An electrocardiogram (ECG) data augmentation method, system, medium, device, and terminal

Abstract

This invention belongs to the field of data augmentation technology and discloses a method, system, medium, device, and terminal for ECG data augmentation. The system constructs a generative adversarial network (GAN) combining a bidirectional long short-term memory network (LSM) and a channel-spatial attention mechanism to augment ECG data. The GAN consists of a generator and a discriminator, which are adversarial. The generator model has two branches: one containing BiLSTM to better capture the dependencies in the temporal sequence, thus preserving key disease features; and another containing CBAM to adaptively refine the input feature map. Finally, the outputs of the two branches are fused and reshaped into a 300x1 one-dimensional sequence through a deconvolution layer. The discriminator model adopts the DCGAN concept and consists of three Conv1D layers, one Dropout layer, one Flatten layer, and one Dense layer. Through adversarial learning, the generator and discriminator ultimately generate realistic ECG heartbeat data, thereby achieving ECG data augmentation.

Description

Technical Field

[0001] This invention belongs to the field of data augmentation technology, and particularly relates to a method, system, medium, device and terminal for electrocardiogram data augmentation. Background Technology

[0002] Currently, electrocardiography (ECG) can objectively reflect the physiological condition and working state of various parts of the heart, and is an important means and primary basis for diagnosing arrhythmias. The MIT-BIH arrhythmia database is one of the world's most recognized ECG signal databases and is widely used in arrhythmia classification and clinical research. However, the ECG data in the MIT-BIH database suffers from severe imbalance, which may pose significant technical challenges to ECG classification models. Most studies tend to apply classification methods to the primary disease class while neglecting the identification of secondary disease classes. Therefore, to achieve better classification performance, it is urgent to use effective data augmentation techniques for secondary disease class samples.

[0003] Based on the above analysis, the problems and shortcomings of existing technologies are as follows: Current ECG classification methods often apply to the primary disease class while neglecting the identification of secondary disease classes, which leads to relatively poor classification accuracy. Classic data augmentation methods to alleviate data imbalance include translation, rotation, and adding noise, which create synthetic data by modifying the original training set. These methods have been proven to alleviate data imbalance, but the processing may also damage relevant ECG features. Generative models aim to learn the distribution of the input data and then generate new samples from the learned distribution. Their advantage is that they can achieve data augmentation without affecting the original dataset. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method, system, medium, device, and terminal for electrocardiogram (ECG) data enhancement, and particularly relates to an ECG data enhancement method, system, medium, device, and terminal using a BiLSTM+CBAM feature fusion generative adversarial network.

[0005] This invention is implemented as follows: An ECG data augmentation method based on a Generative Adversarial Network (GAN) is constructed by integrating a Bi-directional Long Short-Term Memory (BiLSTM) network with a Convolutional Block Attention Module (CBAM). The GAN consists of a generator and a discriminator. The generator contains a BiLSTM branch to capture dependencies in the temporal sequence, thus preserving key disease features; the CBAM branch is used for adaptive feature refinement of the input feature map. Finally, the features from the two branches are fused, and the sequence is reshaped into a 300x1 one-dimensional sequence through a deconvolutional layer. The discriminator employs the concept of a Deep Convolution Generative Adversarial Network (DCGAN), using a network structure with three 1x3 convolutional layers. During the training process of GAN, the generator and discriminator generate realistic ECG heartbeats through adversarial learning, which are then added to the training set of the classification model to achieve ECG data augmentation.

[0006] Furthermore, the electrocardiogram data augmentation method includes the following steps:

[0007] Step 1, Data Preprocessing: Remove high and low frequencies from the acquired ECG signal to eliminate noise, and apply a Butterworth bandpass filter to extract the most valuable information from the ECG signal;

[0008] Step 2, Data partitioning: Based on the R peak positions marked in the MIT-BIH database, the ECG records are divided into multiple ECG heartbeats. The segmented ECG heartbeats are then divided into training and testing sets.

[0009] Step 3: Construct a BiLSTM+CBAM feature fusion generative adversarial network model;

[0010] Step four: Use evaluation metrics to evaluate the performance of the generative adversarial network model.

[0011] Furthermore, in step one, a Butterworth bandpass filter is applied to the original acquired ECG recordings to remove high and low frequencies and eliminate noise.

[0012] Furthermore, in step two, the ECG data augmentation method is based on 14 disease categories from the MIT-BIH database. Following the recommendations of the Association for the Advancement of Medical Instrumentation (AAMI), records 102, 104, 107, and 217 are discarded. Each record is segmented into multiple heartbeats using the R-peak positions marked in the MIT-BIH database. The heartbeat length consists of 100 sampling points before the R-peak and 200 sampling points after the R-peak, for a total of 300 sampling points including the R-peak. Finally, all heartbeat data are divided into a test set and a training set. During the data augmentation phase, only the heartbeat data from the disease categories in the training set are used.

[0013] Furthermore, in step three, random noise of length 100 is input into the generator of the BiLSTM+CBAM feature fusion generative adversarial network model. The generator uses the data distribution learned from the original heartbeat data to generate new ECG heartbeat samples from the input noise. The discriminator's input is the original heartbeat data, i.e., data from various disease categories. Through iterative training against the generator, the generator network gradually learns how to generate more realistic data, while the discriminator network gradually becomes more accurate. Finally, the output of the entire GAN model is the realistic ECG heartbeat data generated by the generator.

[0014] The generator model consists of two branches. The first branch uses an encoder, a BiLSTM, and a decoder. The encoder comprises two deconvolutional layers with different kernel sizes and the same feature map. The decoder has two deconvolutional layers with different kernel sizes. The second branch uses an encoder, a CBAM, and a decoder. The BiLSTM preserves key disease features, and the CBAM performs adaptive feature refinement on the input feature map. Finally, the outputs of the two branches are fused, and the sequence is then reshaped into a 300x1 one-dimensional sequence through a deconvolutional layer. The Tanh activation function is then used to process the network output.

[0015] Adopting the concept of a DCGAN discriminator, the discriminator model consists of three Conv1D layers, one Dropout layer, one Flatten layer, and one Dense layer. The convolutional kernel size is 1x3, and the feature map sizes are 32, 128, and 256, with a stride of 2. The output of the convolutional layers is mapped to a 1x1 value by a fully connected layer and processed by a sigmoid activation function, serving as the output value of the discriminator.

[0016] For the training process of GAN, the generator uses the Adam optimizer with a learning rate of 0.0005; the discriminator uses the RMSprop optimizer with a learning rate of 0.0001. This invention sets the number of iterations to 500 epochs for most disease categories. It is worth mentioning that more training times for GAN are not necessarily better. Usually, the quality of GAN model training is judged by whether the training of the two networks reaches Nash equilibrium.

[0017] Furthermore, in step four, the percentage root mean square error (PRD), root mean square error (RMSE), and discrete Fréchet distance (FD) are used to measure the generation performance of the BiLSTM+CBAM feature fusion generative adversarial network model. The smaller the value of each indicator, the better the generation performance of the GAN. The formulas for each evaluation indicator are as follows:

[0018]

[0019]

[0020] FD(P,Q)=min{||d||}.

[0021] Another object of the present invention is to provide an electrocardiogram (ECG) data enhancement system applying the aforementioned ECG data enhancement method, the ECG data enhancement system comprising:

[0022] The data preprocessing module applies a Butterworth bandpass filter to the raw ECG records to remove high and low frequencies and eliminate noise.

[0023] The data partitioning module is used to divide ECG heartbeats from the lead II segmentation module of the MIT-BIH database into training and test sets.

[0024] The model building module is used to build a BiLSTM+CBAM feature fusion generative adversarial network model;

[0025] The performance evaluation module is used to evaluate the performance of generative adversarial network models using evaluation metrics.

[0026] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the electrocardiogram data enhancement method.

[0027] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the electrocardiogram data enhancement method.

[0028] Another objective of this invention is to provide an information data processing terminal for implementing the electrocardiogram data enhancement system.

[0029] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0030] First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0031] This invention proposes an ECG data augmentation method using a BiLSTM+CBAM feature fusion generative adversarial network. The GAN consists of a generator and a discriminator, which operate in an adversarial relationship. The generator model in this method comprises two branches: a branch containing BiLSTM better captures the dependencies in the temporal sequence, thus preserving key disease features; and a branch containing CBAM adaptively refines the input feature map. Finally, the outputs of the two branches are fused, and the sequence is reshaped into a 300x1 one-dimensional sequence through a deconvolutional layer. Adopting the DCGAN discriminator concept, the discriminator model consists of three Conv1D layers, one Dropout layer, one Flatten layer, and one Dense layer, with a 1x3 kernel size, feature map sizes of 32, 128, and 256, and a stride of 2. During GAN training, the generator and discriminator learn adversarially, ultimately generating realistic ECG heartbeat data. ECG heart rate data generated by GANs can be added to the training set of classification models to increase the number of training samples, thereby alleviating the problem of data imbalance.

[0032] This invention proposes a BiLSTM+CBAM feature fusion generative adversarial network model. Compared to traditional CGAN and DCGAN models, it has a shorter training time, alleviates the mode collapse problem of traditional GAN ​​models, and reduces the convergence of the losses of the two networks. Mode collapse refers to the phenomenon that the model generates samples that are homogeneous and have poor diversity. In addition, this invention uses 14 disease categories from the MIT-BIH database (https: / / www.physionet.org / content / mitdb / 1.0.0 / ) as input data for the GAN model. The percentage root mean square error (PRD), root mean square error (RMSE), and discrete Fréchet distance (FD) all achieve good results. The PRD values ​​for all 14 disease categories are all below 10, the RMSE values ​​are all below 0.4, and the FD values ​​are all below 1.32.

[0033] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0034] The ECG data augmentation method provided in this invention addresses two key issues: the lack of clear features in one-dimensional ECG data and the need for robust disease-specific features. It combines the advantages of BiLSTM (BiLSTM) to better capture the dependencies between time-series sequences, thus preserving key disease features, and CBAM (CBAM-based feature map adaptive refinement) to construct a feature fusion generative adversarial network. The generator of the proposed BiLSTM+CBAM feature fusion generative adversarial network model can learn from different key aspects, thereby generating realistic ECG heartbeat data. Submitting this realistic heartbeat data to the classification step can alleviate the data imbalance problem.

[0035] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0036] The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0037] The main purpose of this invention is to alleviate the problem of data imbalance in the MIT-BIH database, thereby improving the accuracy of electrocardiogram disease classification. The trained disease classification model can be deployed in real-time lightweight wearable devices and real-time monitoring using ECG equipment in hospitals. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0039] Figure 1 This is a flowchart of the electrocardiogram data enhancement method provided in the embodiments of the present invention;

[0040] Figure 2 This is a schematic diagram of the generator model structure provided in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the discriminator model structure provided in an embodiment of the present invention;

[0042] Figure 4 This is an information interaction diagram of the electrocardiogram data enhancement system provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of 14 types of ECG heartbeats generated by GAN provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] To address the problems existing in the prior art, the present invention provides a method, system, medium, device, and terminal for enhancing electrocardiogram data. The present invention will be described in detail below with reference to the accompanying drawings.

[0046] like Figure 1 As shown, the electrocardiogram data enhancement method provided in this embodiment of the invention includes the following steps:

[0047] S101, Data preprocessing: Apply a Butterworth bandpass filter to the raw acquired ECG records to remove high and low frequencies and eliminate noise;

[0048] S102, Data partitioning: ECG heartbeats segmented from lead II in the MIT-BIH database are divided into training and test sets;

[0049] S103, Construct a BiLSTM+CBAM feature fusion generative adversarial network model;

[0050] S104 uses evaluation metrics to evaluate the performance of generative adversarial network models.

[0051] Preferably, the electrocardiogram data enhancement method provided in this embodiment of the invention specifically includes the following steps:

[0052] 1) Data preprocessing

[0053] The raw ECG data is susceptible to noise interference. This method uses a Butterworth bandpass filter to remove high and low frequencies to eliminate noise. Following AAMI recommendations, records 102, 104, 107, and 217 are discarded. Each record is segmented into multiple beats using the R-peak positions marked in the MIT-BIH database. The beat length consists of 100 sampling points before the R-peak and 200 sampling points after the R-peak, for a total of 300 sampling points including the R-peak.

[0054] 2) Data partitioning

[0055] This invention is based on 14 disease categories in the MIT-BIH database, using only recorded heartbeats from lead II, divided into training and test sets. During the data augmentation phase, only data from the training set was used; detailed data partitioning is shown in Table 1.

[0056] Table 1 Dataset Partitioning

[0057]

[0058]

[0059] 3) BiLSTM+CBAM Feature Fusion Generative Adversarial Network Model

[0060] The generator of a GAN takes 100 units of random noise as input, while the discriminator takes the original heartbeat data as input. The heartbeat data generated by the generator is used as the output of the generative adversarial network, and the discriminator judges whether the generated data is real or fake. During the game, the generator is continuously improved to generate more realistic data, and the discriminator continuously improves its ability to distinguish between real and fake data.

[0061] Regarding the generator architecture, it consists of two branches. The first branch uses an encoder, a BiLSTM, and a decoder. The encoder has two deconvolutional layers with different kernel sizes and the same feature map; the decoder has two deconvolutional layers with different kernel sizes. Multiple deconvolutional layers are used to progressively expand the data dimension, gradually enlarging the image to recover richer image details. The BiLSTM helps capture longer-distance dependencies, thus preserving disease features. The second branch uses an encoder, a CBAM, and a decoder. CBAM performs adaptive feature refinement on the input feature map. Finally, the outputs of the two branches are fused, and then a deconvolutional layer reshapes the sequence into a 300x1 one-dimensional sequence. The Tanh activation function is then used to process the network output.

[0062] The discriminator structure adopts the DCGAN discriminator concept, consisting of three Conv1D layers, one Dropout layer, one Flatten layer, and one Dense layer. The convolutional kernel size is 1x3, and the feature map sizes are 32, 128, and 256, with a stride of 2. The output of the convolutional layers is mapped to a 1x1 value by a fully connected layer and processed by a sigmoid activation function, serving as the output value of the discriminator.

[0063] For the training process of GAN, the generator uses the Adam optimizer with a learning rate of 0.0005; the discriminator uses the RMSprop optimizer with a learning rate of 0.0001. This invention sets the number of iterations to 500 epochs for most disease categories. It is worth mentioning that more training times for GAN are not necessarily better. Usually, the quality of GAN model training is judged by whether the training of the two networks reaches Nash equilibrium.

[0064] It is worth noting that this invention addresses the data augmentation problem. For GANs, the percentage root mean square error (PRD), root mean square error (RMSE), and discrete Fréchet distance (FD) can be used to measure the quality of the generation. The smaller the values ​​of each indicator, the better the generation performance of the GAN.

[0065] like Figure 2 The generator model shown outputs as output = [batch size, convolutional dimension, number of channels], and the deconvolutional layer parameters are Deconv1 = [filter, kernel size, stride]. The generator's input is random noise of length 100, which is first stretched to 576 through a fully connected layer. The one-dimensional sequence is reshaped to map to a 9x64 feature sequence. Then, the sequence length is expanded to 75 through two deconvolutional layers in an encoder structure, followed by a BiLSTM to capture longer-distance dependencies, thus preserving disease features; and then CBAM for adaptive feature refinement of the input feature map. Finally, the sequence length is expanded to 300 through two deconvolutional layers in a decoder structure, resulting in an output dimension of 300x8. After two branches, feature fusion is performed, and the fused output is passed through a deconvolutional layer and a permutation layer to finally obtain a 1x300 dimension feature output. To accelerate network convergence, a batch normalization (BN) layer is added after each deconvolutional layer, followed by a LeakyReLU activation function with a slope of 0.2. The final deconvolutional layer after feature fusion uses the Tanh activation function to process the network output.

[0066] like Figure 3The discriminator model shown outputs as output = [batch size, convolutional dimension, number of channels], and the convolutional layer parameters are Conv1 = [filter, kernel size, stride]. The discriminator's input is 1x300 ECG heart rate data. First, the shape is reshaped to fit the convolutional layer's dimensionality requirements. Then, three convolutional layers with kernel sizes of 1x3, feature map sizes of 32, 28, and 256, and a stride of 2 are used to downsample the data. The output after the convolutional layers is a 38x256 dimensional sequence. A batch normalization (BN) layer is added after each convolutional layer, followed by a LeakyReLU activation function with a slope of 0.2. Finally, the convolutional output is flattened and mapped to 1x1 values ​​through a fully connected layer. The discriminator ultimately uses a sigmoid activation function to map the output value of the fully connected layer to a probability value within the range [0, 1].

[0067] like Figure 4 As shown, the electrocardiogram data enhancement system provided in this embodiment of the invention includes:

[0068] The data preprocessing module is used to apply a Butterworth bandpass filter to the raw ECG records to remove high and low frequencies and eliminate noise.

[0069] The data partitioning module is used to divide ECG heartbeats from the lead II segmentation module of the MIT-BIH database into training and test sets.

[0070] The model building module is used to build a BiLSTM+CBAM feature fusion generative adversarial network model;

[0071] The performance evaluation module is used to evaluate the performance of the generative adversarial network model using evaluation metrics. To demonstrate the inventiveness and technical value of the present invention, this section provides specific product or related technology application examples of the claimed technical solutions.

[0072] Applicable devices: The data-enhanced disease classification model can be used for real-time monitoring in lightweight wearable devices and ECG equipment used in hospitals.

[0073] 1. The data used in the experiment were obtained from the MIT-BIH arrhythmia database (https: / / www.physionet.org / content / mitdb / 1.0.0 / ). For noise reduction preprocessing, the ECG recordings were first passed through a Butterworth bandpass filter.

[0074] 2. 21,077 processed ECG beats were used as training samples, and the remaining 81,021 records were used as test samples. The original training set had a total of 21,077 beats. Data augmentation was performed based on the class imbalance ratio (IR).

[0075] 3. The overall process from data augmentation to classification can be summarized in two stages. The first stage involves inputting disease categories other than the normal class from the original training set into a GAN to generate fake data. Then, a corresponding label is added to each class of fake data and combined with the original training set to create an augmented training set. The original dataset and the augmented dataset are then input into the classification model for training. Finally, the generalization error is evaluated using a test set.

[0076] ECG heart rate data for 14 disease categories generated by the BiLSTM+CBAM feature fusion generative adversarial network provided in this embodiment of the invention are shown below. Figure 5 The evaluation results of the BiLSTM+CBAM feature fusion generative adversarial network are shown in Table 2.

[0077] Table 2 Evaluation Index Results of Generative Adversarial Networks

[0078]

[0079]

[0080] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for enhancing electrocardiogram (ECG) data, characterized in that, The ECG data augmentation method includes: constructing a generative adversarial network (GAN) model that fuses features from a bidirectional long short-term memory (BiLSTM) network and a channel-space attention mechanism (CBAM); the GAN consists of a generator and a discriminator. The generator model contains a branch of BiLSTM to capture the dependencies of time sequences, thereby preserving key disease features; a branch containing CBAM is used to adaptively refine the input feature map; the discriminator uses a network structure with three 1x3 convolutional layers; during the training of the GAN, the generator and discriminator learn adversarially to ultimately generate realistic ECG heartbeat data; finally, the generated ECG heartbeat data can be added to the training set of the classification model to achieve ECG data augmentation. Electrocardiogram (ECG) data augmentation methods include the following steps: Step 1, Data Preprocessing: Apply a Butterworth bandpass filter to the raw ECG recordings to remove high and low frequencies from the raw ECG signals to eliminate noise; Step 2, Data Splitting: The ECG heartbeats segmented from lead II in the MIT-BIH database are divided into training and test sets; Step 3: Construct a BiLSTM+CBAM feature fusion generative adversarial network model; Step four: Evaluate the performance of the generative adversarial network model using evaluation metrics; In step one, a Butterworth bandpass filter is used to remove noise interference from the original ECG data; Based on AAMI's recommendations, records 102, 104, 107, and 217 were discarded; the R-wave has been marked in the MIT-BIH database, and 100 sampling points before the R-wave and 200 sampling points after the R-wave were taken for each heartbeat, for a total of 300 sampling points including the R-wave. In step three, the generator input of the BiLSTM+CBAM feature fusion generative adversarial network model is random noise of length 100, the input of the discriminator is the original heartbeat, the data generated by the generator is used as the output, and the discriminator is used to determine the authenticity of the generated data. The generator model consists of two branches. The first branch uses an encoder, BiLSTM, and decoder. The encoder has two deconvolutional layers with different kernel sizes and the same feature map. The decoder has two deconvolutional layers with different kernel sizes. The second branch uses an encoder, CBAM, and decoder. The BiLSTM preserves the key features of the disease, and the CBAM performs adaptive feature refinement on the input feature map. The outputs of the two branches are then fused. After passing through a deconvolutional layer, the sequence is reshaped into a one-dimensional sequence of length 300x1, and the Tanh activation function is used to process the network output. The discriminator model adopts the idea of ​​DCGAN discriminator, which consists of three Conv1D layers, one Dropout layer, one Flatten layer, and one Dense layer. The convolutional kernel size is 1x3, the feature map sizes are 32, 128, and 256, and the stride is 2. The output of the convolutional layer is mapped to a 1×1 value by the fully connected layer and processed by the sigmoid activation function as the output value of the discriminator. For the training process of GAN, the generator uses the Adam optimizer with a learning rate of 0.0005; the discriminator uses the RMSprop optimizer with a learning rate of 0.0001.

2. The electrocardiogram data enhancement method as described in claim 1, characterized in that, In step two, the ECG data augmentation method is based on 14 disease categories in the MIT-BIH database. Only heartbeats from lead II are used, which are divided into training and test sets. During the data augmentation phase, only heartbeat data from the training set are used.

3. The electrocardiogram data enhancement method as described in claim 1, characterized in that, In step four, the percentage root mean square error (PRD), root mean square error (RMSE), and discrete Fréchet distance (FD) are used to measure the generation performance of the BiLSTM+CBAM feature fusion generative adversarial network model. The smaller the value of each index, the better the generation performance of the GAN.

4. An electrocardiogram (ECG) data enhancement system applying the ECG data enhancement method as described in any one of claims 1 to 3, characterized in that, The electrocardiogram data augmentation system includes: The data preprocessing module is used to apply a Butterworth bandpass filter to the raw ECG records to remove high and low frequencies and eliminate noise. The data partitioning module is used to divide ECG heartbeats from the lead II segmentation module of the MIT-BIH database into training and test sets. The model building module is used to build a BiLSTM+CBAM feature fusion generative adversarial network model; The performance evaluation module is used to evaluate the performance of generative adversarial network models using evaluation metrics.

5. A computer device comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the electrocardiogram data enhancement method as described in any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the electrocardiogram data enhancement method as described in any one of claims 1 to 3.

7. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the electrocardiogram data enhancement system as described in claim 4.

Images