram FECG signal morphology graph.

[0044] The above description of the disclosed examples enables professionals and technicians in the field to use this invention. It is obvious to professionals and technicians in the field that multiple modifications to these embodiments are obvious, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features disclosed herein. Those skilled in the art can also make various equivalent deformations or substitutions without violating the spirit of the present invention, and these equivalent deformations or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A fetal electrocardiogram signal extraction system based on an adversarial deep learning network, comprising the following steps: S1, collect one-dimensional abdominal electrical signals of laboring and delivering women in the hospital, as well as electrocardiogram signals of newborns; S2, preprocessing and denoising the original one-dimensional abdominal electrical signals, and using the preprocessed signals as inputs of a binary classification signal quality assessment network to screen out high-quality abdominal electrical signals; S3. Use the screened high-quality abdominal electrical signals as input, use the adversarial deep learning network to establish a FECG reconstruction model; S4. Use the reconstructed FECG signals as input, use the adversarial deep learning network to build a FECG post-processing denoising model.

2. According to claim 1, the fetal ECG signal extraction system based on the adversarial deep learning network is technically characterized in that, in step S1, the transabdominal electrical signals of all selected pregnant women are collected, and the signal quality is marked by experts, and the ECG signals of the corresponding newborns of the parturients are collected, and the R peak positions are marked by experts. It is assumed that the human physiological signals are highly robust and do not change significantly over a period of time, and that the difference between the ECG signals of the fetus inside and outside the uterus is small, and the ECG signals of the newborns can be used as the gold standard for the abdominal electrical signals of the parturients.

3. According to claim 1, the fetal ECG signal extraction system based on the adversarial deep learning network is technically characterized in that, in step S2, signal preprocessing and noise reduction are performed, and the AECG signal 20 minutes before delivery is used, Butterworth filtering and Z-score normalization are performed on it, and then a model including CNN and a fully connected network is used, CNN extracts the spatial features of the signal, 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. Due to the small number of model parameters, it can be deployed on an embedded platform to improve hardware performance. As a general signal quality evaluation method, it increases the application scope of the evaluation model.

4. According to claim 1, the fetal ECG signal extraction system based on the adversarial deep learning network is technically characterized in that the signal reconstruction in step S3 is integrated with the signal post-processing denoising algorithm in step S4, and FECG extraction is performed based on the generative adversarial network by mapping the abdominal electrical signals of parturient women and the ECG signals of newborns, and post-processing denoising is performed by mapping the reconstructed noisy FECG signals and the corresponding ECG signals of newborns, and a loss function is defined based on the discriminator result, and three loss components are added on the basis of the discriminator: spectral loss, time loss, and power loss, so that the FECG morphological features are retained.

5. The fetal electrocardiogram signal extraction system based on adversarial deep learning according to any one of claims 1 to 4, wherein the technical feature is that: A data input unit, used to obtain transabdominal electrical signal data of laboring women and delivery women, as well as electrocardiogram signal data of newborns born by delivery women, and use the transabdominal electrical signal of the delivery women as an input variable of the data quality assessment model; The data quality assessment model unit is used to construct a transabdominal electrical signal quality assessment model based on a deep learning method; wherein the input is a single-channel transabdominal electrical signal and the output is a signal quality binary classification label.

6. According to the fetal ECG signal extraction system based on the adversarial deep learning network described in claim 5, its technical feature is that it also includes a FECG reconstruction model unit, which is used to construct a FECG reconstruction model based on a deep learning method; the input is the four-channel abdominal electrocardiogram AECG data of the pregnant woman, and the output is the morphological features of the FECG signal; the FECG post-processing denoising model unit is used to construct a FECG post-processing denoising model based on a deep learning method; the input is the reconstructed FECG signal containing noise, and the output is a clean FECG signal with a high signal-to-noise ratio (SNR); the extraction result output unit is used to import the model input into the system for calculation to obtain the final pure fetal ECG FECG signal morphological graph.

Images