Mitral valve regurgitation severity assessment method based on PCG signal

Abstract

The invention provides a mitral valve regurgitation severity assessment method based on PCG signals, which comprises the following steps: firstly, synchronously acquiring signals from a heart sound area and an environmental background by adopting a dual-channel audio acquisition technology, eliminating environmental noise and operation interference (such as clothes friction and non-uniform pressing force of a stethoscope) through a self-adaptive noise elimination algorithm, and calculating the severity of the mitral valve regurgitation severity; and carrying out segmentation processing on continuous heart sound signals of the de-noised heart sound by adopting a dynamic time window segmentation strategy to generate short-time heart sound fragments, evaluating the quality of the short-time heart sound fragments in real time by using a model based on a lightweight convolutional neural network, and filtering low-quality fragments. Then, nonlinear features are directly extracted from the short-time heart sound fragments through a deep neural network model based on a clique block, after each effective fragment is subjected to four classifications, classification results of all the fragments are integrated through a majority voting mechanism, and final mitral valve regurgitation severity judgment is generated. The classification robustness is effectively improved through multi-fragment information fusion, and the random error of single-fragment analysis is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of heart sound signal processing, and in particular relates to a method for assessing the severity of mitral regurgitation based on PCG signals. Background Art

[0002] Mitral regurgitation (MR) is a common heart valve disease, and its severity assessment is crucial for clinical decision-making. Although echocardiography is currently the gold standard, it has the limitations of strong equipment dependence, complex operation, and difficulty in large-scale promotion. Traditional cardiac auscultation is limited by physician experience, and the quantitative analysis of murmur characteristics has low consistency, especially when distinguishing mild from severe mitral regurgitation. In recent years, the combination of electronic stethoscopes and artificial intelligence has provided a new approach for the automated analysis of cardiac sound signals (PCG), but existing technologies still face multiple challenges. For example, cardiac sound signals are easily affected by environmental noise and stethoscope operation, and traditional noise reduction algorithms have difficulty in ensuring signal fidelity. Most AI models only support the detection of the presence of mitral regurgitation and lack the ability to classify multi-level severity.

[0003] Existing heart sound analysis methods mainly include traditional pattern recognition methods based on signal processing and artificial intelligence methods based on deep learning. Traditional methods segment heart sound cycles through wavelet denoising and hidden Markov models, and extract manual features such as MFCC combined with support vector machine classification. However, their feature expression is limited and noise resistance is poor. Deep learning methods use neural networks to learn heart sound features end-to-end. Although they avoid the limitations of manual design, they still rely on precise cycle segmentation or time-frequency graph conversion. Although some studies have attempted to directly process the raw signal and focus on disease presence detection, they are limited by low-quality segment interference and lack multi-level classification for the severity of mitral regurgitation MR, making them insufficiently adaptable for practical applications. Summary of the Invention

[0004] To address the challenges of existing technologies, the present invention provides a method for assessing mitral regurgitation severity based on PCG signals. This method addresses the shortcomings of traditional heart sound analysis, which relies on manual experience, suffers from significant noise interference, and cannot accurately quantify MR severity. By integrating noise cancellation, filtering low-quality heart sound segments, and deep neural networks, the present invention proposes a fully automated solution. First, dual-channel audio acquisition technology is used to synchronously acquire signals from the heart sound region and the ambient background. An adaptive noise cancellation algorithm is used to eliminate environmental noise and operational interference (such as clothing friction and uneven stethoscope pressure). A dynamic time window segmentation strategy is applied to the de-noised heart sound signals, segmenting the continuous heart sound signal with a 2-second window length and a 1-second sliding step size to generate short-duration heart sound segments. A lightweight convolutional neural network model is used to evaluate the quality of these short-duration heart sound segments in real time and filter out low-quality segments. Subsequently, a deep neural network model based on clique blocks extracts nonlinear features directly from the short-duration heart sound segments. Each valid segment is classified into four categories (no MR, mild, moderate, and severe). The classification results of all segments are then combined using a majority voting mechanism to generate a final MR severity assessment. Short heart sound segments effectively reduce the amount of real-time computation required, ensuring efficient operation on mobile devices. Multi-segment information fusion effectively improves classification robustness and reduces the random error associated with single-segment analysis.

[0005] The technical solutions of the present invention are as follows:

[0006] A method for assessing the severity of mitral regurgitation based on PCG signals comprises the following steps:

[0007] Obtaining PCG signals of the patient to be evaluated and performing noise elimination processing on them;

[0008] The PCG signal after noise elimination is segmented into short-term heart sound segments using the sliding window method;

[0009] Based on the convolutional neural network, the availability of short-term heart sound segments is judged to obtain valid heart sound segments;

[0010] A deep neural network based on Clique blocks was used to extract features and classify valid heart sound segments to obtain classification results of different severity levels of mitral regurgitation.

[0011] The majority voting strategy was used to integrate the statistical frequencies of the different severity classification results of mitral regurgitation, and the highest frequency category was taken as the evaluation result.

[0012] Preferably, in the method for assessing the severity of mitral regurgitation based on PCG signals, obtaining the PCG signal of the patient to be assessed and performing noise elimination processing on the signal comprises the following steps:

[0013] An electronic stethoscope equipped with dual-channel acquisition is used to collect heart sound signals. The main channel collects target heart sounds, and the reference channel collects environmental background noise.

[0014] The ear clipping adaptive noise cancellation algorithm is used to suppress the interference of non-heart sound components. Assume that the main channel signal is s main (n), the reference channel signal is s ref (n), where n is the discrete time index;

[0015] Since the channels have different physical positions, the signals need to be aligned in the time domain:

[0016] s ref (n) = DelayAlign(s main (n),s ref (n))

[0017] After aligning the signals in the time domain, the background noise is dynamically estimated using the LMS algorithm:

[0018]

[0019] where w k (n) is the adaptive filter coefficient, M = 128 is the filter order;

[0020] The estimated noise component is taken out from the main channel, and the preliminary denoising result is:

[0021]

[0022] Then, a finite impulse response (FIR) bandpass filter is designed to retain the frequency band from 20 Hz to 1000 Hz. The filter transfer function is:

[0023]

[0024] Where h(k) is the filter coefficient designed based on the window function method (using Kaiser window, β = 5, order L = 256), and the final output signal is

[0025]

[0026] Output signal Eliminate high-frequency noise and low-frequency interference.

[0027] Preferably, in the method for assessing the severity of mitral regurgitation based on PCG signals, the PCG signals after noise elimination are segmented into short-duration heart sound segments using a sliding window method, comprising the following steps:

[0028] After denoising The sliding window method is used to segment the heart sound into short segments to balance the continuity of the signal and the computational efficiency. The parameter of the sliding window is set to the window length T. w =2s, the sampling point should be of length N w =T w ×f s , where f s =8000 is the sampling frequency of the stethoscope;

[0029] Therefore, N w =16000, sliding step size T s =1s, corresponding to N s =8000, for a signal of length L, generate fragments. The kth fragment is represented as:

[0030]

[0031] This design divides the continuous heart sound signal into several short heart sound segments of 2 seconds in length, and adjacent short heart sound segments overlap by 50% to avoid truncation of key features.

[0032] Preferably, in the method for assessing the severity of mitral regurgitation based on PCG signals, the availability of short-term segments is judged based on a convolutional neural network to obtain valid heart sound segments, comprising the following steps:

[0033] Short heart sound segments x k (n) To standardize:

[0034]

[0035] in is the segment mean, is the standard deviation of the segment, ∈=10 -6 is a very small integer to prevent division by zero. After normalization, the value range of the signal is

[0036] Construct a one-dimensional convolutional neural network (CNN) to classify the standardized signal; the input layer of the CNN receives the standardized one-dimensional signal where N w =16000, the convolutional layer structure is as follows:

[0037] Convolutional layer 1: The width of the convolution kernel is K1=64, the number of channels in this layer is C1=32, the stride is 4, and the zero padding is The activation function is ReLU, and after convolution, maximum pooling is applied with a pooling kernel width of 2 and a stride of 2. The convolution process can be expressed as:

[0038]

[0039] Convolutional layer 2: The width of the convolution kernel is K2=32, the number of channels is C2=64, and the stride and padding method are the same as those of convolutional layer 1.

[0040] Convolutional layer 3: The width of the convolution kernel is K3 = 16, the number of channels is C3 = 128, and the stride and padding method are the same as those of convolutional layer 1;

[0041] Next, after passing through the fully connected layer, a binary classification result is finally output. The activation function uses Softmax, and the output result is classified into two categories, that is, usable or unusable:

[0042]

[0043] Use labeled datasets in the training phase where l k ∈{available, unavailable}, available positive samples are clear heart sound segments, and negative samples are segments containing operational noise or disconnection;

[0044] The loss function uses cross entropy loss:

[0045]

[0046] The Adam algorithm is selected as the optimizer, the learning rate is set to η = 0.001, and β1 = 0.9 and β2 = 0.999 are used as momentum parameters. In the inference stage, if the probability of the "available" category output by Softmax is greater than 0.5, the segment is retained as valid data input into the subsequent classification model; if the output probability is lower than 0.5, the segment is judged to be a low-quality segment and discarded.

[0047] Preferably, in the PCG signal-based mitral regurgitation severity assessment method, a deep neural network based on a Clique block is used to extract and classify effective heart sound segments to obtain classification results of different mitral regurgitation severity levels, including the following steps:

[0048] The Clique block group consists of 4 cascaded Clique blocks. Each Clique block contains C = 4 one-dimensional convolution sub-blocks. The operation of the c-th sub-block of the b-th Clique block is defined as:

[0049]

[0050] where w b,c is the convolution weight with kernel width K = 3, number of channels C = 64, and stride S = 1. Concat(·) means connecting the outputs of the first three adjacent sub-blocks to form span feature reuse;

[0051] The classification head then receives the output of the Clique block in the last layer and projects it into the pathology category space after global average pooling. The classification process is expressed as:

[0052]

[0053] Where i∈{0,1,2,3} corresponds to four categories: no abnormality, mild, moderate, and severe. F is the feature map output by the Clique block, and w is the full connection weight.

[0054] Dataset used in the training phase Available heart sound segments with annotations and the corresponding pathological label l k , namely normal, mild, moderate, and severe. The probability distribution of pathological categories output by the network uses cross entropy to calculate the loss:

[0055]

[0056] in Is the label l k In the indicator symbol of the i-th category, p(·) is the model’s response to the input segment. The predicted probability of .

[0057] Preferably, in the method for assessing the severity of mitral regurgitation based on PCG signals,

[0058] For the K available segments of the same patient, the classification results are integrated. Specifically, let the output probability distribution of the kth segment be P k =[p k0 ,p k1 ,p k2 ,p k3 ], counting the total votes for the four categories:

[0059]

[0060] in is a conditional function that takes 1 when the internal condition is met, otherwise it takes 0;

[0061] c is the category, and the output is the category with the highest number of votes

[0062] If the highest number of votes is the same, choose the category with the largest weighted probability

[0063] Will The final mitral regurgitation degree results are output corresponding to the labels {0, 1, 2, 3}, namely, no abnormality, mild, moderate, and severe.

[0064] The present invention has the following beneficial effects:

[0065] To address the shortcomings of traditional heart sound analysis, such as reliance on manual experience, significant noise interference, and the inability of existing technologies to accurately quantify the severity of MR, the present invention proposes a fully automatic solution by integrating noise elimination, low-quality heart sound segment filtering, and deep neural networks.

[0066] Short heart sound segments effectively reduce the amount of real-time computation required, ensuring efficient operation on mobile devices. Multi-segment information fusion effectively improves classification robustness and reduces the random error associated with single-segment analysis.

[0067] Fully automated and highly accurate severity grading, dual-channel noise cancellation, and available data filtering modules effectively cope with complex environmental interference. Lightweight data analysis can be integrated into a portable electronic stethoscope, making it suitable for large-scale screening and telemedicine.

[0068] It provides an efficient and economical calculation tool for early screening of mitral regurgitation and has significant clinical application value.

[0069] Other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The present invention provides a flowchart of an embodiment of a method for assessing the severity of mitral regurgitation based on PCG signals. DETAILED DESCRIPTION

[0071] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0072] like Figure 1 As shown, the present invention provides a method for assessing the severity of mitral regurgitation based on PCG signals, which comprises the following steps:

[0073] Step 1: noise elimination processing;

[0074] An electronic stethoscope equipped with two channels is used to collect heart sound signals. The first channel collects heart sound signals, and the second channel collects background noise. Adaptive noise cancellation technology is used to suppress non-heart sound components (such as clothing friction and operating noise), retaining the core heart sound frequency band in the range of 20Hz to 1000Hz (including S1, S2 and noise), thereby improving the signal-to-noise ratio.

[0075] Step 2: Signal segmentation and sliding window interception;

[0076] After noise removal, the continuous heart sound signal is segmented into short segments using a sliding window method. Specifically, the window length is set to 2 seconds, the sliding step is set to 1 second, and a sequence of overlapping segments is generated. This design balances signal continuity with computational efficiency.

[0077] Step 3: Usability determination and low-quality segment filtering;

[0078] A usability assessment module based on a convolutional neural network (CNN) assesses the quality of each heart sound segment. After preprocessing the input segments, the model outputs a "usable" or "unusable" label, automatically filtering out low-quality segments caused by improper operation (such as excessive pressure or the stethoscope being removed from the body), retaining only valid data for input into the subsequent classification model.

[0079] Step 4: Feature extraction and mitral regurgitation classification;

[0080] A deep neural network (DNN) based on Clique blocks automatically extracts and classifies available heart sound segments. The network architecture comprises multiple levels of Clique blocks, each of which extracts temporal features through 1D convolution. These features are then reused and concatenated across levels to produce a four-category classification (none, mild, moderate, and severe mitral regurgitation).

[0081] Step 5: Integration of multiple fragment results and final decision;

[0082] The classification results of multiple available heart sound segments from the same patient are combined using a majority voting strategy. The frequency of occurrence of the four categories in all segments is counted, and the category with the highest frequency is used as the final classification result. This method reduces the impact of single segment misclassification and improves overall classification stability.

[0083] Noise cancellation processing

[0084] The specific implementation method is as follows:

[0085] The present invention uses an electronic stethoscope equipped with dual-channel acquisition to collect heart sound signals. The first channel (main channel) collects the target heart sound, and the second channel (reference channel) synchronously collects the environmental background noise. The adaptive noise elimination algorithm is used to suppress the interference of non-heart sound components. Assume that the main channel signal is S main (n), the reference channel signal is s ref (n), where n is the discrete time index. Due to the physical location differences of the sound channels, the signals need to be aligned in the time domain:

[0086] s ref (n) = DelayAlign(s main (n),s ref (n))

[0087] After aligning the signals in the time domain, the background noise is dynamically estimated using the LMS algorithm:

[0088]

[0089] where w k (n) is the adaptive filter coefficient, and M=128 is the filter order.

[0090] The estimated noise component is taken out from the main channel, and the preliminary denoising result is:

[0091]

[0092] Then, a finite impulse response (FIR) bandpass filter is designed to retain the frequency band from 20 Hz to 1000 Hz (covering S1, S2, and the core components of the noise). The filter transfer function is:

[0093]

[0094] Where h(k) is the filter coefficient designed based on the window function method (using Kaiser window, β = 5, order L = 256), and the final output signal is

[0095]

[0096] Output signal Eliminate high-frequency noise (such as clothing friction) and low-frequency interference (such as breathing sounds) to improve the signal-to-noise ratio.

[0097] Signal segmentation and sliding window interception

[0098] The specific implementation method is as follows:

[0099] After denoising The sliding window method is used to split the signal into short time segments to balance the continuity of the signal and the computational efficiency. The parameter of the sliding window is set to the window length T. w =2s, the sampling point should be of length N w =T w ×f s , where f s =8000 is the sampling frequency of the stethoscope. So N w =16000, sliding step size T s =1s, corresponding to N s =8000, for a signal of length L, generate fragments. The kth fragment is represented as:

[0100]

[0101] This design divides the continuous heart sound signal into several 2-second data segments, and adjacent data segments overlap by 50% to avoid truncation of key features.

[0102] Usability judgment and low-quality clip filtering

[0103] The specific implementation method is as follows:

[0104] Construct a lightweight convolutional neural network (CNN) module to evaluate the quality of each heart sound data segment. First, segment x k (n) To standardize:

[0105]

[0106] in is the segment mean, is the standard deviation of the segment, ∈=10 -6 is a very small integer to prevent division by zero. After normalization, the value range of the signal is Next, a lightweight one-dimensional convolutional neural network (CNN) is designed to classify the standardized signal. The input layer of the CNN receives the standardized one-dimensional signal where N w =16000, the convolutional layer structure is as follows:

[0107] Convolutional layer 1: The width of the convolution kernel is K1=64, the number of channels in this layer is C1=32, the stride is 4, and the zero padding is The activation function is ReLU, and after convolution, maximum pooling is applied with a pooling kernel width of 2 and a stride of 2. The convolution process can be expressed as:

[0108]

[0109] Convolutional layer 2: The width of the convolution kernel is K2=32, the number of channels is C2=64, and the stride and padding method are the same as those of convolutional layer 1.

[0110] Convolutional layer 3: The width of the convolution kernel is K3 = 16, the number of channels is C3 = 128, and the stride and padding method are the same as those of convolutional layer 1.

[0111] Next, after passing through the fully connected layer, a binary classification result is finally output. The activation function uses Softmax, and the output result is classified into two categories, that is, usable or unusable:

[0112]

[0113] Use labeled datasets in the training phase where l k∈{available, unavailable}, available positive samples are clear heart sound segments (complete S1, S2 components and heart murmurs), and negative samples are segments containing operational noise (such as stethoscope sliding, clothing friction) or disconnection (such as insufficient contact pressure). The loss function uses cross entropy loss:

[0114]

[0115] The Adam algorithm is selected as the optimizer, the learning rate is set to η = 0.001, and β1 = 0.9 and β2 = 0.999 are used as momentum parameters. In the inference stage, if the probability of the "available" category output by Softmax is greater than 0.5, the segment is retained as valid data input into the subsequent classification model; if the output probability is lower than 0.5, the segment is judged to be a low-quality segment and discarded.

[0116] Feature extraction and mitral regurgitation classification

[0117] The specific implementation method is as follows:

[0118] The available fragments x extracted by the CNN module k (n), a deep neural network (DNN) based on Clique blocks is used to achieve end-to-end classification. The input layer of the network model receives the standardized available fragments x k (n), and then the signal is input into the Clique block group for feature extraction. The Clique block group consists of 4 cascaded Clique blocks, each of which contains C = 4 one-dimensional convolution sub-blocks. The c-th sub-block operation of the b-th Clique block is defined as:

[0119]

[0120] where w b,c is the convolution weight with kernel width K = 3, number of channels C = 64, and stride S = 1. Concat(·) represents the concatenation of the outputs of the first three adjacent sub-blocks to form span feature reuse. Through this structure, the network can effectively reuse the features extracted by multiple convolution sub-blocks and enhance the information expression capability.

[0121] The classification head then receives the output of the Clique block in the last layer, performs global average pooling, and projects it into the pathology category space. The classification process is expressed as:

[0122]

[0123] Among them, i∈{0,1,2,3} corresponds to the four categories of no abnormality, mild, moderate, and severe, F is the feature map output by the Clique block, and w is the full connection weight.

[0124] Dataset used in the training phase Available heart sound segments with annotations and the corresponding pathological label l k , namely normal, mild, moderate, and severe. The probability distribution of pathological categories output by the network uses cross entropy to calculate the loss:

[0125]

[0126] in Is the label l k In the indicator symbol of the i-th category, p(·) is the model’s response to the input segment. The optimization process improves the accuracy of the model by minimizing this loss function.

[0127] Multi-segment result integration and final decision

[0128] The specific implementation method is as follows:

[0129] For the K available segments of the same patient, the classification results are integrated. Specifically, let the output probability distribution of the kth segment be P k =[p k0 ,p k1 ,p k2 ,p k3 ], counting the total votes for the four categories:

[0130]

[0131] in is a conditional function that takes 1 when the internal condition is met, otherwise it takes 0. c is the category, and the output is the category with the highest number of votes If the highest number of votes is the same, choose the category with the largest weighted probability

[0132] Will The final mitral regurgitation degree results are output corresponding to the labels {0, 1, 2, 3}, namely, no abnormality, mild, moderate, and severe.

[0133] The present invention provides a method for assessing the severity of mitral regurgitation based on PCG signals. First, dual-channel audio acquisition technology is used to synchronously acquire signals from the heart sound region and the ambient background. An adaptive noise cancellation algorithm is used to eliminate environmental noise and operational interference (such as clothing friction and uneven stethoscope pressure). A dynamic time window segmentation strategy is then applied to the de-noised heart sounds, segmenting the continuous heart sound signal with a 2-second window length and a 1-second sliding step size to generate short-duration heart sound segments. A lightweight convolutional neural network model is used to evaluate the quality of these short-duration heart sound segments in real time, filtering out low-quality segments. Subsequently, a deep neural network model based on clique blocks is used to directly extract nonlinear features from the short-duration heart sound segments. Each valid segment is classified into four categories (no MR, mild, moderate, and severe). A majority voting mechanism is then used to integrate the classification results of all segments to generate a final MR severity assessment.

[0134] The present invention aims to improve the accuracy of heart sound signal analysis and the automation level of heart disease diagnosis. This technical solution constructs an end-to-end heart sound pathology analysis system through dual-channel collaborative noise reduction, dynamic quality screening and deep neural network classification. Specifically, it includes: using a dual-channel electronic stethoscope to synchronously collect the main heart sound signal and environmental noise, and realizing two-stage noise elimination through adaptive noise cancellation algorithm and Kaiser window bandpass filtering, effectively suppressing interference signals outside the 20Hz-1000Hz frequency band; using a sliding window strategy to divide the continuous signal into overlapping short-time segments, combining a lightweight convolutional neural network to build a quality discrimination module, and realizing automatic filtering of low-quality segments through standardization processing and a three-layer convolution structure; innovatively designing a deep neural network model based on Clique block cascade, extracting the nonlinear features of heart sounds through a cross-level feature reuse mechanism, and combining global average pooling to complete the four-class probability prediction; finally, using a majority voting strategy to integrate the results of multiple segments and output the mitral regurgitation grading conclusion.

[0135] Compared with traditional heart sound analysis methods, the present invention realizes the full-process automation of noise elimination, quality judgment and pathological grading. The signal-to-noise ratio of the heart sound signal is effectively improved through dual-channel adaptive noise reduction technology. The filtered short-term high-quality heart sound segments are combined with the feature capture capability of the deep learning model to ensure computational efficiency while making the accuracy of mitral regurgitation grading reach a clinical practical level. This solution effectively solves the pain points of traditional methods such as high dependence on manual experience, weak anti-interference ability, and difficulty in extracting long-term heart sound features. It is particularly suitable for medical institutions to carry out large-scale heart disease screening and provide a highly reliable auxiliary decision-making basis for clinical diagnosis.

[0136] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A method for assessing the severity of mitral regurgitation based on PCG signals, characterized in that: The following steps are involved: Obtaining PCG signals of the patient to be evaluated and performing noise elimination processing on them; The PCG signal after noise elimination is segmented into short-term heart sound segments using the sliding window method; Based on the convolutional neural network, the availability of short-term heart sound segments is judged to obtain valid heart sound segments; A deep neural network based on Clique blocks was used to extract features and classify valid heart sound segments to obtain classification results of different severity levels of mitral regurgitation. The majority voting strategy was used to integrate the statistical frequencies of the different severity classification results of mitral regurgitation, and the highest frequency category was taken as the evaluation result.

2. The method for assessing the severity of mitral regurgitation based on PCG signals according to claim 1, wherein: Obtaining the PCG signal of the patient to be evaluated and performing noise removal processing on it includes the following steps: An electronic stethoscope equipped with dual-channel acquisition is used to collect heart sound signals. The main channel collects target heart sounds, and the reference channel collects environmental background noise. The ear clipping adaptive noise cancellation algorithm is used to suppress the interference of non-heart sound components. Assume that the main channel signal is s main (n), the reference channel signal is s ref (n), where n is the discrete time index; Since the channels have different physical positions, the signals need to be aligned in the time domain: s ref (n)=DelayAlign(s main (n),s ref (n)) After aligning the signals in the time domain, the background noise is dynamically estimated using the LMS algorithm: where w k (n) is the adaptive filter coefficient, M = 128 is the filter order; The estimated noise component is taken out from the main channel, and the preliminary denoising result is: Then, a finite impulse response (FIR) bandpass filter is designed to retain the frequency band from 20 Hz to 1000 Hz. The filter transfer function is: Where h(k) is the filter coefficient designed based on the window function method (using Kaiser window, β = 5, order L = 256), and the final output signal is Output signal Eliminate high-frequency noise and low-frequency interference.

3. The method for assessing the severity of mitral regurgitation based on PCG signals according to claim 2, wherein: The PCG signal after noise elimination is segmented into short-duration heart sound segments using a sliding window method, including the following steps: After denoising The sliding window method is used to segment the heart sound into short segments to balance the continuity of the signal and the computational efficiency. The parameter of the sliding window is set to the window length T. w =2s, the sampling point should be of length N w =T w ×f s , where f s =8000 is the sampling frequency of the stethoscope; Therefore, N w =16000, sliding step size T s =1s, corresponding to N s =8000, for a signal of length L, generate fragments. The kth fragment is represented as: This design divides the continuous heart sound signal into several short heart sound segments of 2 seconds in length, and adjacent short heart sound segments overlap by 50% to avoid truncation of key features.

4. The method for assessing the severity of mitral regurgitation based on PCG signals according to claim 3, wherein: The availability of short-term segments is judged based on a convolutional neural network to obtain valid heart sound segments, including the following steps: Short heart sound segments x k (n) To standardize: in is the segment mean, is the standard deviation of the segment, ∈=10 -6 is a very small integer to prevent division by zero. After normalization, the value range of the signal is Construct a one-dimensional convolutional neural network (CNN) to classify the standardized signal; the input layer of the CNN receives the standardized one-dimensional signal where N w =16000, the convolutional layer structure is as follows: Convolutional layer 1: The width of the convolution kernel is K1=64, the number of channels in this layer is C1=32, the stride is 4, and the zero padding is The activation function is ReLU. After convolution, maximum pooling is applied with a pooling kernel width of 2 and a stride of 2. The convolution process can be expressed as: Convolutional layer 2: The width of the convolution kernel is K2=32, the number of channels is C2=64, and the step size and padding method are the same as those of convolutional layer 1. Convolutional layer 3: The width of the convolution kernel is K3 = 16, the number of channels is C3 = 128, and the stride and padding method are the same as those of convolutional layer 1; Next, after passing through the fully connected layer, a binary classification result is finally output. The activation function uses Softmax, and the output result is classified into two categories, that is, usable or unusable: Use labeled datasets in the training phase where l k ∈{available, unavailable}, available positive samples are clear heart sound segments, and negative samples are segments containing operational noise or disconnection; The loss function uses cross entropy loss: The Adam algorithm was selected as the optimizer, with the learning rate set to η = 0.001, and β1 = 0.9 and β2 = 0.999 as momentum parameters. During the inference phase, if the probability of the "usable" category output by Softmax is greater than 0.5, the segment is retained as valid data and input into the subsequent classification model; if the output probability is less than 0.5, the segment is judged to be low-quality and discarded.

5. The method for assessing the severity of mitral regurgitation based on PCG signals according to claim 2, wherein: A deep neural network based on Clique blocks is used to extract and classify the features of valid heart sound segments to obtain classification results of different severity levels of mitral regurgitation, including the following steps: The Clique block group consists of 4 cascaded Clique blocks. Each Clique block contains C = 4 one-dimensional convolution sub-blocks. The operation of the c-th sub-block of the b-th Clique block is defined as: where w b,c is the convolution weight with kernel width K = 3, number of channels C = 64, and stride S = 1. Concat(·) represents the concatenation of the outputs of the first three adjacent sub-blocks to form span feature reuse. The classification head then receives the output of the Clique block in the last layer and projects it into the pathology category space after global average pooling. The classification process is expressed as: Where i∈{0,1,2,3} corresponds to four categories: no abnormality, mild, moderate, and severe. F is the feature map output by the Clique block, and w is the full connection weight. Dataset used in the training phase Available heart sound segments with annotations and the corresponding pathological label l k , that is, no abnormality, mild, moderate, severe, the probability distribution of pathological categories output by the network uses cross entropy to calculate the loss: in Is the label l k In the indicator symbol of the i-th category, p(·) is the model’s response to the input segment. The predicted probability of .

6. The method for assessing the severity of mitral regurgitation based on PCG signals according to claim 5, wherein: For the K available segments of the same patient, the classification results are integrated. Specifically, let the output probability distribution of the kth segment be P k =[px0,p k1 ,p k2 ,p k3 ], counting the total votes for the four categories: in is a conditional function that takes 1 when the internal condition is met, otherwise it takes 0; c is the category, and the output is the category with the highest number of votes If the highest number of votes is the same, choose the category with the largest weighted probability Will The final mitral regurgitation degree results are output corresponding to the labels {0, 1, 2, 3}, namely, no abnormality, mild, moderate, and severe.

Images