A method for detecting hypertension-mediated cardiac damage based on ECG data
Through multi-branch multi-scale long and short-term memory neural network and ECG-XAI framework, multi-scale waveform features and time features are extracted, which solves the problems of low sensitivity and low diagnostic accuracy in early detection of hypertension-mediated cardiac injury, and achieves more Accurate hypertension detection and auxiliary diagnosis and treatment.
Patent Information
- Application Number
- CN202410303780.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-18
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-03-18
AI Technical Summary
The prior art has problems such as low sensitivity, poor anti-noise ability, difficult feature extraction and low diagnostic accuracy in early detection of hypertension-mediated cardiac injury, especially when relying on ECG characteristics.
Using multi-branch multi-scale long and short-term memory neural network MML-Net, combined with LSTM and residual convolutional neural network, multi-scale waveform features and time features are extracted, and high-level through feature extraction module, feature fusion module and multi-layer perceptron classifier Blood pressure diagnosis and ECG-XAI interpretable framework for significance analysis.
It improves the accuracy and robustness of hypertension diagnosis, can detect hypertension-mediated cardiac damage earlier and more accurately, assist in diagnosis and treatment results judgment, and shorten doctors' decision-making time.
Smart Images

Figure CN117883091B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of health big data processing, and mainly to a deep learning-based method for early detection of hypertension-mediated cardiac damage on ECG multi-lead time series data. Background Art
[0002] Hypertension is the main cause of cardiovascular disease, and some patients do not have any symptoms in the early stage of the disease. Hypertension will silently cause damage to the patient's heart, including left ventricular hypertrophy, dysfunction and heart failure, and left atrial remodeling. Therefore, the detection of hypertension-mediated cardiac damage is crucial. At present, cardiac ultrasound is mainly used to evaluate the geometry, mass and relative wall thickness of the left ventricle, and then screen for hypertension-mediated left ventricular structural changes. However, cardiac ultrasound is less sensitive to early or mild damage. Non-patent literature Circulation: Cardiovascular Genetics, 2013, 6 (3): 230-7. It is mentioned that cardiac magnetic resonance imaging (CMR) can increase the sensitivity by 10%, but the examination cost is high and it is not suitable for early routine screening. In short, the image detection method is mainly based on structural damage visible to the naked eye, and usually ignores the early subtle damage caused by chronic high pressure.
[0003] Electrocardiograms can measure the electrical activity signals of the myocardium and reflect the electrophysiological activity state of the heart. Studies have shown that potential structural changes in left ventricular hypertrophy will be reflected in the electrophysiological activity of the heart, and even changes in electrophysiological activity will precede changes in the heart structure visible to the naked eye. However, because the ECG features related to left ventricular hypertrophy, such as voltage, morphology, and repolarization, are too small, the non-patent literature Expert Consensus on the Diagnosis and Treatment of Hypertension and Left Ventricular Hypertrophy. Chinese Journal of Cardiology, 2019, (01): 1-5. mentioned that cardiologists have low sensitivity (<30%) when relying solely on ECG to screen for left ventricular hypertrophy. With the development of deep learning in recent years, a large number of studies have shown that deep learning models can capture more subtle changes in ECG signals than human visual detection, and can help identify ECG features related to hypertension-mediated damage, thereby identifying hypertension through ECG.
[0004] Existing methods have many limitations in diagnosing cardiovascular diseases with less obvious characteristics, such as hypertension-mediated damage, including weak noise resistance, difficulty in feature extraction, low diagnostic accuracy and many other deficiencies. These problems have greatly hindered the progress of related research and clinical applications.
[0005] Therefore, it is very meaningful to propose a method for detecting hypertension-mediated cardiac damage based on ECG multi-lead time series data. Summary of the invention
[0006] In view of the above problems existing in the prior art, this application proposes a method of integrating clinical ECG knowledge into the model design. The multi-branch multi-scale LSTM neural network MML-Net (Multi-branch Multi-scale LSTM neural network) adopts a multi-scale multi-branch design concept, fully combining the advantages of LSTM (Long Short Term Memory) and residual convolutional neural network, including different scales from various waveform forms, and the timing of ECG. LSTM is introduced to extract timing-related features, and a multi-scale multi-branch network structure is designed to extract multi-scale waveform features, which improves the accuracy and robustness of hypertension diagnosis, thereby better meeting the needs of medical and research fields.
[0007] According to one aspect of the present invention, a method for detecting hypertension-mediated cardiac damage based on ECG data is proposed, comprising:
[0008] S1. Collect 12-lead ECG data and clinical diagnosis data of the patient to be tested, perform low-frequency filtering and 50 Hz power frequency filtering on the ECG data, remove ECG samples containing abnormalities, limit the ECG data to ±1500 μV, and calculate the variance, mean value and quartile of each lead of each sample in the ECG data;
[0009] S2. constructing and training an MML-Net network based on the ECG data, wherein the MML-Net network includes an interconnected feature extraction module, a feature fusion module, and a multi-layer perceptron MLP classifier, and extracting multi-scale waveform feature vectors and time feature vectors using the feature extraction module and the feature fusion module. The feature extraction module integrates the 12-lead ECG data using a two-layer squeeze-excitation residual block, and uses the multi-layer perceptron MLP classifier to predict the probability of the case suffering from hypertension;
[0010] S3. Perform interpretable analysis based on the MML-Net network, build an ECG-XAI interpretable framework, calculate the significance of each sampling point of the input ECG for hypertension detection, and segment and align the significance map according to the ECG characteristic waveform to obtain a beat-level significance map aligned with the standard ECG characteristic waveform, and analyze and obtain important ECG segments for hypertension detection.
[0011] Furthermore, the feature extraction module is specifically used to perform the following steps: integrating the 12-lead ECG data using a two-layer squeeze-excitation residual block, extracting waveform features through the feature extraction module with 32 channels containing 1×21 and 1×15 cores, traversing the 12 leads row by row, extracting common waveform features of the 12 leads, and compressing the length of the extracted common waveform features to extract a 32-channel, 12-lead feature vector of 32×12×312.
[0012] Furthermore, the feature fusion module is specifically used to perform the following steps: extracting the time-related features memorized by the memory layer in the LSTM as a time feature vector, and then using three convolution branches of different sizes, 3×3, 5×5, and 7×7, to respectively extract waveform morphological features of different widths of the 32×12×312 32-channel, 12-lead feature vectors from different fine-grainednesses, to obtain three waveform feature vectors of different scales and one of the time feature vectors, to fuse the waveform feature vectors and flatten and connect them with the time feature vector, to obtain the output result of the feature fusion module.
[0013] Furthermore, the multilayer perceptron MLP classifier is specifically used to perform the following steps: the output result of the feature fusion module is input into the multilayer perceptron MLP classifier, there are 512 input nodes in the multilayer perceptron MLP classifier, all connected to 1 output node, and the probability of the case suffering from hypertension is obtained through softmax.
[0014] Furthermore, the explainable analysis includes:
[0015] Step 3-1: Based on the LayerCAM algorithm, quantify the contribution value of each feature in the input of the specified layer in the MML-Net network to hypertension detection, and obtain a class activation map;
[0016] Step 3-2: Obtaining hypertension class activation maps of different scales at different layers of the MML-Net network, upsampling and averaging the hypertension class activation maps, and obtaining a saliency map with the same size and corresponding position as the original input ECG data;
[0017] Step 3-3: Align, segment and average the saliency map according to the ECG characteristic waveform to obtain a beat-level saliency map aligned with the standard ECG characteristic waveform.
[0018] Furthermore, the 12-lead ECG data is collected in a manner of continuous 10s 12-lead ECG data with a sampling rate of 500 Hz and is saved as a 12×5000 matrix.
[0019] According to one aspect of the present invention, a computer-readable medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method described in any one of the embodiments of the first aspect is implemented.
[0020] The present invention improves the accuracy and robustness of hypertension diagnosis by integrating clinical ECG knowledge into model design, including different scales from various waveform forms and the timing of ECG, introducing LSTM to extract timing-related features, and designing a multi-scale and multi-branch network structure to extract multi-scale waveform features, thereby better meeting the needs of the medical and research fields; at the same time, the present invention proposes an interpretable framework ECG-XAI designed for standard ECG feature waveform alignment, thereby obtaining a beat-level significance map aligned with the standard ECG feature waveform, obtaining a more specific manifestation of hypertension-mediated damage in the electrocardiogram, providing the possibility for early detection and pathological interpretation of hypertension-mediated heart damage, improving the accuracy of early detection of hypertension-mediated heart damage, helping to assist in the determination of diagnosis and treatment results, and shortening the decision-making time of doctors. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and are used together with the description to explain the principles of the present invention. It will be easy to recognize other embodiments and many expected advantages of the embodiments because they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with each other. The same reference numerals refer to corresponding similar parts.
[0022] Figure 1 A schematic diagram of an analysis process according to an embodiment of the present invention is shown;
[0023] Figure 2 A schematic diagram of an LSTM structure according to an embodiment of the present invention is shown;
[0024] Figure 3 shows a hypertension class activation map generated in Block 1 according to a specific embodiment of the present invention;
[0025] Figure 4 shows a hypertension class activation map generated in Block 2 according to a specific embodiment of the present invention;
[0026] Figure 5 shows a hypertension class activation map generated in Block3A according to a specific embodiment of the present invention;
[0027] Figure 6 shows a hypertension class activation map generated in Block5A according to a specific embodiment of the present invention;
[0028] Figure 7 shows a hypertension class activation map generated in Block 7A according to a specific embodiment of the present invention;
[0029] Figure 8 shows a significance map for hypertension detection according to a specific embodiment of the present invention;
[0030] Fig. 9 A flow chart of ECG segmentation, waveform alignment and scale transformation according to a specific embodiment of the present invention is shown;
[0031] Fig.10 shows a 12-lead heart rate ECG and significance diagram according to a specific embodiment of the present invention;
[0032] Fig.11 It is a schematic diagram of the structure of a computer device suitable for implementing an electronic device of an embodiment of the present invention. DETAILED DESCRIPTION
[0033] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.
[0034] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0035] Figure 1 FIG. 4 shows a schematic diagram of an analysis process according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:
[0036] In a specific embodiment, 12-lead ECG data of the detected patient with a sampling rate of 500 Hz is obtained and saved as a 12×5000 ECG matrix, and the ECG data is quality controlled and preprocessed: the ECG data is low-frequency filtered and 50 Hz power frequency filtered; the variance, mean, and quartile of each lead of each sample in the ECG data are calculated; ECG samples containing abnormalities are eliminated; and the ECG data is limited to ±1500 μV.
[0037] like Figure 1As shown in the MML-Net network structure diagram in part A and the Block detail structure diagram in part B, the 12×5000 ECG matrix is input into the MML-Net network, analyzed by the feature extraction module, and the 32-channel feature extraction module containing 1×21 and 1×15 cores traverses all the leads row by row to extract the common features of the 12 leads, compresses the output common waveform feature length, and finally extracts a 32-channel, 12-lead feature vector of 32×12×312;
[0038] The feature fusion module of the multi-scale multi-branch LSTM is used to extract and integrate the temporal features of the feature vector output by the feature extraction module, and the waveform morphological features of different widths in the 32×12×312 32-channel, 12-lead feature vector are extracted. The LSTM structure diagram is shown in the figure. Figure 2 As shown:
[0039] The amount of information retained in the long memory layer is determined by the forget gate, and the retention of c is controlled by the forget gate sigmoid activation function. t-1 The proportion of past information, the forget gate is calculated using the following formula:
[0040] f t =σ(W hf h t-1 +W xf x t +b f )
[0041] In the above formula, f t is the output of the forget gate, σ is the sigmoid activation function, whose output is in the interval [0,1], W hf and W xf The hidden layers are h t-1 and input x t To the forget gate output f t The weight of b f is the bias of the forget gate;
[0042] Will be x t and h t-1 The short-term memory composed of the data is updated to the long memory layer as input data, and the newly input short-term memory data is converted into candidate information of the long memory layer through the tanh activation function. Then the candidate information is controlled by the sigmoid activation function of the input gate Update to long memory layer c t The ratio of steps is calculated using the following formula:
[0043] i t =σ(W hi h t-1 +Wxi x t +b i )
[0044]
[0045] In the above formula, i t is the input gate output of the forget gate, is the candidate information of the long memory layer, c t is the long memory layer information, W hi and W xi The hidden layers are h t-1 and input x t To input gate output i t The weight of , tanh is the hyperbolic tangent activation function, W hc and W xc The hidden layers are h t-1 and input x t Candidate information to long memory layer The weight of
[0046] The long memory layer information c is activated by the tanh activation function t It is converted into candidate information of the output layer, and then the output ratio of the candidate information is controlled by the sigmoid activation function of the output gate, which is calculated using the following formula:
[0047] o t =σ(W ho h t-1 +W xo x t +b o )
[0048] h t =o t ·tanh(c t )
[0049] In the above formula, o t is the output gate output, W ho and W xo The hidden layers are h t-1 and input x t To output o t The weight of .
[0050] After flattening and connecting the fused waveform feature vector and time feature vector, they are input into the multi-layer perceptron classifier to predict the probability of the case suffering from hypertension. There are 512 input nodes in the multi-layer perceptron classifier module, all connected to 1 output node, and the probability of illness is obtained through softmax;
[0051] like Figure 1As shown in the interpretable analysis flow chart in part C, based on the LayerCAM algorithm, class activation maps of different depths and granularities are obtained from Block1, Block2, Block3A, Block5A, and Block7A in the trained MML-Net network. The sizes of the class activation maps are 32×12×1250, 32×12×312, 64×12×78, 64×12×78, and 64×12×78, respectively. By obtaining class activation maps from layers of different depths in the MML-Net network, a more comprehensive explanation of the MML-Net network is provided.
[0052] For a given ECG data I, when I is input into the MML-Net network, the predicted probability y of the target category c (hypertension or non-hypertension) is obtained c :
[0053] y c =f c (I,θ)
[0054] In the above formula, f represents the MML-Net network, and θ represents its parameters;
[0055] Let A be the output feature map of a layer in the MML-Net network, A k is the kth channel feature map in A, and the predicted probability y c Relative to the k-th channel feature map A k The gradient at the spatial position (i, j) in can be expressed as:
[0056]
[0057] In the above formula, To find the sign of the partial derivative, Represents the point with coordinates (i, j) on the kth feature map;
[0058] According to the LayerCAM algorithm, the weight assignment of the spatial position (i, j) in the kth feature map is
[0059]
[0060] Weight With the corresponding feature map Multiply and add multiple channels to obtain the class activation map M of the target category c of this layer c
[0061]
[0062] M cIt refers to the class activation map of the target category c (hypertension) in this layer of the MML-Net network. Five class activation maps are generated in layers of different depths and scales of the MML-Net network. The sizes of the class activation maps are 32×12×1250, 32×12×312, 64×12×78, 64×12×78, and 64×12×78, respectively.
[0063] The salient features captured in the shallow layers of the MML-Net network usually provide detailed granular positioning information, while the features extracted from the deep layers often provide general positioning information of the target object. To ensure consistency, these class activation maps are upsampled to the same size as the original input ECG (12×5000) using linear interpolation. The linear interpolation formula is as follows:
[0064]
[0065] In the above formula, x src Represents the horizontal coordinate of a point on the original class activation map, x dst The horizontal coordinate corresponding to the point after the class activation map interpolation, Width src and Width dst Represents the width before and after interpolation respectively.
[0066] Normalize the interpolated class activation map to its maximum and minimum values, such as Figure 3-7 As shown in the figure, they are the class activation maps from Block1, Block2, Block3A, Block5A, and Block7A in the MML-Net network. It can be seen from the figure that the feature granularity captured by the MML-Net network gradually increases, and the captured features are gradually concentrated. Then, by averaging the 5 class activation maps, a saliency map (12×5000) with the same scale and corresponding position as the input ECG data is obtained, as shown in Figure 8 As shown, it represents the significance of each sampling point of the input ECG for hypertension detection.
[0067] In order to provide clear ECG waveform feature information for the segmentation and alignment of the saliency map, each input ECG data (12×5000) is extracted with ECG feature waveforms such as P, Q, R, S and T waves. The accuracy of waveform position determination is improved by synchronizing the time of the ECG cycle of each lead.
[0068] For each electrocardiogram, the R wave positions of the 12 leads are detected respectively, and the R wave at the jth position of the ith lead is recorded as R ij , the total number of R waves in the ith lead is recorded as N Ri , the mode of the R wave number of the 12 leads is expressed as and In comparison, leads with different total numbers of R waves will be removed from further analysis, and on the remaining leads, the P, Q, S, and T waves corresponding to the R wave position are detected according to the R wave position of each lead; to avoid incomplete heartbeats, the first and last heartbeats of each ECG will be eliminated; the average position of the P, Q, R, S, and T waves of each remaining heartbeat on the lead is calculated, which effectively reduces the waveform detection error caused by lead-specific noise and ensures more reliable waveform analysis results.
[0069] like Fig. 9 As shown in , a complete ECG cycle is called a heartbeat. A complete heartbeat includes five characteristic waveforms, namely P, Q, R, S, and T waves. The ECG and its significance map are divided into heartbeat levels according to the ECG cycle; the heartbeats of the same lead are aligned according to the peak positions of the P, Q, R, S, and T waves. Each interval (ECG and significance map) between the peaks of the P, Q, R, S, and T waves in the same heartbeat is upsampled or downsampled to the same length to ensure that the peak positions of each heartbeat are aligned, and beat-level ECG and significance map segments with uniform length and aligned waveforms are obtained; the beat-level segments in the same lead are then accumulated and averaged to obtain complete beat-level ECG and significance maps for each of the 12 leads. In order to maintain the original ratio between the waveforms and conform to the reading habits of clinicians, the beat-level ECG and significance map are stretched and scaled according to the average ratio between all waveforms, as shown in Fig.10 As shown in (A) 12-lead average heart rate ECG, (B) 12-lead average heart rate ECG significance map, and (C) 12-lead individual heart rate ECG and significance map, it is possible to accurately locate important positions corresponding to clinical electrocardiogram waveforms and quantify the significance of each ECG segment for hypertension detection.
[0070] Reference below Fig.11 , which shows a schematic diagram of the structure of a computer system suitable for implementing an electronic device of an embodiment of the present application. Fig.11 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
[0071] like Fig.11 As shown, the computer system includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to the programs stored in read-only memories (ROM) 502 and 503 or the programs loaded from the storage part 509 into the random access memory (RAM) 504. In the RAM 504, various programs and data required for the operation of the system are also stored. The CPU 501, the ROMs 502 and 503, and the RAM 504 are connected to each other through a bus 505. An input / output (I / O) interface 506 is also connected to the bus 505.
[0072] The following components are connected to the I / O interface 506: an input section 507 including a keyboard, a mouse, etc.; an output section 508 including a liquid crystal display (LCD), etc. and a speaker, etc.; a storage section 509 including a hard disk, etc.; and a communication section 510 including a network interface card such as a LAN card, a modem, etc. The communication section 510 performs communication processing via a network such as the Internet. A drive 511 is also connected to the I / O interface 506 as needed. A removable medium 512, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 511 as needed so that a computer program read therefrom is installed into the storage section 509 as needed.
[0073] The above describes the specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0074] Computer program code for performing the operations of the present application may be written in one or more programming languages or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0075] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0076] In the description of the present application, it should be understood that the terms "upper", "lower", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. The wording 'comprising' does not exclude the presence of elements or steps not listed in the claims. The wording 'one' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The simple fact that certain measures are recorded in mutually different dependent claims does not indicate that a combination of these measures cannot be used for improvement. Any reference symbols in the claims should not be interpreted as limiting the scope.
Claims
1. A method for detecting hypertension-mediated cardiac damage based on ECG data, characterized in that: include: S1. Collect 12-lead ECG data and clinical diagnosis data of the case to be tested, perform low-frequency filtering and 50 Hz power frequency filtering on the ECG data, remove ECG samples containing abnormalities, limit the ECG data to ±1500 μV, and calculate the variance, mean value and quartile of each lead of each sample in the ECG data; S2. constructing and training an MML-Net network based on the ECG data, wherein the MML-Net network includes an interconnected feature extraction module, a feature fusion module, and a multi-layer perceptron MLP classifier, and extracting multi-scale waveform feature vectors and time feature vectors using the feature extraction module and the feature fusion module. The feature extraction module integrates the 12-lead ECG data using a two-layer squeeze-excitation residual block, and uses the multi-layer perceptron MLP classifier to predict the probability of the case suffering from hypertension; S3. Perform an interpretable analysis based on the MML-Net network, wherein the interpretable analysis includes: For the given 12-lead ECG data I, when I is input into the MML-Net network, the probability y of the case of target category c suffering from hypertension is obtained c :y c =f c (I, θ), f represents the MML-Net network, θ represents its parameters; let A be the output feature map of one layer of the MML-Net network, A k is the kth channel feature map in A, and the y c Relative to the k-th channel feature map A k The gradient at the spatial position (i, j) in can be expressed as: To find the sign of the partial derivative, represents the point with coordinates (i, j) on the kth feature map; according to the LayerCAM algorithm, the weight assignment of the spatial position (i, j) in the kth feature map is Will With the corresponding Multiply and add multiple channels to obtain the class activation map M of the target category c of the corresponding layer c : Five class activation maps were generated in layers of different depths and scales of the MML-Net network, with sizes of 32×12×1250, 32×12×312, 64×12×78, 64×12×78, and 64×12×78, respectively; the class activation maps were upsampled to the same size as the original input 12×5000 ECG data using linear interpolation. The linear interpolation: x src Represents the horizontal coordinate of the point on the original class activation map, x dst is the horizontal coordinate corresponding to the interpolated activation graph of the class, Width src and Width dst Represent the width before and after interpolation respectively; the class activation map after interpolation is normalized to the maximum and minimum values, which are the class activation maps from Block1, Block2, Block3A, Block5A, and Block7A in the MML-Net network. The feature granularity captured by the MML-Net network gradually increases, and the captured features are gradually concentrated. By averaging 5 class activation maps, a saliency map with the same scale and corresponding position as the input 12×5000 ECG data is obtained, which represents the high The significance of blood pressure detection; the ECG and the corresponding significance map are divided into beat levels according to the electrocardiogram cycle, and the beats of the same lead are aligned according to the peak positions of the P, Q, R, S, and T waves. Each interval between the peaks of the P, Q, R, S, and T waves in the same beat is upsampled or downsampled to the same length to ensure that the peak positions of each beat are aligned, and beat-level ECG and significance map segments with uniform length and aligned waveforms are obtained. The beat-level segments in the same lead are accumulated and averaged to obtain a complete beat-level ECG and beat-level significance map of 12 leads.
2. A method for detecting hypertension-mediated cardiac damage based on ECG data according to claim 1, characterized in that: The feature extraction module is specifically used to perform the following steps: The 12-lead ECG data are integrated using a two-layer squeeze-excitation residual block, and waveform features are extracted through the feature extraction module with 32 channels containing 1×21 and 1×15 cores. The 12 leads are traversed row by row to extract the common waveform features of the 12 leads, and the length of the extracted common waveform features is compressed to extract a 32-channel, 12-lead feature vector of 32×12×312.
3. A method for detecting hypertension-mediated cardiac damage based on ECG data according to claim 2, characterized in that: The feature fusion module is specifically used to perform the following steps: The time-related features memorized by the memory layer in the LSTM are extracted as time feature vectors, and then three convolution branches of different sizes, 3×3, 5×5, and 7×7, are used to extract waveform morphological features of different widths of the 32×12×312 32-channel, 12-lead feature vectors from different fine-grainednesses, to obtain three waveform feature vectors of different scales and one time feature vector, and the waveform feature vectors are fused and flattened with the time feature vector to obtain the output result of the feature fusion module.
4. A method for detecting hypertension-mediated cardiac damage based on ECG data according to claim 3, characterized in that: The multi-layer perceptron MLP classifier is specifically used to perform the following steps: The output result of the feature fusion module is input into the multi-layer perceptron MLP classifier. There are 512 input nodes in the multi-layer perceptron MLP classifier, which are all connected to 1 output node. The probability of the case suffering from hypertension is obtained through softmax.
5. The method for detecting hypertension-mediated cardiac damage based on ECG data according to claim 1, characterized in that: The 12-lead ECG data is collected in a continuous 10 s 12-lead ECG data sampling rate of 500 Hz and is saved as a 12×5000 matrix.
6. A computer readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method according to any one of claims 1 to 5.
