WO2022193312A1 - 基于多导联的心电信号识别方法和心电信号识别装置 - Google Patents

基于多导联的心电信号识别方法和心电信号识别装置 Download PDF

Info

Publication number
WO2022193312A1
WO2022193312A1 PCT/CN2021/081878 CN2021081878W WO2022193312A1 WO 2022193312 A1 WO2022193312 A1 WO 2022193312A1 CN 2021081878 W CN2021081878 W CN 2021081878W WO 2022193312 A1 WO2022193312 A1 WO 2022193312A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
target
neural network
feature
learning model
Prior art date
Application number
PCT/CN2021/081878
Other languages
English (en)
French (fr)
Inventor
陈雪
柳锦女
李玉德
倪先强
梁烁斌
周莉
Original Assignee
京东方科技集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东方科技集团股份有限公司 filed Critical 京东方科技集团股份有限公司
Priority to PCT/CN2021/081878 priority Critical patent/WO2022193312A1/zh
Priority to CN202180000519.8A priority patent/CN115666387A/zh
Publication of WO2022193312A1 publication Critical patent/WO2022193312A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/25—Bioelectric electrodes therefor
    • A61B5/279—Bioelectric electrodes therefor specially adapted for particular uses
    • A61B5/28—Bioelectric electrodes therefor specially adapted for particular uses for electrocardiography [ECG]

Definitions

  • the present application relates to the field of ECG signal identification, and in particular, to a multi-lead-based ECG signal identification method, an ECG signal identification device, an electronic device and a storage medium.
  • the waveforms in the ECG include P wave, Q wave, R wave, S wave and T wave, etc.
  • the flat line from the end of the QRS complex to the beginning of the T wave is called the ST segment.
  • the diagnosis of myocardial ischemia can be made by analyzing the ST segment in the electrocardiogram.
  • the ST segment abnormality is judged mainly by means of single lead or genetic engineering.
  • the above-mentioned diagnostic methods are easy to cause missed diagnosis, and the diagnostic accuracy rate is low.
  • Embodiments of the present application provide a multi-lead-based ECG signal identification method, an ECG signal identification device, an electronic device, and a storage medium.
  • the electrocardiographic signal identification method includes the following steps:
  • Feature extraction is performed on the target signals of multiple leads respectively by using multiple neural networks to obtain corresponding feature labels, and the multiple neural networks are obtained by training according to a mutual learning model.
  • the preprocessing of the ECG signals of the plurality of leads to obtain the target signal comprises:
  • the down-sampled signal is divided into the target signal of preset length.
  • the neural network includes convolutional layers and pooling layers,
  • the convolutional layer includes a sub-convolutional layer and a fully connected layer
  • the pooling layer includes a maximum pooling layer and an average pooling layer, the maximum pooling layer is used for maximum pooling of the signal feature values of the sub-convolutional layers, and the average pooling layer is used for all The signal eigenvalues of the syphon convolutional layer are averagely pooled,
  • the fully connected layer is used to fuse the signal feature values of the average pooling layer to obtain the target signal feature value.
  • the feature extraction using a plurality of neural networks to respectively perform feature extraction on the target signals of a plurality of leads to obtain corresponding feature labels includes:
  • the feature values of the target signal are classified using a Softmax function to obtain the feature labels.
  • the method for identifying ECG signals includes:
  • the ECG signal recognition effect of the mutual learning model is evaluated by using the test set.
  • the optimizing the mutual learning model using the validation set includes:
  • the loss functions are passed to each other in the plurality of neural networks to optimize the mutual learning model.
  • the target probability is obtained by the following conditional formula:
  • x i is the i-th target signal
  • M is the total number of feature labels
  • the m-th feature label of the target signal of the k-th lead in the k-th neural network is the weight of the true label of the target signal.
  • the loss function is calculated by the following conditional expression:
  • the ECG signal identification device includes:
  • a preprocessing module for preprocessing the ECG signals of multiple leads to obtain target signals
  • the feature extraction module is configured to perform feature extraction on the target signals of multiple leads respectively by using multiple neural networks to obtain corresponding feature labels, and the multiple neural networks are trained according to the mutual learning model.
  • the electronic device includes one or more processors, and the processors are configured to:
  • Feature extraction is performed on the target signals of multiple leads respectively by using multiple neural networks to obtain corresponding feature labels, and the multiple neural networks are obtained by training according to a mutual learning model.
  • the processor is used to:
  • the down-sampled signal is divided into the target signal of preset length.
  • the neural network includes convolutional layers and pooling layers,
  • the convolutional layer includes a sub-convolutional layer and a fully connected layer
  • the pooling layer includes a maximum pooling layer and an average pooling layer, the maximum pooling layer is used for maximum pooling of the signal feature values of the sub-convolutional layers, and the average pooling layer is used for all The signal eigenvalues of the syphon convolutional layer are averagely pooled,
  • the fully connected layer is used to fuse the signal feature values of the average pooling layer to obtain the target signal feature value.
  • the feature extraction using a plurality of neural networks to respectively perform feature extraction on the target signals of a plurality of leads to obtain corresponding feature labels includes:
  • the feature values of the target signal are classified using a Softmax function to obtain the feature labels.
  • the method for identifying ECG signals includes:
  • the ECG signal recognition effect of the mutual learning model is evaluated by using the test set.
  • the optimizing the mutual learning model using the validation set includes:
  • the loss functions are passed to each other in the plurality of neural networks to optimize the mutual learning model.
  • the target probability is obtained by the following conditional formula:
  • x i is the i-th target signal
  • M is the total number of feature labels
  • the m-th feature label of the target signal of the k-th lead in the k-th neural network is the weight of the true label of the target signal.
  • the loss function is calculated by the following conditional expression:
  • the non-volatile computer-readable storage medium of the embodiments of the present application stores a computer program, and when the computer program is executed by one or more processors, implements the electrocardiographic signal identification method described in any of the foregoing embodiments.
  • FIG. 1 is a schematic flowchart of a multi-lead-based ECG signal identification method according to an embodiment of the present application.
  • FIG. 2 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
  • FIG. 3 is a schematic block diagram of a multi-lead-based ECG signal identification device according to an embodiment of the present application.
  • FIG. 4 is another schematic flowchart of a multi-lead-based ECG signal identification method according to an embodiment of the present application.
  • FIG. 5 is a schematic diagram of ECG signal preprocessing according to an embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of a neural network according to an embodiment of the present application.
  • FIG. 7 is another schematic flowchart of a multi-lead-based electrocardiographic signal identification method according to an embodiment of the present application.
  • FIG. 8 is another schematic flowchart of a multi-lead-based electrocardiographic signal identification method according to an embodiment of the present application.
  • FIG. 9 is another schematic flowchart of a multi-lead-based electrocardiographic signal identification method according to an embodiment of the present application.
  • the multi-lead-based electrocardiographic signal identification method includes the following steps:
  • S20 Use multiple neural networks to perform feature extraction on the target signals of multiple leads respectively to obtain corresponding feature labels.
  • an embodiment of the present application further provides an electronic device 100 .
  • the electronic device 100 includes a processor 102 and a memory 104, and the memory 104 stores a computer program 106.
  • the computer program 106 is executed by the processor 102, the multi-lead-based electrocardiographic signal identification method according to the embodiment of the present application is implemented, that is to say,
  • the processor 102 is configured to preprocess the ECG signals of multiple leads to obtain target signals, and to perform feature extraction on the target signals of multiple leads respectively by using multiple neural networks to obtain corresponding feature labels.
  • an embodiment of the present application further provides a multi-lead-based ECG signal identification device 110 .
  • the multi-lead-based ECG signal identification method may be implemented by the ECG signal identification device 110 .
  • the ECG signal identification device 110 includes a preprocessing module 112 and a feature extraction module 114 .
  • S10 may be implemented by the preprocessing module 112
  • S20 may be implemented by the feature extraction module 114 . That is to say, the preprocessing module 112 can be used to preprocess the ECG signals of multiple leads to obtain the target signal, and the feature extraction module 114 can be used to use multiple neural networks to respectively feature the target signals of multiple leads. Extract to get the corresponding feature labels.
  • detections for diagnosing ST segment abnormalities are mostly based on single-lead or genetic engineering. Because the heart is a three-dimensional structure, there is no obvious change in the single-lead ECG signal in the occurrence of myocardial ischemia or myocardial infarction in some parts, and the diagnosis method based on single lead is easy to cause missed diagnosis. Diagnostic methods based on genetic engineering identify ST segment abnormalities by locating ST segments and extracting relevant features and inputting them into classifiers.
  • the ECG signal identification device 110 and the electronic device 100 according to the embodiments of the present application, a plurality of neural networks are trained according to the mutual learning model, and the target signals are characterized by using the plurality of neural networks Extraction to obtain the feature label corresponding to the target signal.
  • the accuracy of the feature label can be improved. For leads with a low positive rate and less obvious features, it can also improve the detection accuracy of ST segment abnormalities and achieve higher sensitivity. sex and specificity.
  • the number of leads may be 3, 6, 8, 12, etc., which is not specifically limited.
  • the pressured unipolar limb lead aVR records the voltage difference between the right arm as the positive pole and the left arm and the left leg as the negative pole
  • the pressured unipolar limb lead aVL records the left arm as the positive pole and the right arm and the left
  • the voltage difference between the positive and negative poles after compression with the leg as the negative pole, and the pressured unipolar limb lead aVF records the voltage difference between the left leg as the positive pole and the left arm and the right arm as the negative pole.
  • the positive electrodes that directly reflect the potential changes are placed in 6 specific parts of the chest, and the actual potential of each part of the heart under the electrodes can be recorded through the positions of the leads.
  • a neural network can be established for each lead, and various feature labels can be obtained through feature extraction of multiple neural networks, such as "ST segment elevation”, “ST segment depression”, “ST segment normal”, “ST segment level” “Depression”, “Horizontal elevation of ST segment”, “Upslope ST segment depression”, “ST segment downslope depression”, “Other”, etc.
  • “Other” may refer to a situation where there is too much noise in the ECG signal, which is not specifically limited.
  • Sensitivity may refer to the probability that the real situation of the ECG signal is ST segment abnormality when the feature label obtained by the neural network is "ST segment abnormality”.
  • the specificity may refer to the probability that the real state of the ECG signal is the normal ST segment when the feature label obtained by the neural network is "normal ST segment”.
  • S10 includes:
  • S11 and S12 may be implemented by the preprocessing module 112 . That is to say, the preprocessing module 112 can be used for down-sampling the ECG signal to obtain a down-sampled signal, and for dividing the down-sampled signal into target signals of preset lengths.
  • the processor 102 is configured to perform down-sampling processing on the ECG signal to obtain a down-sampled signal, and to divide the down-sampled signal into target signals of preset lengths.
  • the ECG signals of the predetermined duration of the plurality of leads are collected, and the down-sampling process is performed on the ECG signals to obtain the down-sampled signals.
  • the down-sampling process may be to reduce the predetermined frequency of sampling, so that the ECG signal with the same sampling duration contains relatively few signal points.
  • the down-sampled signal is divided into target signals of preset lengths. In this way, the amount of data processing can be reduced.
  • the predetermined frequency may be the frequency of collecting ECG signals, and the predetermined frequency may be determined according to the performance and function settings of each lead, as well as parameters such as the heartbeat cycle and physical condition of the person under test, for example, it may be 200 Hz, 256 Hz, 300 Hz , 500 Hz, etc., which are not specifically limited.
  • the predetermined duration can be the duration of collecting ECG signals, and the predetermined duration can be determined according to the performance and function settings of each lead, as well as the measured person's heartbeat cycle, physical condition and other parameters, for example, it can be 5 seconds, 8 seconds, 10 seconds, 15 seconds, etc. seconds, 18 seconds, 20 seconds, 50 seconds, etc., there is no specific limitation.
  • the preset length can be the scale for segmenting the ECG signal, that is, the length of the target signal.
  • the preset length is determined according to the performance and function settings of each lead, as well as the measured person's heartbeat cycle, physical condition and other parameters, for example Can be 300, 350, 400, etc.
  • the preset length generally needs to be greater than the predetermined frequency of sampling and less than 1.75 times the sampling frequency.
  • the target signal of preset length includes R waves, and the position of the R waves in the target signal is not limited, for example, it may be in the middle of the signal segment.
  • the sampling frequency of the ECG signal is 500 Hz and the duration is 15 seconds.
  • a down-sampling signal is obtained after down-sampling processing of the ECG signal, and the sampling frequency of the down-sampling signal is 250 Hz.
  • Set the preset length L of the target signal to 300 signal points, use the differential threshold method to identify the position of each R wave in the ECG signal, take the R wave as the midpoint of the signal segment, and intercept about 150 signal points before and after the R wave . For example, if the position of the i-th R wave is R i , the intercepted signal segment is [R i -149, R i +150].
  • the length of the signal segment may be less than the preset length, and at this time, the signal segment can be filled with zeros to make the length of the signal segment equal to the preset length.
  • the target signal is obtained by preprocessing the ECG signal, which is convenient for the neural network to perform feature extraction on the target signal to obtain the feature label, thereby improving the accuracy of the feature label.
  • the neural network includes convolutional layers and pooling layers.
  • the convolutional layer includes sub-convolutional layer and fully connected layer
  • the pooling layer includes maximum pooling layer and average pooling layer
  • the maximum pooling layer is used to obtain maximum pooling for the signal eigenvalues of the sub-convolutional layer
  • the average is used to average pool the signal eigenvalues of the subconvolutional layer
  • the fully connected layer is used to fuse the signal eigenvalues of the average pooling layer to obtain the target signal eigenvalues.
  • the neural network can use a 14-layer residual network, namely ResNet14.
  • ResNet14 consists of 13 subconvolutional layers and 1 fully connected layer FC.
  • the convolutional layer includes a one-dimensional subconvolutional layer 1D_Conv, a first convolutional block Block1, a second convolutional block Block2, a third convolutional block Block3, a fourth convolutional block Block4 and a full
  • the connection layer FC has a total of 14 layers of convolution.
  • the stride of the 14-layer convolution is 2.
  • the first convolutional block Block1 includes a first subconvolutional layer, a second subconvolutional layer and a third subconvolutional layer.
  • the second convolution block Block2 includes a fourth subconvolutional layer, a fifth subconvolutional layer, and a sixth subconvolutional layer.
  • the third convolution block Block3 includes a seventh subconvolutional layer, an eighth subconvolutional layer, and a ninth subconvolutional layer.
  • the fourth convolutional block Block4 includes a tenth subconvolutional layer, an eleventh subconvolutional layer, and a twelfth subconvolutional layer.
  • the convolution kernel of the one-dimensional subconvolution layer 1D_Conv is 1*4, including 8 convolution channels.
  • the convolution kernel of the first sub-convolutional layer is 1*1, including 8 convolution channels.
  • the convolution kernel of the second sub-convolution layer is 1*3, including 8 convolution channels.
  • the convolution kernel of the third sub-convolutional layer is 1*1, including 16 convolution channels.
  • the convolution kernel of the fourth sub-convolutional layer is 1*1, including 16 convolution channels.
  • the convolution kernel of the fifth sub-convolutional layer is 1*3, including 16 convolution channels.
  • the convolution kernel of the sixth sub-convolutional layer is 1*1, including 32 convolution channels.
  • the convolution kernel of the seventh sub-convolutional layer is 1*1, including 32 convolution channels.
  • the convolution kernel of the eighth subconvolutional layer is 1*3, including 32 convolution channels.
  • the convolution kernel of the ninth sub-convolutional layer is 1*1, including 64 convolution channels.
  • the convolution kernel of the tenth sub-convolutional layer is 1*1, including 64 convolution channels.
  • the convolution kernel of the eleventh sub-convolution layer is 1*3, including 64 convolution channels.
  • the convolution kernel of the twelfth subconvolutional layer is 1*1, including 128 convolution channels.
  • the maximum pooling layer MaxPooling has a convolution kernel of 1*3 and a stride of 2.
  • the convolution kernel of the average pooling layer MeanPooling is 1*3 and the stride is 2.
  • Input the target signal into the neural network first extract the features of the target signal by the one-dimensional convolution 1D_Conv, and use the maximum pooling layer MaxPooling to obtain the maximum pooling of the signal eigenvalues of the one-dimensional sub-convolution layer 1D_Conv, and then go through the first A convolution block Block1, a second convolution block Block2, a third convolution block Block3, and a fourth convolution block Block4 obtain the signal feature values of the twelfth sub-convolutional layer.
  • the signal eigenvalues of the twelfth subconvolutional layer are averagely pooled by the average pooling layer MeanPooling, and the signal eigenvalues of the average pooling layer are fused by the fully connected layer FC to obtain the target signal eigenvalues.
  • the neural network may also use residual networks with 10 layers, 18 layers, 24 layers, 26 layers, and 54 layers, etc., which are not specifically limited.
  • the ResNet14 used in this embodiment on the one hand, because the number of network layers is relatively small, the residual network requires relatively few calculation examples and the calculation time is relatively short, on the other hand, the number of layers of ResNet14 can also ensure Feature extraction achieves high accuracy.
  • the neural network may also be of types such as FCNN, CNN, RNN, CRNN, RCR-net, etc., which is not specifically limited.
  • S20 includes:
  • S21 may be implemented by the feature extraction module 114, that is, the feature extraction module 114 may be used to classify the target signal feature values using the Softmax function to obtain feature labels.
  • the processor 102 is further configured to use the Softmax function to classify the target signal feature values to obtain feature labels.
  • the Softmax function to classify the feature values of the target signal enables the neural network to learn the relevant information between different feature labels, and finally map to different types of feature labels, so that the final output feature labels are closer to the real labels.
  • the accuracy of the feature label can be improved, that is, the detection accuracy of ST segment abnormalities can be improved, and the misdiagnosis rate can be reduced.
  • the electrocardiographic signal identification method includes:
  • S01-S04 can be implemented by the preprocessing module 112, that is to say, the preprocessing module 112 can be used to divide the ECG signal into a training set, a verification set and a test set according to a specific ratio, which can be used to utilize
  • the training set is used to train the mutual learning model, which can be used to optimize the mutual learning model by using the validation set, and can be used to evaluate the electrocardiographic signal recognition effect of the mutual learning model by using the test set.
  • the processor 102 is further configured to divide the ECG signals into a training set, a validation set and a test set according to a specific ratio, so as to use the training set to train the mutual learning model, which can be used to use the validation set to train the mutual learning model.
  • the learning model is optimized and used to evaluate the ECG signal recognition effect of the mutual learning model using the test set.
  • the ECG signal is divided into a training set, a validation set and a test set according to a specific ratio, and the specific ratio can be determined according to parameters such as the type and structure of the mutual learning model, for example, it can be 6:2:2, or 6:2:2. 3:1, 5:3:2, 5:2.5:2.5, etc., there is no specific limitation.
  • the training set can be used to train the mutual learning model
  • the verification set can be used to optimize the mutual learning model
  • the test set can be used to evaluate the electrocardiographic signal recognition effect of the mutual learning model.
  • the mutual learning network can be trained in the way of training, verification and testing, that is to say, all the ECG signals in the training set are input into the mutual learning network to train the mutual learning model, and then the verification set is put into the mutual learning network. All ECG signals are input into the mutual learning network to optimize the mutual learning model, and then all the ECG signals in the test set are input into the mutual learning network to evaluate the ECG signal recognition effect of the mutual learning model.
  • the above process can be single-procedural, that is, only one training, one validation and one test are performed for the same batch of samples. It can also be iterative, and can be trained, verified and tested multiple times for the same batch of samples.
  • the mutual learning network can be trained in the manner of training, verification, training, verification, training, verification...testing, that is to say, the ECG signals in the training set and the verification set are divided into multiple batches and input to each other.
  • the learning model after many times of training and verification, all the ECG signals in the test set are input into the mutual learning network to evaluate the ECG signal recognition effect of the mutual learning model.
  • S03 includes:
  • S031 Calculate the target probability that the feature labels of the target signals of multiple leads in the multiple neural networks are the true labels of the target signal
  • S032 Calculate the loss function of each neural network and the peer neural network according to the target probability
  • S033 Pass the loss function to each other in multiple neural networks to optimize the mutual learning model.
  • S031-S033 can be implemented by the preprocessing module 112, that is to say, the preprocessing module 112 can be used to calculate the feature labels of the target signals of multiple leads in the multiple neural networks respectively as the target signal
  • the target probability of the ground truth label can be used to calculate the loss function of each neural network and the peer neural network according to the target probability, and to pass the loss function to each other in multiple neural networks to optimize the mutual learning model.
  • the processor 102 is further configured to calculate the target probability that the feature labels of the target signals of the multiple leads in the multiple neural networks are the true labels of the target signal, and to calculate each neural network according to the target probability.
  • the process of optimizing the mutual learning model in the validation set first calculate the target probability that the target signals of the target signals of multiple leads in the multiple neural networks are the true labels of the target signal, and calculate the target probability according to the target probability.
  • the loss function of each neural network and the peer neural network are passed to each other in multiple neural networks to optimize the mutual learning model.
  • each neural network can be used as a teacher network for other neural networks in the mutual learning model, and multiple neural networks learn from each other through the transfer loss function, without pre-defining the teacher network, which can improve the detection accuracy of ST segment anomalies, and achieve higher accuracy. High sensitivity and specificity.
  • the target probability is obtained by the following conditional expression:
  • x i is the ith target signal
  • M is the total number of feature labels
  • the m-th feature label of the target signal of the k-th lead in the k-th neural network is the weight of the true label of the target signal.
  • N is the total number of target signals x
  • the target signal x can be expressed as xi is the ith target signal.
  • the true label y corresponding to the target signal x can be expressed as where y i ⁇ ⁇ 1,2,3,...,M ⁇ .
  • Each lead corresponds to establish a neural network, that is, each lead corresponds to a corresponding output. is the output of the Softmax function
  • target probability can be understood as The exponential operation of , which can smooth As a result, it avoids the calculated probability of zero and causes the operation to have a breakpoint, and ensures the continuity of the gradient descent.
  • the total number of feature labels M 4, namely "ST-segment elevation”, “ST-segment depression”, “ST-segment normal” and “other”.
  • the probability p that the first feature label of is the true label y i of the i-th target signal xi is:
  • the calculation methods of other feature labels are analogous, and will not be repeated here.
  • the target probability Calculate the loss function of each neural network and the peer neural network, and transfer the obtained loss functions to each other in multiple neural networks. In this way, the continuity of the gradient descent of the neural network can be ensured, the detection accuracy of ST segment anomalies can be improved, and the accuracy of ST segment anomaly detection can be improved. High sensitivity and specificity.
  • the loss function is calculated by the following conditional expression:
  • y i can represent the true label corresponding to the i -th target signal xi
  • m is the predicted label of the k-th neural network ⁇ k .
  • K is the total number of leads, and K can be 3, 6, 8, 12, etc., which is not specifically limited.
  • the kth neural network ⁇ k can be learned from the posterior probability of the peer neural network ⁇ l
  • the network ⁇ k is constrained so that the kth neural network ⁇ k and the peer neural network ⁇ l learn from each other.
  • D KL can be used to measure the difference between the probability distributions predicted by the kth neural network ⁇ k and the peer neural network ⁇ l for the same event, and the parameters transferred between the kth neural network ⁇ k and the peer neural network ⁇ l can be used D KL is to represent.
  • D KL is the matching degree between the kth neural network ⁇ k and the peer neural network ⁇ l .
  • the average can be used as a regularization term to constrain the kth neural network ⁇ k .
  • the total loss is supervision loss Plus the loss mean of multiple neural networks learning from each other in the mutual learning model.
  • multiple neural networks can be optimized through the mutual learning of multiple neural networks in the mutual learning model without the need to pre-define the teacher network, thereby improving the detection accuracy of ST segment abnormalities and achieving higher sensitivity and specificity.
  • Each lead corresponds to a neural network
  • the 12 neural networks correspond to standard lead I, standard lead II, standard lead III, pressure unipolar limb lead aVR, pressure unipolar limb lead aVL, plus Unipolar limb leads aVF, chest V1, chest V2, chest V3, chest V4, chest V5, and chest V6 are pressed.
  • the neural network was evaluated using the above-mentioned multi-lead-based ECG signal identification method, and the calculated sensitivity and specificity data are shown in the table below.
  • the multi-lead-based ECG signal identification method of the embodiment of the present application enables multiple neural networks in the mutual learning model to learn from each other, which can improve the detection accuracy of ST segment abnormalities.
  • High sensitivity and specificity can be achieved even for the pressurized unipolar limb lead aVR and chest lead V1, which have low sensitivity and specificity detected by the single-lead ECG recognition model.
  • the embodiments of the present application also provide a non-volatile computer-readable storage medium storing a computer program.
  • the computer program is executed by one or more processors, the electrocardiographic signal identification method described in any one of the above embodiments is implemented.
  • each functional unit in each embodiment of the present application may be integrated into one processing module, or each unit may exist physically alone, or two or more units may be integrated into one module.
  • the above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.
  • the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, and the like.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Cardiology (AREA)
  • Physics & Mathematics (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

一种基于多导联的心电信号识别方法，包括以下步骤：(S10)对多个导联的心电信号进行预处理以得到目标信号；(S20)利用多个神经网络分别对多个导联的目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。还公开了一种心电信号识别装置、电子设备和存储介质。

Description

基于多导联的心电信号识别方法和心电信号识别装置 技术领域
本申请涉及心电信号识别领域，特别涉及一种基于多导联的心电信号识别方法、心电信号识别装置、电子设备和存储介质。
背景技术
心电图中的波形包括P波、Q波、R波、S波和T波等，其中由QRS波群结束到T波开始的平线称为ST段。通过分析心电图中的ST段可以对心肌缺血做出相关诊断。相关技术中，主要借助单导联或基因工程等方式对ST段异常进行判断。然而，上述诊断方式容易造成漏诊，诊断准确率较低。
发明内容
本申请实施方式提供了一种基于多导联的心电信号识别方法、心电信号识别装置、电子设备和存储介质。
本申请实施方式的基于多导联的心电信号识别方法，所述心电信号识别方法包括以下步骤：
对多个导联的心电信号进行预处理以得到目标信号；
利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
在某些实施方式中，所述对多个导联的心电信号进行预处理以得到目标信号包括：
对所述心电信号进行降采样处理以得到降采样信号；
将所述降采样信号切分为预设长度的所述目标信号。
在某些实施方式中，所述神经网络包括卷积层和池化层，
其中，所述卷积层包括子卷积层和全连接层，
所述池化层包括最大池化层和平均池化层，所述最大池化层用于对所述子卷积层的信号特征值求最大池化，所述平均池化层用于对所述子卷积层的信号特征值求平均池化，
所述全连接层用于对所述平均池化层的信号特征值进行融合以得到目标信号特征值。
在某些实施方式中，所述利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签包括：
利用Softmax函数对所述目标信号特征值进行分类以得到所述特征标签。
在某些实施方式中，所述心电信号识别方法包括：
将心电信号按照特定比例分为训练集、验证集和测试集；
利用所述训练集对所述相互学习模型进行训练；
利用所述验证集对所述相互学习模型进行优化；
利用所述测试集评估所述相互学习模型的心电信号识别效果。
在某些实施方式中，所述利用所述验证集对所述相互学习模型进行优化包括：
计算多个导联的所述目标信号分别在多个所述神经网络中的特征标签为所述目标信号的真实标签的目标概率；
根据所述目标概率计算每个所述神经网络与同伴神经网络的损失函数；
将所述损失函数在所述多个神经网络中相互传递以优化所述相互学习模型。
在某些实施方式中，所述目标概率通过下列条件式得到：
Figure PCTCN2021081878-appb-000001
其中，x i为第i个所述目标信号，
Figure PCTCN2021081878-appb-000002
为第i个所述目标信号在所述相互学习模型的第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的目标概率，M为特 征标签总数，
Figure PCTCN2021081878-appb-000003
为第k个导联的所述目标信号在第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的权重。
在某些实施方式中，所述损失函数通过下列条件式计算：
Figure PCTCN2021081878-appb-000004
Figure PCTCN2021081878-appb-000005
Figure PCTCN2021081878-appb-000006
其中，
Figure PCTCN2021081878-appb-000007
为第k个所述神经网络的总损失，
Figure PCTCN2021081878-appb-000008
为第k个所述神经网络的监督损失，D KL为第k个所述神经网络与所述同伴神经网络的匹配度，K为所述导联的总数。
本申请实施方式的基于多导联的心电信号识别装置，所述心电信号识别装置包括：
预处理模块，用于对多个导联的心电信号进行预处理以得到目标信号；
特征提取模块，用于利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
本申请实施方式的电子设备，所述电子设备包括一个或多个处理器，所述处理器用于：
对多个导联的心电信号进行预处理以得到目标信号；
利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
在某些实施方式中，所述处理器用于：
对所述心电信号进行降采样处理以得到降采样信号；
将所述降采样信号切分为预设长度的所述目标信号。
在某些实施方式中，所述神经网络包括卷积层和池化层，
其中，所述卷积层包括子卷积层和全连接层，
所述池化层包括最大池化层和平均池化层，所述最大池化层用于对所述子卷积层的信号特征值求最大池化，所述平均池化层用于对所述子卷积层的信号特征值求平均池化，
所述全连接层用于对所述平均池化层的信号特征值进行融合以得到目标信号特征值。
在某些实施方式中，所述利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签包括：
利用Softmax函数对所述目标信号特征值进行分类以得到所述特征标签。
在某些实施方式中，所述心电信号识别方法包括：
将心电信号按照特定比例分为训练集、验证集和测试集；
利用所述训练集对所述相互学习模型进行训练；
利用所述验证集对所述相互学习模型进行优化；
利用所述测试集评估所述相互学习模型的心电信号识别效果。
在某些实施方式中，所述利用所述验证集对所述相互学习模型进行优化包括：
计算多个导联的所述目标信号分别在多个所述神经网络中的特征标签为所述目标信号的真实标签的目标概率；
根据所述目标概率计算每个所述神经网络与同伴神经网络的损失函数；
将所述损失函数在所述多个神经网络中相互传递以优化所述相互学习模型。
在某些实施方式中，所述目标概率通过下列条件式得到：
Figure PCTCN2021081878-appb-000009
其中，x i为第i个所述目标信号，
Figure PCTCN2021081878-appb-000010
为第i个所述目标信号在所述相互学习模型的第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的目标概率，M为特征标签总数，
Figure PCTCN2021081878-appb-000011
为第k个导联的所述目标信号在第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的权重。
在某些实施方式中，所述损失函数通过下列条件式计算：
Figure PCTCN2021081878-appb-000012
Figure PCTCN2021081878-appb-000013
Figure PCTCN2021081878-appb-000014
其中，
Figure PCTCN2021081878-appb-000015
为第k个所述神经网络的总损失，
Figure PCTCN2021081878-appb-000016
为第k个所述神经网络的监督损失，D KL为第k个所述神经网络与所述同伴神经网络的匹配度，K为所述导联的总数。
本申请实施方式的非易失性计算机可读存储介质存储有计算机程序，当所述计算机程序被一个或多个处理器执行时，实现上述任意实施方式所述的心电信号识别方法。
本申请的附加方面和优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本申请的实践了解到。
附图说明
本申请的上述和/或附加的方面和优点从结合下面附图对实施方式的描述中将变得明显和容易理解，其中：
图1是本申请实施方式的基于多导联的心电信号识别方法的流程示意图。
图2是本申请实施方式的电子设备的结构示意图。
图3是本申请实施方式的基于多导联的心电信号识别装置的模块示意图。
图4是本申请实施方式的基于多导联的心电信号识别方法的又一流程示意图。
图5是本申请实施方式的心电信号预处理的示意图。
图6是本申请实施方式的神经网络的结构示意图。
图7是本申请实施方式的基于多导联的心电信号识别方法的又一流程示意图。
图8是本申请实施方式的基于多导联的心电信号识别方法的又一流程示意图。
图9是本申请实施方式的基于多导联的心电信号识别方法的又一流程示意图。
具体实施方式
下面详细描述本申请的实施方式，所述实施方式的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施方式是示例性的，仅用于解释本申请，而不能理解为对本申请的限制。
请参阅图1，本申请实施方式的基于多导联的心电信号识别方法，包括以下步骤：
S10：对多个导联的心电信号进行预处理以得到目标信号；
S20：利用多个神经网络分别对多个导联的目标信号进行特征提取以得到对应的特征标签。
请参阅图2，本申请实施方式还提供了一种电子设备100。电子设备100包括处理器102和存储器104，存储器104存储有计算机程序106，计算机程序106被处理器102执行时实现本申请实施方式的基于多导联的心电信号识别方法，也即是说，处理器102用于对多个导联的心电信号进行预处理以得到目标信号，以及用于利用多个神经网络分别对多个 导联的目标信号进行特征提取以得到对应的特征标签。
请参阅图3，本申请实施方式还提供了一种基于多导联的心电信号识别装置110。本申请实施方式的基于多导联的心电信号识别方法可以由心电信号识别装置110实现。心电信号识别装置110包括预处理模块112和特征提取模块114。S10可以由预处理模块112实现，S20可以由特征提取模块114实现。也即是说，预处理模块112可用于对多个导联的心电信号进行预处理以得到目标信号，特征提取模块114可用于利用多个神经网络分别对多个导联的目标信号进行特征提取以得到对应的特征标签。
具体地，相关技术中，针对诊断ST段异常的检测多是基于单导联或基因工程。由于心脏是立体的结构，某些部位发生心肌缺血或者心肌梗死的情况在单导联的心电信号上并没有较为明显的变化，基于单导联的诊断方式容易造成漏诊。基于基因工程的诊断方式通过定位ST段并提取相关特征输入分类器进行ST段异常识别，然而基于基因工程的诊断方式过程比较复杂，基因工程模型的建立存在一定难度且识别精度不高。
本申请实施方式的基于多导联的心电信号识别方法、心电信号识别装置110和电子设备100中，根据相互学习模型对多个神经网络进行训练，利用多个神经网络对目标信号进行特征提取，得到与目标信号对应的特征标签，如此，能够提高特征标签的准确率，对于阳性率较低、特征较为不明显的导联，也能够提高ST段异常的检测精度，达到较高的敏感性和特异性。
进一步地，导联的数量可以是3个、6个、8个、12个等，具体不做限定。
在一些实施例中，导联的数量为12个，分别是标准导联Ⅰ、标准导联Ⅱ、标准导联Ⅲ、加压单极肢体导联aVR、加压单极肢体导联aVL、加压单极肢体导联aVF、胸导联V1、胸导联V2、胸导联V3、胸导联V4、胸导联V5和胸导联V6。其中，标准导联Ⅰ记录左臂与右臂电极间的电压差、标准导联Ⅱ记录左腿与右臂电极间的电压差、标准导联Ⅲ记录左腿与左臂电极间的电压差。加压单极肢体导联aVR记录右臂作为正极与左臂和左腿作为负极的加压后的正负极电压差、加压单极肢体导联aVL记录左臂作为正极与右臂和左腿作为负极的加压后的正负极电压差、加压单极肢体导联aVF记录左腿作为正极与左臂和右臂作为负极的加压后的正负极电压差。胸导联V1-V6将直接反映电势变化的正极放置于胸部的6个特定部位，通过导联的位置可以记录位于电极下方心脏各个部位的实际电位。
每一个导联可以对应建立一个神经网络，通过多个神经网络的特征提取可以得到多种特征标签，例如“ST段抬高”、“ST段压低”、“ST段正常”、“ST段水平压低”、“ST段水平抬高”、“ST段上斜型压低”、“ST段下斜型压低”、“其他”等。“其他”可以是指心电信号中噪声过多的情况，具体不做限定。敏感性可以是指在神经网络得到的特征标签为“ST段异常”的情况下，心电信号的真实情况为ST段异常的概率。特异性可以是指在神经网络得到的特征标签为“ST段正常”的情况下，心电信号的真实情况为ST段正常的概率。
请参阅图4，在某些实施方式中，S10包括：
S11：对心电信号进行降采样处理以得到降采样信号；
S12：将降采样信号切分为预设长度的目标信号。
在某些实施方式中，S11和S12可以由预处理模块112实现。也即是说，预处理模块112可用于对心电信号进行降采样处理以得到降采样信号，以及用于将降采样信号切分为预设长度的目标信号。
在某些实施方式中，处理器102用于对心电信号进行降采样处理以得到降采样信号，以及用于将降采样信号切分为预设长度的目标信号。
具体地，根据预定频率、采集多个导联的预定时长的心电信号，并对心电信号进行降采样处理，得到降采样信号。降采样处理可以是降低采样的预定频率，使得相同采样时长的心电信号中包含的信号点相对较少。经过降采样处理后，将降采样信号切分为预设长度的目标信号。如此，能够减少数据处理量。
其中，预定频率可以是采集心电信号的频率，预定频率可以根据各个导联的性能、功能设置以及被测人物的心跳周期、身体状况等参数确定，例如可以是200赫兹、256赫兹，300赫兹、500赫兹等，具体不做限定。
预定时长可以是采集心电信号的时长，预定时长可以根据各个导联的性能、功能设置以及被测人物的心跳周期、身体状况等参数确定，例如可以是5秒、8秒、10秒、15秒、18秒、20秒、50秒等，具体不做限定。
预设长度可以是对心电信号进行切分的尺度，也即是目标信号的长度，预设长度根据各个导联的性能、功能设置以及被测人物的心跳周期、身体状况等参数确定，例如可以是300、350、400等。预设长度通常需要大于采样的预定频率，且小于1.75倍的采样频率。
预设长度的目标信号中包含R波，R波在目标信号中的位置不做限定，例如可以在信号段的中间。
请参阅图5，在一些实施例中，心电信号的采样频率为500赫兹，持续时间为15秒。对心电信号进行降采样处理后得到降采样信号，降采样信号的采样频率为250赫兹。设定目标信号的预设长度L为300个信号点，使用差分阈值法识别心电信号中各个R波所在的位置，以R波为信号段的中点，截取R波前后约150个信号点。例如，第i个R波所在的位置为R i，则截取的信号段为[R i-149,R i+150]。
此外，在所截取的R波为心电信号中的第一个或最后一个R波的情况下，信号段长度可能小于预设长度，此时可以对信号段进行补零，以使得信号段长度等于预设长度。
如此，对心电信号进行预处理得到目标信号，便于神经网络对目标信号进行特征提取得出特征标签，进而提高特征标签的准确率。
在某些实施方式中，神经网络包括卷积层和池化层。其中，卷积层包括子卷积层和全连接层，池化层包括最大池化层和平均池化层，最大池化层用于对子卷积层的信号特征值求最大池化，平均池化层用于对子卷积层的信号特征值求平均池化，全连接层用于对平均池化层的信号特征值进行融合以得到目标信号特征值。
具体地，神经网络可以使用14层的残差网络，即ResNet14。ResNet14包括13个子卷积层和1个全连接层FC。请参阅图6，在ResNet14中，卷积层包括一维子卷积层1D_Conv、第一卷积块Block1、第二卷积块Block2、第三卷积块Block3、第四卷积块Block4和全连接层FC，共计14层卷积。其中，14层卷积的步长均为2。第一卷积块Block1包括第一子卷积层、第二子卷积层和第三子卷积层。第二卷积块Block2包括第四子卷积层、第五子卷积层和第六子卷积层。第三卷积块Block3包括第七子卷积层、第八子卷积层和第九子卷积层。第四卷积块Block4包括第十子卷积层、第十一子卷积层和第十二子卷积层。
一维子卷积层1D_Conv的卷积核为1*4，包括8个卷积通道。第一子卷积层的卷积核为1*1，包括8个卷积通道。第二子卷积层的卷积核为1*3，包括8个卷积通道。第三子卷积层的卷积核为1*1，包括16个卷积通道。第四子卷积层的卷积核为1*1，包括16个卷积通道。第五子卷积层的卷积核为1*3，包括16个卷积通道。第六子卷积层的卷积核为1*1，包括32个卷积通道。第七子卷积层的卷积核为1*1，包括32个卷积通道。第八子卷积层的卷积核为1*3，包括32个卷积通道。第九子卷积层的卷积核为1*1，包括64个卷积通道。第十子卷积层的卷积核为1*1，包括64个卷积通道。第十一子卷积层的卷积核为1*3，包括64个卷积通道。第十二子卷积层的卷积核为1*1，包括128个卷积通道。最大池化层MaxPooling的卷积核为1*3，步长为2。平均池化层MeanPooling的卷积核为1*3，步长为2。
将目标信号输入神经网络中，先由一维卷积1D_Conv对目标信号进行特征提取，并由最大池化层MaxPooling对一维子卷积层1D_Conv的信号特征值求最大池化，再依次经过第一卷积块Block1、第二卷积块Block2、第三卷积块Block3和第四卷积块Block4，得到 第十二子卷积层的信号特征值。由平均池化层MeanPooling对第十二子卷积层的信号特征值求平均池化，并由全连接层FC对平均池化层的信号特征值进行融合，得到目标信号特征值。
如此，使用残差网络对目标信号进行特征提取，能够减小神经网络的计算误差，提高特征标签的准确率。
需要说明地，神经网络还可以使用10层、18层、24层、26层、54层的残差网络等，具体不做限定。本实施方式中采用的ResNet14，一方面，由于网络层数相对较少，因此残差网络所需要的算例相对较少、运算时间也相对较短，另一方面，ResNet14的层数也能够确保特征提取达到较高的精确度。进一步地，神经网络还可以是FCNN、CNN、RNN、CRNN、RCR-net等类型，具体不做限定。
请参阅图7，在某些实施方式中，S20包括：
S21：利用Softmax函数对目标信号特征值进行分类以得到特征标签。
在某些实施方式中，S21可以由特征提取模块114实现，也即是说，特征提取模块114可用于利用Softmax函数对目标信号特征值进行分类以得到特征标签。
在某些实施方式中，处理器102还用于利用Softmax函数对目标信号特征值进行分类以得到特征标签。
具体地，使用Softmax函数对目标信号特征值进行分类，能够使神经网络学习到不同特征标签之间的相关信息，最终映射到不同种类的特征标签，使得最终输出的特征标签更加接近真实标签。如此，能够提高特征标签的准确率，也即是提高ST段异常的检测精度，降低误诊率。
请参阅图8，在某些实施方式中，心电信号识别方法包括：
S01：将心电信号按照特定比例分为训练集、验证集和测试集；
S02：利用训练集对相互学习模型进行训练；
S03：利用验证集对相互学习模型进行优化；
S04：利用测试集评估相互学习模型的心电信号识别效果。
在某些实施方式中，S01-S04可以由预处理模块112实现，也即是说，预处理模块112可用于将心电信号按照特定比例分为训练集、验证集和测试集，可用于利用训练集对相互学习模型进行训练，可用于利用验证集对相互学习模型进行优化，以及用于利用测试集评估相互学习模型的心电信号识别效果。
在某些实施方式中，处理器102还用于将心电信号按照特定比例分为训练集、验证集和测试集，用于利用训练集对相互学习模型进行训练，可用于利用验证集对相互学习模型进行优化，以及用于利用测试集评估相互学习模型的心电信号识别效果。
具体地，将心电信号按照特定比例分为训练集、验证集和测试集，特定比例可以根据相互学习模型的类型、结构等参数确定，例如可以是6:2:2，也可以是6:3:1，5:3:2，5:2.5:2.5等，具体不做限定。其中，训练集可以用于训练相互学习模型，验证集可以用于对相互学习模型进行优化，测试集可以用于评估相互学习模型的心电信号识别效果。
进一步地，对相互学习网络进行训练的方式可以是多种。
例如，可以按照训练、验证、测试的方式对相互学习网络进行训练，也即是说，将训练集中所有的心电信号一并输入相互学习网络中以对相互学习模型进行训练，后将验证集中所有的心电信号一并输入相互学习网络中以对相互学习模型进行优化，再将测试集中所有的心电信号一并输入相互学习网络中以评估相互学习模型的心电信号识别效果。上述过程可以是单程式的，即对于同一批样本只经过一次训练、一次验证和一次测试。也可以是反复式的，及对于同一批样本可以经过多次训练、验证和测试。
又如，可以按照训练、验证、训练、验证、训练、验证……测试的方式对相互学习网络进行训练，也即是说，训练集和验证集中的心电信号分为多个批次输入相互学习模型，经过多次训练和验证后，再将测试集中所有的心电信号一并输入相互学习网络中以评估相 互学习模型的心电信号识别效果。
请参阅图9，在某些实施方式中，S03包括：
S031：计算多个导联的目标信号分别在多个神经网络中的特征标签为目标信号的真实标签的目标概率；
S032：根据目标概率计算每个神经网络与同伴神经网络的损失函数；
S033：将损失函数在多个神经网络中相互传递以优化相互学习模型。
在某些实施方式中，S031-S033可以由预处理模块112实现，也即是说，预处理模块112可用于计算多个导联的目标信号分别在多个神经网络中的特征标签为目标信号的真实标签的目标概率，可用于根据目标概率计算每个神经网络与同伴神经网络的损失函数，以及用于将损失函数在多个神经网络中相互传递以优化相互学习模型。
在某些实施方式中，处理器102还用于计算多个导联的目标信号分别在多个神经网络中的特征标签为目标信号的真实标签的目标概率，用于根据目标概率计算每个神经网络与同伴神经网络的损失函数，以及用于将损失函数在多个神经网络中相互传递以优化相互学习模型。
具体地，在验证集对相互学习模型进行优化的过程中，先计算多个导联的目标信号分别在多个神经网络中的特征标签为目标信号的真实标签的目标概率，并根据目标概率计算每个神经网络与同伴神经网络的损失函数。随后，将得到的损失函数在多个神经网络中相互传递以优化相互学习模型。
如此，每个神经网络都可以作为相互学习模型中其他神经网络的教师网络，多个神经网络通过传递损失函数进行相互学习，无需预定义教师网络，即能够提高ST段异常的检测精度，达到较高的敏感性和特异性。
在某些实施方式中，目标概率通过下列条件式得到：
Figure PCTCN2021081878-appb-000017
其中，x i为第i个目标信号，
Figure PCTCN2021081878-appb-000018
为第i个目标信号在相互学习模型的第k个神经网络中的第m个特征标签为目标信号的真实标签的目标概率，M为特征标签总数，
Figure PCTCN2021081878-appb-000019
为第k个导联的目标信号在第k个神经网络中的第m个特征标签为目标信号的真实标签的权重。
具体地，N为目标信号x的总数，目标信号x可表示为
Figure PCTCN2021081878-appb-000020
x i为第i个目标信号。目标信号x对应的真实标签y可表示为
Figure PCTCN2021081878-appb-000021
其中y i∈{1,2,3,…,M}。
每一导联对应建立一个神经网络，也即是每一导联对应输出一个
Figure PCTCN2021081878-appb-000022
Figure PCTCN2021081878-appb-000023
为Softmax函数的输出，
Figure PCTCN2021081878-appb-000024
可表示第k个导联的目标信号在第k个神经网络θ k中的第m个特征标签为目标信号x的真实标签y的权重，其中m∈{1,2,3,…,M}。可以理解地，在同一神经网络θ k中，Z 1+Z 2+Z 3+…+Z m＝1。
目标概率
Figure PCTCN2021081878-appb-000025
可以理解为对
Figure PCTCN2021081878-appb-000026
的指数化运算，能够平滑
Figure PCTCN2021081878-appb-000027
的结果，避免算出的概率出现零而导致运算出现断点，确保梯度下降的连续性。
在一些实施例中，特征标签总数M＝4，即“ST段抬高”、“ST段压低”、“ST段正常”和“其他”。m＝1时，表示第1个特征标签为“ST段抬高”。m＝2时，表示第2个特征标签为“ST段压低”。m＝3时，表示第3个特征标签为“ST段正常。m＝4时，表示第4个特征标签为“其他”。则第i个目标信号x i在第k个神经网络θ k中的第1个特征标签为第i个目标信号x i的真实标签y i的概率p为：
Figure PCTCN2021081878-appb-000028
其他特征标签的计算方式以此类推，此处不再赘述。
得到目标概率
Figure PCTCN2021081878-appb-000029
后，根据目标概率
Figure PCTCN2021081878-appb-000030
计算每个神经网络与同伴神经网络的损失函数，并将得到的损失函数在多个神经网络中相互传递，如此，能够确保神经网络梯度下降的连续性，提高ST段异常的检测精度，达到较高的敏感性和特异性。
在某些实施方式中，损失函数通过下列条件式计算：
Figure PCTCN2021081878-appb-000031
Figure PCTCN2021081878-appb-000032
Figure PCTCN2021081878-appb-000033
其中，
Figure PCTCN2021081878-appb-000034
为第k个神经网络θ k的总损失，
Figure PCTCN2021081878-appb-000035
为第k个神经网络θ k的监督损失，D KL为第k个神经网络θ k与同伴神经网络θ l的匹配度，K为导联的总数。
具体地，对于第k个神经网络θ k，可以选用交叉熵损失函数作为θ k的监督损失，记为：
Figure PCTCN2021081878-appb-000036
其中
Figure PCTCN2021081878-appb-000037
k＝1,2,3,…,K。y i可表示第i个目标信号x i对应的真实标签，m为第k个神经网络θ k的预测标签。K为导联的总数，K可以是3个、6个、8个、12个等，具体不做限定。
为了提高第k个神经网络θ k在测试集上的泛化能力，第k个神经网络θ k可以从同伴神经网络θ l的后验概率中学习，以作为一种正则方式对第k个神经网络θ k进行约束，使得第k个神经网络θ k与同伴神经网络θ l进行相互学习。D KL可用于衡量第k个神经网络θ k与同伴神经网络θ l对于同一事件预测出的概率分布的差异，第k个神经网络θ k与同伴神经网络θ l之间相互传递的参数可以用D KL为来表示。
可以理解地，第k个神经网络θ k与同伴神经网络θ l越相似，匹配度越高，D KL越小。D KL为第k个神经网络θ k与同伴神经网络θ l的匹配度，
Figure PCTCN2021081878-appb-000038
为对多个D KL求均值，均值可以作为正则项约束第k个神经网络θ k。
综上，对于第k个神经网络θ k，总损失为
Figure PCTCN2021081878-appb-000039
即监督损失
Figure PCTCN2021081878-appb-000040
加上相互学习模型中多个神经网络相互学习的损失均值。
如此，无需预定义教师网络，通过相互学习模型中多个神经网络的相互学习，即能够对多个神经网络进行优化，进而提高ST段异常的检测精度，达到较高的敏感性和特异性。
在一些实施例中，导联总数为12个，即K＝12。每一个导联对应建立一个神经网络，12个神经网络分别对应标准导联Ⅰ、标准导联Ⅱ、标准导联Ⅲ、加压单极肢体导联aVR、加压单极肢体导联aVL、加压单极肢体导联aVF、胸导联V1、胸导联V2、胸导联V3、胸导联V4、胸导联V5和胸导联V6。使用上述基于多导联的心电信号识别方法对神经网络进行评估，所算出的敏感性和特异性数据如下表。
由下表可以看出，本申请实施方式的基于多导联的心电信号识别方法，使得相互学习模型中的多个神经网络进行相互学习，能够提高ST段异常的检测精度。即便是对于单导联的心电识别模型检测出的敏感性和特异性较低的加压单极肢体导联aVR和胸导联V1，也能够达到较高的敏感性和特异性。
Figure PCTCN2021081878-appb-000041
Figure PCTCN2021081878-appb-000042
本申请实施方式还提供了一种存储有计算机程序的非易失性计算机可读存储介质。当所述计算机程序被一个或多个处理器执行时，实现上述任一实施方式所述的心电信号识别方法。
在本说明书的描述中，参考术语“一个实施方式”、“一些实施方式”、“示意性实施方式”、“示例”、“具体示例”、或“一些示例”等的描述意指结合实施方式或示例描述的具体特征、结构、材料或者特点包含于本申请的至少一个实施方式或示例中。在本说明书中，对上述术语的示意性表述不一定指的是相同的实施方式或示例。而且，描述的具体特征、结构、材料或者特点可以在任何的一个或多个实施方式或示例中以合适的方式结合。
本技术领域的普通技术人员可以理解实现上述实施例方法携带的全部或部分步骤是可以通过程序来指令相关的硬件完成，所述的程序可以存储于一种计算机可读存储介质中，该程序在执行时，包括方法实施例的步骤之一或其组合。
此外，在本申请的各个实施例中的各功能单元可以集成在一个处理模块中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时，也可以存储在一个计算机可读取存储介质中。
上述提到的存储介质可以是只读存储器，磁盘或光盘等。
尽管已经示出和描述了本申请的实施方式，本领域的普通技术人员可以理解：在不脱离本申请的原理和宗旨的情况下可以对这些实施方式进行多种变化、修改、替换和变形，本申请的范围由权利要求及其等同物限定。

Claims (18)

  1. 一种基于多导联的心电信号识别方法，其特征在于，所述心电信号识别方法包括以下步骤：
    对多个导联的心电信号进行预处理以得到目标信号；
    利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
  2. 根据权利要求1所述的心电信号识别方法，其特征在于，所述对多个导联的心电信号进行预处理以得到目标信号包括：
    对所述心电信号进行降采样处理以得到降采样信号；
    将所述降采样信号切分为预设长度的所述目标信号。
  3. 根据权利要求1所述的心电信号识别方法，其特征在于，所述神经网络包括卷积层和池化层，
    其中，所述卷积层包括子卷积层和全连接层，
    所述池化层包括最大池化层和平均池化层，所述最大池化层用于对所述卷积层的信号特征值求最大池化，所述平均池化层用于对所述卷积层的信号特征值求平均池化，
    所述全连接层用于对所述平均池化层的信号特征值进行融合以得到目标信号特征值。
  4. 根据权利要求3所述的心电信号识别方法，其特征在于，所述利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签包括：
    利用Softmax函数对所述目标信号特征值进行分类以得到所述特征标签。
  5. 根据权利要求1所述的心电信号识别方法，其特征在于，所述心电信号识别方法包括：
    将心电信号按照特定比例分为训练集、验证集和测试集；
    利用所述训练集对所述相互学习模型进行训练；
    利用所述验证集对所述相互学习模型进行优化；
    利用所述测试集评估所述相互学习模型的心电信号识别效果。
  6. 根据权利要求5所述的心电信号识别方法，其特征在于，所述利用所述验证集对所述相互学习模型进行优化包括：
    计算多个导联的所述目标信号分别在多个所述神经网络中的特征标签为所述目标信号的真实标签的目标概率；
    根据所述目标概率计算每个所述神经网络与同伴神经网络的损失函数；
    将所述损失函数在所述多个神经网络中相互传递以优化所述相互学习模型。
  7. 根据权利要求6所述的心电信号识别方法，其特征在于，所述目标概率通过下列条件式得到：
    Figure PCTCN2021081878-appb-100001
    其中，x i为第i个所述目标信号，
    Figure PCTCN2021081878-appb-100002
    为第i个所述目标信号在所述相互学习模型的第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的目标概率，M为特征标签总数，
    Figure PCTCN2021081878-appb-100003
    为第k个导联的所述目标信号在第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的权重。
  8. 根据权利要求7所述的心电信号识别方法，其特征在于，所述损失函数通过下列条件式计算：
    Figure PCTCN2021081878-appb-100004
    Figure PCTCN2021081878-appb-100005
    Figure PCTCN2021081878-appb-100006
    其中，
    Figure PCTCN2021081878-appb-100007
    为第k个所述神经网络的总损失，
    Figure PCTCN2021081878-appb-100008
    为第k个所述神经网络的监督损失，D KL为第k个所述神经网络与所述同伴神经网络的匹配度，K为所述导联的总数。
  9. 一种基于多导联的心电信号识别装置，其特征在于，所述心电信号识别装置包括：
    预处理模块，用于对多个导联的心电信号进行预处理以得到目标信号；
    特征提取模块，用于利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
  10. 一种电子设备，其特征在于，所述电子设备包括一个或多个处理器，所述处理器用于：
    对多个导联的心电信号进行预处理以得到目标信号；
    利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签，所述多个神经网络根据相互学习模型训练得到。
  11. 根据权利要求10所述的电子设备，其特征在于，所述处理器用于：
    对所述心电信号进行降采样处理以得到降采样信号；
    将所述降采样信号切分为预设长度的所述目标信号。
  12. 根据权利要求10所述的电子设备，其特征在于，所述神经网络包括卷积层和池化层，
    其中，所述卷积层包括子卷积层和全连接层，
    所述池化层包括最大池化层和平均池化层，所述最大池化层用于对所述卷积层的信号特征值求最大池化，所述平均池化层用于对所述卷积层的信号特征值求平均池化，
    所述全连接层用于对所述平均池化层的信号特征值进行融合以得到目标信号特征值。
  13. 根据权利要求12所述的电子设备，其特征在于，所述利用多个神经网络分别对多个导联的所述目标信号进行特征提取以得到对应的特征标签包括：
    利用Softmax函数对所述目标信号特征值进行分类以得到所述特征标签。
  14. 根据权利要求10所述的电子设备，其特征在于，所述心电信号识别方法包括：
    将心电信号按照特定比例分为训练集、验证集和测试集；
    利用所述训练集对所述相互学习模型进行训练；
    利用所述验证集对所述相互学习模型进行优化；
    利用所述测试集评估所述相互学习模型的心电信号识别效果。
  15. 根据权利要求14所述的电子设备，其特征在于，所述利用所述验证集对所述相 互学习模型进行优化包括：
    计算多个导联的所述目标信号分别在多个所述神经网络中的特征标签为所述目标信号的真实标签的目标概率；
    根据所述目标概率计算每个所述神经网络与同伴神经网络的损失函数；
    将所述损失函数在所述多个神经网络中相互传递以优化所述相互学习模型。
  16. 根据权利要求15所述的电子设备，其特征在于，所述目标概率通过下列条件式得到：
    Figure PCTCN2021081878-appb-100009
    其中，x i为第i个所述目标信号，
    Figure PCTCN2021081878-appb-100010
    为第i个所述目标信号在所述相互学习模型的第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的目标概率，M为特征标签总数，
    Figure PCTCN2021081878-appb-100011
    为第k个导联的所述目标信号在第k个所述神经网络中的第m个特征标签为所述目标信号的真实标签的权重。
  17. 根据权利要求16所述的电子设备，其特征在于，所述损失函数通过下列条件式计算：
    Figure PCTCN2021081878-appb-100012
    Figure PCTCN2021081878-appb-100013
    Figure PCTCN2021081878-appb-100014
    其中，
    Figure PCTCN2021081878-appb-100015
    为第k个所述神经网络的总损失，
    Figure PCTCN2021081878-appb-100016
    为第k个所述神经网络的监督损失，D KL为第k个所述神经网络与所述同伴神经网络的匹配度，K为所述导联的总数。
  18. 一种存储有计算机程序的非易失性计算机可读存储介质，其特征在于，当所述计算机程序被一个或多个处理器执行时，实现权利要求1-8中任一项所述的心电信号识别方法。
PCT/CN2021/081878 2021-03-19 2021-03-19 基于多导联的心电信号识别方法和心电信号识别装置 WO2022193312A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/081878 WO2022193312A1 (zh) 2021-03-19 2021-03-19 基于多导联的心电信号识别方法和心电信号识别装置
CN202180000519.8A CN115666387A (zh) 2021-03-19 2021-03-19 基于多导联的心电信号识别方法和心电信号识别装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/081878 WO2022193312A1 (zh) 2021-03-19 2021-03-19 基于多导联的心电信号识别方法和心电信号识别装置

Publications (1)

Publication Number Publication Date
WO2022193312A1 true WO2022193312A1 (zh) 2022-09-22

Family

ID=83321453

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/081878 WO2022193312A1 (zh) 2021-03-19 2021-03-19 基于多导联的心电信号识别方法和心电信号识别装置

Country Status (2)

Country Link
CN (1) CN115666387A (zh)
WO (1) WO2022193312A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115813401A (zh) * 2022-09-28 2023-03-21 山东大学 一种基于动态学习的心电图分类方法及系统
CN116869546A (zh) * 2023-08-11 2023-10-13 河北医科大学第一医院 一种通过人工智能判读心电信号的方法、心电图机
CN117338310A (zh) * 2023-11-16 2024-01-05 齐鲁工业大学(山东省科学院) 一种基于对比学习和多尺度特征提取的心电信号分类方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116584951B (zh) * 2023-04-23 2023-12-12 山东省人工智能研究院 一种基于弱监督学习的心电信号检测和定位方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106560848A (zh) * 2016-10-09 2017-04-12 辽宁工程技术大学 模拟生物双向认知能力的新型神经网络模型及训练方法
CN107491735A (zh) * 2017-07-20 2017-12-19 浙江工业大学 人类行为识别的标签和交互关系联合学习方法
US20190287026A1 (en) * 2018-03-14 2019-09-19 International Business Machines Corporation Learning service blockchain
CN110522440A (zh) * 2019-08-12 2019-12-03 广州视源电子科技股份有限公司 基于分组卷积神经网络的心电信号识别装置
CN110751214A (zh) * 2019-10-21 2020-02-04 山东大学 一种基于轻量级可变形卷积的目标检测方法及系统
CN111539320A (zh) * 2020-04-22 2020-08-14 山东大学 基于互相学习网络策略的多视角步态识别方法及系统
CN111657926A (zh) * 2020-07-08 2020-09-15 中国科学技术大学 一种基于多导联信息融合的心律失常分类方法
CN112257526A (zh) * 2020-10-10 2021-01-22 中国科学院深圳先进技术研究院 一种基于特征交互学习的动作识别方法及终端设备

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106560848A (zh) * 2016-10-09 2017-04-12 辽宁工程技术大学 模拟生物双向认知能力的新型神经网络模型及训练方法
CN107491735A (zh) * 2017-07-20 2017-12-19 浙江工业大学 人类行为识别的标签和交互关系联合学习方法
US20190287026A1 (en) * 2018-03-14 2019-09-19 International Business Machines Corporation Learning service blockchain
CN110522440A (zh) * 2019-08-12 2019-12-03 广州视源电子科技股份有限公司 基于分组卷积神经网络的心电信号识别装置
CN110751214A (zh) * 2019-10-21 2020-02-04 山东大学 一种基于轻量级可变形卷积的目标检测方法及系统
CN111539320A (zh) * 2020-04-22 2020-08-14 山东大学 基于互相学习网络策略的多视角步态识别方法及系统
CN111657926A (zh) * 2020-07-08 2020-09-15 中国科学技术大学 一种基于多导联信息融合的心律失常分类方法
CN112257526A (zh) * 2020-10-10 2021-01-22 中国科学院深圳先进技术研究院 一种基于特征交互学习的动作识别方法及终端设备

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115813401A (zh) * 2022-09-28 2023-03-21 山东大学 一种基于动态学习的心电图分类方法及系统
CN116869546A (zh) * 2023-08-11 2023-10-13 河北医科大学第一医院 一种通过人工智能判读心电信号的方法、心电图机
CN117338310A (zh) * 2023-11-16 2024-01-05 齐鲁工业大学(山东省科学院) 一种基于对比学习和多尺度特征提取的心电信号分类方法
CN117338310B (zh) * 2023-11-16 2024-04-09 齐鲁工业大学(山东省科学院) 一种基于对比学习和多尺度特征提取的心电信号分类方法

Also Published As

Publication number Publication date
CN115666387A (zh) 2023-01-31

Similar Documents

Publication Publication Date Title
WO2022193312A1 (zh) 基于多导联的心电信号识别方法和心电信号识别装置
CN109620205B (zh) 心电数据分类方法、装置、计算机设备和存储介质
Kiranyaz et al. Convolutional neural networks for patient-specific ECG classification
KR102451795B1 (ko) 심전도 신호의 검출 방법
US11324455B2 (en) Artificial intelligence-based interference recognition method for electrocardiogram
WO2021017313A1 (zh) 房颤的检测方法、装置、计算机设备及存储介质
CN111626114B (zh) 基于卷积神经网络的心电信号心律失常分类系统
WO2019161611A1 (zh) 心电信息处理方法和心电工作站
CN110638430A (zh) 多任务级联神经网络ecg信号心律失常疾病分类模型和方法
CN112690802B (zh) 一种检测心电信号的方法、装置、终端及存储介质
CN110020636B (zh) 一种基于异常特征值的室性早搏智能分析方法
Zhang et al. [Retracted] An ECG Heartbeat Classification Method Based on Deep Convolutional Neural Network
CN111657925A (zh) 基于机器学习的心电信号分类方法、系统、终端以及存储介质
CN111956214A (zh) 基于U-net端对端神经网络模型的QRS波自动检测方法
CN111803059A (zh) 一种基于时域卷积网络的心电信号分类方法及装置
CN111053552B (zh) 一种基于深度学习的qrs波检测方法
Li et al. DeepECG: Image-based electrocardiogram interpretation with deep convolutional neural networks
Chen et al. SE-ECGNet: multi-scale SE-Net for multi-lead ECG data
CN115470832A (zh) 一种基于区块链的心电信号数据处理方法
CN113180688B (zh) 基于残差神经网络的冠心病心电图筛查系统及方法
CN113229798B (zh) 模型迁移训练方法、装置、计算机设备及可读存储介质
CN113171102B (zh) 基于持续深度学习的ecg数据分类方法
CN115414049A (zh) 一种基于深度神经网络的可穿戴心电图实时诊断系统
He et al. Automatic classification of arrhythmias by residual network and BiGRU with attention mechanism
CN113593698A (zh) 一种基于图注意网络的中医证型识别方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21930902

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE