A method and system for identifying ventricular premature beat based on deep learning mask mechanism

Abstract

The application provides a ventricular premature beat recognition method and system based on a deep learning mask mechanism. The method comprises: training an initial model to obtain a reference electrocardiogram signal generation model, the training sample of the initial model being an electrocardiogram signal in which a ventricular premature beat feature interval obtained based on a mask operation is hidden, and the expected output of the initial model being an electrocardiogram signal in which the ventricular premature beat feature interval is restored; inputting a target electrocardiogram signal in which a target interval is hidden into the reference electrocardiogram signal generation model to obtain a reference electrocardiogram signal, the target interval covering an R wave to be detected; and comparing the target interval of the target electrocardiogram signal with a ventricular premature beat feature interval of the reference electrocardiogram signal to identify whether a heart beat corresponding to the corresponding R wave belongs to ventricular premature beat. The system is used to implement the method. The application can solve the problem that the ventricular premature beat recognition method based on a recognition model has insufficient recognition result accuracy due to the influence of the comprehensiveness of the model training sample.

Description

Technical Field

[0001] This invention belongs to the field of electrocardiogram signal detection, and more specifically, relates to a method and system for identifying premature ventricular contractions based on a deep learning masking mechanism. Background Technology

[0002] In related technologies, one type of ventricular premature beat (PVC) identification method based on a recognition model first trains an initial model using ECG signal samples of PVC cases to obtain a PVC identification model. Then, based on the obtained PVC identification model, it detects whether a given input ECG signal contains PVCs. Currently, the initial model used in these recognition model-based PVC identification methods is typically a pattern recognition model, a deep learning model, or a machine learning model. The accuracy of the output results of these models is closely positively correlated with the comprehensiveness of the training samples. However, due to the highly individualized and nonlinear nature of each individual's ECG signal, the waveform morphology of ECG signals of PVC cases varies significantly between different individuals. Obtaining comprehensive training samples is virtually impossible. Therefore, these recognition model-based PVC identification methods typically use ECG signal samples of typical PVC cases to train the initial model. The lack of comprehensiveness in these training samples leads to insufficient accuracy in the identification results of these recognition model-based PVC identification methods. Summary of the Invention

[0003] The purpose of this invention is to solve the problem that existing ventricular premature beat identification methods based on identification models suffer from insufficient accuracy due to the influence of the comprehensiveness of the model training samples.

[0004] To achieve the above objectives, the present invention provides a method and system for identifying premature ventricular contractions (PVCs) based on a deep learning masking mechanism.

[0005] According to a first aspect of the present invention, a method for identifying premature ventricular contractions (PVCs) based on a deep learning masking mechanism is provided, the method comprising the following steps:

[0006] The initial model is trained to obtain a reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature regions obtained based on deep learning masking operations. The expected output of the initial model is an ECG signal with the ventricular premature beat feature regions restored.

[0007] The target ECG signal hidden in the target interval is input into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected;

[0008] The target interval of the target electrocardiogram signal is compared with the ventricular premature beat characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to ventricular premature beat.

[0009] Optionally, the method for obtaining the training samples includes:

[0010] Acquire a first number of ECG signals of a first length that include R waves in cases of premature ventricular contractions;

[0011] Determine the characteristic interval of ventricular premature beats in each electrocardiogram signal, wherein the characteristic interval of ventricular premature beats covers the R wave in the corresponding electrocardiogram signal under the condition of ventricular premature beats;

[0012] The characteristic regions of premature ventricular contractions in each electrocardiogram signal are hidden.

[0013] Optionally, acquiring a first number of ECG signals of a first length that include R waves in cases of premature ventricular contractions includes:

[0014] A second number of raw ECG signals were acquired, each of which was obtained from a different ECG signal provider.

[0015] Each raw electrocardiogram signal is preprocessed;

[0016] Determine whether there is an R wave in the case of premature ventricular contractions in each preprocessed ECG signal; if so, mark the corresponding R wave.

[0017] From the preprocessed ECG signals of R waves in various cases of premature ventricular contractions, a first number of ECG signals of a first length containing R waves in cases of premature ventricular contractions are randomly selected.

[0018] Alternatively, the initial model may be a bidirectional long short-term memory recurrent neural network.

[0019] Optionally, the bidirectional long short-term memory recurrent neural network includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a bidirectional long short-term memory network layer, a first fully connected layer, and a second fully connected layer connected in sequence.

[0020] Optionally, the first convolutional layer includes 16 one-dimensional convolutional kernels with a stride of 1 and a size of 8.

[0021] And / or, the second convolutional layer comprises 32 one-dimensional convolutional kernels with a stride of 1 and a size of 8;

[0022] And / or, the third convolutional layer comprises 64 one-dimensional convolutional kernels with a stride of 1 and a size of 4;

[0023] And / or, the fourth convolutional layer comprises 128 one-dimensional convolutional kernels with a stride of 1 and a size of 2;

[0024] And / or, the first pooling layer, the second pooling layer, the third pooling layer and the fourth pooling layer are all max pooling layers, and the size and step size of their respective pooling kernels are all 2;

[0025] And / or, the data vector dimension of the bidirectional long short-term memory network layer is 10, and the hidden element dimension is 20;

[0026] And / or, the first fully connected layer has 128 neurons and uses the ReLU function as the activation function;

[0027] And / or, the number of neurons in the second fully connected layer is L, L = A × B, the length of the training sample is A seconds, the sampling frequency of the training sample is B Hz, and L is rounded up.

[0028] Optionally, the method for obtaining the target ECG signal hidden in the target interval includes:

[0029] Preprocess the electrocardiogram signals to be tested;

[0030] Identify each valid R wave in the preprocessed ECG signal;

[0031] For the current valid R wave to be detected, obtain its corresponding target interval and obtain the target ECG signal corresponding to the target interval;

[0032] The target region in the target electrocardiogram signal is hidden.

[0033] Optionally, the criteria for determining an effective R-wave are:

[0034] If the length from the start point of an R wave to the start point of the ECG signal to be detected, and the length from its end point to the end point of the ECG signal to be detected, are both not less than a predetermined second length, then the R wave is determined to be a valid R wave.

[0035] And / or, the target interval is equal in length to the characteristic interval of the premature ventricular contraction, the target interval uniformly covers the corresponding effective R wave, and the characteristic interval of the premature ventricular contraction uniformly covers the R wave in the corresponding premature ventricular contraction case;

[0036] And / or, the target ECG signal has the same length as the training sample, and the target ECG signal before the target interval is hidden uniformly covers its target interval.

[0037] Optionally, comparing the target interval of the target ECG signal with the ventricular premature beat characteristic interval of the reference ECG signal to identify whether the heartbeat corresponding to the R wave belongs to a ventricular premature beat includes:

[0038] Obtain the absolute value C of the difference between the target interval and the characteristic interval of premature ventricular contractions;

[0039] Obtain the absolute value D of the difference between the target ECG signal and the reference ECG signal;

[0040] If the ratio of the absolute value C to the absolute value D is less than a predetermined ratio threshold, then the heartbeat corresponding to the R wave is determined to be a premature ventricular contraction.

[0041] According to a second aspect of the present invention, a system for identifying premature ventricular contractions (PVCs) based on a deep learning masking mechanism is provided, the system comprising the following functional modules:

[0042] The reference ECG signal generation model acquisition module is used to train the initial model to obtain the reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature intervals obtained based on deep learning masking operations. The expected output of the initial model is the ECG signal with the ventricular premature beat feature intervals restored.

[0043] The reference ECG signal acquisition module is used to input the target ECG signal hidden in the target interval into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected;

[0044] The premature ventricular contraction (PVC) identification module is used to compare the target interval of the target electrocardiogram signal with the PVC characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to a premature ventricular contraction (PVC).

[0045] The beneficial effects of this invention are as follows:

[0046] The present invention discloses a method for identifying premature ventricular contractions (PVCs) based on a deep learning masking mechanism. First, an initial model is trained to obtain a reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden PVC feature intervals obtained through deep learning masking operations. The expected output of the initial model is an ECG signal with the restored PVC feature intervals. Second, a target ECG signal with a hidden target interval is input into the reference ECG signal generation model to obtain a reference ECG signal. The target interval covers the R wave to be detected. Finally, the target interval of the target ECG signal is compared with the PVC feature interval of the reference ECG signal to identify whether the heartbeat corresponding to the R wave belongs to a premature ventricular contraction (PVC).

[0047] Unlike existing methods for identifying premature ventricular contractions (PVCs) that directly determine the presence of PVCs in the ECG signal to be detected based on the corresponding PVC identification model, the PVC identification method based on a deep learning masking mechanism of this invention generates a reference ECG signal corresponding to the target ECG signal based on a reference ECG signal generation model. The difference between the reference ECG signal and the target ECG signal is that the portion of the reference ECG signal corresponding to the target interval of the target ECG signal is a PVC characteristic interval, that is, the portion of the generated signal with PVC characteristic waveforms. After generating the reference ECG signal, the PVC identification method based on a deep learning masking mechanism of this invention compares the target interval with the corresponding PVC characteristic interval, that is, compares the object to be detected with the predicted abnormal object, and then determines whether the heartbeat corresponding to the target interval belongs to a PVC by the degree of difference between the two.

[0048] As can be seen from the above, compared with existing ventricular premature beat (PVC) identification methods based on recognition models, the PVC identification method based on deep learning masking mechanism of the present invention has a lower dependence on training samples and is less affected by the comprehensiveness of training samples. Under the same comprehensiveness of the training samples used, the PVC identification method based on deep learning masking mechanism of the present invention has higher accuracy in identification results, thus effectively solving the problem that the existing PVC identification methods based on recognition models suffer from insufficient accuracy due to the influence of the comprehensiveness of the model training samples.

[0049] The ventricular premature beat identification system based on deep learning masking mechanism of the present invention belongs to the same general inventive concept as the above-mentioned ventricular premature beat identification method based on deep learning masking mechanism. It has at least the same beneficial effects as the above-mentioned ventricular premature beat identification method based on deep learning masking mechanism, and its beneficial effects will not be repeated here.

[0050] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0051] The present invention can be better understood by referring to the following description taken in conjunction with the accompanying drawings, in which the same or similar reference numerals are used throughout the drawings to denote the same or similar parts.

[0052] Figure 1 A flowchart illustrating the implementation of a method for identifying premature ventricular contractions based on a deep learning masking mechanism according to an embodiment of the present invention is shown.

[0053] Figure 2 A block diagram of a ventricular premature beat identification system based on a deep learning masking mechanism according to an embodiment of the present invention is shown. Detailed Implementation

[0054] To enable those skilled in the art to more fully understand the technical solutions of the present invention, exemplary embodiments of the present invention will be described more comprehensively and in detail below with reference to the accompanying drawings. Obviously, the one or more embodiments of the present invention described below are merely one or more specific ways to implement the technical solutions of the present invention, and are not exhaustive. It should be understood that other ways belonging to a general inventive concept can be used to implement the technical solutions of the present invention, and should not be limited to the embodiments described exemplary. Based on one or more embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0055] Example: Figure 1 A flowchart illustrating the implementation of a deep learning masking mechanism-based method for identifying premature ventricular contractions (PVCs) according to an embodiment of the present invention is shown. (Refer to...) Figure 1 The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to embodiments of the present invention includes the following steps:

[0056] Step S100: Train the initial model to obtain a reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature intervals obtained based on deep learning masking operations. The expected output of the initial model is an ECG signal with the ventricular premature beat feature intervals restored.

[0057] Step S200: Input the target ECG signal hidden in the target interval into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected;

[0058] Step S300: Compare the target interval of the target ECG signal with the ventricular premature beat characteristic interval of the reference ECG signal to identify whether the heartbeat corresponding to the R wave belongs to ventricular premature beat.

[0059] Specifically, the deep learning masking mechanism used in the ventricular premature beat identification method of this invention is as follows: after obtaining the electrocardiogram (ECG) signal containing ventricular premature beat features, a masking operation is performed on the ECG signal containing ventricular premature beat features, that is, the interval containing ventricular premature beat features on the ECG signal is hidden, and it is used as the training sample of the initial model to obtain the reference ECG signal generation model. The expected output of the reference ECG signal generation model is the ECG signal that has not undergone the masking operation, that is, the original ECG signal containing ventricular premature beat features.

[0060] When using a reference ECG signal generation model to detect the target ECG signal, the target region containing the R wave in the target ECG signal is hidden. The target ECG signal that has undergone the hiding operation is then input into the reference ECG signal generation model to obtain a reference ECG signal. The R wave in the reference ECG signal corresponding to the target region in the target ECG signal is an R wave with ventricular premature beat characteristics. By comparing the difference between the R wave in the target ECG signal and the R wave with ventricular premature beat characteristics in the reference ECG signal, it can be determined whether the R wave in the target ECG signal has ventricular premature beat characteristics.

[0061] Further, in step S100 of this embodiment of the invention, the method for obtaining the training samples includes:

[0062] Acquire a first number of ECG signals of a first length that include R waves in cases of premature ventricular contractions;

[0063] Determine the characteristic interval of ventricular premature beats in each electrocardiogram signal, wherein the characteristic interval of ventricular premature beats covers the R wave in the corresponding electrocardiogram signal under the condition of ventricular premature beats;

[0064] The characteristic regions of premature ventricular contractions in each electrocardiogram signal are hidden.

[0065] Furthermore, in this embodiment of the invention, acquiring a first number of ECG signals of a first length that include R waves in the case of premature ventricular contractions includes:

[0066] A second number of raw ECG signals were acquired, each of which was obtained from a different ECG signal provider.

[0067] Each raw electrocardiogram signal is preprocessed;

[0068] Determine whether there is an R wave in the case of premature ventricular contractions in each preprocessed ECG signal; if so, mark the corresponding R wave.

[0069] From the preprocessed ECG signals of R waves in various cases of premature ventricular contractions, a first number of ECG signals of a first length containing R waves in cases of premature ventricular contractions are randomly selected.

[0070] Specifically, in this embodiment of the invention, the specific process for obtaining the training samples of the initial model is as follows:

[0071] Collect raw electrocardiogram (ECG) signals in the form of dynamic ECGs from more than 5,000 people. The ECG acquisition time for each person is no less than 2 hours, and the sampling frequency is 250Hz.

[0072] A high-pass filter with a cutoff frequency of 0.5 Hz was used to filter each raw ECG signal to remove drift and respiratory interference from the raw ECG signal;

[0073] For each filtered ECG signal, mark the location of each R wave and determine whether the heartbeat corresponding to each R wave is a premature ventricular contraction. If yes, mark the R wave as 1; otherwise, mark the R wave as 0.

[0074] From all filtered ECG signals containing an R wave labeled 1, at least 20,000 ECG segments of 20 seconds each containing an R wave labeled 1 are randomly selected. These are training samples where the ventricular premature beat feature intervals are not hidden, and these ECG segments are recorded as X = {x1, x2, ..., xN}.

[0075] For an electrocardiogram segment xn (n = 1, 2, 3… N), the peak position of the R wave marked as 1 is marked as rloc. The portion of electrocardiogram segment xn from rloc-0.5 seconds to rloc+0.5 seconds is zeroed out, and the portion of electrocardiogram segment with zeroed out is denoted as xxn. The portion of electrocardiogram segment xn from rloc-0.5 seconds to rloc+0.5 seconds is the characteristic interval of premature ventricular contractions.

[0076] The ECG image segment xnn is the training sample of the initial model, and the ECG image segment xn is the expected output of the initial model.

[0077] Furthermore, in step S100 of this embodiment of the invention, the initial model is a bidirectional long short-term memory recurrent neural network.

[0078] Furthermore, in this embodiment of the invention, the bidirectional long short-term memory recurrent neural network includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a bidirectional long short-term memory network layer, a first fully connected layer, and a second fully connected layer connected in sequence.

[0079] Furthermore, in this embodiment of the invention, the first convolutional layer includes 16 one-dimensional convolutional kernels with a stride of 1 and a size of 8;

[0080] The second convolutional layer consists of 32 one-dimensional convolutional kernels with a stride of 1 and a size of 8;

[0081] The third convolutional layer consists of 64 one-dimensional convolutional kernels with a stride of 1 and a size of 4;

[0082] The fourth convolutional layer consists of 128 one-dimensional convolutional kernels with a stride of 1 and a size of 2;

[0083] The first, second, third, and fourth pooling layers are all max pooling layers, and the size and step size of their respective pooling kernels are all 2.

[0084] The data vector dimension of the bidirectional long short-term memory network layer is 10, and the hidden element dimension is 20.

[0085] The first fully connected layer has 128 neurons and uses the ReLU function as the activation function;

[0086] The second fully connected layer has L neurons, where L = A × B. The length of the training samples is A seconds, and the sampling frequency of the training samples is B Hz. L is rounded up.

[0087] Specifically, in this embodiment of the invention, the length of the training sample on the time axis is 20 seconds, the sampling frequency of the training sample is 250Hz, and the number of neurons in the second fully connected layer is L = 20 × 250 = 5000.

[0088] Specifically, in this embodiment of the invention, any one of the following algorithms is used to train the initial model: stochastic gradient descent, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm, and Adamax algorithm.

[0089] Furthermore, in step S200 of this embodiment of the invention, the method for obtaining the target electrocardiogram signal hidden in the target interval includes:

[0090] Preprocess the electrocardiogram signals to be tested;

[0091] Identify each valid R wave in the preprocessed ECG signal;

[0092] For the current valid R wave to be detected, obtain its corresponding target interval and obtain the target ECG signal corresponding to the target interval;

[0093] The target region in the target electrocardiogram signal is hidden.

[0094] Furthermore, in this embodiment of the invention, the criterion for determining the effective R-wave is:

[0095] If the length from the start point of an R wave to the start point of the ECG signal to be detected, and the length from its end point to the end point of the ECG signal to be detected, are both not less than a predetermined second length, then the R wave is determined to be a valid R wave.

[0096] Furthermore, in this embodiment of the invention, the target interval is equal in length to the ventricular premature beat characteristic interval, the target interval uniformly covers the corresponding effective R wave, and the ventricular premature beat characteristic interval uniformly covers the R wave in the corresponding ventricular premature beat situation.

[0097] Furthermore, in this embodiment of the invention, the target ECG signal has the same length as the training sample, and the target ECG signal before the target interval is hidden uniformly covers the target interval.

[0098] Specifically, in this embodiment of the invention, uniform coverage of the effective R wave in the target interval means that the length from the start of the target interval to the start of the effective R wave is equal to the length from the end of the target interval to the end of the effective R wave. Uniform coverage of the R wave in the case of premature ventricular contraction (PVC) in the characteristic interval means that the length from the start of the characteristic interval to the start of the R wave in the case of PVC is equal to the length from the end of the characteristic interval to the end of the R wave in the case of PVC. The above lengths are all lengths on the signal time axis.

[0099] Specifically, in this embodiment of the invention, the specific process for obtaining the target electrocardiogram signal hidden in the target interval is as follows:

[0100] Acquire the ECG signal to be detected in the form of a dynamic electrocardiogram;

[0101] The ECG signal to be detected is preprocessed, specifically by filtering it with a high-pass filter with a cutoff frequency of 0.5Hz.

[0102] An R-wave extraction algorithm is used to identify each R-wave in the electrocardiogram signal to be detected;

[0103] For the i-th R wave, if the length from the start of the R wave to the start of the ECG signal to be detected is no more than 10 seconds, or the length from the end of the R wave to the end of the ECG signal to be detected is no more than 10 seconds, then the R wave is considered an invalid R wave, meaning the heartbeat corresponding to this R wave is of an undetermined type, due to the dynamic ECG... Figure 1 The recording time is generally quite long, and not judging the R wave within 10 seconds before or after does not affect the analysis results;

[0104] For the i-th R wave, if it is determined to be a valid R wave, the peak position of the R wave is marked as rloci, and the portion from rloci-10 seconds to rloci+10 seconds in the ECG signal to be detected is taken as ECG image segment xi, i.e., the target ECG signal. The portion from rloci-0.5 seconds to rloci+0.5 seconds in ECG image segment xi is zeroed out, and the ECG image segment with the zeroed-out portion is recorded as xxi, i.e., the target ECG signal before the target interval is hidden. The portion from rloci-0.5 seconds to rloci+0.5 seconds in ECG image segment xi is the target interval.

[0105] Furthermore, in this embodiment of the invention, step S300, which compares the target interval of the target electrocardiogram signal with the ventricular premature beat characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to a ventricular premature beat, includes:

[0106] Obtain the absolute value C of the difference between the target interval and the characteristic interval of premature ventricular contractions;

[0107] Obtain the absolute value D of the difference between the target ECG signal and the reference ECG signal;

[0108] If the ratio of the absolute value C to the absolute value D is less than a predetermined ratio threshold, then the heartbeat corresponding to the R wave is determined to be a premature ventricular contraction.

[0109] Specifically, in this embodiment of the invention, an electrocardiogram (ECG) image segment xi is generated based on the input ECG image segment xxi, referring to the ECG signal generation model. ′ ;

[0110] The following formula is used to identify whether the heartbeat corresponding to the R wave is a premature ventricular contraction:

[0111]

[0112] In the above formula, val is the quantified result of the difference between the target interval and the characteristic interval of premature ventricular contractions, that is, the ratio of the absolute value C to the absolute value D, and xi ′ [-0.5, rloci+0.5 represents the electrocardiogram segment xi] ′ The portion from rloci-0.5 seconds to rloci+0.5 seconds in the ECG segment xi is the characteristic interval of the generated premature ventricular contractions (PVCs). xi[-0.5, rloci+0.5] is the portion from rloci-0.5 seconds to rloci+0.5 seconds in the ECG segment xi, which is the target interval.

[0113] If val is less than 0.2, the heartbeat corresponding to the R wave is considered to be a premature ventricular contraction (PVC); otherwise, the heartbeat corresponding to the R wave is considered not to be a PVC.

[0114] Accordingly, based on the ventricular premature beat identification method based on deep learning masking mechanism proposed in the embodiments of the present invention, the embodiments of the present invention also propose a ventricular premature beat identification system.

[0115] Figure 2 A block diagram of a ventricular premature beat identification system based on a deep learning masking mechanism according to an embodiment of the present invention is shown. (Refer to...) Figure 2 The ventricular premature beat identification system based on deep learning masking mechanism in this embodiment of the invention includes the following functional modules;

[0116] The reference ECG signal generation model acquisition module is used to train the initial model to obtain the reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature intervals obtained based on deep learning masking operations. The expected output of the initial model is the ECG signal with the ventricular premature beat feature intervals restored.

[0117] The reference ECG signal acquisition module is used to input the target ECG signal hidden in the target interval into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected;

[0118] The premature ventricular contraction (PVC) identification module is used to compare the target interval of the target electrocardiogram signal with the PVC characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to a premature ventricular contraction (PVC).

[0119] While one or more embodiments of the present invention have been described above, those skilled in the art will recognize that the present invention can be implemented in any other form without departing from its spirit and scope. Therefore, the embodiments described above are illustrative and not restrictive, and many modifications and substitutions will be apparent to those skilled in the art without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for identifying premature ventricular contractions (PVCs) based on a deep learning masking mechanism, characterized in that, include: The initial model is trained to obtain a reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature regions obtained based on deep learning masking operations. The expected output of the initial model is an ECG signal with the ventricular premature beat feature regions restored. The target ECG signal hidden in the target interval is input into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected; The target interval of the target electrocardiogram signal is compared with the ventricular premature beat characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to ventricular premature beat.

2. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 1, characterized in that, The methods for obtaining the training samples include: Acquire a first number of ECG signals of a first length that include R waves in cases of premature ventricular contractions; Determine the characteristic interval of ventricular premature beats in each electrocardiogram signal, wherein the characteristic interval of ventricular premature beats covers the R wave in the corresponding electrocardiogram signal under the condition of ventricular premature beats; The characteristic regions of premature ventricular contractions in each electrocardiogram signal are hidden.

3. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 2, characterized in that, The acquisition of a first number of ECG signals of a first length that include R waves in cases of premature ventricular contractions includes: A second number of raw ECG signals were acquired, each of which was obtained from a different ECG signal provider. Each raw electrocardiogram signal is preprocessed; Determine whether there is an R wave in the case of premature ventricular contractions in each preprocessed ECG signal; if so, mark the corresponding R wave. From the preprocessed ECG signals of R waves in various cases of premature ventricular contractions, a first number of ECG signals of a first length containing R waves in cases of premature ventricular contractions are randomly selected.

4. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 1, characterized in that, The initial model is a bidirectional long short-term memory recurrent neural network.

5. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 4, characterized in that, The bidirectional long short-term memory recurrent neural network includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a bidirectional long short-term memory network layer, a first fully connected layer, and a second fully connected layer, connected in sequence.

6. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 5, characterized in that, The first convolutional layer includes 16 one-dimensional convolutional kernels with a stride of 1 and a size of 8; And / or, the second convolutional layer comprises 32 one-dimensional convolutional kernels with a stride of 1 and a size of 8; And / or, the third convolutional layer comprises 64 one-dimensional convolutional kernels with a stride of 1 and a size of 4; And / or, the fourth convolutional layer comprises 128 one-dimensional convolutional kernels with a stride of 1 and a size of 2; And / or, the first pooling layer, the second pooling layer, the third pooling layer and the fourth pooling layer are all max pooling layers, and the size and step size of their respective pooling kernels are all 2; And / or, the data vector dimension of the bidirectional long short-term memory network layer is 10, and the hidden element dimension is 20; And / or, the first fully connected layer has 128 neurons and uses the ReLU function as the activation function; And / or, the number of neurons in the second fully connected layer is L, L = A × B, the length of the training sample is A seconds, the sampling frequency of the training sample is B Hz, and L is rounded up.

7. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 1, characterized in that, The method for obtaining the target ECG signal hidden in the target interval includes: Preprocess the electrocardiogram signals to be tested; Identify each valid R wave in the preprocessed ECG signal; For the current valid R wave to be detected, obtain its corresponding target interval and obtain the target ECG signal corresponding to the target interval; The target region in the target electrocardiogram signal is hidden.

8. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 7, characterized in that, The criteria for determining an effective R-wave are as follows: If the length from the start point of an R wave to the start point of the ECG signal to be detected, and the length from its end point to the end point of the ECG signal to be detected, are both not less than a predetermined second length, then the R wave is determined to be a valid R wave. And / or, the target interval is equal in length to the characteristic interval of the premature ventricular contraction, the target interval uniformly covers the corresponding effective R wave, and the characteristic interval of the premature ventricular contraction uniformly covers the R wave in the corresponding premature ventricular contraction case; And / or, the target ECG signal has the same length as the training sample, and the target ECG signal before the target interval is hidden uniformly covers its target interval.

9. The method for identifying premature ventricular contractions based on a deep learning masking mechanism according to claim 1, characterized in that, The step of comparing the target interval of the target ECG signal with the ventricular premature beat characteristic interval of the reference ECG signal to identify whether the heartbeat corresponding to the R wave belongs to ventricular premature beats includes: Obtain the absolute value C of the difference between the target interval and the characteristic interval of premature ventricular contractions; Obtain the absolute value D of the difference between the target ECG signal and the reference ECG signal; If the ratio of the absolute value C to the absolute value D is less than a predetermined ratio threshold, then the heartbeat corresponding to the R wave is determined to be a premature ventricular contraction.

10. A ventricular premature beat identification system based on a deep learning masking mechanism, characterized in that, include: The reference ECG signal generation model acquisition module is used to train the initial model to obtain the reference ECG signal generation model. The training samples of the initial model are ECG signals with hidden ventricular premature beat feature intervals obtained based on deep learning masking operations. The expected output of the initial model is the ECG signal with the ventricular premature beat feature intervals restored. The reference ECG signal acquisition module is used to input the target ECG signal hidden in the target interval into the reference ECG signal generation model to obtain the reference ECG signal, wherein the target interval covers the R wave to be detected; The premature ventricular contraction (PVC) identification module is used to compare the target interval of the target electrocardiogram signal with the PVC characteristic interval of the reference electrocardiogram signal to identify whether the heartbeat corresponding to the R wave belongs to a premature ventricular contraction (PVC).

Images