Electrocardiogram detection and analysis method, device, electronic equipment and storage medium

Abstract

The present application provides an electrocardiogram detection and analysis method, device, electronic device and storage medium, including: performing key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine the first electrocardiogram feature; performing frequency domain processing on the electrocardiogram signal based on discrete wavelet transform to determine the second electrocardiogram feature; performing deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine the third electrocardiogram feature; performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature and the third electrocardiogram feature, and determining the detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram feature. By integrating multiple electrocardiogram features to perform multi-source signal enhancement and adaptive fusion, the fused features can be used to assist doctors in locating pathological features, thereby improving the accuracy of pathological feature positioning.

Description

Technical Field

[0001] The present application relates to the technical field of automatic electrocardiogram detection, and in particular to an electrocardiogram detection and analysis method, device, electronic equipment and storage medium. Background Art

[0002] As the most common diagnostic tool for heart disease in clinical practice, the electrocardiogram (ECG) plays a crucial role in the diagnosis and treatment of various heart diseases. Traditional ECG analysis methods rely on manually designed static features, such as the PR interval, QRS amplitude, and ST segment slope. Because abnormal electrophysiological activity in children with CHD often manifests as multi-lead coordinated dynamic waveform distortion, traditional ECG analysis methods have significant flaws in the detection of pediatric congenital heart disease. Therefore, improving the accuracy of ECG detection and analysis has become a technical issue that cannot be underestimated. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide an electrocardiogram detection and analysis method, device, electronic device and storage medium, which integrates multiple electrocardiogram features to perform multi-source signal enhancement and adaptive fusion, so that the fused features can be used to assist doctors in locating pathological features and improve the accuracy of pathological feature positioning.

[0004] The present invention provides an electrocardiogram (ECG) detection and analysis method, which includes:

[0005] performing key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal;

[0006] Processing the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal;

[0007] Deep learning implicit feature extraction is performed on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes;

[0008] The first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature are subjected to feature fusion, so that a detection and analysis result corresponding to the electrocardiogram is determined based on the fused electrocardiogram features.

[0009] In one possible implementation, the performing of deep learning implicit feature extraction on the electrocardiogram signal based on the multi-scale one-dimensional temporal convolution residual topology unit and the dynamic temporal attention mechanism to determine the third electrocardiogram feature of the electrocardiogram signal includes:

[0010] Performing time domain dimension processing on the electrocardiogram signal based on multiple one-dimensional time series convolution modules in a multi-scale one-dimensional time series convolution residual topology unit, so that each of the one-dimensional time series convolution modules outputs a time domain feature of the electrocardiogram signal;

[0011] Performing temporal convolution processing on the multiple temporal features based on a dynamic temporal attention mechanism to generate an attention weight matrix;

[0012] Performing probability space mapping processing on the attention weight matrix based on the activation function to determine the probability value of each of the time domain features;

[0013] Each of the domain features is fused based on the probability value to determine a third electrocardiogram feature of the electrocardiogram signal.

[0014] In a possible implementation, for any of the one-dimensional temporal convolution modules, the multiple one-dimensional temporal convolution modules in the multi-scale one-dimensional temporal convolution residual topology unit perform time domain dimension processing on the electrocardiogram signal, so that each of the one-dimensional temporal convolution modules outputs the time domain features of the electrocardiogram signal, including:

[0015] Based on the first temporal convolution subunit in the one-dimensional temporal convolution module, performing convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature;

[0016] Continue to process the target convolution feature based on the next time series convolution subunit until each time series convolution subunit in the one-dimensional time series convolution module has completed processing, and then output the time domain feature of the electrocardiogram signal.

[0017] In a possible implementation, the first temporal convolution subunit in the one-dimensional temporal convolution module performs convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature, including:

[0018] Performing one-dimensional convolution processing of a first kernel size and a second kernel size on the electrocardiogram signal in a time domain dimension to determine a first convolution feature and a second convolution feature;

[0019] Normalizing the first convolution feature to determine a normalized first convolution feature, and further performing one-dimensional convolution and normalization on the normalized first convolution feature to determine a third convolution feature;

[0020] Performing a flat pooling process on the second convolution feature, and processing the second convolution feature after the flat pooling process and the third convolution feature based on an activation function to determine the target convolution feature.

[0021] In a possible implementation manner, after determining the third electrocardiogram feature of the electrocardiogram signal, the detection and analysis method further includes:

[0022] The third electrocardiogram feature is processed based on the spatial correlation relationship between the feature gradient extracted by back propagation and the attention weight matrix to generate a time-sensitive heat map.

[0023] In a possible implementation, the performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature includes:

[0024] performing normalization processing on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature to determine a normalized electrocardiogram feature;

[0025] Based on the dual-path feature Transformer network layer, attention Transformer network layer and activation function, the normalized ECG features are fused and the fused ECG features are output.

[0026] In one possible implementation, the dual-path-based feature Transformer network layer, the attention Transformer network layer, and the activation function perform feature fusion processing on the normalized electrocardiogram features and output the fused electrocardiogram features, including:

[0027] Based on the first feature Transformer network layer, key feature selection is performed on the normalized ECG features to determine multiple first target ECG features;

[0028] performing nonlinear correlation processing on the plurality of first target ECG features based on the attention Transformer network layer to determine nonlinear spatial correlation features between the plurality of target ECG features;

[0029] performing a dot product operation on the plurality of nonlinear spatial correlation features and the plurality of target ECG features to determine a plurality of second target ECG features;

[0030] Processing the plurality of second target ECG features based on the second feature Transformer network layer to determine a plurality of third target ECG features, and processing the plurality of third target ECG features based on the activation function to determine a plurality of fourth target ECG features;

[0031] Based on the attention Transformer network layer, the second feature Transformer network layer and the activation function, the multiple fourth target ECG features are further processed to determine the multiple fifth target ECG features, and the multiple fifth target ECG features are fused to determine the fused ECG features.

[0032] The present application also provides an electrocardiogram detection and analysis device, the detection and analysis device comprising:

[0033] a spatial positioning module, configured to perform key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template, and determine a first electrocardiogram feature of the electrocardiogram signal;

[0034] a frequency domain feature extraction module, configured to process the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal;

[0035] an implicit feature extraction module, configured to perform deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes;

[0036] The feature fusion module is used to perform multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature and the third electrocardiogram feature, so as to determine the detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features.

[0037] An embodiment of the present application also provides an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of the electrocardiogram detection and analysis method as described above are performed.

[0038] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the electrocardiogram detection and analysis method as described above are executed.

[0039] An embodiment of the present application provides an electrocardiogram detection and analysis method, device, electronic device and storage medium, the detection and analysis method comprising: performing key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal; performing frequency domain processing on the electrocardiogram signal based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal; performing deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes; performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature and the third electrocardiogram feature to determine the detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features. By integrating multiple electrocardiogram features to enhance multi-source signals and adaptively fusing them, the fused features can be used to assist doctors in locating pathological features, thereby improving the accuracy of pathological feature positioning.

[0040] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 A flowchart of an electrocardiogram detection and analysis method provided in an embodiment of the present application;

[0043] Figure 2 A processing diagram of the temporal convolution subunit provided in an embodiment of the present application;

[0044] Figure 3 A schematic diagram of the processing of multi-source feature fusion provided in an embodiment of the present application;

[0045] Figure 4 This is one of the structural schematic diagrams of an electrocardiogram detection and analysis device provided in an embodiment of the present application;

[0046] Figure 5 This is a second structural diagram of an electrocardiogram detection and analysis device provided in an embodiment of the present application;

[0047] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work falls within the scope of protection of the present application.

[0049] First, the application scenarios to which this application is applicable are introduced. This application can be applied in the field of electrocardiogram automatic detection technology.

[0050] Research has found that the electrocardiogram (ECG), the most common diagnostic tool for heart disease in clinical practice, plays a crucial role in the diagnosis and treatment of various heart diseases. Traditional ECG analysis methods rely on manually designed static features, such as the PR interval, QRS amplitude, and ST segment slope. Because abnormal electrophysiological activity in children with CHD often manifests as multi-lead coordinated dynamic waveform distortion, traditional ECG analysis methods have significant flaws in the detection of pediatric congenital heart disease. Therefore, improving the accuracy of ECG detection and analysis has become a technical issue that cannot be underestimated.

[0051] Based on this, an embodiment of the present application provides an electrocardiogram detection and analysis method, which integrates multiple electrocardiogram features to perform multi-source signal enhancement and adaptive fusion, so that the fused features can be used to assist doctors in locating pathological features and improve the accuracy of pathological feature positioning.

[0052] See also Figure 1 , Figure 1 This is a flow chart of an electrocardiogram detection and analysis method provided in an embodiment of the present application. Figure 1 As shown in , the detection and analysis method provided in the embodiment of the present application includes:

[0053] S101: performing key feature positioning processing on an electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal.

[0054] In this step, a key feature positioning process is performed on the electrocardiogram signal using a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal.

[0055] Among them, the eleven-point ECG spatial positioning template adds the starting and ending point information of three bands compared to the traditional five key point recognition of P, QRS, and T, and baselines more standard information to reduce the calculation complexity of other relative features, including R wave peak point, Q wave peak point, S wave peak point, QRS complex starting point, QRS complex end point, P wave peak point, P wave starting point, P wave starting point, P wave end point, T wave peak point, T wave starting point, and T wave end point.

[0056] Specifically, the slope is highlighted by differential operation , highlighting the peak value by square operation , sliding window integral smoothing energy envelope First, the QRS wave group candidate area is marked by threshold detection. The first time the slope significantly exceeds the critical value of significant change in the candidate area is the QRS starting point. The first time the slope falls below the critical value of significant change is the end point of QRS. , The maximum value position is the peak position of the R wave Based on the located R peak, search for P wave in the fixed window before the R peak, where the peak position is the position of the P wave peak. Similarly, search for T wave in the fixed window after the R peak, where the peak position is the position of the T wave peak. The starting and ending positions are determined by the slope flat threshold. , the site of the regression baseline threshold and local baseline levels for positioning.

[0057] Here, the eleven-point ECG spatial positioning template performs key feature location processing on the ECG signal to determine the first ECG feature of the ECG signal. This first ECG feature includes the average P wave peak value, the average Q wave peak value, the average R wave peak value, the average S wave peak value, the average T wave peak value, the average QRS complex integral value, the average T wave integral value, the average P wave integral value, the ratio of the average P wave peak value to the average R wave peak value, and other features.

[0058] In this application, 188 characteristic indicators (lead distinction calculation) were ultimately defined and extracted, comprehensively covering multiple dimensions such as waveform amplitude, time interval, morphological changes, and generating some indirect comparison indicators. The goal is to deeply characterize the physiological significance and pathological characteristics of ECG signals from a clinical perspective, providing solid support for the interpretability and application of subsequent diagnostic models. In addition, in the subsequent model, multiple first ECG features are integrated with implicit deep learning features derived from the original ECG signal and frequency domain features derived from wavelet decomposition, thereby forming a comprehensive outline of ECG information.

[0059] S102: Process the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal.

[0060] In this step, the discrete wavelet transform is used to extract frequency domain features of the ECG signal to capture the energy distribution characteristics of different frequency bands. Specifically, the Daubechies wavelet function is used to perform a 9-layer wavelet decomposition of the ECG signal, decomposing the original signal into approximate coefficients and detail coefficients at different scales. To more effectively characterize the frequency domain characteristics of the signal, the wavelet coefficients of layers 4 to 8 and the energy value E of the approximate coefficients of the 9th layer are extracted as features. These coefficients can cover the most diagnostically valuable frequency bands in the ECG signal (such as the frequency range corresponding to the P wave, QRS complex, and T wave). Frequency domain feature extraction is performed in the most diagnostically valuable frequency bands, ultimately obtaining 72-dimensional features for 12 leads, which are used to describe the energy distribution of the signal at various scales. These features provide a solid frequency domain information foundation for subsequent classification or pattern recognition tasks.

[0061] S103: Based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism, deep learning implicit feature extraction is performed on the electrocardiogram signal to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of multiple one-dimensional temporal convolution modules with different kernel sizes.

[0062] In this step, the multi-scale one-dimensional temporal convolution residual topology unit and the dynamic temporal attention mechanism are used to perform deep learning implicit feature extraction on the ECG signal to determine the third ECG feature of the ECG signal.

[0063] The kernel size can be set to 3, 5, 7 or other sizes.

[0064] In one possible implementation, the performing of deep learning implicit feature extraction on the electrocardiogram signal based on the multi-scale one-dimensional temporal convolution residual topology unit and the dynamic temporal attention mechanism to determine the third electrocardiogram feature of the electrocardiogram signal includes:

[0065] A: Based on multiple one-dimensional temporal convolution modules in a multi-scale one-dimensional temporal convolution residual topology unit, the electrocardiogram signal is processed in the time domain dimension so that each of the one-dimensional temporal convolution modules outputs the time domain features of the electrocardiogram signal.

[0066] Here, multiple one-dimensional time series convolution modules are used to perform time domain dimension processing on the electrocardiogram signal, so that each one-dimensional time series convolution module outputs the time domain features of the electrocardiogram signal.

[0067] In a possible implementation, for any of the one-dimensional temporal convolution modules, the multiple one-dimensional temporal convolution modules in the multi-scale one-dimensional temporal convolution residual topology unit perform time domain dimension processing on the electrocardiogram signal, so that each of the one-dimensional temporal convolution modules outputs the time domain features of the electrocardiogram signal, including:

[0068] (1): Based on the first time series convolution subunit in the one-dimensional time series convolution module, the electrocardiogram signal is subjected to convolution processing, normalization processing, average pooling processing and activation function processing in the time domain dimension to determine the target convolution feature.

[0069] Here, the first temporal convolution subunit in the one-dimensional temporal convolution module performs convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature.

[0070] In a possible implementation, the first temporal convolution subunit in the one-dimensional temporal convolution module performs convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature, including:

[0071] a: performing one-dimensional convolution processing of a first kernel size and a second kernel size on the electrocardiogram signal in the time domain dimension to determine a first convolution feature and a second convolution feature.

[0072] Here, one-dimensional convolution processing with a first kernel size and a second kernel size is performed on the electrocardiogram signal in the time domain dimension to determine a first convolution feature and a second convolution feature.

[0073] The first kernel size is 3 and the second kernel size is 1.

[0074] b: performing normalization processing on the first convolution feature to determine the normalized first convolution feature, and further performing one-dimensional convolution processing and normalization processing on the normalized first convolution feature to determine the third convolution feature.

[0075] Here, the first convolution feature is normalized to determine the normalized first convolution feature, and the normalized first convolution feature is further subjected to one-dimensional convolution and normalization to determine the third convolution feature.

[0076] c: Performing flat pooling on the second convolution feature, processing the second convolution feature after the flat pooling and the third convolution feature based on the activation function to determine the target convolution feature.

[0077] Here, the second convolution feature is subjected to a flat pooling process, and the second convolution feature and the third convolution feature after the flat pooling process are processed according to the activation function to determine the target convolution feature.

[0078] (2): Continue to process the target convolution feature based on the next time series convolution subunit until each time series convolution subunit in the one-dimensional time series convolution module is processed and the time domain feature of the electrocardiogram signal is output.

[0079] Here, the target convolution feature is processed according to the next temporal convolution subunit until each temporal convolution subunit in the one-dimensional temporal convolution module is processed and multiple time domain features of the electrocardiogram signal are output.

[0080] For further information, see Figure 2 , Figure 2 This is a schematic diagram of the processing of the time series convolution subunit provided in the embodiment of the present application. Figure 2 As shown in , the ECG signal is processed in the time domain dimension using one-dimensional convolution and batch normalization / linear rectification activation function to determine the target convolution feature.

[0081] In the present application, each first temporal convolution subunit is configured with a 3D feature receptive field and a 2-step downsampling mechanism, which gradually expands the feature receptive field through different kernel sizes to realize the extraction and fusion of global unstructured patterns in the deep feature space. In the forward propagation path of the residual structure, the feature flow of the backbone network will be synchronously diverted to the feature aggregation layer of the shortcut path before entering the residual module. This path uses the mean pooling operator to extract statistical characteristics of local temporal segments, and realizes feature dimensionality reduction by mapping adjacent feature vectors to regional statistics. This compression strategy based on regional feature representation not only retains the distribution characteristics of the original data, but also reduces the complexity of the model through dimensionality reduction processing. Finally, the deep abstract features of the three paths are fused by tensor concatenation. This multi-flow feature integration mechanism not only enhances the model's ability to express multi-scale features, but its identity mapping characteristics are more conducive to the stable propagation of gradients in deep networks, thereby significantly improving the convergence efficiency and generalization performance of the classification model during parameter optimization.

[0082] B: Based on the dynamic temporal attention mechanism, the multiple temporal features are subjected to temporal convolution processing to generate an attention weight matrix.

[0083] Here, the dynamic temporal attention mechanism is used to perform temporal convolution processing on multiple temporal features to generate an attention weight matrix.

[0084] C: Performing probability space mapping processing on the attention weight matrix based on the activation function to determine the probability value of each of the time domain features.

[0085] Here, the attention weight matrix is ​​mapped into a probability space according to the activation function to determine the probability value of each time domain feature.

[0086] D: performing fusion processing on each of the domain features based on the probability value to determine a third electrocardiogram feature of the electrocardiogram signal.

[0087] Here, each domain feature is fused according to the probability value to determine the third electrocardiogram feature of the electrocardiogram signal.

[0088] In this application, a dynamic temporal attention mechanism is embedded in a multi-scale one-dimensional temporal convolutional residual topology unit, enhancing the semantic expressiveness of deep features through a feature recalibration strategy. The core processing flow consists of a multi-stage feature interaction operation: first, a temporal convolutional unit (receptive field dimension 10, stride 2) is used to model the spatial correlation of the input feature sequence, generating an attention coefficient matrix that matches the original feature dimension. A feature normalization unit dynamically adjusts the distribution parameters to achieve stability control of the feature space, ensuring alignment with the feature dimension of the backbone network. A nonlinear gating mechanism is then introduced, using the sigmoid function to map the attention weights to the probability space [0, 1], establishing a feature selection channel based on a probability distribution. Finally, a tensor inner product operation is used to achieve adaptive fusion of feature vectors, where the weight parameters also participate in the mean-covariance moment estimation of the feature distribution. This joint optimization mechanism of bimodal statistics not only captures second-order correlations between features but also maintains gradient integrity during end-to-end training through differential properties.

[0089] In one possible implementation, after determining the third electrocardiogram feature of the electrocardiogram signal, the detection and analysis method further includes:

[0090] The third electrocardiogram feature is processed based on the spatial correlation relationship between the feature gradient extracted by back propagation and the attention weight matrix to generate a time-sensitive heat map.

[0091] Here, we employ the Gradient Weighted Class Activation Mapping (Grad-CAM) method to visualize the location of key segments in ECG features processed by a dynamic temporal attention mechanism. Backpropagation is used to extract the spatial correlation between feature gradients and attention weights, generating a temporal-sensitive heatmap that visually demonstrates the model's attention intensity on pathological feature regions such as QRS complexes and ST segments, enabling alignment verification of feature response patterns with ECG physiological semantics.

[0092] S104: performing feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature, so as to determine a detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features.

[0093] In this step, feature fusion is performed on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature, so that the detection and analysis results corresponding to the electrocardiogram are determined based on the fused electrocardiogram features.

[0094] In a possible implementation manner, the performing feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature includes:

[0095] I: performing normalization processing on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature to determine a normalized electrocardiogram feature.

[0096] Here, the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature are normalized to determine the normalized electrocardiogram feature.

[0097] II: Based on the dual-path feature Transformer network layer, attention Transformer network layer and activation function, the normalized ECG features are fused and the fused ECG features are output.

[0098] Here, the normalized ECG features are subjected to multi-source feature fusion processing based on the dual-path feature Transformer network layer, the attention Transformer network layer, and the activation function, and the fused ECG features are output.

[0099] In one possible implementation, the dual-path-based feature Transformer network layer, the attention Transformer network layer, and the activation function perform multi-source feature fusion processing on the normalized electrocardiogram features and output the fused electrocardiogram features, including:

[0100] i: Based on the first feature Transformer network layer, key feature selection is performed on the normalized ECG features to determine multiple first target ECG features.

[0101] Here, key feature selection is performed on the normalized electrocardiogram features according to the first feature Transformer network layer to determine multiple first target electrocardiogram features.

[0102] ii: performing nonlinear association processing on the plurality of first target ECG features based on the attention Transformer network layer to determine nonlinear spatial association features between the plurality of target ECG features.

[0103] Here, nonlinear correlation processing is performed on multiple first target ECG features according to the attention Transformer network layer to determine the nonlinear spatial correlation features between the multiple target ECG features.

[0104] iii: performing a dot product operation on the plurality of the nonlinear spatial correlation features and the plurality of the target ECG features to determine a plurality of second target ECG features.

[0105] iv. Based on the second feature Transformer network layer, the plurality of the second target ECG features are processed to determine a plurality of third target ECG features; based on the activation function, the plurality of the third target ECG features are processed to determine a plurality of fourth target ECG features.

[0106] Here, the second target ECG features are processed according to the second feature Transformer network layer to determine a plurality of third target ECG features, and the third target ECG features are processed using an activation function to determine a plurality of fourth target ECG features.

[0107] iv: Based on the attention Transformer network layer, the second feature Transformer network layer and the activation function, the multiple fourth target ECG features are further processed to determine multiple fifth target ECG features, and the multiple fifth target ECG features are fused to determine the fused ECG features.

[0108] For further information, see Figure 3 , Figure 3 This is a schematic diagram of the processing of multi-source feature fusion provided in the embodiment of this application. Figure 3 As shown in the figure, batch normalization, dual-path feature Transformer network layer, attention Transformer network layer, dot product operation and linear rectification activation function are used to perform multi-source feature fusion processing on multiple ECG features, and the fused ECG features are output.

[0109] In this application, multi-source signal enhancement is performed by integrating ECG waveform data, wavelet transform features, and features determined by an eleven-point ECG spatial positioning template. Automatic feature extraction and adaptive fusion are achieved using an end-to-end deep neural network model. To address the asynchrony of data acquisition from different leads, independent input channels are designed. A feature-level fusion module is used at the end of the network to integrate multi-dimensional information, and a spatiotemporal attention mechanism is used to dynamically focus on key ECG segments. The model also embeds an interpretability module to visualize abnormal areas in the output ECG signal, assisting physicians in locating pathological features.

[0110] The present application provides an electrocardiogram detection and analysis method, which includes: performing key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine the first electrocardiogram feature of the electrocardiogram signal; performing frequency domain processing on the electrocardiogram signal based on discrete wavelet transform to determine the second electrocardiogram feature of the electrocardiogram signal; performing deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine the third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of multiple one-dimensional temporal convolution modules with different kernel sizes; performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature, so that the detection and analysis result corresponding to the electrocardiogram is determined based on the fused electrocardiogram feature. By integrating multiple electrocardiogram features to perform multi-source signal enhancement and adaptive fusion, the fused features are used to assist doctors in locating pathological features, thereby improving the accuracy of pathological feature positioning.

[0111] See also Figure 4 、 Figure 5 , Figure 4 This is one of the structural schematic diagrams of an electrocardiogram detection and analysis device provided in an embodiment of the present application; Figure 5 This is a second structural diagram of an electrocardiogram detection and analysis device provided in an embodiment of the present application. Figure 4 As shown in , the detection and analysis device 400 includes:

[0112] A spatial positioning module 410 is configured to perform key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal;

[0113] A frequency domain feature extraction module 420 is configured to process the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal;

[0114] an implicit feature extraction module 430 for performing deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes;

[0115] The feature fusion module 440 is used to perform multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature and the third electrocardiogram feature, so as to determine the detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features.

[0116] Furthermore, when the implicit feature extraction module 430 is used to perform deep learning implicit feature extraction on the electrocardiogram signal based on the multi-scale one-dimensional temporal convolution residual topology unit and the dynamic temporal attention mechanism to determine the third electrocardiogram feature of the electrocardiogram signal, the implicit feature extraction module 430 is specifically used to:

[0117] Performing time domain dimension processing on the electrocardiogram signal based on multiple one-dimensional time series convolution modules in a multi-scale one-dimensional time series convolution residual topology unit, so that each of the one-dimensional time series convolution modules outputs a time domain feature of the electrocardiogram signal;

[0118] Performing temporal convolution processing on the multiple temporal features based on a dynamic temporal attention mechanism to generate an attention weight matrix;

[0119] Performing probability space mapping processing on the attention weight matrix based on the activation function to determine the probability value of each of the time domain features;

[0120] Each of the domain features is fused based on the probability value to determine a third electrocardiogram feature of the electrocardiogram signal.

[0121] Furthermore, the implicit feature extraction module 430 is used for any one-dimensional temporal convolution module, and the multiple one-dimensional temporal convolution modules in the multi-scale one-dimensional temporal convolution residual topology unit perform time domain dimension processing on the electrocardiogram signal so that each one-dimensional temporal convolution module outputs the time domain features of the electrocardiogram signal. The implicit feature extraction module 430 is specifically used to:

[0122] Based on the first temporal convolution subunit in the one-dimensional temporal convolution module, performing convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature;

[0123] Continue to process the target convolution feature based on the next time series convolution subunit until each time series convolution subunit in the one-dimensional time series convolution module has completed processing, and then output the time domain feature of the electrocardiogram signal.

[0124] Furthermore, when the implicit feature extraction module 430 is used to perform convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension based on the first temporal convolution subunit in the one-dimensional temporal convolution module to determine the target convolution feature, the implicit feature extraction module 430 is specifically used to:

[0125] Performing one-dimensional convolution processing of a first kernel size and a second kernel size on the electrocardiogram signal in a time domain dimension to determine a first convolution feature and a second convolution feature;

[0126] Normalizing the first convolution feature to determine a normalized first convolution feature, and further performing one-dimensional convolution and normalization on the normalized first convolution feature to determine a third convolution feature;

[0127] Performing a flat pooling process on the second convolution feature, and processing the second convolution feature after the flat pooling process and the third convolution feature based on an activation function to determine the target convolution feature.

[0128] Further, such as Figure 5 As shown, the detection and analysis device 400 further includes a display module 450, which is used to:

[0129] The third electrocardiogram feature is processed based on the spatial correlation relationship between the feature gradient extracted by back propagation and the attention weight matrix to generate a time-sensitive heat map.

[0130] Furthermore, performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature includes:

[0131] performing normalization processing on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature to determine a normalized electrocardiogram feature;

[0132] Based on the dual-path feature Transformer network layer, attention Transformer network layer and activation function, the normalized ECG features are fused and the fused ECG features are output.

[0133] Furthermore, when the feature fusion module 440 is used to perform feature fusion processing on the normalized electrocardiogram features using the dual-path-based feature Transformer network layer, the attention Transformer network layer, and the activation function and output the fused electrocardiogram features, the feature fusion module 440 is specifically used to:

[0134] Based on the first feature Transformer network layer, key feature selection is performed on the normalized ECG features to determine multiple first target ECG features;

[0135] performing nonlinear correlation processing on the plurality of first target ECG features based on the attention Transformer network layer to determine nonlinear spatial correlation features between the plurality of target ECG features;

[0136] performing a dot product operation on the plurality of nonlinear spatial correlation features and the plurality of target ECG features to determine a plurality of second target ECG features;

[0137] Processing the plurality of second target ECG features based on the second feature Transformer network layer to determine a plurality of third target ECG features, and processing the plurality of third target ECG features based on the activation function to determine a plurality of fourth target ECG features;

[0138] Based on the attention Transformer network layer, the second feature Transformer network layer and the activation function, the multiple fourth target ECG features are further processed to determine the multiple fifth target ECG features, and the multiple fifth target ECG features are fused to determine the fused ECG features.

[0139] An embodiment of the present application provides an electrocardiogram detection and analysis device, which includes: a spatial positioning module, which is used to perform key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal; a frequency domain feature extraction module, which is used to process the electrocardiogram signal in the frequency domain dimension based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal; an implicit feature extraction module, which is used to perform deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of multiple one-dimensional temporal convolution modules with different kernel sizes; a feature fusion module, which is used to perform multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature and the third electrocardiogram feature, so as to determine the detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features. By integrating multiple electrocardiogram features to enhance multi-source signals and adaptively fusing them, the fused features can be used to assist doctors in locating pathological features, thereby improving the accuracy of pathological feature positioning.

[0140] See also Figure 6 , Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 6 As shown in FIG, the electronic device 600 includes a processor 610 , a memory 620 and a bus 630 .

[0141] The memory 620 stores machine-readable instructions executable by the processor 610. When the electronic device 600 is running, the processor 610 communicates with the memory 620 via the bus 630. When the machine-readable instructions are executed by the processor 610, the above-mentioned Figure 1 The specific implementation of the steps of the electrocardiogram detection and analysis method in the method embodiment shown can be found in the method embodiment and will not be repeated here.

[0142] The embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The specific implementation of the steps of the electrocardiogram detection and analysis method in the method embodiment shown can be found in the method embodiment and will not be repeated here.

[0143] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0144] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. There may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interface, indirect coupling or communication connection of devices or units, which may be electrical, mechanical or other forms.

[0145] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0146] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0147] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0148] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-mentioned embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting and analyzing an electrocardiogram, characterized in that: The detection and analysis method comprises: performing key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template to determine a first electrocardiogram feature of the electrocardiogram signal; Processing the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal; Deep learning implicit feature extraction is performed on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes; performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature, so as to determine a detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features; The performing multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature includes: performing normalization processing on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature to determine a normalized electrocardiogram feature; Based on the dual-path feature Transformer network layer, attention Transformer network layer and activation function, the normalized ECG features are fused and the fused ECG features are output; The dual-path-based feature Transformer network layer, the attention Transformer network layer, and the activation function perform feature fusion processing on the normalized electrocardiogram features and output the fused electrocardiogram features, including: Based on the first feature Transformer network layer, key feature selection is performed on the normalized ECG features to determine multiple first target ECG features; performing nonlinear correlation processing on the plurality of first target ECG features based on the attention Transformer network layer to determine nonlinear spatial correlation features between the plurality of first target ECG features; performing a dot product operation on the plurality of the nonlinear spatial correlation features and the plurality of the first target ECG features to determine a plurality of second target ECG features; Processing the plurality of second target ECG features based on the second feature Transformer network layer to determine a plurality of third target ECG features, and processing the plurality of third target ECG features based on the activation function to determine a plurality of fourth target ECG features; Based on the attention Transformer network layer, the second feature Transformer network layer and the activation function, the multiple fourth target ECG features are further processed to determine the multiple fifth target ECG features, and the multiple fifth target ECG features are fused to determine the fused ECG features.

2. The detection and analysis method according to claim 1, characterized in that: The method of performing deep learning implicit feature extraction on the electrocardiogram signal based on the multi-scale one-dimensional temporal convolution residual topology unit and the dynamic temporal attention mechanism to determine the third electrocardiogram feature of the electrocardiogram signal includes: Performing time domain dimension processing on the electrocardiogram signal based on multiple one-dimensional time series convolution modules in a multi-scale one-dimensional time series convolution residual topology unit, so that each of the one-dimensional time series convolution modules outputs a time domain feature of the electrocardiogram signal; Performing temporal convolution processing on the multiple temporal features based on a dynamic temporal attention mechanism to generate an attention weight matrix; Performing probability space mapping processing on the attention weight matrix based on the activation function to determine the probability value of each of the time domain features; Each of the domain features is fused based on the probability value to determine a third electrocardiogram feature of the electrocardiogram signal.

3. The detection and analysis method according to claim 2, characterized in that: For any of the one-dimensional time series convolution modules, the multiple one-dimensional time series convolution modules in the multi-scale one-dimensional time series convolution residual topology unit perform time domain dimension processing on the electrocardiogram signal, so that each of the one-dimensional time series convolution modules outputs the time domain features of the electrocardiogram signal, including: Based on the first temporal convolution subunit in the one-dimensional temporal convolution module, performing convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature; Continue to process the target convolution feature based on the next time series convolution subunit until each time series convolution subunit in the one-dimensional time series convolution module has completed processing, and then output the time domain feature of the electrocardiogram signal.

4. The detection and analysis method according to claim 3, characterized in that: The first temporal convolution subunit in the one-dimensional temporal convolution module performs convolution processing, normalization processing, average pooling processing, and activation function processing on the electrocardiogram signal in the time domain dimension to determine the target convolution feature, including: Performing one-dimensional convolution processing of a first kernel size and a second kernel size on the electrocardiogram signal in a time domain dimension to determine a first convolution feature and a second convolution feature; Normalizing the first convolution feature to determine a normalized first convolution feature, and further performing one-dimensional convolution and normalization on the normalized first convolution feature to determine a third convolution feature; Performing a flat pooling process on the second convolution feature, and processing the second convolution feature after the flat pooling process and the third convolution feature based on an activation function to determine the target convolution feature.

5. The detection and analysis method according to claim 3, characterized in that: After determining the third electrocardiogram feature of the electrocardiogram signal, the detection and analysis method further includes: The third electrocardiogram feature is processed based on the spatial correlation relationship between the feature gradient extracted by back propagation and the attention weight matrix to generate a time-sensitive heat map.

6. An electrocardiogram detection and analysis device, characterized in that: The detection and analysis device comprises: a spatial positioning module, configured to perform key feature positioning processing on the electrocardiogram signal based on a preset eleven-point electrocardiogram spatial positioning template, and determine a first electrocardiogram feature of the electrocardiogram signal; a frequency domain feature extraction module, configured to process the electrocardiogram signal in the frequency domain based on discrete wavelet transform to determine a second electrocardiogram feature of the electrocardiogram signal; an implicit feature extraction module, configured to perform deep learning implicit feature extraction on the electrocardiogram signal based on a multi-scale one-dimensional temporal convolution residual topology unit and a dynamic temporal attention mechanism to determine a third electrocardiogram feature of the electrocardiogram signal; wherein the multi-scale one-dimensional temporal convolution residual topology unit is composed of a plurality of one-dimensional temporal convolution modules with different kernel sizes; a feature fusion module, configured to perform multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature, so as to determine a detection and analysis result corresponding to the electrocardiogram based on the fused electrocardiogram features; The feature fusion module is used to perform multi-source feature fusion on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature: performing normalization processing on the first electrocardiogram feature, the second electrocardiogram feature, and the third electrocardiogram feature to determine a normalized electrocardiogram feature; Based on the dual-path feature Transformer network layer, attention Transformer network layer and activation function, the normalized ECG features are fused and the fused ECG features are output; The feature fusion module is used to perform feature fusion processing on the normalized ECG features based on the dual-path feature Transformer network layer, the attention Transformer network layer, and the activation function, and output the fused ECG features: Based on the first feature Transformer network layer, key feature selection is performed on the normalized ECG features to determine multiple first target ECG features; performing nonlinear correlation processing on the plurality of first target ECG features based on the attention Transformer network layer to determine nonlinear spatial correlation features between the plurality of first target ECG features; performing a dot product operation on the plurality of the nonlinear spatial correlation features and the plurality of the first target ECG features to determine a plurality of second target ECG features; Processing the plurality of second target ECG features based on the second feature Transformer network layer to determine a plurality of third target ECG features, and processing the plurality of third target ECG features based on the activation function to determine a plurality of fourth target ECG features; Based on the attention Transformer network layer, the second feature Transformer network layer and the activation function, the multiple fourth target ECG features are further processed to determine the multiple fifth target ECG features, and the multiple fifth target ECG features are fused to determine the fused ECG features.

7. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus. When the processor is running, the machine-readable instructions execute the steps of the electrocardiogram detection and analysis method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the electrocardiogram detection and analysis method according to any one of claims 1 to 5 are executed.

Images