Multi-feature fusion heart sound classification method based on channel attention mechanism

Abstract

The invention discloses a multi-feature fusion heart sound classification method based on a channel attention mechanism, and belongs to the technical field of signal processing and heart disease diagnosis. The method comprises the steps that collected original heart sound records are preprocessed, a self-adaptive two-stage filtering method based on local variance and global variance is adopted, a wavelet denoising technology is combined, background noise is effectively removed, high-quality heart sound signals are extracted, and meanwhile heart sound events are accurately detected through a double-threshold segmentation method; fusing information of different dimensions to enrich the expression of the heart sound signal; a deep learning architecture fusing a convolutional recurrent neural network and a convolutional neural network is constructed, a channel attention mechanism is introduced, the fusion weight of each sub-network is dynamically adjusted according to the importance of features, and the classification accuracy is improved. According to the method, the quality of heart sound signals can be effectively improved, multi-domain features are fully fused, meanwhile, the discrimination ability of the model for heart sound categories is enhanced, and therefore the heart sound diagnosis precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of signal processing and heart disease diagnosis, and particularly relates to a multi-feature fusion heart sound classification method based on channel attention mechanism. Background Art

[0002] Signal processing technology is ubiquitous. Signals can be sounds, images, electromagnetic waves, etc., covering many fields such as audio processing, image analysis, communication, etc. The goal of signal processing is to eliminate noise, enhance signal quality, extract useful information, compress data, etc.

[0003] In recent years, the incidence and mortality of cardiovascular diseases (CVD) have remained high globally, becoming one of the major challenges faced by global public health. Its high incidence makes early screening and diagnosis particularly important. Heart sound, as an important physiological signal representing the state of the cardiovascular system, contains rich pathological information. By analyzing and processing heart sound signals, non-invasive detection of CVD can be achieved, providing an important basis for early detection and personalized intervention of diseases. However, traditional heart sound classification usually relies on doctors' experience, and the results of heart sound auscultation largely depend on doctors' auscultation skills and experience levels.

[0004] To overcome the limitations of time-consuming and laborious manual auscultation, some methods based on machine learning and deep learning have been proposed. Deep networks such as CNN and CRNN have shown superior performance in heart sound auscultation. These methods can reduce the dependence on manual experience and significantly improve the level of automated heart sound auscultation and diagnostic accuracy. However, the above methods still have the following problems:

[0005] Poor applicability of data acquisition equipment: Most methods for heart sound auscultation study heart sounds collected by digital stethoscopes. In non-clinical environments, the purchase of digital stethoscopes requires additional expenses, and their portability is also insufficient. At the same time, using a digital stethoscope for heart sound collection usually requires repeated operations at multiple auscultation points, such as the aortic valve auscultation area, pulmonary valve auscultation area, etc., increasing the complexity of the operation and the discomfort of patients.

[0006] Insufficient multi-network collaboration ability: Existing models mostly adopt a single network structure or directly fuse the outputs of multiple networks, making it difficult to fully utilize the complementary information of different features, resulting in limited classification performance. A single network structure may have strong adaptability to a certain type of feature, but it is difficult to extract multi-level information of heart sounds. And simply stacking the outputs of multiple networks lacks pertinence and may introduce irrelevant or redundant information, affecting the final classification effect. In addition, different networks have different sensitivities to features, and a fixed fusion method is difficult to dynamically adapt to the feature distributions of different heart sound signals, further limiting the generalization ability of the model. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a multi - feature fusion heart sound classification method based on channel attention mechanism, which combines the pathological features, bispectrum features of heart sound signals with the MFCC feature maps commonly used in heart sound signal analysis, constructs a heart sound classification network based on channel attention mechanism, and thus realizes heart sound classification.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A multi - feature fusion heart sound classification method based on channel attention mechanism includes the following steps:

[0010] Step 1: For the collected original heart sound signal, after removing the DC component of the signal, median filtering is performed using an adaptive two - stage filtering method, and Butterworth band - pass filtering is used to retain the effective components of the heart sound signal. High - quality heart sound signals are obtained through wavelet denoising. Heart sound events are detected by the double - threshold segmentation method to obtain the start and end positions of each heart sound event in the high - quality heart sound signal.

[0011] Step 2: Construct a feature extraction module. Based on the high - quality heart sound signal and the start and end positions of heart sound events, artificial features representing different heart sound patterns are extracted, namely IBI features, bispectrum features, and MFCC feature maps.

[0012] Step 3: Construct a heart sound classification network based on channel attention mechanism. By optimizing feature selection, it automatically focuses on key features, enhances the model's learning ability for heart sound signals. The channel attention mechanism dynamically adjusts the weights of each channel, improves the representation of important features, suppresses redundant information, and thus improves the accuracy of the model in heart sound classification, especially the recognition ability of abnormal heart sounds.

[0013] Step 4: Input the high - quality heart sound signal and the different artificial features into the heart sound classification network based on channel attention mechanism for training to obtain a trained multi - feature channel attention model MFCA.

[0014] Step 5: After processing the heart sound signal to be classified through Steps 1 - 2, input it into the trained model to obtain a heart sound diagnosis result.

[0015] According to a preferred embodiment of the present invention, Step 1 specifically includes:

[0016] Step 1.1: Place the mobile device on the chest of the subject, use the built - in microphone to collect the original heart sound signal of the subject, and perform mean normalization on the original heart sound signal to eliminate the DC component of the signal.

[0017] Step 1.2: Design an adaptive two-stage filtering method. First, calculate the global variance of the original heart sound signal after removing the DC component, and use it as a threshold. Set a sliding window with a size of 10 seconds, perform median filtering within each sliding window, and calculate the local variance of the window. If the local variance of the window exceeds the threshold, perform quadratic median filtering;

[0018] Step 1.3: To retain the effective frequency components of the heart sound signal, use a Butterworth filter with a frequency range of 25 Hz to 400 Hz to filter out noise and retain the effective frequency components of the heart sound signal;

[0019] Step 1.4: Use multi-Bessel bases to perform 5-level wavelet decomposition on the heart sound signal retaining the effective frequency components. The original signal is decomposed into wavelet coefficients in multiple different frequency bands, including high-frequency and low-frequency components. Use the soft threshold denoising method to process the obtained wavelet coefficients, set the threshold to 1.5 times the standard deviation of the noise, and through the wavelet reconstruction process, recombine the denoised coefficients to obtain a high-quality heart sound signal;

[0020] Step 1.5: For the high-quality heart sound signal, use the double-threshold segmentation method to detect the heart sound events existing therein. Heart sound events refer to specific sound waveforms generated by the cardiac cycle in the heart sound signal. Set the high threshold to be 20% of the average amplitude value of the high-quality heart sound signal, and the low threshold to be 0.1% of the maximum amplitude value of the high-quality heart sound signal. After detecting the local peaks of the high-quality heart sound signal, mark the start and end points of each heart sound event through the double threshold, and define the time range of the effective high-quality heart sound signal.

[0021] According to a preferred embodiment of the present invention, step 2 specifically includes:

[0022] Step 2.1: Use the start and end positions of each heart sound event to calculate the number of heartbeat cycles and the duration of each heartbeat cycle, and obtain the indicators used as IBI features. The indicators include the average heartbeat interval, the median heartbeat interval, the standard deviation of the heartbeat interval, the 50 ms interval ratio, the root mean square difference of the heartbeat interval, the heart rate variability time index, the standard deviation of the short axis and the long axis of the Poincaré plot, and the ratio of the two;

[0023] Step 2.2: Perform Fourier transform on the high-quality heart sound signal to obtain the frequency domain representation , select the frequency pair and , perform complex conjugate operations on the spectral components corresponding to the frequency composite components and , and calculate the expected value of the non-linear interaction between the frequency points to obtain the bispectral features;

[0024] Step 2.3: After performing Fourier transform on the high-quality signal, process the spectrum using a Mel filter bank, then take the logarithm of the output value of each filter, and apply discrete cosine transform (DCT) to extract the first 13 MFCC coefficients as MFCC features.

[0025] According to a preferred embodiment of the present invention, step 3 specifically includes: The heart sound classification network includes a parallel one-dimensional CNN module, a one-dimensional CRNN module, and a two-dimensional CNN module. Among them, the one-dimensional CNN module and the one-dimensional CRNN module respectively process the inter-beat interval (IBI) feature and the high-quality heart sound signal to obtain one-dimensional features, and the two-dimensional CNN module is used to process the bispectrum feature and the Mel frequency cepstral coefficient (MFCC) feature map to obtain two-dimensional features; The one-dimensional CNN module, the one-dimensional CRNN module, and the two-dimensional CNN module include a plurality of serial convolutional layers, batch normalization layers, and max pooling layers. The one-dimensional CRNN module also includes a bidirectional long short-term memory network at the output end.

[0026] The heart sound classification network also includes a one-dimensional channel attention module for processing one-dimensional features and a two-dimensional channel attention module for processing two-dimensional features.

[0027] The channel attention module generates attention weights through global average pooling, a fully connected layer, and a ReLu activation function, and multiplies them with the original input features to achieve weighting. When the original features are two-dimensional feature data, use of the input format, where represents the number of channels, represents the height, represents the width. When the original features are one-dimensional data, use of the format, where represents the length of the data. Through the channel attention mechanism, the network can dynamically adjust the weights according to the importance of different channels, enhance the model's attention to important information in the heart sound signal, and suppress redundant or irrelevant information.

[0028] According to a preferred embodiment of the present invention, step 4 specifically includes:

[0029] Input the high-quality heart sound signal obtained in step 1 and the three artificial features obtained in step 2, namely the inter-beat interval feature, the bispectrum feature, and the MFCC feature map, as training samples into the heart sound classification network based on the channel attention mechanism constructed in step 3. The high-quality heart sound signal corresponds to the one-dimensional CRNN module in the heart sound classification network, the inter-beat interval feature corresponds to the one-dimensional CNN module, and the bispectrum feature and the MFCC feature map correspond to the two-dimensional CNN module. Use the Adam optimizer to optimize the network, and finally obtain the trained heart sound classification model MFCA.

[0030] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the foregoing multi-feature fusion heart sound classification method based on a channel attention mechanism.

[0031] In a third aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which when executed by a processor can cause the processor to implement the foregoing multi-feature fusion heart sound classification method based on a channel attention mechanism.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention designs a multi-feature fusion network based on a channel attention mechanism for heart sound classification. By integrating multi-dimensional CNN and CRNN, this model effectively fuses multi-domain features and captures deep and comprehensive information of heart sounds, thereby improving the classification accuracy.

[0034] Signal processing methods such as adaptive two-stage filtering and wavelet denoising are used to preprocess the heart sound signal, which can effectively remove environmental noise and improve the quality of the heart sound signal, thereby providing a higher-quality data basis to meet actual clinical needs.

[0035] A heart sound dataset collected by a mobile device is constructed. At the same time, compared with traditional methods, the method of the present invention can achieve higher accuracy in both binary classification and multi-classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a framework diagram of the multi-feature fusion heart sound classification method based on a channel attention mechanism of the present invention;

[0037] Figure 2 is a schematic diagram of the multi-feature fusion heart sound classification network architecture based on a channel attention mechanism of the present invention;

[0038] Figure 3 is a schematic diagram of the channel attention structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present invention will be further described below with reference to the drawings and embodiments.

[0040] As Figure 1 shown, the process of the multi-feature fusion heart sound classification method based on a channel attention mechanism of the present invention is as follows:

[0041] Step 1: In the preprocessing stage, first perform a mean normalization operation on the original heart sound signal to ensure that the signal fluctuations are near zero and avoid the interference of the DC component on subsequent processing. An adaptive two-stage filtering method is used to remove extreme outliers in a short time. In the first stage, calculate the global variance of the entire original heart sound signal as the threshold, perform median filtering on each sliding window set to 10 seconds, and calculate the local variance of each window of the original heart sound signal. If the local variance of the window exceeds the threshold, perform a secondary median filtering. To retain the effective frequency components of the heart sound signal, use a Butterworth band-pass filter with a frequency range of 25 Hz to 400 Hz to filter out noise. Use the multi-Bessel wavelet basis to perform 5-level wavelet decomposition on the signal, and reconstruct the wavelet coefficients into a high-quality heart sound signal based on a soft threshold, where the soft threshold is set to 1.5 times the standard deviation of the noise. After denoising, use a double-threshold segmentation method to detect heart sound events, and set the high threshold to be 20% of the average amplitude value of the high-quality heart sound signal, which is used to detect the potential range of heart sound events, and set the low threshold to be 0.1% of the maximum amplitude value of the high-quality heart sound signal to further find the precise endpoints of heart sound events and complete the annotation of the start and end times of heart sound events.

[0042] Step 2: In the feature extraction stage, a total of three artificial features are extracted, namely IBI features, bispectrum features, and MFCC feature maps.

[0043] Step 2.1: Extract IBI features: Use the start and end times of the heart sound events obtained above to obtain the number of heartbeats and the duration of each heartbeat cycle, and extract the following IBI features to evaluate heart health:

[0044] Time domain: average heart rate interval, median heart rate interval, standard deviation of heart rate intervals, 50 ms interval ratio, root mean square difference of heart rate intervals, heart rate variability time index;

[0045] Nonlinear: standard deviation of the short axis and long axis of the Poincaré plot and their ratio.

[0046] Step 2.2: Extract bispectrum features: Perform a short-time Fourier transform (STFT) on the heart sound signal, calculate the bispectrum to quantify the phase coupling information of different frequency components, so as to capture the nonlinear characteristics of the heart sound signal. Specifically: perform a Fourier transform on the signal to obtain the frequency domain representation , select the frequency pair and , for the frequency composite component corresponding spectral components and perform complex conjugate operations, and calculate the expected value of the nonlinear interaction between frequency points to obtain the bispectrum features.

[0047] Step 2.3, extract MFCC feature map: use the Mel filter bank to process the STFT spectrum, extract the Mel spectrum and perform DCT, obtain MFCC features, simulate the nonlinear perception characteristics of the human ear, and finally extract a MFCC feature map of size A two-dimensional matrix, where is the number of time frames.

[0048] Step 3: Construct a heart sound classification network. The network architecture combines multiple deep learning modules to comprehensively improve the classification effect of heart sound signals. The entire network consists of a one-dimensional CNN module, a one-dimensional CRNN module, and a two-dimensional CNN module. The network architecture details are as follows: Figure 2 As shown in the figure, each network module processes the original heart sound signal, IBI feature, bispectral feature and MFCC feature map respectively. The heart sound signal and the features extracted from it are input into different networks. The heartbeat interval feature is a series of one-dimensional data calculated based on the heartbeat interval, so it is input into the one-dimensional CNN module to mine the hidden rules and patterns of different pathological features. The heart sound signal contains rich time domain and frequency domain information, so it is input into the one-dimensional CRNN module. While extracting local features using the convolution layer, it can capture long-term correlation through RNN, which helps to accurately model the dynamic characteristics of the heart sound signal. The two-dimensional CNN module is suitable for processing two-dimensional data with complex structures, such as MFCC feature maps and bispectral features. Specifically: the two-dimensional CNN module is composed of multiple serial convolution layers, batch normalization layers and maximum pooling layers, and the important channels are weighted by the two-dimensional channel attention mechanism; the one-dimensional CNN module and the one-dimensional CRNN module containing the bidirectional long short-term memory network both contain serial convolution layers, batch normalization layers and maximum pooling layers, and the important channels are weighted by the one-dimensional channel attention mechanism. The one-dimensional and two-dimensional features processed by the attention module are fused and input into the classifier to obtain the classification result. The channel attention structure generates attention weights through global average pooling, fully connected layers and activation functions, and multiplies them with the original features to achieve weighting. Each part works closely together to complete the heart sound classification task.

[0049] To further improve network performance, design Figure 3 The channel attention mechanism module shown enables the network to dynamically adjust the contribution of features in different network modules through channel weighting. The channel attention structure generates attention weights through global average pooling, fully connected layers, and ReLu activation functions, and multiplies them with the original features to achieve weighting. When the original features are two-dimensional feature data, use The input format is: Indicates the number of channels, Indicates height, Indicates width. When the original feature is one-dimensional feature data, use The format is Indicates the length of the data.

[0050] Step 4: Train the constructed heart sound classification model. Input the high-quality heart sound signals and the different artificial features into the constructed network model for training to obtain the trained model MFCA.

[0051] Step 5: Input the heart sound signals to be classified, after being processed in Step 1 and Step 2, into the trained heart sound classification model MFCA to obtain the heart sound diagnosis result.

[0052] The experimental results of MFCA on the collected data set are accuracy 0.9954, specificity 0.9938, sensitivity 0.9969, and average accuracy 0.9954, which fully prove the effectiveness of the method of the present invention.

[0053] In a second aspect, the present invention provides an electronic device, including: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the foregoing multi-feature fusion heart sound classification method based on the channel attention mechanism.

[0054] In a third aspect, the present invention provides a computer-readable storage medium, on which executable instructions are stored, and when the instructions are executed by a processor, the processor can implement the foregoing multi-feature fusion heart sound classification method based on the channel attention mechanism.

[0055] The specific embodiments described above have further elaborated on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-feature fusion heart sound classification method based on channel attention mechanism, characterized in that, It includes the following steps: Step 1: Preprocess the collected original heart sound signal to obtain a high-quality heart sound signal. Detect the heart sound events of the high-quality heart sound signal through the double-threshold segmentation method to obtain the start and end positions of each heart sound event; Step 2: Construct a feature extraction module. Based on the obtained high-quality heart sound signal and the start and end positions of each heart sound event, extract artificial features representing different heart sound patterns; Step 3: Construct a heart sound classification network based on the channel attention mechanism, and extract the one-dimensional and two-dimensional features of the high-quality heart sound signal and the artificial features of the different heart sound patterns; Step 4: Input the high-quality heart sound signal and the artificial features of the different heart sound patterns into the heart sound classification network based on the channel attention mechanism for training to obtain a trained multi-feature channel attention model MFCA; Step 5: After processing the heart sound signal to be classified through Steps 1 - 2, input it into the trained model MFCA to obtain a heart sound diagnosis result.

2. The multi-feature fusion heart sound classification method based on channel attention mechanism according to claim 1, characterized in that In Step 1, preprocessing the collected original heart sound signal to obtain a high-quality heart sound signal includes removing the DC component of the original heart sound signal, performing median filtering using an adaptive two-stage filtering method, and using Butterworth band-pass filtering to retain the effective components of the heart sound signal, and obtaining a high-quality heart sound signal through wavelet denoising.

3. The multi-feature fusion heart sound classification method based on channel attention mechanism according to claim 2, characterized in that Step 1 specifically includes: Step 1.1: Use a mobile device to collect the original heart sound signal of the subject, and perform mean normalization on the original heart sound signal to eliminate the DC component; Step 1.2: Design an adaptive two-stage filtering method, calculate the global variance of the original heart sound signal after removing the DC component, use it as a threshold, set a sliding window, perform median filtering within each sliding window, and calculate the local variance of the sliding window. If the local variance of the sliding window exceeds the threshold, perform secondary median filtering; Step 1.3: Use a Butterworth filter with a frequency range of 25Hz to 400Hz to filter out noise and retain the effective frequency components of the heart sound signal; Step 1.4: Use the multi-Bessel wavelet basis to perform 5-level wavelet decomposition on the heart sound signal retaining the effective frequency components, decompose the signal into wavelet coefficients of different frequencies, use the soft threshold denoising method to process and obtain the wavelet coefficients, set the threshold to 1.5 times the noise standard deviation, and through the wavelet reconstruction process, recombine the denoised coefficients into a complete signal to obtain a high-quality heart sound signal; Step 1.

5. For high-quality heart sound signals, use the double-threshold segmentation method to detect the heart sound events therein. The heart sound events refer to specific sound waveforms generated by the cardiac cycle in high-quality heart sound signals. Set the high threshold to be 20% of the average amplitude of the high-quality heart sound signal, and the low threshold to be 0.1% of the maximum amplitude value of the high-quality heart sound signal. After detecting the local peaks of the high-quality heart sound signal, mark the start and end points of each heart sound event through the double thresholds to define the time range of the effective high-quality heart sound signal.

4. A multi-feature fusion heart sound classification method based on channel attention mechanism according to claim 1, characterized in that, In Step 2, the artificial features representing different heart sound patterns include the inter-beat interval IBI feature, the bispectrum feature, and the Mel-frequency cepstral coefficient MFCC feature map.

5. The multi-feature fusion heart sound classification method based on channel attention mechanism according to claim 4, characterized in that, Step 2 specifically includes: Step 2.1: Use the start and end positions of each heart sound event to calculate the number of heart beats and the duration of each heart beat cycle, and obtain the indicators as the IBI feature. The indicators include the average heart beat interval, the median heart beat interval, the standard deviation of the heart beat interval, the 50ms interval ratio, the root mean square difference of the heart beat intervals, the heart rate variability time index, the standard deviation of the short axis of the Poincaré plot and the standard deviation of the long axis and their ratio; Step 2.2: Perform Fourier transform on the high-quality heart sound signal to obtain frequency domain representation , select the frequency pair and , for the frequency composite component The corresponding spectral components and Perform complex conjugate operation and calculate the expected value of the nonlinear interaction between frequency points to obtain the bispectral characteristics; Step 2.3: After performing Fourier transform on the high-quality heart sound signal, use the Mel filter bank to process the spectrum, then take the logarithm of the output value of each filter, and apply discrete cosine transform to extract the first 13 MFCC coefficients as MFCC features.

6. The multi - feature fusion heart sound classification method based on channel attention mechanism according to claim 1, characterized in that The said step 3 includes: The heart sound classification network includes a parallel one-dimensional CNN module, a one-dimensional CRNN module, and a two-dimensional CNN module. Among them, the one-dimensional CNN module and the one-dimensional CRNN module respectively process the inter-beat interval (IBI) feature and the high-quality heart sound signal to obtain one-dimensional features, and the two-dimensional CNN module is used to process the bispectrum feature and the Mel frequency cepstral coefficient (MFCC) feature map to obtain two-dimensional features; the one-dimensional CNN module, the one-dimensional CRNN module, and the two-dimensional CNN module include a plurality of convolutional layers, batch normalization layers, and max pooling layers in series, and the one-dimensional CRNN module also includes a bidirectional long short-term memory network at the output end.

7. A multi - feature fusion heart sound classification method based on channel attention mechanism according to claim 6, characterized in that, The heart sound classification network also includes a one-dimensional channel attention module for processing one-dimensional features and a two-dimensional channel attention module for processing two-dimensional features.

8. A multi-feature fusion heart sound classification method based on channel attention mechanism according to claim 7, characterized in that The channel attention module generates attention weights through global average pooling, a fully connected layer, and a ReLu activation function, and multiplies them with the original input features to achieve weighting. When the original features are two-dimensional feature data, the input format is used, where represents the number of channels, represents the height, represents the width. When the original features are one-dimensional data, the format is used, where represents the length of the data.

9. An electronic device, characterized in that, It includes: One or more processors; A memory for storing one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement a multi-feature fusion heart sound classification method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, Stored thereon are executable instructions, which when executed by a processor can enable the processor to implement a multi-feature fusion heart sound classification method according to any one of claims 1-8.

Images