Left ventricular hypertrophy classification method and device, electronic equipment, storage medium and product

Abstract

The invention provides a left ventricular hypertrophy classification method and device, electronic equipment, a storage medium and a product, and the method comprises the steps: determining a median waveform of an electrocardiosignal, and extracting an amplitude feature from the median waveform; inputting the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; wherein the left ventricular hypertrophy classification model comprises a plurality of parallel feature extraction modules used for analyzing different electrocardiogram leads; and the feature extraction module comprises a feature fusion module which is used for performing normalization processing on the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature so as to realize amplitude reduction. According to the invention, based on deep learning, efficient and high-precision automatic classification of left ventricular hypertrophy diseases based on electrocardiosignals is realized; the median waveform and the amplitude feature are subjected to element-by-element multiplication operation, so that amplitude reduction is realized, loss of amplitude information is avoided, and the classification accuracy of left ventricular hypertrophy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrocardiogram signal processing, and in particular to a left ventricular hypertrophy classification method, device, electronic equipment, storage medium and product. Background Art

[0002] Left ventricular hypertrophy (LVH) is a common heart disease that may lead to serious complications such as heart failure and ventricular arrhythmias. LVH is defined as an anatomical increase in the size of the left ventricle and can be assessed by electrocardiography, two-dimensional (2D) and three-dimensional (3D) echocardiography, speckle tracking echocardiography (STE), or cardiac magnetic resonance imaging (CMR).

[0003] Electrocardiograms (ECGs) are widely used as a low-cost and easy-to-use cardiac monitoring method, providing a wider range of diagnostic services. Currently, there are 37 ECG criteria for diagnosing left ventricular hypertrophy (LVH). The most common ones are the Cornell voltage criterion, the Sokolov-Lyon criterion, and the R-aVL voltage criterion. Doctors can use a combination of these criteria to determine whether a patient has LVH.

[0004] However, traditional ECG-based judgment criteria are limited by sensitivity and specificity. The characteristics of mild or early left ventricular hypertrophy may not be obvious in the ECG, leading to misleading results. Secondly, due to certain differences in heart structure between different individuals, some individuals with left ventricular hypertrophy may present a nearly normal image on the ECG, and this change is difficult to detect with the naked eye, affecting the accuracy of the diagnosis. In addition, factors such as age, gender, and weight can affect the interpretation of the ECG, resulting in variations in ECG results, making the criteria for judging left ventricular hypertrophy difficult to apply to all populations.

[0005] Therefore, how to improve the classification accuracy of left ventricular hypertrophy is an urgent problem to be solved. Summary of the Invention

[0006] The present invention provides a left ventricular hypertrophy classification method, device, electronic device, storage medium and product, which are used to solve the defect of poor classification accuracy of left ventricular hypertrophy diseases in the existing technology and realize efficient and high-precision automatic classification of left ventricular hypertrophy diseases.

[0007] The present invention provides a left ventricular hypertrophy classification method, comprising: determining a median waveform of the electrocardiogram signal and extracting amplitude features from the median waveform; Inputting the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; Among them, the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

[0008] According to a left ventricular hypertrophy classification method provided by the present invention, the feature extraction module includes multiple feature fusion modules and maximum pooling layers arranged in an interlaced manner, which are used to gradually extract features of specific electrocardiogram leads; the feature fusion module includes: a normalization layer, configured to perform normalization processing on the first local feature of the median waveform to obtain a second local feature; a multiplication operation layer, configured to perform an element-by-element multiplication operation on the second local feature and the amplitude feature to generate an intermediate feature; A convolutional layer, configured to extract features from the intermediate features to obtain third local features; a batch normalization layer, configured to perform normalization on the third local feature to obtain a fourth local feature; an addition layer, configured to add the fourth local feature to the first local feature to obtain an added feature; The output layer is used to output the added features.

[0009] According to a left ventricular hypertrophy classification method provided by the present invention, the left ventricular hypertrophy classification model further includes a mixed expert module; the mixed expert module includes: A gating network is used to integrate the output features of each feature extraction module to obtain an integrated feature; based on the integrated feature, the weights corresponding to each expert submodule are calculated, and a preset number of expert submodules with the highest weights are activated; An expert submodule, configured to perform a linear transformation on the integrated features; The computation layer is used to perform weighted summation of the outputs of all activated expert submodules and the corresponding weights to obtain a comprehensive feature representation; The expert submodule is trained using a differentiated expert mask strategy.

[0010] According to a left ventricular hypertrophy classification method provided by the present invention, the left ventricular hypertrophy classification model is trained in the following manner: Randomly dividing a sample data set into multiple mutually exclusive subsets; the sample data set includes sample electrocardiogram signals of multiple user objects and corresponding classification labels; In the current iteration process, a mutually exclusive subset is determined as a test data set, and other mutually exclusive subsets except the test data set are determined as training data sets; Determining a median waveform of the sample ECG signals in the training data set, and extracting amplitude features from the median waveform of the sample ECG signals; Inputting the amplitude feature and the median waveform of the sample electrocardiogram signal into a classification model to be trained, and obtaining a probability classification training result output by the classification model to be trained; Based on the loss between the classification label and the probability classification training result, the parameters of the classification model to be trained are adjusted, and the left ventricular hypertrophy classification model is obtained after the model training is completed.

[0011] According to a left ventricular hypertrophy classification method provided by the present invention, obtaining the median waveform of the electrocardiogram signal includes: Acquire an original electrocardiogram signal, filter out baseline interference and high-frequency interference in the original electrocardiogram signal, and obtain an electrocardiogram signal after interference filtering; Detecting an R-wave position, dividing the interference-filtered ECG signal into a plurality of cardiac cycles based on the R-wave position, and truncating the ECG signal within each cardiac cycle to obtain a new cardiac cycle; All new heartbeat cycles are aligned to the same time axis and the median of each time point is calculated to obtain the median waveform.

[0012] According to a left ventricular hypertrophy classification method provided by the present invention, extracting amplitude features from the median waveform includes: An electrocardiogram signal waveform analysis tool is used to extract the amplitude characteristics of a target waveform from the median waveform; the target waveform includes at least one of a P wave, a Q wave, an R wave, an S wave, a T wave, a P1 wave, an R1 wave, an S1 wave, a T1 wave, a J wave, a JX wave, and a JXE wave.

[0013] The present invention also provides a left ventricular hypertrophy classification device, comprising: an extraction module, configured to determine a median waveform of the electrocardiogram signal and extract amplitude features from the median waveform; a classification module, configured to input the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; Among them, the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any of the above-described left ventricular hypertrophy classification methods is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for classifying left ventricular hypertrophy as described above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned left ventricular hypertrophy classification methods.

[0017] The left ventricular hypertrophy classification method, device, electronic device, storage medium and product provided by the present invention input amplitude features and median waveforms into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result of left ventricular hypertrophy output by the model, thereby realizing efficient and high-precision automatic classification of left ventricular hypertrophy diseases based on electrocardiogram signals based on deep learning; by designing the left ventricular hypertrophy classification model to include multiple parallel feature extraction modules for analyzing different electrocardiogram leads, the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature, thereby realizing amplitude restoration, solving the problem of amplitude information loss caused by the ECG signal standardization process, and helping to further improve the classification accuracy of left ventricular hypertrophy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the left ventricular hypertrophy classification method provided by an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of the structure of the left ventricular hypertrophy classification model provided by an embodiment of the present invention.

[0021] Figure 3 It is a structural diagram of a feature extraction module provided by an embodiment of the present invention.

[0022] Figure 4 It is a structural diagram of the feature fusion module provided in an embodiment of the present invention.

[0023] Figure 5 It is a structural diagram of the hybrid expert module provided by an embodiment of the present invention.

[0024] Figure 6 This is the second flow chart of the left ventricular hypertrophy classification method provided by an embodiment of the present invention.

[0025] Figure 7 Schematic diagram of a confusion matrix provided by an embodiment of the present invention.

[0026] Figure 8 3 is a schematic diagram comparing the ROC curves of the EFFNet model and the ECG rule provided in an embodiment of the present invention.

[0027] Figure 9 Schematic diagram of the structure of a left ventricular hypertrophy classification device provided by an embodiment of the present invention.

[0028] Figure 10 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0030] In the description of the embodiments of the present invention, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0031] Currently, left ventricular hypertrophy can be assessed by electrocardiography, two-dimensional (2D) and three-dimensional (3D) echocardiography, speckle tracking echocardiography (STE), or cardiac magnetic resonance imaging (CMR).

[0032] Cardiac magnetic resonance imaging (CMR) is a commonly used method for diagnosing left ventricular hypertrophy. The total volume of the left ventricular muscle is estimated by segmenting it, and then multiplied by the myocardial density to obtain the left ventricular mass index, thereby assessing the degree of left ventricular hypertrophy. The specific calculation steps are as follows: 1. Acquire a short-axis cine SSFP (Steady-State Free Precession) sequence covering the entire left ventricle (LV), from base to apex.

[0033] 2. Segment the left ventricular endocardial border and epicardial border manually or automatically using artificial intelligence methods on each short-axis slice. The left ventricular myocardial volume (LVV) is calculated as: ; in, is the adventitial area on the i-th short-axis slice, is the intimal area on the i-th short-axis slice. S is the slice thickness (usually 6-8 mm), and G is the gap between adjacent slices (usually 0-2 mm).

[0034] 3. The biological tissue density of myocardium is 1.05g / cm 3 , the left ventricular mass can be obtained by multiplying the volume by the density: ; 4. Divide the left ventricular mass by the body surface area to get the left ventricular mass index: ; BSA is body surface area, which is calculated using the DuBois formula.

[0035] In practice, it is usually possible to achieve a left ventricular mass index greater than 55 g / m 2 Female subjects and left ventricular mass index greater than 72 g / m 2 of the male subjects were diagnosed with left ventricular hypertrophy.

[0036] Although cardiac magnetic resonance imaging (CMR) is a mature and accurate technique for diagnosing left ventricular hypertrophy (LVH), the high cost of the equipment makes it difficult to disseminate in areas with less developed medical conditions. Furthermore, CMR quality is affected by the operator's experience and skill level, requiring significant time and technical training. Electrocardiography (ECG), on the other hand, is widely used as a low-cost and easy-to-use cardiac monitoring method, providing a wider range of diagnostic services. Currently, there are 37 ECG criteria for diagnosing LVH, the most common of which are the Cornell voltage criterion, the Sokolov-Lyon criterion, and the R-aVL voltage criterion. Physicians can combine these criteria to determine whether a patient has LVH.

[0037] However, traditional ECG-based judgment criteria are limited by sensitivity and specificity. The characteristics of mild or early left ventricular hypertrophy may not be obvious in the ECG, leading to misleading results. Secondly, due to certain differences in heart structure between different individuals, some individuals with left ventricular hypertrophy may present a nearly normal image on the ECG, and this change is difficult to detect with the naked eye, affecting the accuracy of the diagnosis. In addition, factors such as age, gender, and weight can affect the interpretation of the ECG, resulting in variations in ECG results, making the criteria for judging left ventricular hypertrophy difficult to apply to all populations.

[0038] In response to the above problems, an embodiment of the present invention inputs amplitude features and median waveforms into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result of left ventricular hypertrophy output by the model, thereby realizing efficient and high-precision automatic classification of left ventricular hypertrophy diseases based on electrocardiogram signals based on deep learning; by designing a left ventricular hypertrophy classification model including multiple parallel feature extraction modules for analyzing different electrocardiogram leads, the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature, thereby realizing amplitude restoration, solving the problem of amplitude information loss caused by the ECG signal standardization process, and helping to further improve the classification accuracy of left ventricular hypertrophy.

[0039] Figure 1 This is one of the flow charts of the left ventricular hypertrophy classification method provided by the embodiment of the present invention. Figure 1 An embodiment of the present invention provides a method for classifying left ventricular hypertrophy, which may include the following steps: Step 101: determine the median waveform of the electrocardiogram signal, and extract amplitude features from the median waveform.

[0040] It should be noted that the execution entity of the left ventricular hypertrophy classification method provided in the embodiments of the present invention can be an electronic device, a component of an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, a mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA). The non-mobile electronic device can be a server, network attached storage (NAS), personal computer (PC), television, ATM, or self-service machine, etc., which is not specifically limited in the embodiments of the present invention. The following description of the embodiments of the present invention uses a server as the execution entity.

[0041] In some embodiments, the original ECG signal can be obtained from an ECG device (such as an electrocardiograph, a wearable device, etc.), and interference components such as baseline drift and high-frequency noise in the original ECG signal can be removed by filtering, and the R-wave position can be detected to align the heartbeats; then, the median waveform can be calculated for the aligned heartbeats to eliminate inter-individual variations and retain morphological features.

[0042] The median waveform is obtained by applying a median filter to the aligned heartbeat signals. This waveform can be used to reduce noise and extract a representative ECG waveform (i.e., highlight the typical shape of the signal). Specifically, the median waveform is generated by sorting multiple heartbeat signals at each time point and taking the median value of the signals at that time point.

[0043] In some embodiments, amplitude features of multiple waveforms can be extracted from the median waveform for subsequent input into a left ventricular hypertrophy classification model for classification. For example, amplitude features of the P wave, QRS complex, and T wave can be extracted from the median waveform.

[0044] Step 102: Input the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result of left ventricular hypertrophy; wherein the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication of the normalized median waveform with the amplitude feature to achieve amplitude restoration.

[0045] Figure 2Schematic diagram of the structure of the left ventricular hypertrophy classification model provided by the embodiment of the present invention. Figure 2 In some embodiments, a left ventricular hypertrophy classification model (EFFNet, i.e., an electrocardiogram-based feature fusion network) may include multiple feature extraction modules (Branches) for analyzing different electrocardiogram leads, a mixture of experts (MoE) module, an activation layer (ReLU), a fully connected layer (Linear), and an output layer (Output) using a parallel architecture. Amplitude features and median waveforms may be input into each feature extraction module for feature extraction. After feature extraction, the output features of all feature extraction modules may be combined and processed through the mixture of experts module and the fully connected layer to generate a final left ventricular hypertrophy classification probability. Gender information (i.e., gender) may be input into the model as an additional feature to increase the sensitivity of the left ventricular hypertrophy classification model to gender differences, thereby improving classification accuracy.

[0046] Specifically, the output features of all feature extraction modules can be input into the hybrid expert module, which can dynamically select and combine the outputs of different expert sub-modules according to the input features to generate a comprehensive feature representation; then, the comprehensive feature representation can be input into the activation layer, which is conducive to the model learning the complex relationship between the input data features and enhancing the expressive ability of the left ventricular hypertrophy model; finally, the output of the activation layer can be mapped to the final output space through the fully connected layer, and the classification probability result of left ventricular hypertrophy can be output through the output layer.

[0047] The probability classification result of left ventricular hypertrophy may include the probability of left ventricular hypertrophy and the probability of non-left ventricular hypertrophy.

[0048] In some embodiments, the ECG signal may include multiple ECG leads of varying dimensions, and each feature extraction module may be dedicated to analyzing a single ECG lead. For example, a left ventricular hypertrophy classification model may employ a parallel architecture, consisting of 12 independent processing branches (i.e., feature extraction modules); for each of the 12 ECG leads, each branch may be dedicated to analyzing a single ECG lead.

[0049] In some embodiments, before inputting the amplitude features and median waveform into the trained left ventricular hypertrophy classification model, the median waveform and amplitude features may be standardized, and then the standardized median waveform and amplitude features may be input into the left ventricular hypertrophy classification model for classification.

[0050] Since different features often have differences in numerical magnitude, this scale difference will cause some weights to update much faster than other weights during the gradient descent optimization process, thereby hindering model convergence; the embodiment of the present invention can effectively alleviate the impact of scale differences by standardizing the data input to the model, making it easier for the model to find the optimal solution. The optimization algorithm in deep learning is particularly sensitive to the distribution of input data. When the data distribution deviates from the standard normal distribution, the convergence speed of the model may be greatly reduced; through standardization, the data input to the model can present a more symmetrical and centralized distribution characteristic, which can significantly accelerate the convergence speed of the optimization process. In addition, the standardized data achieves a more balanced distribution, which not only reduces the excessive influence of specific features on model training, but also improves the generalization ability of the model.

[0051] However, normalization can result in the loss of original signal amplitude information, which is crucial for diagnosing left ventricular hypertrophy (LVH). In conventional network architectures, convolutional layers are typically used to extract signal morphological features, while amplitude information is often not integrated until fully connected layers. This approach only achieves shallow information fusion, limiting the ability of convolutional neural networks (CNNs) to effectively capture characteristic patterns in ECG signals.

[0052] To address this issue, embodiments of the present invention design a feature fusion module that integrates morphological features (i.e., those of the median waveform) and amplitude features at the shallow convolutional network stage, thereby enabling a deeper fusion of the two types of information and achieving accurate amplitude restoration. The core structure of each feature extraction module (i.e., processing branch) may include a feature fusion module, which employs a sophisticated multi-feature integration mechanism. The median composite signal (i.e., median waveform) is normalized using a nominal Sigmoid layer within the feature fusion module, scaling its value to the [0, 1] range. By performing an element-by-element multiplication of the normalized median waveform with the extracted amplitude features, accurate amplitude restoration can be achieved, thereby resolving the issue of amplitude information loss caused by the ECG signal normalization process in existing techniques and improving the accuracy of left ventricular hypertrophy classification.

[0053] The embodiment of the present invention inputs the amplitude features and median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy output by the model, thereby realizing efficient and high-precision automatic classification of left ventricular hypertrophy diseases based on electrocardiogram signals based on deep learning; by designing the left ventricular hypertrophy classification model to include multiple parallel feature extraction modules for analyzing different electrocardiogram leads, the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude features, thereby realizing amplitude restoration, solving the problem of amplitude information loss caused by the ECG signal standardization process, and helping to further improve the classification accuracy of left ventricular hypertrophy.

[0054] In an optional embodiment, obtaining the median waveform of the electrocardiogram signal may specifically include: Step S11 , obtaining an original ECG signal, filtering out baseline interference and high-frequency interference in the original ECG signal, and obtaining an interference-filtered ECG signal.

[0055] Raw ECG signal data typically contains significant noise, requiring preprocessing. In some embodiments, baseline interference in the raw ECG signal can be first filtered out using a 3rd-order Butterworth high-pass filter with a cutoff frequency of 0.5 Hz. High-frequency interference in the signal can then be filtered out using a 3rd-order Butterworth low-pass filter with a cutoff frequency of 40 Hz. This eliminates interference components from the raw ECG signal and produces an ECG signal after interference filtering. This helps prevent the impact of interference components in the signal on classification accuracy and improves subsequent classification accuracy for left ventricular hypertrophy.

[0056] Step S12: detecting the R wave position, dividing the interference-filtered ECG signal into multiple heartbeat cycles based on the R wave position, and truncating the ECG signal in each heartbeat cycle to obtain a new heartbeat cycle.

[0057] In some embodiments, the Pan-Tompkins algorithm can be used to extract the R-wave position. However, the Pan-Tompkins algorithm can only extract the R-wave position of a single lead. By clustering and integrating information from 12 ECG leads, the R-wave position can be more accurately determined.

[0058] In some embodiments, the ECG signal after interference filtering can be divided into a single heartbeat cycle by taking the 400ms before the R wave and the 600ms after the R wave according to the R wave position; then, the ECG signal data in each heartbeat cycle can be truncated according to the -0.3×RR interval to 0.7×RR interval to obtain a new heartbeat cycle.

[0059] The RR interval refers to the time interval between two R waves.

[0060] Step S13: align all new heartbeat cycles onto the same time axis, and calculate the median of each time point to obtain a median waveform.

[0061] In some embodiments, all cardiac cycles can be aligned to the same time axis. For each sampling point, the median of all cardiac cycles at that time point (i.e., the sampling point) can be calculated based on the ECG signal values ​​of all cardiac cycles at that sampling point to obtain a median waveform.

[0062] In an optional embodiment, extracting the amplitude feature from the median waveform may specifically include: using an ECG signal waveform analysis tool to extract the amplitude feature of the target waveform from the median waveform; the target waveform includes at least one of a P wave, a Q wave, an R wave, an S wave, a T wave, a P1 wave, an R1 wave, an S1 wave, a T1 wave, a J wave, a JX wave and a JXE wave.

[0063] In some embodiments, the ECGdeil toolbox can be used to extract the amplitude features of the ECG signal from the median waveform. The ECGdeil toolbox is an open source MATLAB tool that can be used for ECG waveform analysis.

[0064] For each ECG lead, the amplitude features (12 amplitude features) of the following waveforms (12 waveforms) can be extracted: P wave, Q wave, R wave, S wave, T wave, P1 wave, R1 wave, S1 wave, T1 wave, J wave, JX wave and JXE wave, totaling 144 features.

[0065] In an optional embodiment, the feature extraction module includes a plurality of feature fusion modules and a maximum pooling layer arranged in an interlaced manner, for gradually extracting features of specific electrocardiogram leads; the feature fusion module may specifically include: a normalization layer, for normalizing the first local feature of the median waveform to obtain a second local feature; a multiplication operation layer, for performing an element-by-element multiplication operation on the second local feature and the amplitude feature to generate an intermediate feature; a convolution layer, for extracting features from the intermediate feature to obtain a third local feature; a batch normalization layer, for normalizing the third local feature to obtain a fourth local feature; an addition layer, for adding the fourth local feature to the first local feature to obtain the added feature; and an output layer, for outputting the added feature.

[0066] Figure 3 Schematic diagram of the structure of the feature extraction module provided by the embodiment of the present invention. Figure 3In some embodiments, the core structure of each independent processing branch (i.e., Branch, feature extraction module) can include a feature extraction sequence of five feature fusion modules (i.e., Fusion Modules). These five feature fusion modules can be interleaved with four maximum pooling layers (MaxPool) to gradually extract and refine the features of a specific ECG lead. The feature extraction module can also include a one-dimensional convolutional layer (Conv1), an activation layer (ReLU), an adaptive average pooling layer (AdaptiveAvgPool), and an output layer (Output).

[0067] Specifically, the median waveform can be input into a one-dimensional convolutional layer Conv1 (16, 71, 1, 35) to capture local patterns and results in the signal and obtain local features. The parameters indicate that there are 16 convolution kernels, each with a size of 71, an input channel of 1, and a stride of 35, thus generating 16 initial channels through the one-dimensional convolution layer. Inputting the local features extracted by the one-dimensional convolution layer into the activation layer can help the model learn more complex features. Through the interleaved arrangement of multiple feature fusion modules and maximum pooling layers, the features of specific leads can be gradually extracted and refined. Through the adaptive average pooling layer, each channel of the feature map can be compressed into a single value, thereby unifying the size of the feature map and facilitating subsequent classification tasks. Finally, the comprehensive features of a single ECG lead processed by this feature extraction module can be output through the output layer.

[0068] Figure 4 Schematic diagram of the structure of the feature fusion module provided by the embodiment of the present invention. Figure 4 In some embodiments, the feature fusion module may include a normalization layer (Sigmoid), a multiplication layer (multiply), a first convolution layer (Conv1 (192, 9, 1, 4)), a first batch normalization layer (Batch Norm), a first activation layer (ReLU), a second convolution layer (Conv1 (16, 9, 1, 4)), a second batch normalization layer (Batch Norm), an addition layer (Add), a second activation layer (ReLU) and an output layer (Output).

[0069] Specifically, the M first local features of the median waveform can be normalized using a normalization layer to obtain M second local features. The M second local features can be element-wise multiplied with the N amplitude features using a multiplication layer to generate M×N intermediate features. The intermediate features can be extracted using a first convolutional layer to obtain third local features. These features are processed using a batch normalization layer and an activation layer to obtain fourth local features. The channel dimension is compressed using a second convolutional layer, thereby compressing the fourth local features to obtain compressed features. The compressed features are normalized using a second batch normalization layer to obtain fifth local features. The fifth local features can be added to the first local features of the median waveform using an addition layer to obtain the added features, which are then output via an activation layer and an output layer.

[0070] Specifically, in the feature fusion module, each of the 16 initial channels can be required to interact with the 12 amplitude features, generating 192 intermediate channels (i.e., intermediate features). After extracting features from these expanded channels through the first convolutional layer, the channel dimension is compressed from 192 to 16 through a second convolutional layer to address the potential channel explosion problem and maintain computational efficiency. Taking into account the feature compression effect of this layer, batch normalization is introduced to enhance model convergence and reduce the risk of overfitting while preserving the network's representational capabilities.

[0071] The feature fusion module designed in the embodiment of the present invention can realize the integration of morphological features (i.e., the morphological features of the median waveform, and the local features of the median waveform can refer to local morphological features) and amplitude features in the shallow convolutional network stage (i.e., the multiplication operation multiply layer before the convolution layer Conv1 (192, 9, 1, 4)), thereby promoting the deeper fusion of the two types of information to achieve accurate amplitude restoration.

[0072] In an optional embodiment, the left ventricular hypertrophy classification model may also include a hybrid expert module; the hybrid expert module includes: a gating network for integrating the output features of each feature extraction module to obtain an integrated feature; based on the integrated feature, calculating the weights corresponding to each expert sub-module, and activating a preset number of expert sub-modules with the highest weights; an expert sub-module for performing a linear transformation on the integrated feature; an operation layer for performing a weighted summation of the outputs of all activated expert sub-modules and the corresponding weights to obtain a comprehensive feature representation; wherein, the expert sub-module is trained using a differentiated expert mask strategy.

[0073] A mixture of experts (MoE) is an advanced architectural paradigm that improves the overall performance of large-scale models by integrating multiple specialized sub-models (referred to as "experts"). Its core design principle is to dynamically decompose input tasks and route them to different expert sub-modules, thereby improving the accuracy of left ventricular hypertrophy classification models while maintaining computational efficiency. It is important to note that the MoE framework used in this paper shares conceptual similarities with left ventricular hypertrophy (LVH) diagnostic guidelines, namely, both use differentiated classification thresholds based on gender.

[0074] Figure 5 Schematic diagram of the structure of the hybrid expert module provided by the embodiment of the present invention. Figure 5 In some embodiments, the gating network may include a convolutional layer (CNN Features), a fully connected layer (Linear), and a classification layer (Softmax), and the operation layer may include a multiplication operation layer (multiply) and an addition layer (add).

[0075] Specifically, the output features of each feature extraction module can be combined to form an integrated feature. A convolutional layer can then be used to extract useful feature representations from the integrated feature. A fully connected layer can then be used to map the features extracted by the convolutional layer to a space with other dimensions. Finally, a softmax layer can be used to convert the output of the fully connected layer into a probability distribution, outputting the weights corresponding to each expert submodule. Gender information (i.e., gender) can be input into the model as an additional feature to increase the sensitivity of the left ventricular hypertrophy classification model to gender differences, thereby improving classification accuracy.

[0076] Specifically, during inference, a Top-8 gating mechanism can be used to activate only the eight expert submodules with the highest weights to optimize computational throughput. Each expert submodule performs a linear transformation on the input data (i.e., integrated features). The outputs of the activated expert submodules are then weighted using the weights corresponding to each expert submodule output by the Softmax layer through a multiply layer. Finally, all weighted data is summed through an add layer, achieving a weighted summation of the outputs of all activated expert submodules and their corresponding weights to obtain a comprehensive feature representation. This comprehensive feature representation is then output through the output layer for subsequent classification tasks.

[0077] It should be noted that the eight expert submodules with the highest weights may also be different depending on the current input data.

[0078] In some implementations, the MoE module can include 24 expert submodules, each composed of 16 neurons. To simulate gender-specific feature processing, the MoE module can employ a differentiated expert masking strategy during training and testing: for male subject data, only expert submodules 11-24 are invoked (disabling expert submodules 1-10), while for female subject data, expert submodules 1-10 and 21-24 are activated (disabling expert submodules 10-20). Expert submodules 20-24 can be designated as shared modules, participating in all data discrimination tasks.

[0079] In an optional embodiment, the left ventricular hypertrophy classification model is trained by the following steps: Step S21, randomly dividing the sample data set into multiple mutually exclusive subsets; the sample data set includes sample ECG signals of multiple user objects and corresponding classification labels; Step S22: in the current iteration process, determining a mutually exclusive subset as a test data set, and determining other mutually exclusive subsets except the test data set as training data sets; Step S23, determining the median waveform of the sample ECG signals in the training data set, and extracting amplitude features from the median waveform of the sample ECG signals; Step S24, inputting the amplitude feature and the median waveform of the sample ECG signal into the classification model to be trained, and obtaining a probability classification training result output by the classification model to be trained; Step S25 , adjusting the parameters of the classification model to be trained based on the loss between the classification label and the probability classification training result, and obtaining the left ventricular hypertrophy classification model after the model training is completed.

[0080] In some embodiments, a comprehensive biomedical database containing comprehensive health information and biological sample data for over 500,000 participants can be obtained. In this database, 54,962 subjects have available electrocardiogram (ECG) records, of which 38,290 participants have both cardiac magnetic resonance (CMR) data and corresponding ECG records. Therefore, a sample dataset can be constructed based on data from this comprehensive biomedical database.

[0081] The ECG data in this database is 10 seconds long, 12 leads, and the sampling frequency is 500Hz. The database sample distribution statistics can be shown in Table 1: Table 1

[0082] Except for gender, all other values ​​are expressed as mean ± variance.

[0083] In some embodiments, to eliminate biases that may be introduced by data partitioning and enhance model robustness, embodiments of the present invention employ a five-fold cross-validation strategy during model training and evaluation. Specifically, the sample dataset is randomly divided into five mutually exclusive subsets. In each iteration, four of these mutually exclusive subsets (approximately 24,505 samples) are used as the training dataset, and the remaining subset (approximately 7,658 samples) is used as the test dataset.

[0084] In an embodiment of the present invention, after determining a training dataset and a test dataset, the median waveform of the sample ECG signals in the training dataset can be determined, and amplitude features can be extracted from the median waveform of the sample ECG signals. The amplitude features and the median waveform of the sample ECG signals are then input into a classification model to be trained, and a probabilistic classification training result output by the classification model to be trained is obtained. Based on the loss between the classification label and the probabilistic classification training result, the parameters of the classification model to be trained are adjusted, and after model training is completed, the left ventricular hypertrophy classification model is obtained. The classification label can be negative / positive.

[0085] The present invention adopts a five-fold cross-validation strategy to train and evaluate the left ventricular hypertrophy classification model, which can minimize the impact of data partitioning deviation on model accuracy and is conducive to improving the classification accuracy of the left ventricular hypertrophy classification model.

[0086] In order to better understand the embodiments of the present invention, an embodiment is provided below for illustration.

[0087] Figure 6 This is the second flow chart of the left ventricular hypertrophy classification method provided by the embodiment of the present invention. Figure 6 In one embodiment, the present invention can implement the classification of left ventricular hypertrophy through an ECG signal classification method based on feature fusion, thereby solving the problem of amplitude information loss caused by the ECG signal normalization process in the prior art and improving the classification accuracy of left ventricular hypertrophy. The method may include the following steps: 1. Signal preprocessing: Obtain the raw ECG signal, remove baseline drift and high-frequency noise through filtering, and detect the R-wave position to align the heartbeat; 2. Median waveform generation: Calculate the median waveform of the aligned heartbeats to eliminate inter-individual variation and preserve morphological features; 3. Amplitude feature extraction: Extract the amplitude features of the P wave, QRS complex, and T wave from the median waveform to form an amplitude feature vector; 4. Multimodal feature fusion: The amplitude feature vector and ECG signal are input into a convolutional neural network (CNN) and feature fusion is performed in the CNN. 5. Classification decision: Output the probability classification result of left ventricular hypertrophy through the fully connected layer.

[0088] To comprehensively evaluate the performance of left ventricular hypertrophy classification models, a confusion matrix can be used as a fundamental tool for accuracy assessment. The confusion matrix can be used to calculate four core performance metrics: accuracy, precision, recall, and F1-score. Together, these metrics constitute a multi-dimensional quantitative system for model prediction accuracy. For validation of binary classification tasks, the receiver operating characteristic (ROC) area under the curve (AUC) metric can be additionally introduced. This metric measures the left ventricular hypertrophy classification model's ability to discriminate between classes at different thresholds, providing a classification performance evaluation standard that is unaffected by class distribution.

[0089] Figure 7 Schematic diagram of the confusion matrix provided by the embodiment of the present invention. Figure 7 True Positives (TP) refers to the number of positive classes correctly predicted by the model, located in the lower right corner of the figure, with a value of 160.0; True Negatives (TN) refers to the number of negative classes correctly predicted by the model, located in the upper left corner of the figure, with a value of 37693.0; False Positives (FP) refers to the number of negative classes incorrectly predicted as positive by the model, located in the upper right corner of the figure, with a value of 60.0; False Negatives (FN) refers to the number of positive classes incorrectly predicted as negative by the model, located in the lower left corner of the figure, with a value of 376.0. Therefore, based on TP, TN, FP, and FN, the following metrics can be calculated: accuracy 0.989, precision 0.727, recall 0.299, and F1 score 0.423.

[0090] In comparison experiments with traditional ECG diagnostic standards, the Cornell voltage standard, the Sokolow-Lyon standard, and the aVL lead standard can be selected as references. These three standards are among the most commonly used ECG judgment rules for clinical diagnosis of left ventricular hypertrophy (LVH). Specifically: 1) The Cornell voltage standard is based on the combination of the S wave amplitude in lead V3 and the R wave amplitude in lead aVL. When the sum of the two exceeds the sex-specific threshold, it indicates LVH. 2) The Sokolow-Lyon criteria are based on the sum of the amplitudes of the S wave in lead V1 and the R wave in leads V5 / V6, using a preset cutoff value as the basis for diagnosis; 3) The aVL lead standard examines the R wave amplitude in the aVL lead alone, and when it reaches a certain height, it is judged to be positive for LVH.

[0091] As shown in Table 2, the comparative experimental results show that the F1 score of the left ventricular hypertrophy classification model (i.e., the EFFNet model) is 67.4% higher than the best-performing traditional electrocardiogram standard.

[0092] Table 2

[0093] Figure 8 : This is a schematic diagram comparing the ROC curves of the EFFNet model and the ECG rule provided by the embodiment of the present invention. Figure 8 To further demonstrate the performance differences, ROC curve comparisons clearly demonstrate the EFFNet model's performance advantages over traditional diagnostic standards. These comparisons not only validate the model's effectiveness in identifying LVH, but also reveal its complementary and enhanced role in traditional diagnostic methods.

[0094] The left ventricular hypertrophy classification device provided by the present invention is described below. The left ventricular hypertrophy classification device described below and the left ventricular hypertrophy classification method described above can be referred to in correspondence with each other.

[0095] Figure 9 Schematic diagram of the structure of the left ventricular hypertrophy classification device provided by the embodiment of the present invention. Figure 9 An embodiment of the present invention provides a left ventricular hypertrophy classification device, which may include the following modules: Extraction module 910, for determining the median waveform of the ECG signal and extracting amplitude features from the median waveform; a classification module 920 for inputting the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; Among them, the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

[0096] The embodiment of the present invention inputs the amplitude features and median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy output by the model, thereby realizing efficient and high-precision automatic classification of left ventricular hypertrophy diseases based on electrocardiogram signals based on deep learning; by designing the left ventricular hypertrophy classification model to include multiple parallel feature extraction modules for analyzing different electrocardiogram leads, the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude features, thereby realizing amplitude restoration, solving the problem of amplitude information loss caused by the ECG signal standardization process, and helping to further improve the classification accuracy of left ventricular hypertrophy.

[0097] Figure 10 An example of a physical structure diagram of an electronic device is shown below. Figure 10 As shown, the electronic device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communications bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other via the communications bus 1040. The processor 1010 may call logic instructions in the memory 1030 to execute a left ventricular hypertrophy classification method, which includes: determining a median waveform of an electrocardiogram signal and extracting amplitude features from the median waveform; inputting the amplitude features and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result for left ventricular hypertrophy; wherein the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction modules include a feature fusion module for normalizing the median waveform and performing an element-by-element multiplication operation on the normalized median waveform and the amplitude features to achieve amplitude restoration.

[0098] Furthermore, the logic instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0099] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the left ventricular hypertrophy classification method provided by the above methods, the method including: determining the median waveform of the electrocardiogram signal and extracting amplitude features from the median waveform; inputting the amplitude features and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result of left ventricular hypertrophy; wherein the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

[0100] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the left ventricular hypertrophy classification method provided by the above-mentioned methods, the method comprising: determining the median waveform of the electrocardiogram signal, and extracting amplitude features from the median waveform; inputting the amplitude features and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probabilistic classification result of left ventricular hypertrophy; wherein the left ventricular hypertrophy classification model comprises multiple parallel feature extraction modules for analyzing different electrocardiogram leads; the feature extraction module comprises a feature fusion module for normalizing the median waveform, and performing element-by-element multiplication operation on the normalized median waveform and the amplitude features to achieve amplitude restoration.

[0101] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0102] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for classifying left ventricular hypertrophy, characterized in that: include: determining a median waveform of the electrocardiogram signal and extracting amplitude features from the median waveform; Inputting the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; Wherein, the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; The feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

2. The left ventricular hypertrophy classification method according to claim 1, characterized in that: The feature extraction module includes multiple feature fusion modules and maximum pooling layers arranged in an interlaced manner, which are used to gradually extract features of specific electrocardiogram leads; the feature fusion module includes: a normalization layer, configured to perform normalization processing on the first local feature of the median waveform to obtain a second local feature; a multiplication operation layer, configured to perform an element-by-element multiplication operation on the second local feature and the amplitude feature to generate an intermediate feature; A convolutional layer, configured to extract features from the intermediate features to obtain third local features; a batch normalization layer, configured to perform normalization on the third local feature to obtain a fourth local feature; an addition layer, configured to add the fourth local feature to the first local feature to obtain an added feature; The output layer is used to output the added features.

3. The left ventricular hypertrophy classification method according to claim 1, characterized in that: The left ventricular hypertrophy classification model further includes a mixed expert module; the mixed expert module includes: A gating network is used to integrate the output features of each feature extraction module to obtain an integrated feature; based on the integrated feature, the weights corresponding to each expert submodule are calculated, and a preset number of expert submodules with the highest weights are activated; An expert submodule, configured to perform a linear transformation on the integrated features; The computation layer is used to perform weighted summation of the outputs of all activated expert submodules and the corresponding weights to obtain a comprehensive feature representation; The expert submodule is trained using a differentiated expert mask strategy.

4. The left ventricular hypertrophy classification method according to claim 1, characterized in that: The left ventricular hypertrophy classification model is trained in the following way: Randomly dividing a sample data set into multiple mutually exclusive subsets; the sample data set includes sample electrocardiogram signals of multiple user objects and corresponding classification labels; In the current iteration process, a mutually exclusive subset is determined as a test data set, and other mutually exclusive subsets except the test data set are determined as training data sets; Determining a median waveform of the sample ECG signals in the training data set, and extracting amplitude features from the median waveform of the sample ECG signals; Inputting the amplitude feature and the median waveform of the sample electrocardiogram signal into a classification model to be trained, and obtaining a probability classification training result output by the classification model to be trained; Based on the loss between the classification label and the probability classification training result, the parameters of the classification model to be trained are adjusted, and the left ventricular hypertrophy classification model is obtained after the model training is completed.

5. The left ventricular hypertrophy classification method according to claim 1, characterized in that: The step of obtaining a median waveform of an electrocardiogram signal includes: Acquire an original electrocardiogram signal, filter out baseline interference and high-frequency interference in the original electrocardiogram signal, and obtain an electrocardiogram signal after interference filtering; Detecting an R-wave position, dividing the interference-filtered ECG signal into a plurality of cardiac cycles based on the R-wave position, and truncating the ECG signal within each cardiac cycle to obtain a new cardiac cycle; All new heartbeat cycles are aligned to the same time axis and the median of each time point is calculated to obtain the median waveform.

6. The left ventricular hypertrophy classification method according to claim 1, characterized in that: The extracting the amplitude feature from the median waveform includes: An electrocardiogram signal waveform analysis tool is used to extract the amplitude characteristics of a target waveform from the median waveform; the target waveform includes at least one of a P wave, a Q wave, an R wave, an S wave, a T wave, a P1 wave, an R1 wave, an S1 wave, a T1 wave, a J wave, a JX wave, and a JXE wave.

7. A left ventricular hypertrophy classification device, characterized in that: include: an extraction module, configured to determine a median waveform of the electrocardiogram signal and extract amplitude features from the median waveform; a classification module, configured to input the amplitude feature and the median waveform into a trained left ventricular hypertrophy classification model to obtain a probability classification result of left ventricular hypertrophy; Wherein, the left ventricular hypertrophy classification model includes multiple parallel feature extraction modules for analyzing different electrocardiogram leads; The feature extraction module includes a feature fusion module for normalizing the median waveform and performing element-by-element multiplication operation on the normalized median waveform and the amplitude feature to achieve amplitude restoration.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the left ventricular hypertrophy classification method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the left ventricular hypertrophy classification method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the left ventricular hypertrophy classification method according to any one of claims 1 to 6 is implemented.

Images