Heartbeat Division Method and System Based on Sub-Waveform Relationship Extraction

Abstract

The present disclosure provides a heartbeat division method and system based on sub-waveform relationship extraction, which relates to the technical field of electrocardiogram signal processing. The method includes obtaining a 12-lead electrocardiogram data tensor and annotation data, and performing preprocessing to obtain an electrocardiogram signal and a heartbeat division label matrix; inputting the electrocardiogram signal into a deep learning network to obtain alignment features, adding an electrocardiogram sequence position encoding to the alignment features to obtain new alignment features; generating a sub-waveform feature extraction matrix through the waveform types of sub-waveforms in the annotation data, combining the sub-waveform feature extraction matrix with the new alignment features, performing dimensional change and sorting, and obtaining a sub-waveform-based feature tensor; extracting sub-waveform time series features from the sub-waveform-based feature tensor and then generating a sub-waveform relationship matrix using matrix multiplication, dividing by sub-waveform units to obtain the final heartbeat division result, and determining the heartbeat to which the sub-waveform belongs.

Description

Technical Field

[0001] The present disclosure relates to the technical field of electrocardiogram signal processing, and in particular, to a heartbeat division method and system based on sub-waveform relationship extraction. Background Art

[0002] The statements in this part only provide background technical information related to the present disclosure, and do not necessarily constitute prior art.

[0003] With the wide application of electrocardiogram technology in recent decades, it plays an irreplaceable role in the examination and diagnosis of cardiovascular diseases. At present, 12-lead electrocardiogram is the most basic and common electrocardiogram acquisition method.

[0004] The high-speed iteration of machine learning and deep learning has driven the development of electrocardiogram-assisted diagnosis algorithms including electrocardiogram noise reduction, feature extraction, and electrocardiogram classification. However, most of the existing work focuses on the classification of the overall segment. However, a heartbeat is a cycle of cardiac activity. Therefore, each heartbeat should have its own type, and in the dynamic electrocardiogram scenario with a time of more than 24 hours, there are obvious limitations for precise positioning and the like.

[0005] At present, the classification method for a single heartbeat only makes a judgment based on the QRS complex, or takes the QRS as the center, searches for the p wave forward and the t wave backward. There is no problem in a normal electrocardiogram. However, in cases such as severe atrioventricular block and non-conduction of the p wave, there is no corresponding QRS wave for the p wave or pacing spike, and it needs to be classified separately; in addition, the combination of different sub-waveforms also represents the important features of this heartbeat. The current solutions do not consider the relationship between the combinations of different sub-waveforms, and there are still many limitations in precise positioning. Summary of the Invention

[0006] In order to solve the above problems, the present disclosure proposes a heartbeat division method based on sub-waveform relationship extraction. Waveform alignment features are extracted through deep learning, and position embedding is added to enhance the influence of position features. Then, combined with the feature extraction matrix, a feature set with sub-waveforms as units is obtained. Two recurrent neural networks are used to respectively generalize and add the temporal features of the heartbeat. The sub-waveform correlation matrix is obtained through matrix multiplication, and the final heartbeat division result is obtained by dividing with sub-waveforms as units.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] A heartbeat division method based on sub-waveform relationship extraction, comprising:

[0009] Obtain a 12-lead electrocardiogram data tensor and annotation data, and after preprocessing, obtain an electrocardiogram signal and a heartbeat division label matrix;

[0010] Input the electrocardiogram (ECG) signal into a deep learning network to obtain alignment features, add ECG sequence position encoding to the alignment features to obtain new alignment features; generate a sub-waveform feature extraction matrix based on the waveform types of sub-waveforms in the labeled data, combine the sub-waveform feature extraction matrix with the new alignment features, and perform dimensional change and sorting to obtain a sub-waveform-based feature tensor.

[0011] Extract sub-waveform time series features from the sub-waveform-based feature tensor, and then use matrix multiplication to generate a sub-waveform relationship matrix. Divide by sub-waveform to obtain the final heartbeat division result and determine the heartbeat to which the sub-waveform belongs.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] A heartbeat division system based on sub-waveform relationship extraction, comprising:

[0014] A preprocessing module for obtaining a 12-lead ECG data tensor and performing preprocessing on the labeled data to obtain an ECG signal and a heartbeat division label matrix.

[0015] A feature extraction module for inputting the ECG signal into a deep learning network to obtain alignment features, adding ECG sequence position encoding to the alignment features to obtain new alignment features.

[0016] A relationship matrix generation module for generating a sub-waveform feature extraction matrix based on the waveform types of sub-waveforms in the labeled data, combining the sub-waveform feature extraction matrix with the new alignment features, and performing dimensional change and sorting to obtain a sub-waveform-based feature tensor; extracting sub-waveform time series features from the sub-waveform-based feature tensor and then using matrix multiplication to generate a sub-waveform relationship matrix.

[0017] A division module for dividing by sub-waveform to obtain the final heartbeat division result and determining the heartbeat to which the sub-waveform belongs.

[0018] Compared with the prior art, the beneficial effects of the present disclosure are:

[0019] The method of the present disclosure extracts waveform alignment features through deep learning, adds position embedding to enhance the influence of position features, and then combines with the feature extraction matrix to obtain a feature set with sub-waveforms as units. The direct tensor multiplication operation using the calculated matrix can improve the training efficiency. Then, two recurrent neural network modules are used to generalize and add the temporal features of heartbeats respectively. Finally, the sub-waveform correlation matrix is obtained through matrix multiplication, and the final model is obtained through backpropagation training. The final heartbeat division result is obtained by dividing with sub-waveforms as units, and the heartbeat to which the sub-waveform belongs is judged, which can solve the problem that in the identification of severe atrioventricular block, non-conducted P wave, etc., there is no corresponding QRS wave for the sub-waveform and subsequent separate classification is impossible; in addition, the combination of different sub-waveforms also represents the important features of this heartbeat, so it can provide more abundant feature guidance for subsequent classification tasks and can also provide heartbeat status guidance on the doctor's analysis interface. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the disclosure. The illustrative embodiments and descriptions thereof of the disclosure are used to explain the disclosure and do not constitute an improper limitation of the disclosure.

[0021] Figure 1 It is a flowchart of the heartbeat division method based on sub-waveform relationship extraction of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are all illustrative and are intended to provide a further description of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless otherwise clearly specified in the context, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Embodiment 1

[0026] In an embodiment of the present disclosure, a heartbeat division method based on sub-waveform relationship extraction is provided, including:

[0027] Step 1: Obtain the 12-lead electrocardiogram data tensor and the labeled data, and perform preprocessing to obtain the electrocardiogram signal and the heartbeat division label matrix;

[0028] Step 2: Input the electrocardiogram (ECG) signal into the deep learning network to obtain the alignment features, add the ECG sequence position encoding to the alignment features, and obtain the new alignment features;

[0029] Step 3: Generate a sub-waveform feature extraction matrix through the waveform types of the sub-waveforms in the labeled data. After combining the sub-waveform feature extraction matrix with the new alignment features and performing dimensional change and sorting, obtain the feature tensor based on the sub-waveforms;

[0030] Step 4: Extract the sub-waveform time series features from the feature tensor based on the sub-waveforms, then generate the sub-waveform relationship matrix using matrix multiplication, divide by the sub-waveform as a unit to obtain the final heartbeat division result, and determine the heartbeat to which the sub-waveform belongs.

[0031] As an embodiment, in Step 1, after preprocessing the 12-lead ECG data tensor and the labeled data, an ECG signal and a heartbeat division label matrix are obtained, including:

[0032] The experimental data uses the actual hospital ECG data, and the heartbeats are labeled by annotators, mainly including two parts:

[0033] 12-lead ECG data ECG_wave: The tensor is the 12-lead ECG data ECG_wave, stored as a tensor feature of [12, wave_len], where the value of the feature is the voltage; the labeled data anno_data contains the sub-waveform type wave_type and the position annotation wave_loc of the 12-lead ECG data ECG_wave waveform, as well as the heartbeat set to which each sub-waveform belongs. The sub-waveforms belonging to the same heartbeat are in the same heartbeat set.

[0034] Obtain the 12-lead ECG data ECG_wave tensor and the labeled data anno_data. The dimension of ECG_wave is (12, wave_len), and through preprocessing, the preprocessed ECG signal ECG_wave and the heartbeat division label matrix label are obtained. The dimension of label is (w_size_total, w_size_total).

[0035] The steps of the preprocessing are as follows:

[0036] 1. Use filtering to remove out-of-band noise;

[0037] 2. Use the labeled data to generate the sub-waveform correlation matrix label for the training label function. The dimension of label is (w_size_total, w_size_total), where w_size_total is the total number of sub-waveforms in this ECG data. label represents which sub-waveform in the second dimension the sub-waveform in the first dimension belongs to. If there is a sub-waveform i belonging to sub-waveform j, then in the matrix label[w_s_i，w_s_j] = 1. w_s_i and w_s_j represent the corresponding positions of i and j in w_size_total. If there is a QRS, the concentric beat waveform points to the QRS. If the beat is a single waveform, the waveform points to itself. If it is an incomplete beat, the waveform points to the first sub-waveform.

[0038] 3. To enhance the robustness of the model, random noise interference is added to ECG_wave, including random lead replacement, random lead replacement with noise, and random zeroing of leads.

[0039] As an embodiment, in step two, the steps of the alignment feature extraction method are as follows:

[0040] Input the electrocardiogram signal into the deep learning network for feature extraction. First, obtain the electrocardiogram features with different receptive fields through multi-scale convolution, and horizontally splice them to obtain the first alignment feature. Pass the first alignment feature through a multi-layer SE network to model the weight relationship between channels, obtain the second alignment feature, and finally input the second alignment feature into a bidirectional recurrent neural network to obtain the forward and backward temporal relationships, and connect a multi-layer perceptron with unchanged forward and backward dimensions to generalize the features to obtain the final alignment feature.

[0041] Specifically,

[0042] 1. Input the 12-lead electrocardiogram signal ECG_wave data into the feature extraction network. First, use a multi-scale convolution scheme with 5 different sizes of convolutional kernels to obtain the electrocardiogram features with different receptive fields. And horizontally splice them to obtain the first alignment feature ECG feature 1, with the dimension (hide1, wave_len), where hide1 represents the dimension of the multi-scale convolution hidden layer features.

[0043] 2. Pass the first alignment feature BCG feature 1 through a multi-layer SE network to model the weight relationship between channels, and obtain the second alignment feature ECG feature 2, with the dimension (hide2, wave_len), where hide2 represents the dimension of the hidden layer features after the multi-layer SE module.

[0044] 3. Finally, input the second alignment feature ECG feature 2 into the bidirectional recurrent neural network to obtain the forward and backward temporal relationships, and then connect a multi-layer perceptron with unchanged forward and backward dimensions to generalize the features to obtain the final alignment feature ECG feature 。

[0045] Each pixel of the second dimension wave_len of the alignment feature corresponds to the second dimension wave of the electrocardiogram signal ECG_wave len

[0046] Then, add an electrocardiogram (ECG) sequence position encoding to the alignment feature to obtain a new alignment feature. The step of adding the ECG sequence position encoding is as follows: Increase the influence of the waveform order in the waveform, and add an ECG position embedding to each dimension of the feature vector.

[0047] Specifically, increase the influence of the waveform order in the waveform, and add an ECG position embedding ecg pos _emb, with a dimension of (5 * emb_size, wave_len), where:

[0048] For each dimension, the value at odd positions 2i + 1 takes:

[0049] For even dimensions 2i, it takes:

[0050] Among them, ecg pos represents the position index of the ECG sequence, and start random is the randomly initialized value of the position embedding formed to prevent the network from memorizing the position pair division result during learning. Let ECG feature = ECG feature + ecg pos _emb.

[0051] As an embodiment, in step three, generate a sub-waveform feature extraction matrix through the waveform types of the sub-waveforms in the labeled data. After combining the sub-waveform feature extraction matrix with the new alignment feature and performing dimensional transformation and sorting, obtain a feature tensor based on the sub-waveform;

[0052] It includes: transform the dimension of the new alignment feature, perform matrix multiplication of the new alignment feature and the feature extraction matrix in the last two dimensions to generate a first sub-waveform feature tensor, then splice the five waveforms and use tensor dimension transformation to perform dimension transformation on the first sub-waveform feature tensor to obtain a second sub-waveform feature tensor, and sort the 5 * w_size_max sub-waveforms in the second sub-waveform feature tensor in the order of appearance time, and remove the padding part filled for alignment matrix to obtain a feature tensor based on the sub-waveform.

[0053] Specifically, first generate a sub-waveform feature extraction matrix, initialize a feature extraction matrix M with a dimension of (5, wave_len, w_size_max), where the elements in the matrix are initialized to 0, 5 represents 5 waveforms, wave_len represents the waveform length, and w_size_max = max(num P , num QRS , num T , num U, num 起搏钉 ) represents the maximum value of the quantity in 5 sub - waveforms, represents the number of waveforms of wave_type_i type in this electrocardiogram. In order to convert the five waveform matrices into the same size, when w_size_max exceeds , the following part is all set to 0 as padding to make up, and its position is recorded.

[0054] Add elements to the matrix. If there is waveform i, and the starting point in wave_len is [strat i , end i , then M[i type , strat i : end i , i w_size_max = 1, where i type represents the sub - waveform category to which waveform i belongs, strat i : end i represents all corresponding points from the starting point to the ending point of heartbeat i, and i w_size_max represents the number of the sub - waveform corresponding to i in the w_size_max of i type . The purpose is to extract an emb_size - dimensional feature vector from each sub - waveform at the corresponding waveform position.

[0055] Based on the feature tensor wave of the sub - waveform feature The generation steps are as follows:

[0056] ECG feature The feature dimension is transformed from (5 * emb_size, wave_len) to (5, emb_size, wave_len).

[0057] Multiply ECG feature and the feature extraction matrix M in the last two dimensions, so that (5, emb_size, wave_len) and (5, wave_len, w_size_max) generate the first sub - waveform feature tensor wave feature 1, which represents the features of w_size_max sub - waveforms in the five waveforms, the feature dimension is emb_size, and the part where w_size_max exceeds the number of such sub - waveforms is set to 0.

[0058] Concatenate the five waveforms and use tensor dimension transformation to make the first sub - waveform feature tensor wave featureThe dimension of 1 changes from (5, emb_size, w_size_max) to (5 * w_size_max, emb_size) to obtain the second sub-waveform feature tensor wave feature 2.

[0059] For the second sub-waveform feature tensor wave feature Sort the 5 * w_size_max sub-waveforms in wave 2 in the order of their appearance time, and remove the padding part filled for matrix alignment to obtain the new wave feature Feature matrix.

[0060] Furthermore, the step of generating the sub-waveform relationship matrix includes:

[0061] Pass the obtained sub-waveform-based feature tensor through a first recurrent neural network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the first feature; pass the obtained sub-waveform-based feature tensor through a second recurrent neural network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the second feature. After dimension transformation, perform matrix multiplication on the first feature and the second feature to obtain the finally extracted sub-waveform relationship matrix.

[0062] Specifically, pass wave feature through the first recurrent neural network module 1 of a bidirectional recurrent network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the first feature feature1, with dimensions (w_size_total, emb_size)

[0063] Pass wave feature through the second recurrent neural network module 2 of a bidirectional recurrent network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the second feature freature2, with dimensions (w_size_total, emb_size). Invert the dimensions to (emb_size, w_size_total) for subsequent matrix multiplication

[0064] Let the first feature feature1 and the second feature feature2 perform matrix multiplication to obtain the finally extracted sub-waveform relationship matrix wave_correlation.

[0065] As an embodiment, the steps of training the deep network model are:

[0066] a) Obtain the 12-lead ECG data ECG_wave tensor and the annotation data anno_data. The dimension of ECG_wave is (12, wave_len), and through preprocessing, the preprocessed ECG signal ECG_wave and the heartbeat division label matrix label are obtained. The dimension of label is (w_size_total, w_size_total).

[0067] b) Pass the ECG_wave preprocessed in a) through the pairwise feature extraction method to obtain the pairwise feature ECG feature , with the dimension of (5 * emb_size, wave_len). Here, 5 represents 5 sub-waveforms: P, QRS, T, U, and pacing spike. emb_size represents the embedding feature dimension taken for each waveform, and wave_len is the length of the ECG waveform, which corresponds one-to-one with the second dimension position of the ECG_wave tensor.

[0068] c) Add the ECG sequence position encoding to ECG feature to enhance the influence of the sequence on the division result. Obtain the new ECG feature feature, with the dimension unchanged.

[0069] d) Generate the sub-waveform feature extraction matrix M through the wave_type of the sub-waveforms in anno_data. The dimension of M is (5, wave_len, w_size_max).

[0070] e) Combine the feature extraction matrix M and ECG feature , and after dimension transformation and sorting, obtain the sub-waveform-based feature tensor wave feature , where the dimension of wave feature is (s_size_total, emb_size), representing the emb_size-dimensional features of each sub-waveform, where w_size_total represents the total number of sub-waveforms.

[0071] f) Pass wave feature through the sub-waveform correlation matrix generation module to obtain the sub-waveform relationship matrix wave_correlation, with the dimension of (w_size_total, w_size_total). The value in the i-th row and j-th column represents the probability value that the sub-waveform represented by the i-th row belongs to the sub-waveform represented by the j-th column.

[0072] g) Input wave_correlation and label into the loss calculation module to obtain the division loss value loss.

[0073] h) Using an optimization algorithm, according to the partition loss value loss, backpropagate to optimize the deep learning network parameters of the feature extraction module and the correlation matrix generation module.

[0074] i) Repeat steps a) to h) until the loss reaches a stable convergence state.

[0075] j) Obtain the electrocardiogram data to be predicted. Using the trained parameter Θ, obtain the predicted relationship matrix through the process of steps a)-f). Store the obtained heartbeat relationship and display it on the doctor's diagnosis page to achieve the purpose of assisting diagnosis.

[0076] Among them, the steps of the loss calculation module in step g) are as follows:

[0077] g-1) Since each sub-waveform in the correlation matrix can only belong to one heartbeat, therefore, for each row, use the softmax function to represent which sub-waveform this sub-waveform may belong to, and obtain the predicted value

[0078] g-2) The negative sample rate of the relationship matrix label is counted, which exceeds 99%. Therefore, add the positive and negative sample balancing strategy of the matrix, and the loss function is:

[0079]

[0080] Among them, represents the predicted sub-waveform relationship matrix, label represents the actual sub-waveform relationship matrix, and ε is to prevent the value of 0 outside the domain from appearing in the logarithm. Among them, sum represents the summation operation, and log represents the logarithm operation. represents the predicted sub-waveform relationship matrix, label represents the actual sub-waveform relationship matrix, and ε is a very small value to prevent the value of 0 outside the domain from appearing in the logarithm.

[0081] Among them, the optimization algorithm selected is adamW.

[0082] Embodiment 2

[0083] In an embodiment of the present disclosure, a heartbeat partitioning system based on sub-waveform relationship extraction is provided, including:

[0084] A preprocessing module for obtaining a 12-lead electrocardiogram data tensor and preprocessing the annotation data to obtain an electrocardiogram signal and a heartbeat partitioning label matrix;

[0085] A feature extraction module for inputting the electrocardiogram signal into a deep learning network to obtain alignment features, and adding electrocardiogram sequence position encoding to the alignment features to obtain new alignment features;

[0086] A relationship matrix generation module, configured to generate a sub-waveform feature extraction matrix through the waveform types of sub-waveforms in the labeled data, combine the sub-waveform feature extraction matrix with new alignment features, perform dimensional change sorting, and obtain a sub-waveform-based feature tensor; extract sub-waveform timing features from the sub-waveform-based feature tensor and generate a sub-waveform relationship matrix using matrix multiplication;

[0087] A division module, configured to perform division in units of sub-waveforms to obtain a final heartbeat division result and determine the heartbeat to which the sub-waveform belongs.

[0088] This disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate a means for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0090] Although the specific implementation manners of the present disclosure have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A heartbeat division method based on sub-waveform relationship extraction, characterized in that, Including: Acquire the 12-lead ECG data tensor and the annotation data, and after preprocessing, obtain the ECG signal and the heartbeat division label matrix; Input the ECG signal into the deep learning network to obtain the alignment features, add the ECG sequence position encoding to the alignment features, and obtain the new alignment features; Generate the sub-waveform feature extraction matrix through the waveform types of the sub-waveforms in the annotation data, combine the sub-waveform feature extraction matrix with the new alignment features, and after dimension change and sorting, obtain the sub-waveform-based feature tensor; extract the sub-waveform time series features from the sub-waveform-based feature tensor and then generate the sub-waveform relationship matrix using matrix multiplication; Perform division in units of sub-waveforms to obtain the final heartbeat division result and determine the heartbeat to which the sub-waveform belongs; The step of generating the sub-waveform relationship matrix includes: Pass the obtained sub-waveform-based feature tensor through a first recurrent neural network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the first feature; Pass the obtained sub-waveform-based feature tensor through a second recurrent neural network with invariant dimensions, and generalize the features using a multi-layer perceptron to obtain the second feature. After dimension transformation, perform matrix multiplication on the first feature and the second feature to obtain the finally extracted sub-waveform relationship matrix.

2. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, The 12-lead electrocardiogram data tensor stores 12-lead electrocardiogram data as tensor features, indicating the waveform length, where the value of the feature in the tensor feature is voltage; the annotation data includes the sub-waveform type and position annotation of the 12-lead electrocardiogram data waveform, and the heartbeat set to which each sub-waveform belongs. Sub-waveforms belonging to the same heartbeat are in the same heartbeat set.

3. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, The method of the preprocessing is: use filtering to filter out the out-of-band noise, and utilize the annotation data to generate the heartbeat division label matrix with the training label function; randomly add noise interference to the 12-lead data.

4. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, The method steps of the alignment feature extraction are: input the ECG signal into the deep learning network for feature extraction. First, obtain the ECG features with different receptive fields through multi-scale convolution, and horizontally splice to obtain the first alignment feature. Pass the first alignment feature through a multi-layer SE network to model the weight relationship between channels to obtain the second alignment feature. Finally, input the second alignment feature into a bidirectional recurrent neural network to obtain the forward and backward time series relationships, and connect a multi-layer perceptron with invariant forward and backward dimensions to generalize the features to obtain the final alignment features.

5. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, The step of adding the ECG sequence position encoding is: increase the influence of the waveform sequence after the waveform, and add the ECG position embedding to each dimension of the feature vector.

6. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, The generation steps of the sub-waveform-based feature tensor are as follows: transform the dimension of the new alignment feature, perform matrix multiplication on the new alignment feature and the feature extraction matrix in the last two dimensions to generate the first sub-waveform feature tensor, then splice five sub-waveforms and perform dimension transformation on the first sub-waveform feature tensor using tensor dimension transformation to obtain the second sub-waveform feature tensor. Sort the sub-waveforms in the second sub-waveform feature tensor in the order of their appearance time, and remove the padding part filled for matrix alignment to obtain the sub-waveform-based feature tensor, where represents the maximum value among the quantities of the five sub-waveforms.

7. The heartbeat division method based on sub-waveform relationship extraction according to claim 1, characterized in that, Optimize the deep learning model, use the optimization algorithm, and according to the division loss value, backpropagate to optimize the deep learning network parameters in the process of obtaining the alignment features and generating the sub-waveform relationship matrix, and repeat the iteration until the loss reaches a stable convergence state.

8. The heartbeat division method based on sub-waveform relationship extraction according to claim 7, characterized in that, The optimization algorithm adopted is .

9. A heartbeat division system based on sub-waveform relationship extraction, characterized in that, Including: A preprocessing module for acquiring the 12-lead ECG data tensor and the annotation data, and after preprocessing, obtaining the ECG signal and the heartbeat division label matrix; A feature extraction module for inputting the ECG signal into the deep learning network to obtain the alignment features, adding the ECG sequence position encoding to the alignment features, and obtaining the new alignment features; A relationship matrix generation module for generating the sub-waveform feature extraction matrix through the waveform types of the sub-waveforms in the annotation data, combining the sub-waveform feature extraction matrix with the new alignment features, and after dimension change and sorting, obtaining the sub-waveform-based feature tensor; extracting the sub-waveform time series features from the sub-waveform-based feature tensor and then generating the sub-waveform relationship matrix using matrix multiplication; A partitioning module, which is used to partition with sub - waveforms as units to obtain the final heart - beat partitioning result and determine the heart - beat to which the sub - waveform belongs; The steps for generating the sub - waveform relationship matrix include: Pass the obtained feature tensor based on sub - waveforms through a first recurrent neural network with invariant dimensions, and generalize the features using a multi - layer perceptron to obtain a first feature; Pass the obtained feature tensor based on sub - waveforms through a second recurrent neural network with invariant dimensions, and generalize the features using a multi - layer perceptron to obtain a second feature. After dimension transformation, perform matrix multiplication on the first feature and the second feature to obtain the finally extracted sub - waveform relationship matrix.

Images