Electrocardiogram image enhancement processing method and system based on deep learning

Through deep learning technology, combined with wavelet denoising, adversarial enhancement and convolutional model fusion, the problem of insufficient data quality dependence and generalization ability in ECG image enhancement is solved, and the signal quality and diagnostic accuracy of ECG are improved.

CN119066349AInactive Publication Date: 2024-12-03JINING KESHUN BIOTECHNOLOGY CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411059392.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-03
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing ECG image enhancement methods have problems such as strong dependence on data quality and insufficient generalization ability, which affects the accuracy and efficiency of abnormal features of ECG image.

Method used

Using deep learning-based electrocardiogram image enhancement processing method, multi-dimensional electrocardiogram data sets are generated through wavelet denoising, inherent mode decomposition denoising, adversarial enhancement, multi-lead mode data simulation and convolutional model fusion, to improve the robustness of the model Sex and diagnostic abilities.

Benefits of technology

It improves the quality and clarity of the ECG signal, enhances the generalization ability and diagnostic accuracy of the model, and simplifies the ECG image processing and analysis process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066349A_ABST
    Figure CN119066349A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of image processing, in particular to an electrocardiogram image enhancement processing method and system based on deep learning. The method comprises the following steps: acquiring an electrocardiogram training set; performing wavelet denoising and intrinsic mode decomposition denoising on the electrocardiogram training set to obtain a waveform denoising data set; adversarial enhancement of different abnormal types is carried out on the waveform denoising data set, so that an enhanced and expanded data set is obtained; performing multi-lead mode data simulation on the enhanced and expanded data set to obtain a multi-lead simulation data set; performing multi-dimensional data integration on the multi-lead data set to obtain a multi-dimensional electrocardiogram data set; and performing convolution model fusion on the multi-dimensional electrocardiogram data set to obtain an electrocardiogram enhanced convolution model. According to the invention, enhancement effects of various electrocardiogram images can be realized based on deep learning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an electrocardiogram image enhancement processing method and system based on deep learning. Background Art

[0002] Electrocardiogram (ECG) image enhancement processing is one of the important technologies in the field of medical image analysis and is widely used in the diagnosis and monitoring of heart diseases, such as arrhythmia and myocardial infarction. However, existing ECG image enhancement methods have the problems of strong dependence on data quality and insufficient generalization ability, which greatly limits the application effect of ECG images in different devices and environments, and affects the accuracy and efficiency of identifying abnormal features of ECG images. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide an electrocardiogram image enhancement processing method and system based on deep learning to solve at least one of the above technical problems.

[0004] To achieve the above object, the present invention provides an electrocardiogram image enhancement processing method based on deep learning, comprising the following steps:

[0005] Step S1: obtaining an electrocardiogram training set; performing wavelet denoising and intrinsic mode decomposition denoising on the electrocardiogram training set, thereby obtaining a waveform denoising data set;

[0006] Step S2: performing adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set;

[0007] Step S3: performing multi-lead mode data simulation on the enhanced and enlarged data set to obtain a multi-lead simulation data set; performing multi-dimensional data integration on the multi-lead data set to obtain a multi-dimensional ECG data set;

[0008] Step S4: Perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

[0009] The present invention firstly uses the electrocardiogram to often contain various types of noise, such as electrode interference, motion artifacts, etc., which can interfere with the characteristics and accuracy of the signal. Through wavelet denoising and intrinsic mode decomposition denoising, the influence of noise can be effectively reduced, and the quality and clarity of the electrocardiogram signal can be improved. Wavelet denoising and intrinsic mode decomposition denoising methods can retain useful information and features in the electrocardiogram as much as possible while removing noise. This helps to maintain the integrity and accuracy of the electrocardiogram signal and avoid the loss of important physiological features during the denoising process. The obtained waveform denoising data set can be used as input for subsequent processing steps to provide clean and denoised data for the training of the electrocardiogram enhancement model. This helps to improve the learning effect and generalization ability of the model and reduce excessive sensitivity to noise and interference. Through adversarial enhancement, abnormal electrocardiogram data with diversity can be generated. This helps to improve the robustness and generalization ability of the model, so that it can better cope with various abnormal situations and changes in the real world. Adversarial enhancement technology can highlight the characteristics and details of abnormal signals by perturbing and adjusting abnormal electrocardiogram data. This helps to improve the model's ability to identify and analyze abnormal situations and further improve the enhancement effect of the electrocardiogram. The enhanced enlarged dataset generated by adversarial enhancement can increase the number of training samples, which helps to improve the training effect and generalization ability of the model. At the same time, this also reduces the dependence on large-scale annotated ECG data, reducing the cost and workload of data collection and annotation. By simulating multi-lead mode, ECG data of multiple leads can be generated. This can more comprehensively observe the electrical activity of the heart in different directions and improve the information content and diagnostic ability of ECG data. Integrating ECG data of multiple leads into a multidimensional ECG dataset can provide more comprehensive and detailed ECG signal information. Multidimensional data can contain correlations and interactions between different leads, which helps the model better understand the patterns and characteristics of the electrical activity of the heart. Multi-lead simulation datasets and multidimensional ECG datasets provide richer data resources for the training and analysis of ECG enhancement models. This helps to improve the model's ability to understand and interpret ECGs and improve the accuracy of the model's detection and diagnosis of abnormal conditions. The convolutional model can automatically extract features and patterns from ECG data, and understand the laws and abnormalities of cardiac electrical activity through training and learning. The model can learn richer feature representations from multidimensional ECG datasets to improve the ECG enhancement effect and diagnostic accuracy. Through convolutional model fusion, the advantages and characteristics of multiple sub-models can be combined to improve the overall performance and robustness of the model. Different convolutional models can understand ECG data from different angles and scales, complement and enhance each other, and improve the model's ability to diagnose heart diseases. The ECG enhancement convolutional model obtained through training can be applied to real-time ECG image processing and analysis. The model can quickly process and enhance ECG data, provide real-time diagnostic results and feedback, and help doctors make accurate and rapid diagnostic decisions.

[0010] Preferably, the present invention further provides an electrocardiogram image enhancement processing system based on deep learning, which is used to execute the electrocardiogram image enhancement processing method based on deep learning as described above, and the electrocardiogram image enhancement processing system based on deep learning comprises:

[0011] A denoising module is used to obtain an electrocardiogram training set; wavelet denoising and intrinsic mode decomposition denoising are performed on the electrocardiogram training set to obtain a waveform denoising data set;

[0012] The enhancement and expansion module is used to perform adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set;

[0013] The multi-dimensional simulation module is used to perform multi-lead mode data simulation on the enhanced and enlarged data set, thereby obtaining a multi-lead simulation data set; and to perform multi-dimensional data integration on the multi-lead data set, thereby obtaining a multi-dimensional ECG data set;

[0014] The convolution enhancement module is used to perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

[0015] In summary, the present invention provides an electrocardiogram image enhancement processing system based on deep learning. The electrocardiogram image enhancement processing system based on deep learning is composed of a denoising module, an enhancement and expansion module, a multidimensional simulation module and a convolution enhancement module. It can realize any one of the electrocardiogram image enhancement processing methods based on deep learning described in the present invention. The internal structures of the system cooperate with each other. A large amount of image data is divided into blocks by adopting key electrocardiogram abnormality features, and then key features are clustered and enhanced for different data blocks respectively, thereby ensuring the reliability and identifiability of the data. Moreover, by associating the enhanced data with a few lead modes by multiple lead modes, and simulating the electrocardiogram data obtained by monitoring a few lead modes, the available data set is greatly increased. Finally, the noise and key features in the image are enhanced by deep learning convolution, thereby improving the clarity and reliability of the electrocardiogram image, thereby simplifying the operation flow of the electrocardiogram image enhancement processing system based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0017] Figure 1 This is a schematic diagram of the steps of the electrocardiogram image enhancement processing method based on deep learning according to the present invention;

[0018] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0019] Figure 3 for Figure 2 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0020] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0021] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0022] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0023] To achieve this, please refer to Figures 1 to 3 The present invention provides an electrocardiogram image enhancement processing method based on deep learning, comprising the following steps:

[0024] Step S1: obtaining an electrocardiogram training set; performing wavelet denoising and intrinsic mode decomposition denoising on the electrocardiogram training set, thereby obtaining a waveform denoising data set;

[0025] Step S2: performing adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set;

[0026] Step S3: performing multi-lead mode data simulation on the enhanced and enlarged data set to obtain a multi-lead simulation data set; performing multi-dimensional data integration on the multi-lead data set to obtain a multi-dimensional ECG data set;

[0027] Step S4: Perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

[0028] In the embodiment of the present invention, please refer to Figure 1 FIG. 1 is a schematic diagram of the process flow of the electrocardiogram image enhancement processing method based on deep learning of the present invention. In this example, the electrocardiogram image enhancement processing method based on deep learning includes the following steps:

[0029] Step S1: obtaining an electrocardiogram training set; performing wavelet denoising and intrinsic mode decomposition denoising on the electrocardiogram training set, thereby obtaining a waveform denoising data set;

[0030] The embodiment of the present invention first obtains an electrocardiogram data set for training, which contains multiple electrocardiogram signal samples, and the sample data contains different anomalies or noises. Next, the electrocardiogram training set is denoised. Wavelet denoising and intrinsic mode decomposition denoising are used. After such processing, a waveform denoising data set is obtained, in which the electrocardiogram signal has removed some noise, making it cleaner and clearer.

[0031] Step S2: performing adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set;

[0032] The embodiment of the present invention aims to increase the sample diversity and richness of the data set by performing adversarial enhancement on the waveform denoising data set. Different abnormal types, such as arrhythmia, myocardial ischemia, etc., are used to generate electrocardiogram samples with different abnormal characteristics through adversarial generative networks or other enhancement technologies. This can expand the scale of the data set and include more different types of abnormal situations.

[0033] Step S3: performing multi-lead mode data simulation on the enhanced and enlarged data set to obtain a multi-lead simulation data set; performing multi-dimensional data integration on the multi-lead data set to obtain a multi-dimensional ECG data set;

[0034] The embodiment of the present invention simulates multi-lead mode data for the enhanced and enlarged data set. Multi-lead mode refers to the use of multiple electrodes or sensors to record signals at different locations in electrocardiogram acquisition. By simulating the multi-lead mode, a simulated multi-lead electrocardiogram data set can be generated. Then, multi-dimensional data integration is performed on the multi-lead data set. The electrocardiogram signals from different leads are integrated to form a multi-dimensional electrocardiogram data set. In this way, the relationship and interaction between different leads can be comprehensively considered.

[0035] Step S4: Perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

[0036] The embodiment of the present invention performs convolution model fusion on a multi-dimensional ECG data set. The convolution model is a commonly used deep learning model suitable for processing image and sequence data. By combining multiple convolution models, their advantages in different aspects can be fully utilized.

[0037] The present invention firstly uses the electrocardiogram to often contain various types of noise, such as electrode interference, motion artifacts, etc., which can interfere with the characteristics and accuracy of the signal. Through wavelet denoising and intrinsic mode decomposition denoising, the influence of noise can be effectively reduced, and the quality and clarity of the electrocardiogram signal can be improved. Wavelet denoising and intrinsic mode decomposition denoising methods can retain useful information and features in the electrocardiogram as much as possible while removing noise. This helps to maintain the integrity and accuracy of the electrocardiogram signal and avoid the loss of important physiological features during the denoising process. The obtained waveform denoising data set can be used as input for subsequent processing steps to provide clean and denoised data for the training of the electrocardiogram enhancement model. This helps to improve the learning effect and generalization ability of the model and reduce excessive sensitivity to noise and interference. Through adversarial enhancement, abnormal electrocardiogram data with diversity can be generated. This helps to improve the robustness and generalization ability of the model, so that it can better cope with various abnormal situations and changes in the real world. Adversarial enhancement technology can highlight the characteristics and details of abnormal signals by perturbing and adjusting abnormal electrocardiogram data. This helps to improve the model's ability to identify and analyze abnormal situations and further improve the enhancement effect of the electrocardiogram. The enhanced enlarged dataset generated by adversarial enhancement can increase the number of training samples, which helps to improve the training effect and generalization ability of the model. At the same time, this also reduces the dependence on large-scale annotated ECG data, reducing the cost and workload of data collection and annotation. By simulating multi-lead mode, ECG data of multiple leads can be generated. This can more comprehensively observe the electrical activity of the heart in different directions and improve the information content and diagnostic ability of ECG data. Integrating ECG data of multiple leads into a multidimensional ECG dataset can provide more comprehensive and detailed ECG signal information. Multidimensional data can contain correlations and interactions between different leads, which helps the model better understand the patterns and characteristics of the electrical activity of the heart. Multi-lead simulation datasets and multidimensional ECG datasets provide richer data resources for the training and analysis of ECG enhancement models. This helps to improve the model's ability to understand and interpret ECGs and improve the accuracy of the model's detection and diagnosis of abnormal conditions. The convolutional model can automatically extract features and patterns from ECG data, and understand the laws and abnormalities of cardiac electrical activity through training and learning. The model can learn richer feature representations from multidimensional ECG datasets to improve the ECG enhancement effect and diagnostic accuracy. Through convolutional model fusion, the advantages and characteristics of multiple sub-models can be combined to improve the overall performance and robustness of the model. Different convolutional models can understand ECG data from different angles and scales, complement and enhance each other, and improve the model's ability to diagnose heart diseases. The ECG enhancement convolutional model obtained through training can be applied to real-time ECG image processing and analysis. The model can quickly process and enhance ECG data, provide real-time diagnostic results and feedback, and help doctors make accurate and rapid diagnostic decisions.

[0038] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in FIG. 1 , in this embodiment, step S1 includes the following steps:

[0039] Step S11: Obtain an electrocardiogram training set;

[0040] Step S12: extracting waveform data from the electrocardiogram training set to obtain electrocardiogram waveform data;

[0041] Step S13: performing wavelet denoising on the ECG waveform data to obtain wavelet denoised data;

[0042] Step S14: extracting R waves from the wavelet denoising data to obtain R wave data;

[0043] Step S15: performing peak value change curvature analysis on the R wave data, thereby obtaining R wave peak value change data;

[0044] Step S16: performing RR periodic waveform block division on the wavelet denoising data according to the R wave peak value change data, thereby obtaining a waveform block data set;

[0045] Step S17: performing non-stationary ECG threshold screening and marking on the waveform block data set, thereby obtaining a non-stationary marked data set;

[0046] Step S18: performing intrinsic mode decomposition denoising on the non-stationary labeled data set, thereby obtaining a waveform denoising data set.

[0047] The embodiment of the present invention collects electrocardiogram data sets for training. These data sets can come from medical institutions, research databases or other reliable sources. Ensure that the data set contains a sufficient number and diversity of electrocardiogram samples, covering different cases, normal and abnormal conditions. For each sample in the electrocardiogram training set, waveform data extraction is performed. This includes converting the electrocardiogram signal into time series waveform data for subsequent processing and analysis. Wavelet denoising technology is applied to the electrocardiogram waveform data to reduce noise and interference. Wavelet denoising is a signal processing method based on wavelet transform, which can effectively reduce the noise in the electrocardiogram signal and improve the signal quality and clarity. After wavelet denoising, the data obtained is called wavelet denoising data. R wave extraction is performed on the wavelet denoising data. The R wave is the main peak in the QRS complex in the electrocardiogram, representing the characteristic point of the heart contraction. By applying the R wave detection algorithm, the R wave data can be extracted from the wavelet denoising data for subsequent analysis and processing. The peak change curvature analysis is performed on the R wave data to obtain the peak change information of the R wave. Peak change curvature analysis is a method used to study the change of local characteristics of signals, which can extract the change characteristics of R waves in time and amplitude. Through this analysis, the R wave peak change data is obtained, which describes the dynamic change of R waves in the electrocardiogram. According to the R wave peak change data, the wavelet denoising data is divided into RR periodic waveform blocks. The RR period refers to the time interval between two adjacent R waves, reflecting the rhythmicity of the heart. According to the RR period, the wavelet denoising data is divided into a series of waveform blocks, each block corresponding to an RR period. In this way, a waveform block data set is obtained, in which each waveform block contains the ECG waveform data of a complete RR period. The waveform block data set is subjected to non-stationary ECG threshold screening. Non-stationary ECG signals have the characteristics of changing frequency and amplitude, while stationary signals have constant statistical characteristics. By applying the threshold screening method, non-stationary waveform blocks can be identified from the data set and marked accordingly. In this way, a non-stationary labeled data set is obtained, in which each waveform block is marked as a stationary or non-stationary state, providing key information for subsequent processing. The intrinsic mode function (IMF) denoising technique is applied to the non-stationary labeled data in the non-stationary labeled dataset. Intrinsic mode decomposition is a method based on adaptive data decomposition, which can decompose the non-stationary signal into a series of intrinsic mode functions (IMFs). The denoising effect can be achieved by removing the IMFs with higher frequencies or noise components. After the intrinsic mode decomposition denoising process, the obtained data is called a waveform denoising dataset, in which the non-stationary labeled data has been free of certain noise, which improves the signal quality and clarity, and is suitable for subsequent analysis and model training.

[0048] The present invention first collects and obtains an electrocardiogram data set for training. These electrocardiogram data can come from multiple sources, including medical institutions, research institutions or public databases. Obtaining diverse and representative electrocardiogram data is crucial for the training and generalization capabilities of the model. Electrocardiogram waveform data is extracted from the electrocardiogram training set. Electrocardiogram waveform data usually includes the amplitude and time information of the electrocardiogram signal. By extracting the waveform data, the electrocardiogram signal can be converted into numerical data for subsequent processing and analysis. By applying wavelet denoising technology, the electrocardiogram waveform data is denoised. Wavelet denoising can effectively reduce noise interference in the electrocardiogram and improve the quality and accuracy of the signal. The denoised wavelet data can better reflect the electrical activity characteristics and abnormal conditions of the heart. The R wave is a characteristic peak in the QRS complex in the electrocardiogram, which plays an important role in representing and analyzing the electrical activity of the heart. In this step, the R wave is extracted from the wavelet denoising data to obtain the R wave data. This helps to further analyze and identify the rhythm and abnormal conditions of the heart. By performing peak change curvature analysis on the R wave data, the change of the R wave peak value can be obtained. This change can provide important information about heart rhythm and abnormalities. Peak change curvature data can be used to further analyze and identify the type and severity of heart disease. According to the time interval of the R wave peak change data, the wavelet denoised data is segmented into RR periodic waveform blocks. This divides the ECG data into multiple waveform blocks with periodicity. The waveform block data set can be used to analyze and identify the characteristics of heart rhythm and abnormalities. The waveform block data set is screened and marked by applying the non-stationary ECG threshold screening method. Non-stationary ECG threshold screening can identify and mark the non-stationary parts of the heart rhythm, such as heart rate variability. The non-stationary marked data set helps to further analyze and identify abnormalities of the heart. The non-stationary marked data set is denoised using the intrinsic mode decomposition denoising technique. Intrinsic mode decomposition is a signal processing method that decomposes non-stationary signals into multiple intrinsic mode functions (IMFs). By removing the noise component, a waveform denoised data set can be obtained, which contains clearer and more accurate ECG signals, which helps to further analyze and identify abnormalities of the heart.

[0049] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 2 Detailed step flow diagram of step S2 in FIG. 1 , in this embodiment, step S2 includes the following steps:

[0050] Step S21: screening abnormal ECG data for the waveform denoising data set, and merging data with the same ECG source, thereby obtaining an abnormal ECG data set;

[0051] Step S22: clustering abnormal features of different interval bands on the abnormal ECG data set to obtain abnormal ECG clustering data, wherein the abnormal features of different interval bands clustering includes PR interval clustering, ST segment clustering and QT interval clustering;

[0052] Step S23: extracting the features of P waves and T waves between the same clustered image blocks for the abnormal ECG clustering data, thereby obtaining P wave clustering data and T wave clustering data;

[0053] Step S24: performing feature error discrimination between the same cluster image blocks on the abnormal ECG cluster data according to the P wave cluster data and the T wave cluster data, thereby obtaining waveform error discrimination data;

[0054] Step S25: mining relevant waveform features of abnormal ECG clustering data according to the waveform error discrimination data, thereby obtaining error-related ECG data;

[0055] Step S26: performing amplitude time series change analysis on the P wave clustering data and the T wave clustering data, thereby obtaining amplitude time series data;

[0056] Step S27: Performing different abnormality types of countermeasure enhancement on the error-related ECG data according to the amplitude time series data, thereby obtaining an enhanced and expanded data set.

[0057] The embodiment of the present invention screens abnormal ECG data for waveform denoising data sets. An abnormal detection algorithm (e.g., based on statistical methods or machine learning methods) is used to analyze the waveform denoising data sets to identify and mark abnormal ECG data, such as abnormal conditions such as arrhythmia and myocardial ischemia. Then, data with the same cardiac power source are merged. According to the source of the ECG signal (i.e., the same patient), abnormal ECG data with the same cardiac power source are merged together for subsequent analysis and processing. In this way, an abnormal ECG data set is obtained, which contains abnormal ECG data that have been screened and merged, providing a basis for subsequent abnormal feature analysis and model training. Abnormal features of different interval bands are clustered for the abnormal ECG data set. This can be achieved by cluster analysis methods (e.g., K-means clustering, hierarchical clustering, etc.). Specifically, PR interval clustering can be performed, and the PR intervals in the abnormal ECG data set are clustered to identify and distinguish different PR interval patterns. Similarly, ST segment clustering can be performed, and the ST segments in the abnormal ECG data set are clustered to find different ST segment abnormal patterns. In addition, QT interval clustering can be performed to cluster the QT intervals in the abnormal ECG data set to identify different QT interval abnormal patterns. In this way, abnormal ECG clustering data is obtained, which contains the abnormal feature clustering results of different interval bands, which is helpful for further analysis and classification of abnormal ECG data. Feature extraction is performed on abnormal ECG clustering data, focusing on the features of P wave and T wave. By analyzing the ECG waveform in the clustered image block, the features of P wave and T wave, such as amplitude, duration, slope, etc., are extracted. In this way, P wave clustering data and T wave clustering data are obtained, which contain the extracted P wave and T wave features, providing a basis for subsequent feature error discrimination and waveform feature mining. Feature error discrimination of abnormal ECG clustering data is performed using P wave clustering data and T wave clustering data. P wave and T wave features between different clustered image blocks are compared, and the error or difference metric between them is calculated to determine whether there is an abnormality or change in the waveform features of abnormal ECG data. In this way, waveform error discrimination data is obtained, which includes the characteristic error discrimination results calculated based on the characteristics of the P wave and the T wave, providing a basis for subsequent waveform feature mining and related waveform feature mining. According to the waveform error discrimination data, the abnormal ECG clustering data is mined for related waveform features. By analyzing the features in the waveform error discrimination data, the waveform features related to the abnormal ECG data are identified, such as morphological changes, amplitude changes, etc. In this way, error-related ECG data is obtained, which includes ECG data related to the waveform error discrimination data. By observing and analyzing the amplitude changes of the P wave and the T wave over time, such as the increase or decrease trend of the amplitude, the change in duration, etc. In this way, amplitude time series data is obtained, which includes the amplitude change information of the P wave and the T wave. According to the amplitude time series data, the error-related ECG data is enhanced against different abnormal types.By analyzing and identifying the abnormal types in the amplitude time series data, data enhancement is performed for different abnormal types, such as synthesizing new abnormal ECG data or transforming the original data through signal processing technology. Specifically, by mining the error features of abnormal feature points between image blocks, and then generating adversarial training based on this, the features in the error that are helpful in judging ECG abnormalities are amplified to a certain extent. For features that are not obvious, the amplification weight can be appropriately increased, thereby obtaining an enhanced and expanded data set. That is, while enhancing the key features, the overall data set is also amplified according to the combination of the enhancement degrees of different features.

[0058] The present invention first screens abnormal ECG data for waveform denoising data sets. By using a specific anomaly detection algorithm or model, abnormal ECG data can be identified and screened. After screening, data from the same ECG source are merged to obtain a data set containing abnormal ECG data. This data set will be used for further abnormal feature analysis and data mining. Abnormal feature clustering analysis is performed for different interval bands of the abnormal ECG data set, such as PR interval, ST segment and QT interval. Band data with similar abnormal features are classified into one category through a clustering algorithm to obtain abnormal ECG clustering data. In this way, the abnormal features of different bands can be better understood and analyzed, providing a basis for subsequent feature extraction and abnormal type identification. Feature extraction is performed on abnormal ECG clustering data, focusing on P wave and T wave. By extracting the features of P wave and T wave, such as amplitude, duration, slope, etc., P wave clustering data and T wave clustering data can be obtained. These clustering data will help further abnormal type discrimination and diagnosis. According to the P wave clustering data and the T wave clustering data, feature error discrimination is performed on the abnormal ECG clustering data. By calculating the feature errors between different clustered image blocks, the waveform similarities and differences of abnormal ECG data can be evaluated. This will generate waveform error discrimination data for further abnormal type discrimination and abnormal degree assessment. Using waveform error discrimination data, relevant waveform feature mining is performed on abnormal ECG clustering data. By analyzing the relationship between waveform error and ECG data, waveform features related to abnormal types can be found. This will generate error-related ECG data for further abnormal type identification and feature analysis. Amplitude time series change analysis is performed on P wave clustering data and T wave clustering data. By analyzing the amplitude change pattern over time, the dynamic features in ECG signals can be revealed. This will obtain amplitude time series data, which provides time series information of ECG signal amplitude changes, providing support for further abnormal type discrimination and dynamic feature analysis. Amplitude time series data is used to perform adversarial enhancement on error-related ECG data. By introducing different abnormal types and amplitude change patterns, a more diverse and scaled-up abnormal ECG dataset can be generated. This enhanced dataset can be used for tasks such as model training, abnormal type identification, and performance evaluation to improve the robustness and generalization ability of the algorithm.

[0059] Preferably, step S27 includes the following steps:

[0060] Step S271: extracting key features of different abnormal types from the error-related ECG data, thereby obtaining heart rate data, peak data, trough data, and interwave length data;

[0061] Step S272: amplifying the characteristics of the waveform slope change segment of the peak data and the trough data, thereby obtaining fluctuation change amplified data;

[0062] Step S273: marking the amplitude time series data with similar fluctuations according to the peak data and the trough data, thereby obtaining similar trend marking data;

[0063] Step S274: using a waveform similarity superposition algorithm and performing waveform alignment and superposition on the obviously fluctuating amplified data according to similar trend marking data, thereby obtaining synthetic waveform data;

[0064] Step S275: Perform antagonistic enhancement of the synthetic waveform data during the heart rate variability period according to the heart rate data and the interwave interval length data, thereby obtaining an enhanced and enlarged data set.

[0065] The embodiment of the present invention extracts key features from error-related ECG data to obtain heart rate data, peak data, trough data, and interwave length data. Specifically, the value of heart rate (heart beats per minute) is calculated by analyzing the R peak position in the ECG signal. ECG signal processing algorithms (such as peak detection methods) can be used to detect R peaks, calculate heart rate data, identify the peak position in the ECG signal, and extract relevant features, such as peak amplitude, duration, etc. Waveform analysis algorithms (such as peak detection methods) can be used to detect peaks, extract relevant features, identify the trough position in the ECG signal, and extract relevant features, such as trough amplitude, duration, etc. Waveform analysis algorithms (such as valley detection methods) can be used to detect troughs, extract relevant features, and calculate the time interval between adjacent peaks or troughs, that is, the interwave length. Interwave length data can be obtained by measuring the time difference between waves. For the peak data and trough data, the characteristics of the waveform slope change segment are amplified to obtain the fluctuation change amplified data. Specifically, by calculating the slope value between the peak and the trough, the slope change segment in the waveform is detected, and the waveform data is amplified according to the position and characteristics of the slope change segment. Signal processing technology (such as interpolation or filtering) can be used to amplify the slope change segment to highlight the change of fluctuations. The peak data and trough data are used to mark the amplitude time series data for similar fluctuations to obtain similar trend marking data. Specifically, according to the amplitude data of the peak and the trough, the amplitude change value of the corresponding position in the amplitude time series data is calculated, and according to the size and trend of the amplitude change value, the amplitude time series data is marked for similar fluctuations, for example, the areas of similar fluctuations are marked as the same category or binarized. By using the waveform similarity superposition algorithm and combining with the similar trend marking data, the waveform alignment superposition operation is performed on the obviously fluctuating amplified data, so as to obtain the synthesized waveform data, that is, according to the fluctuation change amplified data obtained in step S272, part of the data of obvious fluctuation is selected as input, and the waveform similarity superposition algorithm is used to perform waveform alignment superposition on the obviously fluctuating amplified data. According to the similar trend marking data, the regions with similar fluctuation characteristics are found, and they are aligned and superimposed to enhance the display of fluctuations. According to the heart rate data and the interwave length data, the synthesized waveform data is enhanced during the heart rate variability period to obtain an enhanced and enlarged data set, that is, the heart rate variability information is calculated by combining the heart rate data and the interwave length data, including the variation amplitude and the frequency of the heart rate, and the synthesized waveform data is enhanced according to the information of the heart rate variability. The variation of the heart rate can be simulated by adjusting the amplitude, frequency or shape of the waveform, so as to expand the variation range of the data set.

[0066] The present invention first extracts key features according to different abnormal types for error-related ECG data. These key features may include heart rate data, i.e., information about heart rate frequency; peak data and trough data, which are used to describe the extreme points of ECG waveforms; and interwave length data, which represent the time interval between adjacent peaks. Extracting these key features is helpful for further abnormal type discrimination and ECG signal analysis. For the peak data and trough data, feature amplification of the waveform slope change segment is performed. By amplifying the part where the waveform slope changes, the change of fluctuation can be highlighted and made more significant. The fluctuation change amplified data obtained in this way can better show the change characteristics of the waveform, which is helpful for abnormal detection and feature analysis. The peak data and trough data are used to mark similar fluctuations in amplitude time series data. By observing the positions and characteristics of the peaks and troughs, similar fluctuation trends can be identified. In this way, similar trend marking data can be generated for subsequent waveform analysis and abnormal type identification. The waveform similarity superposition algorithm is used, and the waveform alignment and superposition of the obvious fluctuation amplified data is performed according to the similar trend marking data. By finding waveform segments with similar fluctuation trends and aligning and superimposing them, synthetic waveform data is obtained. This synthetic waveform data can better show the characteristics of fluctuations and provide more accurate abnormal type identification and analysis basis. According to the heart rate data and interwave length data, the synthetic waveform data is enhanced during the heart rate variability period. By introducing changes in heart rate and adjustments to the interwave period, a more diverse and scaled ECG dataset can be generated. This enhanced dataset can be used for tasks such as model training, abnormal type identification, and performance evaluation to improve the robustness and generalization ability of the algorithm.

[0067] Preferably, step S275 includes the following steps:

[0068] Performing data division and abnormal division of adjacent ECG monitoring time series on the synthesized waveform data, thereby obtaining non-abnormal waveform data and abnormal waveform data;

[0069] Performing generative model training on non-abnormal waveform data to obtain a non-abnormal generative model;

[0070] Constructing long and short-term memory models for non-abnormal waveform data and abnormal waveform data, thereby obtaining a memory model;

[0071] Based on the memory model, the data of non-abnormal waveform data and abnormal waveform data are predicted and supplemented with the unmonitored time series between the data, so as to obtain the complete synthetic waveform data;

[0072] Extract abnormal trend features from the complete synthetic waveform data to obtain abnormal trend data;

[0073] According to the abnormal trend data, abnormality-related features are enhanced on the heart rate data and the wave interval length data, so as to obtain trend-related enhanced data;

[0074] The discriminant model is trained on the trend-related enhanced data to obtain a discriminant adversarial model;

[0075] Generate adversarial training on the non-abnormal generative model and the discriminative adversarial model to obtain a generative adversarial network model;

[0076] Based on the generative adversarial network model, adversarial enhancement is performed during heart rate variability to obtain an enhanced and expanded dataset.

[0077] The embodiment of the present invention divides the synthesized waveform data according to adjacent ECG monitoring time series, for example, divides the continuous ECG cycle data into windows or paragraphs, uses an abnormal detection algorithm or a marking method, divides the adjacent time series data according to the abnormal marking or detection result, and separates the abnormal waveform data from the non-abnormal waveform data. The non-abnormal waveform data is used as training data for the generative model. The data can be preprocessed as needed, such as normalization or smoothing, and an appropriate generative model algorithm (such as a generative adversarial network, a variational autoencoder, etc.) is selected, and the non-abnormal waveform data is used for model training. During the training process, the parameters of the model are optimized to restore the characteristics of the input data to the greatest extent. The non-abnormal waveform data and the abnormal waveform data are used as training data for the LSTM model. The data can be preprocessed as needed, such as normalization or smoothing, and the LSTM model is trained using the non-abnormal waveform data and the abnormal waveform data. LSTM is a deep learning model suitable for processing time series data, which can capture long-term dependencies in waveform data. Using a memory model, the non-abnormal waveform data and the abnormal waveform data are predicted and complemented for the unmonitored time series between the data to obtain complete synthesized waveform data. Use the trained memory model to predict and complete the unmonitored time series between the non-abnormal waveform data and the abnormal waveform data. Through model prediction, fill the missing or unmonitored parts between the data to restore the complete waveform data. Use a suitable feature extraction method to extract abnormal trend features from the complete synthetic waveform data. These features can include information related to abnormal trends such as waveform shape, frequency components, amplitude changes, etc. Based on the abnormal trend data, feature enhancement is performed on the heart rate data and interwave length data. The heart rate data and interwave length data can be adjusted and corrected by comparison, correlation, and pattern matching with the abnormal trend data to enhance the abnormal related features. The trend related enhanced data is used as the training data of the discriminant model. The data can be preprocessed as needed, such as normalization or smoothing, and an appropriate discriminant model algorithm (such as support vector machine, random forest, etc.) can be selected, and the trend related enhanced data can be used for model training. During the training process, the parameters of the model are optimized to achieve the discrimination and classification of abnormal data. The non-abnormal generative model and the discriminative adversarial model are combined, and generative adversarial training is performed by alternately training the generator and the discriminator. The generator is responsible for generating synthetic waveform data, and the discriminator is responsible for discriminating and evaluating the generated waveform data. Using the generative adversarial network model, the data during the heart rate variability period is adversarially enhanced to obtain an enhanced and enlarged dataset.

[0078] The present invention first divides the data of adjacent ECG monitoring time series for synthetic waveform data, and divides the data according to a certain time window to form multiple data fragments. Then, according to the abnormal judgment standard, these data fragments are divided into non-abnormal waveform data and abnormal waveform data for subsequent analysis and processing, wherein the data division of adjacent ECG monitoring time series can continuously analyze the ECG data and improve the continuity of the data. For non-abnormal waveform data, a generative model is trained. The generative model can be a model that can generate a distribution that conforms to non-abnormal waveform data, such as a generative adversarial network (GAN) or a variational autoencoder (VAE). By training the generative model, the characteristics and distribution of non-abnormal waveform data can be learned, providing a basis for subsequent data generation and anomaly detection. For non-abnormal waveform data and abnormal waveform data, a long short-term memory model (LSTM) is constructed. LSTM is a recurrent neural network model suitable for time series data modeling and prediction. By training the LSTM model, the temporal dependency of waveform data can be learned, providing a basis for subsequent data prediction and completion. The constructed LSTM model is used to predict and complete the unmonitored time series between the non-abnormal waveform data and the abnormal waveform data. Through the prediction ability of the model, the missing data in the monitoring process can be filled, so as to obtain complete synthetic waveform data with continuous time series information. For the complete synthetic waveform data, the abnormal trend feature is extracted. By analyzing the trend changes of the waveform data, such as slope, peak value, waveform shape and other features, the abnormal trend information is extracted. These abnormal trend data can be used for subsequent abnormal detection and analysis. Using the extracted abnormal trend data, the abnormal related features of the heart rate data and the interwave length data are enhanced. By integrating the abnormal trend information into the heart rate data and the interwave length data, their abnormal related features are enhanced, so that these data can better reflect the abnormal situation and improve the accuracy of abnormal detection. For the trend-related enhanced data, the discriminant model is trained. The discriminant model can be a model that can distinguish between non-abnormal and abnormal data, such as a support vector machine (SVM) or a deep neural network. By training the discriminant model, the distinguishing features of non-abnormal and abnormal data can be learned, providing a basis for subsequent abnormal detection and classification. Generative adversarial training is performed using the trained non-abnormal generative model and discriminative adversarial model. Generative adversarial training uses a game-based approach to allow the generative model and the discriminative model to compete with each other and learn from each other, so as to improve the generative model's ability to generate real non-abnormal waveform data and enable the discriminative model to more accurately distinguish between non-abnormal and abnormal data. The generative adversarial network model can be used to generate more realistic synthetic data that is more consistent with the distribution of non-abnormal waveform data. Using the trained generative adversarial network model, the heart rate data is enhanced during the mutation period. By introducing changes and adjustments in the heart rate, more diverse and scaled-up heart rate data can be generated.The enhanced and expanded dataset obtained in this way can be used for tasks such as model training, anomaly type identification and performance evaluation, improving the robustness and generalization ability of the algorithm.

[0079] Preferably, step S3 comprises the following steps:

[0080] Step S31: dividing the enhanced and enlarged data set into lead modes, thereby obtaining a low-lead data set and a multi-lead data set;

[0081] Step S32: screening the low lead data set and the multi-lead data set for data of the same cardiac source, thereby obtaining a homologous lead data set;

[0082] Step S33: performing data matching of the same ECG monitoring dimension on the homologous lead data set, thereby obtaining the ECG data set of the same dimension;

[0083] Step S34: performing correlation regression analysis on the same-dimensional ECG data set to obtain the same-dimensional ECG related data;

[0084] Step S35: performing multi-dimensional trend correlation analysis on the homologous lead data set according to the same-dimensional ECG correlation data, thereby obtaining multi-dimensional ECG trend data;

[0085] Step S36: constructing an ECG spatial activity model for the multi-lead data set, thereby obtaining an ECG activity model;

[0086] Step S37: performing multi-lead mode data simulation based on the ECG activity model, thereby obtaining a multi-lead simulation data set;

[0087] Step S38: extracting the R wave peak value from the multi-lead data set to obtain the R wave peak value data;

[0088] Step S39: performing amplitude correction on the multi-channel simulation data set according to the R wave peak value data, thereby obtaining amplitude-corrected multi-channel data;

[0089] Step S310: performing multi-dimensional data integration on the amplitude-corrected multi-lead data and the multi-lead data set to obtain a multi-dimensional ECG data set.

[0090] The embodiment of the present invention divides the enhanced and enlarged data set into a low-lead data set and a multi-lead data set according to the lead mode of the ECG monitoring device. The lead mode division can be automatically divided based on the known lead configuration information, or the lead mode of each sample can be manually marked. The low-lead data set, that is, the lead orientation information is less, and the multi-lead data set, that is, the lead orientation information is more. The number of leads used for monitoring in different lead modes is different. For example, a 3-lead ECG uses three leads (usually I, II and III leads), a single-lead ECG uses only one lead, and a 12-lead ECG uses 12 leads (including I, II, III, aVR, aVL, aVF and V1-V6 leads). The low-lead data set and the multi-lead data set are screened for data with the same ECG source. The ECG signal source can be different lead data of the same patient or data from different patients but under the same cardiac pathological conditions. The data from different patients but under the same cardiac pathological conditions can be assisted in identification according to the discriminator adversarial model obtained previously. Perform data matching of the same ECG monitoring dimension on the homologous lead data set to ensure that the ECG monitoring dimension of each sample is the same. The ECG monitoring dimension mainly includes monitoring data of different positions of ECG monitoring. Data matching can be achieved by adjusting the sampling rate of the data, cropping or padding the data to match the same duration, and adjusting the gain. Perform correlation regression analysis on the ECG data set of the same dimension to study the correlation between the leads of the ECG signal. Statistical methods or machine learning algorithms can be used to calculate the correlation between the leads. Based on the ECG related data of the same dimension, perform multi-dimensional trend correlation analysis on the homologous lead data set to explore the multi-dimensional change trend of the ECG signal. Statistical methods, time series analysis or machine learning algorithms can be used to analyze the trend relationship between the lead data. Construct an ECG spatial activity model for the multi-lead data set to establish a spatial distribution model of the ECG signal. The ECG spatial activity model can be based on traditional electrophysiological knowledge or a model learned from the data using a machine learning algorithm. Based on the ECG activity model, perform multi-lead mode data simulation to generate a simulated multi-lead data set. The ECG activity model can be used to interpolate, extrapolate or generate new multi-lead data for the low-lead data set. The multi-lead simulated data set can be used to increase the number of samples, expand the data set, or to verify and compare the performance of the analysis algorithm. The R-wave peak value is extracted from the multi-lead data set, and the R-wave peak data on each lead is extracted. According to the R-wave peak data, the multi-lead simulated data set is amplitude corrected, that is, the amplitude of each lead is adjusted. The amplitude correction can be based on the amplitude difference between different leads or based on a predefined amplitude standard. The multi-lead data after amplitude correction can maintain the same amplitude range.The amplitude-corrected multi-lead data and the multi-lead data sets are integrated into a multi-dimensional ECG data set. The multi-dimensional data integration can correspond and match the data based on the same time index or other identifiers, that is, the correspondence of each monitoring dimension. More specifically, the simulated multi-dimensional ECG data and the multi-dimensional ECG data in the training set are clustered and integrated between dimensions to obtain a complete, multi-dimensional, and sufficient training set.

[0091] The present invention first divides the lead mode by enhancing and expanding the data set. The lead mode refers to the arrangement of electrodes used in ECG monitoring, and the common ones are single lead and multi-lead. By dividing the data set by the lead mode, the data set can be divided into a low-lead data set and a multi-lead data set for subsequent data analysis and processing. For the low-lead data set and the multi-lead data set, data screening of the same cardiac power source is performed. The cardiac power source refers to the source of the ECG signal, such as different leads or electrode positions. By selecting data of the same cardiac power source, a homologous lead data set can be obtained to ensure the consistency and comparability of the data. Data matching of the same ECG monitoring dimension is performed on the homologous lead data set. The ECG monitoring dimension refers to the parameters and features recorded when collecting ECG signals, such as sampling rate, time window length, etc. By matching the same ECG monitoring dimension, a same-dimensional ECG data set can be obtained to ensure the consistency and comparability of the data. Correlation regression analysis is performed on the same-dimensional ECG data set. Correlation regression analysis can explore the relationship between different ECG features, such as the association between heart rate and waveform shape. By analyzing the correlation, the same-dimensional ECG related data can be obtained for subsequent trend analysis and anomaly detection. Using the obtained same-dimensional ECG related data, multi-dimensional trend correlation analysis is performed on the homologous lead data set. Multi-dimensional trend correlation analysis can reveal the trend relationship between different leads, such as the changing trend of waveform shape. By analyzing the multi-dimensional trend, multi-dimensional ECG trend data can be obtained for further anomaly detection and analysis. For the multi-lead data set, the ECG spatial activity model is constructed. The ECG spatial activity model can describe the spatial relationship and activity pattern between different leads. By constructing the ECG activity model, the spatial variation law of ECG signals can be learned, providing a basis for subsequent simulation and analysis. Using the constructed ECG activity model, multi-lead pattern data is simulated. By simulating and generating multi-lead data, a synthetic data set with different lead patterns can be obtained for tasks such as model training, anomaly detection and performance evaluation. For the multi-lead data set, the R wave peak is extracted. The R wave peak is the position of the highest peak in the QRS complex in the ECG signal, which is usually used for ECG feature extraction and heart rate calculation. By extracting the R wave peak value, the R wave peak data on each lead can be obtained. The previously extracted R wave peak data is used to perform amplitude correction on the multi-lead simulation data set. Amplitude correction is to adjust the amplitude of the ECG signal so that the amplitude of the simulation data is consistent with the actual data. Through amplitude correction, corrected multi-lead data can be obtained to improve the authenticity and reliability of the simulation data. Multi-dimensional data integration is performed on the amplitude-corrected multi-lead data and the original multi-lead data set. Multi-dimensional data integration is to merge and integrate ECG data from different leads and different data sources to form a comprehensive multi-dimensional ECG data set. Through multi-dimensional data integration, more comprehensive and rich ECG data can be obtained for further analysis and application.

[0092] Preferably, step S37 includes the following steps:

[0093] Step S371: performing trend dimension reduction on the ECG activity model according to the multi-dimensional ECG trend data, thereby obtaining the ECG activity trend model;

[0094] Step S372: performing time-domain clustering on the low-lead data set to obtain time-domain clustering data;

[0095] Step S373: Predicting the trend of ECG multi-dimensional activity on the time-domain clustering data based on the ECG activity trend model, thereby obtaining multi-dimensional trend prediction data;

[0096] Step S374: shallowly learning the ECG activity model according to the multi-dimensional trend prediction data, thereby obtaining a low-conductivity ECG activity model;

[0097] Step S375: Perform multi-lead mode data simulation based on the low-lead ECG activity model to obtain a multi-lead simulation data set.

[0098] The embodiment of the present invention performs trend dimensionality reduction on the ECG activity model according to the multidimensional ECG trend data to reduce the dimension of the data and extract the main trend information. Dimensionality reduction techniques such as principal component analysis (PCA) or linear discriminant analysis (LDA) can be used to convert the multidimensional ECG trend data into a lower-dimensional ECG activity trend model. The ECG activity trend model will retain the main trend information as the correlation feature between the multi-lead data and the low-lead data. For the same type of multi-lead data and low-lead data, the trend information can represent the same abnormal heart rate characteristics of both. Time domain clustering is performed on the low-lead data set to classify similar ECG signal segments into the same cluster. Time domain clustering can use clustering algorithms such as K-means clustering, DBSCAN, etc. to group ECG signals. The clustering results will obtain different clusters, each cluster representing a class of ECG signal segments with similar characteristics. Specifically, the low-lead data set is clustered according to the multi-lead time domain type. Based on the ECG activity trend model, the trend of ECG multi-dimensional activity is predicted for the time domain clustering data, that is, the future trend of the ECG signal in each cluster is predicted. Time series prediction methods such as ARIMA and LSTM can be used to predict the trend of ECG signals in the future based on historical data. The multi-dimensional trend prediction data will contain the predicted trend of the ECG signal in each cluster for subsequent model construction and analysis. Based on the multi-dimensional trend prediction data, the ECG activity model is shallowly learned. By learning the relationship between the multi-dimensional trend and the ECG signal, a low-conductivity ECG activity model can be constructed. A shallow neural network or other machine learning model can be used for learning and training, with the multi-dimensional trend prediction data as input features and the ECG signal as the target output. Shallow learning will enable the low-conductivity ECG activity model to generate new ECG signals based on the multi-dimensional trend prediction data, capture trends and patterns, and simulate multi-lead pattern data based on the low-conductivity ECG activity model to generate a simulated multi-lead data set.

[0099] The present invention firstly performs trend dimension reduction on the ECG activity model by using the obtained multidimensional ECG trend data. Trend dimension reduction is to convert the multidimensional trend data into a more concise and interpretable form so as to better understand and analyze the evolution trend of the ECG signal. Through the ECG activity trend model, the important trend features in the ECG signal can be captured, providing a basis for subsequent prediction and analysis. For the low-lead data set, time domain clustering analysis is performed. Time domain clustering is a method of grouping data according to similarity or distance, which is used to discover different patterns or categories in the data set. By performing time domain clustering on the low-lead data set, the data can be divided into different groups for subsequent trend prediction and pattern recognition. Using the obtained ECG activity trend model, the trend of ECG multidimensional activity is predicted on the time domain clustering data. The trend prediction of ECG multidimensional activity is based on the historical data and trend model of the ECG signal to predict the ECG activity in a future period of time. By predicting the multidimensional trend, the future development trend of different leads in the ECG signal can be obtained for further analysis and decision-making. The ECG activity model is shallowly learned using the multidimensional trend prediction data. Shallow learning is a data-based learning method that improves the model's ability to fit data and prediction accuracy through model iteration and optimization. By performing shallow learning on multi-dimensional trend prediction data, the ECG activity model can be optimized to better adapt to the characteristics and changes of low-lead data, thereby improving the performance and effect of the model. The optimized low-lead ECG activity model is used to simulate multi-lead pattern data. By simulating and generating multi-lead data, a synthetic data set with different lead patterns can be obtained for tasks such as model training, performance evaluation, and data analysis. Simulation based on the low-lead ECG activity model can improve the authenticity and reliability of the data.

[0100] Preferably, step S4 comprises the following steps:

[0101] Step S41: dividing the multidimensional ECG data set to obtain a training data set and a test data set;

[0102] Step S42: constructing an ECG image convolution layer with different state transition timings for the training data set, thereby obtaining a convolution model set;

[0103] Step S43: performing multi-state image enhancement on the test data set based on the convolution model set, thereby obtaining enhanced ECG data;

[0104] Step S44: extracting QRS wave from the enhanced ECG data to obtain QRS wave data;

[0105] Step S45: performing noise feature discrimination on the QRS wave data to obtain noise discrimination data;

[0106] Step S46: performing convolution model fusion on the convolution model set according to the noise discrimination data to obtain an electrocardiogram enhanced convolution model.

[0107] The embodiment of the present invention divides the multidimensional ECG data set into a training data set and a test data set by dividing the data. The data division can be based on a certain ratio, such as 70% of the data is used for training and 30% of the data is used for testing. It is ensured that the diversity and representativeness of the ECG data are taken into account when dividing the data to ensure the validity and reliability of the training and test data. The ECG image convolution layer with different state transition timings is constructed for the training data set. A convolutional neural network (CNN) or other convolutional models can be used to design a convolutional layer structure suitable for ECG image processing. By constructing multiple convolutional models with different state transition timings, the features and patterns of the ECG data in different states can be captured. Based on the convolutional model set, the test data set is multi-state image enhanced to improve the quality and diversity of the test data. Different state transitions, such as rotation, translation, scaling, etc., can be applied to each test sample to generate enhanced ECG image data. The enhanced ECG data will contain more diverse situations and patterns for subsequent analysis and model evaluation. QRS wave extraction is performed on the enhanced ECG data to obtain QRS wave data. QRS wave detection algorithms such as Pan-Tompkins algorithm, methods based on convolutional neural networks, etc. can be used to extract QRS waves from ECG data. Noise feature discrimination is performed on QRS wave data to distinguish noise from valid signals. Noise feature extraction methods such as wavelet transform, frequency domain analysis, etc. can be used to extract noise features in QRS wave data. Specifically, noise feature discrimination can be performed according to the noise extraction method of step S1. According to the noise discrimination data, convolution model fusion is performed on the convolution model set to obtain an ECG enhanced convolution model. The results of the convolution model set can be fused by weighted averaging, feature fusion or model combination. The ECG enhanced convolution model will comprehensively consider the prediction results and noise features of multiple convolution models.

[0108] The present invention first divides a multidimensional ECG data set into a training data set and a test data set by performing a data partitioning operation on the data set. Data partitioning is to use independent data for training and testing in the process of model development and evaluation to accurately evaluate the performance and generalization ability of the model. By partitioning the data set, the performance of the model on unseen data can be ensured and reliable evaluation results can be provided. For the training data set, a set of convolutional layer models of ECG images is constructed according to the timing of different state transitions. Through the construction of the convolutional layer, the local features and spatial relationships of the ECG image can be extracted, and the important patterns and structures in the ECG signal can be captured. By constructing multiple convolutional models, the features and change patterns of the ECG image can be more accurately captured for different state transition timings. The test data set is enhanced by using the constructed convolutional model set. Multi-state image enhancement is to process the ECG image by applying the convolutional model to enhance the clarity, contrast or other specific features of the image. By enhancing the ECG data, the quality and recognizability of the data can be improved, providing a better data basis for subsequent waveform analysis and feature extraction. The enhanced ECG data is subjected to QRS wave extraction to extract relevant information of the QRS wave. The QRS wave is an important characteristic waveform in the ECG, which contains the contraction and relaxation process of the ventricle and plays a key role in the analysis and identification of ECG signals. Through QRS wave extraction, relevant data such as the time, amplitude and morphology of the QRS wave can be obtained, which provides a basis for subsequent ECG signal analysis and disease diagnosis. The extracted QRS wave data is subjected to noise feature discrimination to distinguish noise from valid signals. Noise feature discrimination is to determine whether there is noise interference in the data by analyzing the noise characteristics of the QRS wave, such as frequency components, amplitude fluctuations, etc. Through noise discrimination, noise interference can be filtered out and pure QRS wave data can be extracted to ensure the accuracy and reliability of subsequent analysis. According to the noise discrimination data, the previously constructed convolution model set is subjected to convolution model fusion. Convolution model fusion is to synthesize and integrate the results of multiple convolution models to improve the performance and robustness of the model. By adjusting and optimizing the convolution model set according to the noise discrimination data, a more accurate and robust ECG enhancement convolution model can be obtained. This model can be applied to the enhanced processing of ECG signals to improve signal quality and reduce noise interference, thereby providing more reliable results for ECG analysis and diagnosis.

[0109] Preferably, step S42 includes the following steps:

[0110] Step S421: performing multi-state feature classification on the training data set to obtain state classification data;

[0111] Step S422: performing time series analysis on the state classification data to obtain state time series data;

[0112] Step S423: performing state change trend analysis on the state time series data, thereby obtaining state change trend data;

[0113] Step S424: Calculate the state change probability of the state classification data according to the state change trend data, thereby obtaining a state probability matrix;

[0114] Step S425: constructing an ECG image convolution layer with different state transition timings for the multidimensional ECG data set according to the state probability matrix, thereby obtaining a convolution model set.

[0115] The embodiment of the present invention classifies each sample into different state categories by performing multi-state feature classification on the training data set. Supervised learning methods such as support vector machine (SVM) and random forest can be used to classify the state according to the characteristics of the sample. Time series analysis is performed on the state classification data to explore the time series relationship and pattern between states. Time series analysis methods such as autocorrelation analysis and moving average can be used to extract and analyze the time series features of the state classification data. Time series analysis will help understand the evolution and periodicity of the state. State change trend analysis is performed on the state time series data to study the change trend and law between states. Trend analysis methods such as linear regression and exponential smoothing can be used to perform trend fitting and analysis on the state time series data. State change trend analysis will reveal the trend characteristics such as growth, decline, and periodicity between states, providing a basis for state probability calculation. According to the state change trend data, the state change probability is calculated for the state classification data to obtain a state probability matrix. According to the change trend and time series relationship between states, the transition probability or transfer probability of the state can be calculated. The state probability matrix will reflect the transition probability between states, which is used for subsequent convolution model construction and electrocardiogram image processing of state transition time series. According to the state probability matrix, the ECG image convolution layer with different state transition timings is constructed for the multidimensional ECG data set. The state transition order and timing of the ECG image can be adjusted according to the state transition probability in the state probability matrix. By constructing a set of convolution models with different state transition timings, the characteristics and patterns of the ECG data in different states can be captured. Specifically, the state probability matrix can be constructed through a hidden Markov model to ensure the integrity of the state and the prediction effect of possible states.

[0116] The present invention firstly classifies the data according to different states by performing multi-state feature classification on the training data set. By applying feature extraction and classification algorithm to the electrocardiogram data, the electrocardiogram signals in different states can be classified into corresponding state categories. In this way, state classification data can be obtained, in which each sample is marked as a corresponding state, providing a basis for subsequent time series analysis and state change trend analysis. Time series analysis is performed on the state classification data to sort and organize the electrocardiogram signals in time order. Time series analysis can reveal the time relationship and change law between the states of the electrocardiogram signals, thereby obtaining state time series data. By analyzing the state time series data, the duration of the state, the conversion frequency and the time series pattern between the states can be understood, providing a data basis for the subsequent state change trend analysis. State change trend analysis is performed on the state time series data to explore the change trend and law between the states. By analyzing the trend of state change, information such as the relative duration, conversion speed and frequency between the states can be understood. In this way, state change trend data can be obtained, which describes the evolution mode and trend characteristics between the states, providing a basis for the subsequent state probability calculation. The state change probability calculation is performed on the state classification data using the state change trend data. By analyzing the duration, transition frequency, and trend characteristics between states, the probability of state transition can be inferred. In this way, a state probability matrix can be obtained, in which each element represents the probability of transitioning from one state to another. The state probability matrix provides the probability information of state transitions and provides guidance for the subsequent construction of convolutional models. According to the state probability matrix, the ECG image convolution layer with different state transition timings is constructed for the multidimensional ECG dataset. According to the transition probability between different states, the weights and parameters of different states in the convolutional model set can be adjusted to better adapt to the state changes of the ECG signal. In this way, a convolutional model set for different state transition timings can be obtained, providing a variety of model choices for subsequent ECG image processing and analysis.

[0117] Preferably, the present invention further provides an electrocardiogram image enhancement processing system based on deep learning, which is used to execute the electrocardiogram image enhancement processing method based on deep learning as described above, and the electrocardiogram image enhancement processing system based on deep learning comprises:

[0118] A denoising module is used to obtain an electrocardiogram training set; wavelet denoising and intrinsic mode decomposition denoising are performed on the electrocardiogram training set to obtain a waveform denoising data set;

[0119] The enhancement and expansion module is used to perform adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set;

[0120] The multi-dimensional simulation module is used to perform multi-lead mode data simulation on the enhanced and enlarged data set, thereby obtaining a multi-lead simulation data set; and to perform multi-dimensional data integration on the multi-lead data set, thereby obtaining a multi-dimensional ECG data set;

[0121] The convolution enhancement module is used to perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

[0122] In summary, the present invention provides an electrocardiogram image enhancement processing system based on deep learning. The electrocardiogram image enhancement processing system based on deep learning is composed of a denoising module, an enhancement and expansion module, a multidimensional simulation module and a convolution enhancement module. It can realize any one of the electrocardiogram image enhancement processing methods based on deep learning described in the present invention. The internal structures of the system cooperate with each other. A large amount of image data is divided into blocks by adopting key electrocardiogram abnormality features, and then key features are clustered and enhanced for different data blocks respectively, thereby ensuring the reliability and identifiability of the data. Moreover, by associating the enhanced data with a few lead modes by multiple lead modes, and simulating the electrocardiogram data obtained by monitoring a few lead modes, the available data set is greatly increased. Finally, the noise and key features in the image are enhanced by deep learning convolution, thereby improving the clarity and reliability of the electrocardiogram image, thereby simplifying the operation flow of the electrocardiogram image enhancement processing system based on deep learning.

[0123] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0124] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for electrocardiogram image enhancement processing based on deep learning, characterized in that: The following steps are involved: Step S1: Obtain an electrocardiogram training set; Perform wavelet denoising and intrinsic mode decomposition denoising on the ECG training set to obtain a waveform denoising data set; Step S2: performing adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set; Step S3: performing multi-lead mode data simulation on the enhanced and enlarged data set to obtain a multi-lead simulation data set; performing multi-dimensional data integration on the multi-lead data set to obtain a multi-dimensional ECG data set; Step S4: Perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

2. The electrocardiogram image enhancement processing method based on deep learning according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Obtain an electrocardiogram training set; Step S12: extracting waveform data from the electrocardiogram training set to obtain electrocardiogram waveform data; Step S13: performing wavelet denoising on the ECG waveform data to obtain wavelet denoised data; Step S14: extracting R waves from the wavelet denoising data to obtain R wave data; Step S15: performing peak value change curvature analysis on the R wave data, thereby obtaining R wave peak value change data; Step S16: performing RR periodic waveform block division on the wavelet denoising data according to the R wave peak value change data, thereby obtaining a waveform block data set; Step S17: performing non-stationary ECG threshold screening and marking on the waveform block data set, thereby obtaining a non-stationary marked data set; Step S18: performing intrinsic mode decomposition denoising on the non-stationary labeled data set, thereby obtaining a waveform denoising data set.

3. The electrocardiogram image enhancement processing method based on deep learning according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: screening abnormal ECG data for the waveform denoising data set, and merging data with the same ECG source, thereby obtaining an abnormal ECG data set; Step S22: clustering abnormal features of different interval bands on the abnormal ECG data set to obtain abnormal ECG clustering data, wherein the abnormal features of different interval bands clustering includes PR interval clustering, ST segment clustering and QT interval clustering; Step S23: extracting the features of P waves and T waves between the same clustered image blocks for the abnormal ECG clustering data, thereby obtaining P wave clustering data and T wave clustering data; Step S24: performing feature error discrimination between the same cluster image blocks on the abnormal ECG cluster data according to the P wave cluster data and the T wave cluster data, thereby obtaining waveform error discrimination data; Step S25: mining relevant waveform features of abnormal ECG clustering data according to the waveform error discrimination data, thereby obtaining error-related ECG data; Step S26: performing amplitude time series change analysis on the P wave clustering data and the T wave clustering data, thereby obtaining amplitude time series data; Step S27: Performing different abnormality types of countermeasure enhancement on the error-related ECG data according to the amplitude time series data, thereby obtaining an enhanced and expanded data set.

4. The electrocardiogram image enhancement processing method based on deep learning according to claim 3 is characterized in that: Step S27 includes the following steps: Step S271: extracting key features of different abnormal types from the error-related ECG data, thereby obtaining heart rate data, peak data, trough data, and interwave length data; Step S272: amplifying the characteristics of the waveform slope change segment of the peak data and the trough data, thereby obtaining fluctuation change amplified data; Step S273: marking the amplitude time series data with similar fluctuations according to the peak data and the trough data, thereby obtaining similar trend marking data; Step S274: using a waveform similarity superposition algorithm and performing waveform alignment and superposition on the obviously fluctuating amplified data according to similar trend marking data, thereby obtaining synthetic waveform data; Step S275: Perform antagonistic enhancement of the synthetic waveform data during the heart rate variability period according to the heart rate data and the interwave interval length data, thereby obtaining an enhanced and enlarged data set.

5. The electrocardiogram image enhancement processing method based on deep learning according to claim 4 is characterized in that: Step S275 includes the following steps: Performing data division and abnormal division of adjacent ECG monitoring time series on the synthesized waveform data, thereby obtaining non-abnormal waveform data and abnormal waveform data; Performing generative model training on non-abnormal waveform data to obtain a non-abnormal generative model; Constructing long and short-term memory models for non-abnormal waveform data and abnormal waveform data, thereby obtaining a memory model; Based on the memory model, the data of non-abnormal waveform data and abnormal waveform data are predicted and supplemented with the unmonitored time series between the data, so as to obtain the complete synthetic waveform data; Extract abnormal trend features from the complete synthetic waveform data to obtain abnormal trend data; According to the abnormal trend data, abnormality-related features are enhanced on the heart rate data and the wave interval length data, so as to obtain trend-related enhanced data; The discriminant model is trained on the trend-related enhanced data to obtain a discriminant adversarial model; Generate adversarial training on the non-abnormal generative model and the discriminative adversarial model to obtain a generative adversarial network model; Based on the generative adversarial network model, adversarial enhancement is performed during heart rate variability to obtain an enhanced and expanded dataset.

6. The electrocardiogram image enhancement processing method based on deep learning according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: dividing the enhanced and enlarged data set into lead modes, thereby obtaining a low-lead data set and a multi-lead data set; Step S32: screening the low lead data set and the multi-lead data set for data of the same cardiac source, thereby obtaining a homologous lead data set; Step S33: performing data matching of the same ECG monitoring dimension on the homologous lead data set, thereby obtaining the ECG data set of the same dimension; Step S34: performing correlation regression analysis on the same-dimensional ECG data set to obtain the same-dimensional ECG related data; Step S35: performing multi-dimensional trend correlation analysis on the homologous lead data set according to the same-dimensional ECG correlation data, thereby obtaining multi-dimensional ECG trend data; Step S36: constructing an ECG spatial activity model for the multi-lead data set, thereby obtaining an ECG activity model; Step S37: performing multi-lead mode data simulation based on the ECG activity model, thereby obtaining a multi-lead simulation data set; Step S38: extracting the R wave peak value from the multi-lead data set to obtain the R wave peak value data; Step S39: performing amplitude correction on the multi-channel simulation data set according to the R wave peak value data, thereby obtaining amplitude-corrected multi-channel data; Step S310: performing multi-dimensional data integration on the amplitude-corrected multi-lead data and the multi-lead data set to obtain a multi-dimensional ECG data set.

7. The electrocardiogram image enhancement processing method based on deep learning according to claim 6 is characterized in that: Step S37 includes the following steps: Step S371: performing trend dimension reduction on the ECG activity model according to the multi-dimensional ECG trend data, thereby obtaining the ECG activity trend model; Step S372: performing time-domain clustering on the low-lead data set to obtain time-domain clustering data; Step S373: Predicting the trend of ECG multi-dimensional activity on the time-domain clustering data based on the ECG activity trend model, thereby obtaining multi-dimensional trend prediction data; Step S374: shallowly learning the ECG activity model according to the multi-dimensional trend prediction data, thereby obtaining a low-conductivity ECG activity model; Step S375: Perform multi-lead mode data simulation based on the low-lead ECG activity model to obtain a multi-lead simulation data set.

8. The electrocardiogram image enhancement processing method based on deep learning according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: dividing the multidimensional ECG data set to obtain a training data set and a test data set; Step S42: constructing an ECG image convolution layer with different state transition timings for the training data set, thereby obtaining a convolution model set; Step S43: performing multi-state image enhancement on the test data set based on the convolution model set, thereby obtaining enhanced ECG data; Step S44: extracting QRS wave from the enhanced ECG data to obtain QRS wave data; Step S45: performing noise feature discrimination on the QRS wave data to obtain noise discrimination data; Step S46: performing convolution model fusion on the convolution model set according to the noise discrimination data to obtain an electrocardiogram enhanced convolution model.

9. The electrocardiogram image enhancement processing method based on deep learning according to claim 8, characterized in that: Step S42 includes the following steps: Step S421: performing multi-state feature classification on the training data set to obtain state classification data; Step S422: performing time series analysis on the state classification data to obtain state time series data; Step S423: performing state change trend analysis on the state time series data, thereby obtaining state change trend data; Step S424: Calculate the state change probability of the state classification data according to the state change trend data, thereby obtaining a state probability matrix; Step S425: constructing an ECG image convolution layer with different state transition timings for the multidimensional ECG data set according to the state probability matrix, thereby obtaining a convolution model set.

10. An electrocardiogram image enhancement processing system based on deep learning, characterized in that: Used to execute the electrocardiogram image enhancement processing method based on deep learning as claimed in claim 1, the electrocardiogram image enhancement processing system based on deep learning comprises: A denoising module is used to obtain an electrocardiogram training set; wavelet denoising and intrinsic mode decomposition denoising are performed on the electrocardiogram training set to obtain a waveform denoising data set; The enhancement and expansion module is used to perform adversarial enhancement of different anomaly types on the waveform denoising data set, thereby obtaining an enhanced and expanded data set; The multi-dimensional simulation module is used to perform multi-lead mode data simulation on the enhanced and enlarged data set, thereby obtaining a multi-lead simulation data set; and to perform multi-dimensional data integration on the multi-lead data set, thereby obtaining a multi-dimensional ECG data set; The convolution enhancement module is used to perform convolution model fusion on the multi-dimensional ECG data set to obtain an ECG enhanced convolution model.

Citation Information

Cited By

  • Waveform superposition optimization method and system for energy surgical equipment

    CN119360134A