A multi-view electrocardiogram hashing encoding method, system, medium and device

Abstract

The present invention discloses a multi-view electrocardiogram hashing encoding method, system, medium and device, which relates to the technical field of intelligent data learning. The method includes steps of obtaining the original electrocardiogram signals of multiple leads and constructing a first view according to the original signals; constructing a second view by using the autocorrelation function of the original signals, and constructing a third view of the original signals by using the fast Fourier transform; extracting the signal feature representations in each view by using a convolutional neural network; and fusing the features of the three views by using a linear projection technique to obtain the hashing encoding of the electrocardiogram. The present invention can overcome the defects of insufficient signal features of a single view of the electrocardiogram and low fusion efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent data learning, and in particular, to a multi-view electrocardiogram hashing encoding method, system, medium and device. Background Art

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

[0003] An electrocardiogram (ECG) is a diagnostic tool used to reflect the periodicity and time-dependence of myocardial electrical activity on the body surface. Due to its cost-effective and efficient characteristics, ECG has become an indispensable means for clinical diagnosis of cardiovascular diseases. In recent years, with the rapid development of intelligent medical technology, machine learning-based electrocardiogram detection methods have become an important auxiliary tool for medical professionals to diagnose electrocardiograms. Although significant progress has been made in this field, developing lightweight models to further improve the accuracy of electrocardiogram abnormality detection remains a major challenge.

[0004] Most of the current multi-view electrocardiogram (ECG) learning in the research field regards multiple leads of the electrocardiogram as different views based on the differences in the heart positions reflected by different leads. For example, a 12-lead electrocardiogram is divided into 6 signal views, and then a feature extractor is designed for each view, and late integration is achieved through fusion technology. However, it is easy to have the problem of insufficient signal features of a single view, resulting in an unsatisfactory final result.

[0005] In addition, the differences in fusion technology also seriously affect the final result of electrocardiogram learning. Therefore, effectively fusing multi-view features is crucial for improving the performance of deep networks. The existing fusion strategies mainly focus on directly performing weighted averaging on the output features of each sub-network or introducing adaptive parameters to achieve the fusion of different view features, aiming to effectively integrate heterogeneous information. However, these methods often fail to fully consider the diversity and differences between different view features, resulting in the fused features not capturing the core discriminant information.

[0006] In summary, how to overcome the deficiencies of insufficient single-view signal features and low fusion efficiency of electrocardiograms has become an urgent problem to be solved in the prior art. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technologies, the purpose of the present invention is to provide a multi-view electrocardiogram hashing encoding method, system, medium and device. By using autocorrelation technology to construct the second view of the ECG signal, and using the fast Fourier transform to construct the third view, and then constructing a convolutional neural network (CNN) feature extraction framework based on these three different perspectives, and finally using a machine learning algorithm to generate a unified ECG hashing encoding to support subsequent classification and detection tasks.

[0008] To achieve the above purpose, the present invention is realized through the following technical solutions:

[0009] The first aspect of the present invention provides a multi-view electrocardiogram hashing encoding method, including the following steps:

[0010] Obtain the original electrocardiogram signals of multiple leads, and construct the first view according to the original signals;

[0011] Use the autocorrelation function of the original signal to construct the second view, and use the fast Fourier transform to construct the third view of the original signal;

[0012] Use a convolutional neural network to extract the signal feature representations in each view, and obtain the first view feature, the second view feature and the third view feature respectively;

[0013] Adopt a linear projection technique to fuse the features of the three views to obtain the hashing encoding of the electrocardiogram.

[0014] Furthermore, the autocorrelation function of the original signal is used to capture the linear correlation degree of the signal at any different moments, find the periodic features submerged by noise and the discarded fundamental frequencies.

[0015] Furthermore, the specific steps of using the fast Fourier transform to construct the third view of the original signal are:

[0016] Perform the fast Fourier transform on the original signal to realize the conversion from the time domain to the frequency domain, and obtain the frequency domain signal;

[0017] Use a filtering technique to perform noise reduction processing on the frequency domain signal;

[0018] Use the inverse fast Fourier transform to re-convert the frequency domain signal after noise reduction processing to the time domain, so as to obtain the electrocardiogram signal after noise reduction processing as the third view.

[0019] Furthermore, the convolutional neural network structure includes four convolutional layers, three pooling layers, three fully connected layers and two Dropout layers.

[0020] Even further, the specific steps of using a convolutional neural network to extract the signal feature representations in each view are:

[0021] Extract the local features of each lead in each view using a convolutional layer;

[0022] Reduce the feature dimension of the local features through a downsampling pooling layer;

[0023] Use a convolutional layer to perform cross-lead feature fusion on the features with reduced dimension;

[0024] Input the features after cross-lead fusion into a convolutional layer and a fully connected layer to complete the feature extraction of the convolutional neural network.

[0025] Furthermore, the specific steps for fusing the features of the three views using the linear projection technique are as follows:

[0026] Use the linear projection technique to map the features of the three views into the Hamming space respectively;

[0027] Construct an objective function, solve the objective function according to the optimization strategy, and obtain the hash code of the electrocardiogram.

[0028] Even further, the specific steps for constructing the objective function are as follows:

[0029] Use a matrix decomposition function to capture the hash code of each view to obtain a preliminary objective function;

[0030] Define a shared hash code and establish a mapping relationship between the hash code of each view and the shared hash code;

[0031] Introduce a residual factor and perform learning using label information to obtain the final objective function.

[0032] The second aspect of the present invention provides a multi-view electrocardiogram hash coding system, including:

[0033] A data acquisition module, configured to acquire the original electrocardiogram signals of multiple leads and construct a first view according to the original signals;

[0034] A heterogeneous view construction module, configured to construct a second view using the autocorrelation function of the original signal and construct a third view of the original signal using the fast Fourier transform;

[0035] A feature extraction module, configured to extract the signal feature representations in each view using a convolutional neural network to obtain the first view features, the second view features, and the third view features respectively;

[0036] A feature fusion and hash coding module, configured to fuse the features of the three views using the linear projection technique to obtain the hash code of the electrocardiogram.

[0037] In a third aspect of the present invention, there is provided a medium having a program stored thereon, which when executed by a processor implements the steps in the multi-view electrocardiogram hashing encoding method as described in the first aspect of the present invention.

[0038] In a fourth aspect of the present invention, there is provided an apparatus including a memory, a processor, and a program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the steps in the multi-view electrocardiogram hashing encoding method as described in the first aspect of the present invention.

[0039] The above one or more technical solutions have the following beneficial effects:

[0040] The present invention discloses a multi-view electrocardiogram hashing encoding method, system, medium, and apparatus. Most of the current research on multi-view electrocardiogram (ECG) learning in the field is based on the differences in the heart positions reflected by different leads, and multiple leads of the electrocardiogram are regarded as different views. For example, a 12-lead electrocardiogram is divided into 6 signal views, and then a feature extractor is designed for each view, and late integration is achieved through fusion technology. Different from the above methods, the present invention proposes a new strategy, that is, starting from the original electrocardiogram signal to construct multiple views, and then combining the original signals to capture key information. This method can effectively supplement the deficiencies of the single-view signal features.

[0041] In the field of electrocardiogram signal preprocessing, noise reduction techniques are widely used to eliminate various artifacts such as baseline drift, power line interference, electrode movement, and muscle artifacts. From the perspective of spectral analysis, autocorrelation processing can be used as an effective strategy to mitigate white noise interference. The present invention uses the autocorrelation function to process the original signal, and through autocorrelation operation, a second view of the signal rich in periodic components is obtained. In addition, the fast Fourier transform (FFT) can not only realize the conversion of the signal from the time domain to the frequency domain, but also comprehensively reveal the characteristics and structure of the signal through spectral analysis. At the same time, with the help of processing means such as filtering, the FFT can effectively remove noise and significantly improve the anti-interference performance of the signal.

[0042] The present invention extracts features from the three views by constructing a convolutional neural network feature extraction framework, and finally uses a machine learning algorithm to generate a unified ECG hashing code. By introducing a learnable residual factor to construct an objective function and using an optimization strategy to solve it, more efficient feature fusion is achieved, greatly improving the stability and accuracy of the hashing code generation.

[0043] The advantages of the additional aspects of the present invention will be partially given in the following description, partially become apparent from the following description, or be understood through the practice of the present invention. Description of the Drawings

[0044] The accompanying drawings of the specification, which form a part of the present invention, are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0045] Figure 1 It is a comparison chart of the classification effects of three different signals of CNN in PhysioNet / CinC Challenge 2017 in the first embodiment of the present invention;

[0046] Figure 2 It is a multi-view ECG hash coding framework diagram in the first embodiment of the present invention;

[0047] Figure 3 It is a chart showing the proportion of the number of samples in each category of the data set in the first embodiment of the present invention;

[0048] Figure 4 It is a comparison result chart of ACC of each method in each category in the first embodiment of the present invention. Detailed implementation manners

[0049] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. 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 invention belongs.

[0050] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof;

[0051] Embodiment 1:

[0052] Embodiment 1 of the present invention provides a multi-view electrocardiogram hashing encoding method. In the field of electrocardiogram signal preprocessing, noise reduction techniques are widely used to eliminate various artifacts such as baseline drift, power line interference, electrode movement, and muscle artifacts. From the perspective of spectral analysis, autocorrelation processing can be used as an effective strategy to mitigate white noise interference. In this embodiment, the autocorrelation function is used to process the original signal, and a second view of the signal rich in periodic components is obtained through autocorrelation operation. In addition, the fast Fourier transform (FFT) can not only transform the signal from the time domain to the frequency domain, but also comprehensively reveal the characteristics and structure of the signal through spectral analysis. At the same time, with the help of processing means such as filtering, FFT can effectively remove noise and significantly improve the anti-interference performance of the signal. Therefore, in this embodiment, FFT is used to construct a third view of the signal. Experiments based on the dataset PhysioNet / CinC Challenge 2017 prove that, as Figure 1 shown, there are significant differences in the contributions of different views to the recognition rates of various categories, so combining multiple views helps to capture more discriminative features.

[0053] The multi-view ECG hashing encoding method in this embodiment includes: First, two new views of the original signal are constructed through the autocorrelation function and fast Fourier transform (FFT) of the signal; Second, the convolutional neural network (CNN) is used to extract the signal feature representations in each view; Third, the linear projection technology is used to fuse the three extracted features; Finally, the denoised and periodic-rich signal hashing encoding is realized, and the mapping relationship between the hash dictionary and the label information is established.

[0054] To achieve the above goal, given an ECG dataset , where represents the training example set containing examples, represents the set of real numbers, represents the th example, represents the corresponding onehot label set, represents the label corresponding to the th example, represents the number of categories; and represents the feature dimension of the training example, represents the number of channels or leads, is the number of training signals. The goal of the present invention is to find a mapping to realize the encoded representation of the ECG signal, and then be able to quickly realize the recognition and detection of the signal. The whole method consists of three parts: the heterogeneous view construction module, the CNN feature extraction module, and the feature fusion and hashing encoding module.

[0055] As Figure 2As shown in the figure, it specifically includes the following steps:

[0056] Step 1: Obtain the original electrocardiogram signals of multiple leads and construct a first view based on the original signals.

[0057] Specifically, the original signals of multiple leads are constructed into a view, namely the first view. The purpose of doing this is to maintain the correlation between multiple leads of the same signal. Before inputting into the convolutional neural network, the first view is normalized.

[0058] Step 2: Construct a second view using the autocorrelation function of the original signal and construct a third view of the original signal using the fast Fourier transform.

[0059] Step 2.1: The autocorrelation function of the original signal is used to capture the linear correlation degree of the signal at any different moments, find the periodic features that may be submerged by noise, and the fundamental frequencies that may be discarded.

[0060] In this embodiment, the definition of the autocorrelation function is:

[0061] (1).

[0062] Where, is the autocorrelation function, is the expectation operator, and represent two different moments. Using the above function, the second view of the original signal can be calculated. It can be seen from the results of Figure 1 that the signal under this view has a particularly obvious noise reduction effect.

[0063] (2).

[0064] Step 2.2: Construct a third view of the original signal using the fast Fourier transform.

[0065] Step 2.2.1: Perform a fast Fourier transform (FFT) on the original signal to achieve the conversion from the time domain to the frequency domain and obtain the frequency domain signal.

[0066] The FFT technology can reveal the features that are difficult to identify in the time domain. By converting the signal to the frequency domain, the key spectral information is captured, thus significantly improving the distinguishability of the features.

[0067] Step 2.2.2: Use filtering technology to perform noise reduction processing on the frequency domain signal to effectively remove the noise components.

[0068] Step 2.2.3: Use the inverse fast Fourier transform (IFFT) to convert the denoised frequency-domain signal back to the time domain, thereby obtaining the denoised electrocardiogram signal as the third view. 。

[0069] (3)。

[0070] where represents the fast Fourier transform operator, represents the filter, represents the inverse fast Fourier transform operator.

[0071] Thus, three views of the ECG signal can be constructed, thereby constructing a three-view dataset , represents the view number, and , represents the training example, represents the corresponding multi-view shared label.

[0072] Step 3: Use a convolutional neural network to extract the signal feature representations in each view, respectively obtaining the first view feature, the second view feature, and the third view feature.

[0073] In this embodiment, there are three feature extraction networks. The three CNN feature extraction networks have the same structure. For each view, the convolutional neural network structure includes four convolutional layers, three pooling layers, three fully connected layers, and two Dropout layers. The network integrates the features of different-lead electrocardiogram (ECG) signals through the following steps:

[0074] Step 3.1: Use convolutional layers to extract the local features of each lead in each view.

[0075] Specifically, the multi-lead ECG heartbeat signals are input into two convolutional layers with a convolutional kernel size of 1*3 to extract the local features of each lead.

[0076] Step 3.2: Reduce the feature dimension of the local features through a downsampling pooling layer to simplify the model and reduce the risk of overfitting.

[0077] Step 3.3: Use convolutional layers to perform cross-lead feature fusion on the features with reduced dimensions.

[0078] Use a convolutional layer with a convolutional kernel size of 2*1 for feature fusion, so that the features of any two leads can be combined.

[0079] Step 3.4: Input the cross-lead fused features into convolutional layers and fully connected layers to complete the convolutional neural network feature extraction.

[0080] The features after cross - lead fusion are input into a 1×1 convolutional layer and the subsequent three fully - connected layers to complete the feature extraction of the convolutional neural network. The features extracted by the CNN are .

[0081] Step 4: Adopt the linear projection technique to fuse the first - view features, second - view features, and third - view features to obtain the hash code of the electrocardiogram.

[0082] Step 4.1: Use the linear projection technique to map the first - view features, second - view features, and third - view features into the Hamming space respectively.

[0083] This embodiment aims to represent the CNN features as a shared hash code. Given the essential differences in the features of each view, directly mapping to the shared hash code will cause energy loss. Therefore, the features of each view are mapped into the Hamming space respectively and then fused.

[0084] Step 4.2: Construct the objective function and solve the objective function according to the optimization strategy to obtain the hash code of the electrocardiogram.

[0085] Step 4.2.1: Use the matrix decomposition function to capture the hash code of each view to obtain the preliminary objective function.

[0086] Use the matrix decomposition function to capture the hash code of each view. The specific objective function to be minimized is:

[0087] (4).

[0088] Where is the identity matrix of order is the orthogonal projection factor, is the hash representation of each view.

[0089] Step 4.2.2: Define the shared hash code and establish a mapping relationship between the hash code of each view and the shared hash code.

[0090] Assume that the shared hash code is , and establish the following mapping relationship between the hash code of each view and the shared hash code:

[0091] (5).

[0092] Where represents the objective function to be minimized, is the correction factor between the shared hash code and the single - view hash code.

[0093] Step 4.2.3: Introduce a residual factor and use label information for learning to obtain the final objective function.

[0094] To further improve the discriminability of the hash code, label information is used for learning. Given that the label matrix is a one-hot matrix with most of its elements being zero, directly establishing a mapping relationship between the label matrix and the hash code may result in significant energy loss and is not conducive to improving the discriminative performance of the hash code. Therefore, a learnable residual factor is introduced to reduce the difference between the two and optimize the discriminative ability of the hash code. The specific method is as follows:

[0095] (6).

[0096] where represents the objective function to be minimized, is an orthogonal basis used to learn the semantic hash code representation. And is the learnable residual factor. To increase the difference between heterogeneous classes and reduce the numerical difference between the Hamming space and the label space, let the initial . To simplify the optimization, let denote the label with the residual factor, then the objective can be rewritten as:

[0097] (7).

[0098] where represents the parameter of the regularization term to prevent overfitting.

[0099] In summary, the overall objective function can be defined as:

[0100] (8).

[0101] where, respectively represent different objective functions to be minimized.

[0102] Step 4.2.4: Solve the objective function according to the optimization strategy.

[0103] Next, the optimization strategy for the objective is given. Due to the binary discrete constraint in the Hamming space, the objective function is non-differentiable, resulting in , both being difficult to optimize, while and both have orthogonal constraints and cannot directly use gradient descent to optimize the objective. Therefore, different optimization schemes are set for each variable to achieve the objective Overall optimization.

[0104] Update : Remove the terms irrelevant to from the objective function, fix and , and the objective can be simplified to the following optimization sub - problem:

[0105] (9).

[0106] Using the gradient of , it is easy to derive its optimal value as . According to the definition of the residual factor, the expected label to be learned is expected to contain the true label elements. Therefore, according to to revise , the update rule of

[0107] (10).

[0108] Update : Remove the terms irrelevant to from the objective function, fix , and the objective can be simplified to the following optimization sub - problem:

[0109] (11).

[0110] Since , so , and thus the objective is further simplified to:

[0111] (12).

[0112] The problem corresponds to the classical orthogonal Procrustes problem. The optimization objective can be achieved by using the singular value decomposition (SVD), , so can be used to calculate :

[0113] (13).

[0114] Update : Remove the terms irrelevant to from the objective function, fix , and , so the objective can be simplified to the following optimization sub - problem:

[0115] (14).

[0116] Since , so , thus the target can be simplified to:

[0117] (15).

[0118] From the inequality relationship, it can be known that:

[0119] (16).

[0120] Update : Eliminate the terms irrelevant to in the objective function, fix , then the target is simplified to the following optimization sub-problem:

[0121] (17).

[0122] The discrete constraint of makes it difficult to optimize the target . Therefore, relax the constraint of so that

[0123] (18).

[0124] Using Equation we can get:

[0125] (19).

[0126] Update : Eliminate the terms irrelevant to in the objective function, fix , then the target can be simplified to the following optimization sub-problem:

[0127] (20).

[0128] Since ， , so , the above equation can be further simplified to:

[0129] (21).

[0130] Through a SVD update strategy similar to , let then there is the following update rule:

[0131] (22).

[0132] Update : Eliminate the terms irrelevant to in the objective function, fix , and according to , the objective can be simplified to the following optimization sub - problem:

[0133] (23).

[0134] From the inequality relationship, it can be seen that:

[0135] (24).

[0136] To verify the effectiveness of the proposed multi - view ECG collaborative hashing coding, the publicly available 2017 PhysioNet / CinC Challenge ECG dataset is used for verification. The dataset consists of 8528 ECG segments of 30 - 60 seconds, including four categories: normal sinus rhythm , atrial fibrillation , other heart rhythms , and noisy ECG .

[0137] To verify the discriminative ability of the ECG collaborative hashing coding, 6000 samples are selected for training, and the rest are used for testing. An SVM classifier and several popular deep models are selected for comparison. The deep models include: BiLSTM, CNN, CNNBiLSTM. Since the experimental dataset classes are extremely imbalanced, as Figure 3 shown, normal sinus rhythm accounts for 59%, atrial fibrillation accounts for 9%, other heart rhythms account for 29%, and noisy ECG accounts for 3%. Therefore, multiple metrics are selected for the experiment. The metrics include: ACC, F1, G_means, Precise, Recall, and Specificity. By comparing the predicted classes and the actual classes, four key parameters are obtained: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). Then the above metrics are defined as follows:

[0138] .

[0139] .

[0140] .

[0141] Table 1 Comparison of each method under multiple evaluation criteria

[0142]

[0143] The experimental results are shown in Table 1, indicating that the multi-view ECG hashing coding method proposed in this embodiment can effectively make up for the deficiencies of single-view signal features by fusing multi-view features. Under multiple evaluation criteria, this method is significantly better than the current popular deep models. The multi-view ECG hashing coding fully considers the diversity and differences between different view features, effectively integrates heterogeneous information, and thus can capture the most core discriminant information. Figure 4 The experimental results of [reference] show that the fusion of multi-view features is particularly significant for improving the recognition effect of low-recognition-rate categories, which is consistent with Figure 1 the experimental results of [reference].

[0144] Embodiment 2:

[0145] Embodiment 2 of the present invention provides a multi-view electrocardiogram hashing coding system, including:

[0146] A data acquisition module, configured to acquire the original electrocardiogram signals of multiple leads and construct a first view according to the original signals;

[0147] A heterogeneous view construction module, configured to construct a second view using the autocorrelation function of the original signal and construct a third view of the original signal using the fast Fourier transform;

[0148] A feature extraction module, configured to extract the signal feature representations in each view using a convolutional neural network to obtain a first view feature, a second view feature, and a third view feature respectively;

[0149] A feature fusion and hashing coding module, configured to fuse the features of the three views using a linear projection technique to obtain the hashing coding of the electrocardiogram.

[0150] Embodiment 3:

[0151] Embodiment 3 of the present invention provides a medium, on which a program is stored, and when the program is executed by a processor, it implements the steps in the multi-view electrocardiogram hashing coding method as described in Embodiment 1 of the present invention.

[0152] Embodiment 4:

[0153] Embodiment 4 of the present invention provides a device, including a memory, a processor, and a program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-view electrocardiogram hashing coding method as described in Embodiment 1 of the present invention.

[0154] The steps involved in Embodiments 2, 3, and 4 above correspond to those in Method Embodiment 1. For specific implementation manners, reference may be made to the relevant description part of Embodiment 1.

[0155] Those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0156] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, 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 invention.

Claims

1. A multi-view electrocardiogram hash coding method, characterized in that: The following steps are involved: Acquire multi-lead electrocardiogram original signals, and construct a first view according to the original signals; A second view is constructed using an autocorrelation function of the original signal, and a third view of the original signal is constructed using a fast Fourier transform, filtering, and an inverse fast Fourier transform; A convolutional neural network is used to extract the signal feature representation in each view, and the first view feature, the second view feature and the third view feature are obtained respectively; The features of the three views are combined to obtain the hash code of the ECG; Specifically: The three view features are mapped to the Hamming space respectively using linear projection technology; Construct the objective function, solve the objective function according to the optimization strategy, and obtain the hash code of the electrocardiogram; the specific steps of constructing the objective function are: Use matrix decomposition function to capture the hash code of each view and get the preliminary objective function; Define a shared hash code and establish a mapping relationship between the hash code of each view and the shared hash code; The residual factor is introduced, and the label information is used for learning to obtain the final objective function.

2. The multi-view electrocardiogram hash coding method according to claim 1, characterized in that: The autocorrelation function of the original signal is used to capture the linear correlation of the signal at any different time, and to find the periodic features submerged by the noise and the discarded fundamental frequency.

3. The multi-view electrocardiogram hash coding method according to claim 1, characterized in that: The specific steps of constructing the third view of the original signal using fast Fourier transform, filtering, and inverse fast Fourier transform are: Perform fast Fourier transform on the original signal to achieve conversion from time domain to frequency domain and obtain frequency domain signal; Use filtering technology to reduce noise on frequency domain signals; The frequency domain signal after noise reduction is converted back to the time domain by inverse fast Fourier transform, so as to obtain the denoised electrocardiogram signal as the third view.

4. The multi-view electrocardiogram hash coding method according to claim 1, characterized in that: The convolutional neural network structure consists of four convolutional layers, three pooling layers, three fully connected layers, and two Dropout layers.

5. The multi-view electrocardiogram hash coding method as claimed in claim 4, characterized in that: The specific steps of using convolutional neural network to extract signal feature representation in each view are: The convolutional layer is used to extract the local features of each lead in each view; Reduce the feature dimension of local features through a downsampling pooling layer; Use convolutional layers to perform cross-channel feature fusion on the features after dimensionality reduction; The cross-lead fused features are input into the convolutional layer and the fully connected layer to complete the convolutional neural network feature extraction.

6. A multi-view electrocardiogram hash coding system, characterized in that: include: A data acquisition module is configured to acquire original electrocardiogram signals of multiple leads and construct a first view according to the original signals; A heterogeneous view construction module is configured to construct a second view using an autocorrelation function of the original signal and to construct a third view of the original signal using a fast Fourier transform, filtering, and an inverse fast Fourier transform; A feature extraction module is configured to extract a signal feature representation in each view using a convolutional neural network to obtain a first view feature, a second view feature, and a third view feature respectively; The feature fusion and hash coding module is configured to fuse the three view features to obtain the hash code of the ECG; Specifically: The three view features are mapped to the Hamming space respectively using linear projection technology; Construct the objective function, solve the objective function according to the optimization strategy, and obtain the hash code of the electrocardiogram; the specific steps of constructing the objective function are: Use matrix decomposition function to capture the hash code of each view and get the preliminary objective function; Define a shared hash code and establish a mapping relationship between the hash code of each view and the shared hash code; The residual factor is introduced, and the label information is used for learning to obtain the final objective function.

7. A computer-readable storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor of a terminal device and executing the multi-view electrocardiogram hash coding method described in any one of claims 1-5.

8. A terminal device, characterized in that: The invention comprises a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store a plurality of instructions, wherein the instructions are suitable for being loaded by the processor and executing the multi-view electrocardiogram hash coding method according to any one of claims 1 to 5.

Images