Heart sound feature recognition method based on wavelet spectral analysis

Abstract

The invention relates to the technical field of heart sound signal processing, and discloses a heart sound feature recognition method based on wavelet spectral analysis. The method comprises the following steps: collecting heart sound signal data and carrying out signal filtering and segmentation processing to obtain segmented cardiac cycle signals; performing n-level discrete wavelet transform decomposition to obtain an approximation coefficient and n detail coefficients; performing multi-dimensional statistical feature calculation to obtain a multi-dimensional feature vector; performing principal component analysis dimension reduction processing to obtain a dimension reduction feature vector; fitting feature distribution by using a Gaussian mixture model and introducing a dynamic adjustment factor to obtain an adaptive classification threshold; the hidden Markov model is constructed through the dimension reduction feature vectors, the likelihood probability is calculated through a forward algorithm, the likelihood probability is compared with a self-adaptive classification threshold value, and the heart sound feature recognition result is output. The time-frequency features of the heart sound signals are fully utilized, the feature distinction degree and the recognition accuracy are improved, and the recognition accuracy is improved. And meanwhile, the method has good robustness and generalization ability.

Description

Technical Field

[0001] The present application relates to the technical field of heart sound signal processing, and in particular to a heart sound feature recognition method based on wavelet spectrum analysis. Background Art

[0002] Heart sounds are important physiological signals for evaluating cardiac function and diagnosing heart diseases. With the advancement of medical technology, the demand for automated heart sound analysis and recognition systems in clinical applications is increasing. Traditional heart sound analysis methods mainly rely on feature extraction in the time domain and frequency domain, but these methods have limitations when processing non-stationary signals. Heart sound signals are non-stationary and nonlinear, and contain complex time-frequency features, which poses challenges to accurate recognition and classification.

[0003] In addition, heart sound signals are easily disturbed by environmental noise, breathing sounds, body movements and other factors, which leads to a decrease in signal quality and affects the accuracy of feature extraction. Existing heart sound feature recognition methods still have deficiencies in noise suppression and signal segmentation, and it is difficult to effectively process heart sound signals collected in complex clinical environments. At the same time, traditional feature extraction and classification algorithms often use fixed parameters and thresholds, lack adaptive capabilities, and are difficult to adapt to changes in heart sound characteristics in different individuals and different physiological states. Summary of the invention

[0004] The present application provides a heart sound feature recognition method based on wavelet spectrum analysis, which is used to make full use of the time-frequency characteristics of heart sound signals, improve the feature discrimination and recognition accuracy, and has good robustness and generalization ability.

[0005] In a first aspect, the present application provides a method for identifying heart sound features based on wavelet spectrum analysis, the method for identifying heart sound features based on wavelet spectrum analysis comprising:

[0006] Collecting heart sound signal data, and performing signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals;

[0007] Performing n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients;

[0008] Performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector;

[0009] Performing principal component analysis and dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector;

[0010] Based on the reduced dimension feature vector, a Gaussian mixture model is used to fit the feature distribution and a dynamic adjustment factor is introduced to obtain an adaptive classification threshold;

[0011] A hidden Markov model is constructed by using the dimension-reduced feature vector, and a forward algorithm is used to calculate the likelihood probability, the likelihood probability is compared with the adaptive classification threshold, and a heart sound feature recognition result is output.

[0012] In a second aspect, the present application provides a heart sound feature recognition device based on wavelet spectrum analysis, and the heart sound feature recognition device based on wavelet spectrum analysis includes:

[0013] An acquisition module is used to acquire heart sound signal data, and perform signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals;

[0014] A decomposition module, used for performing n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients;

[0015] A calculation module, used for performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector;

[0016] A processing module, used for performing principal component analysis and dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector;

[0017] A fitting module, used to fit the feature distribution using a Gaussian mixture model based on the reduced-dimensional feature vector and introduce a dynamic adjustment factor to obtain an adaptive classification threshold;

[0018] The output module is used to construct a hidden Markov model through the dimension reduction feature vector, calculate the likelihood probability using a forward algorithm, compare the likelihood probability with the adaptive classification threshold, and output the heart sound feature recognition result.

[0019] The third aspect of the present application provides a heart sound feature recognition device based on wavelet spectrum analysis, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the heart sound feature recognition device based on wavelet spectrum analysis executes the above-mentioned heart sound feature recognition method based on wavelet spectrum analysis.

[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned heart sound feature recognition method based on wavelet spectrum analysis.

[0021] In the technical solution provided by the present application, a method combining adaptive minimum mean square error noise suppression and Hilbert transform is used for signal preprocessing, which effectively improves the quality and signal-to-noise ratio of heart sound signals. Discrete wavelet transform and multi-dimensional statistical feature calculation are introduced to make full use of the time-frequency characteristics of heart sound signals and improve the distinguishability and characterization ability of features. Through principal component analysis and dimensionality reduction processing, the redundancy of feature vectors is reduced, the calculation efficiency is improved, and key feature information is retained. The Gaussian mixture model is used to fit the feature distribution and the dynamic adjustment factor is introduced to realize the dynamic adjustment of the adaptive classification threshold, and the adaptability of the method to different individuals and physiological states is enhanced. Combined with the hidden Markov model and the forward algorithm, the temporal characteristics of the heart sound signal are fully utilized to improve the accuracy and reliability of feature recognition. By optimizing the selection of wavelet basis functions and the number of decomposition layers, the efficiency of wavelet transform and the pertinence of feature extraction are improved. The energy analysis and information entropy criteria are introduced to realize the adaptive optimization of the number of wavelet decomposition layers and improve the effectiveness of feature extraction. The method combining multiple statistical features and scale-related features is adopted to comprehensively characterize the statistical characteristics and multi-scale features of heart sound signals and enhance the expression ability of features. Through the Bayesian information criterion evaluation and expectation maximization algorithm, the parameter estimation of the Gaussian mixture model is optimized, and the accuracy of feature distribution fitting is improved. The dynamic adjustment factor and threshold adjustment formula are introduced to realize the adaptive update of the classification threshold, and improve the generalization ability of the method to different data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 This is a schematic diagram of an embodiment of a method for identifying heart sound features based on wavelet spectrum analysis in an embodiment of the present application;

[0024] Figure 2 Schematic diagram of an embodiment of a heart sound feature recognition device based on wavelet spectrum analysis in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiment of the present application provides a method for identifying cardiac sound features based on wavelet spectrum analysis. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the heart sound feature recognition method based on wavelet spectrum analysis in the embodiment of the present application includes:

[0027] Step S101, collecting heart sound signal data, and performing signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals;

[0028] It is understandable that the execution subject of the present application may be a heart sound feature recognition device based on wavelet spectrum analysis, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0029] Specifically, the heart sound signal data is collected, and the heart sound signal data is band-pass filtered to remove the low-frequency and high-frequency noise in the signal to obtain the heart sound signal after initial filtering. Band-pass filtering usually selects a specific frequency range to ensure that the effective frequency components of the heart sound signal are retained. The heart sound signal after initial filtering is subjected to adaptive minimum mean square error noise suppression processing to reduce the random noise in the signal and obtain the heart sound signal after noise reduction. The heart sound signal after noise reduction is subjected to Hilbert transform to obtain the envelope of the signal. The envelope is a curve that reflects the change of signal amplitude over time. By analyzing the envelope, the peak point in the heart sound signal is detected and the demarcation point of the heart sound cycle is determined. The heart sound signal after noise reduction is segmented according to the demarcation point to obtain the heart sound signal of the initial segmentation. The segmented signal is subjected to short-time energy analysis to calculate the energy value of each segment. Short-time energy analysis can provide information about the change of signal strength over time. Based on the energy value of each segment, the initial segmented heart sound signal is identified by S1 and S2 sounds. These two sounds mark different stages in the heart cycle and correspond to the sounds emitted when the heart valve closes. By identifying these specific sounds, the first segmented heart sound signal is obtained. The first segmented heart sound signal is subjected to zero-phase filtering. Zero-phase filtering is a filtering method that does not introduce phase delay and can maintain the original form of the signal to obtain the second segmented heart sound signal. The second segmented heart sound signal is resampled using the cubic spline interpolation algorithm to unify the sampling rate of the signal and obtain a segmented cardiac cycle signal.

[0030] Step S102, performing n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients;

[0031] Specifically, the wavelet basis function is selected for the segmented cardiac cycle signal. The wavelet basis function that is most suitable for processing the heart sound signal is selected from a variety of wavelet basis functions. Generally, the selection criteria of the wavelet basis function include evaluation indicators such as orthogonality and compact support, which help to ensure the accuracy and efficiency of the wavelet transform. In this process, the optimal Daubechies wavelet basis function is selected by comparing the performance of different wavelet basis functions in processing heart sound signals. According to the optimal Daubechies wavelet basis function, a wavelet filter group is constructed. The wavelet filter group includes a low-pass decomposition filter h and a high-pass decomposition filter g, which are used to extract low-frequency and high-frequency components in the signal, respectively. The segmented cardiac cycle signal is convolved with the low-pass decomposition filter h, and downsampled by 2 times to obtain the first-level approximate coefficient A1, which contains the low-frequency components in the original signal and represents the general shape of the signal. At the same time, the segmented cardiac cycle signal is convolved with the high-pass decomposition filter g, and downsampled by 2 times to obtain the first-level detail coefficient D1, which contains the high-frequency components in the signal and provides more detailed signal change information. Use the first-level approximation coefficient A1 as the new input signal and perform wavelet transform again. Repeat the previous steps, perform low-pass filtering and high-pass filtering on A1, and obtain the second-level approximation coefficient A2 and detail coefficient D2 respectively. This process can be repeated continuously, and the approximation coefficient of the previous layer is newly decomposed each time until the n-th level approximation coefficient An and detail coefficient Dn are obtained. Energy analysis is performed on the obtained n-level approximation coefficients and detail coefficients. Energy analysis is to calculate the energy proportion of each level of coefficients, that is, the proportion of each level of coefficients in the total energy of the signal. Through energy analysis, the energy distribution characteristics of the signal at different scales are obtained, which can reflect the structure and characteristics of the signal. According to the energy distribution characteristics, the information entropy criterion is used to determine the optimal decomposition level nopt. The information entropy criterion is a method commonly used to determine the complexity of the signal. By quantifying the degree of confusion of the information, it helps to select the most appropriate number of decomposition levels, thereby avoiding too many or too few decomposition levels from capturing the signal characteristics inaccurately. Based on the optimized wavelet decomposition layer number nopt, the segmented cardiac cycle signal is re-discretely transformed, and the signal is finally decomposed using the optimized decomposition layer number to obtain an approximate coefficient and n detail coefficients. These coefficients are the performance of the signal at different frequency components.

[0032] Step S103, performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector;

[0033] Specifically, the mean of the approximate coefficient and the n detail coefficients is calculated respectively, and n+1 mean features are obtained. The mean reflects the overall level of the signal on the frequency component. Based on the mean feature, the approximate coefficient and the n detail coefficients are centered, that is, each coefficient is subtracted from the corresponding mean to obtain a centered coefficient sequence. The variance is calculated for the centered coefficient sequence. The variance is an indicator of the distance of each data point in the signal from the mean, reflecting the discreteness of the signal. n+1 variance features are calculated. The third-order central moment is calculated based on the centered coefficient sequence to obtain n+1 skewness features. The skewness feature is used to describe the symmetry or skewness of the signal and can reveal the asymmetry of the signal distribution. Similarly, the fourth-order central moment is calculated based on the centered coefficient sequence to obtain n+1 kurtosis features. The kurtosis feature measures the peak degree or flatness of the signal and provides further information about the distribution morphology of the signal. The energy is calculated for the approximate coefficient and the n detail coefficients respectively to obtain n+1 energy features. The energy feature represents the energy of the signal at each wavelet scale and is an important measure of signal strength. According to the energy feature, the relative energy proportion of each coefficient is calculated to obtain n+1 energy ratio features, which can reflect the contribution of each scale to the total energy. The Shannon entropy of the approximate coefficient and n detail coefficients is calculated to obtain n+1 entropy value features. The entropy value feature measures the complexity and information content of the signal and is a quantitative description of the uncertainty of the signal. Based on the approximate coefficient and n detail coefficients, the mutual correlation coefficient between adjacent scale coefficients is calculated to obtain n scale-related features. These features reveal the similarity and correlation between different scales, which helps to understand the correlation structure of the signal in different frequency ranges. The mean feature, variance feature, skewness feature, kurtosis feature, energy feature, energy ratio feature, entropy feature and scale-related feature are combined in sequence to obtain a complete multidimensional feature vector.

[0034] Step S104, performing principal component analysis and dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector;

[0035] Specifically, the multidimensional eigenvectors are standardized, and the mean of each feature is adjusted to 0 and the variance is adjusted to 1 to eliminate the influence of different feature dimensions on the results of principal component analysis. The standardized data forms a standardized feature matrix, in which the values ​​of each feature are on a unified scale, which is convenient for subsequent calculations. The feature covariance matrix is ​​calculated based on the standardized feature matrix. The feature covariance matrix is ​​a symmetric matrix in which each element represents the covariance between two features. The calculation of this matrix is ​​to understand the linear relationship between the features. The larger the covariance, the stronger the correlation between the two features. The feature covariance matrix is ​​subjected to eigenvalue decomposition, and the matrix is ​​decomposed into eigenvalues ​​and corresponding eigenvectors. The eigenvalue represents the variance of the data in the direction of the corresponding eigenvector, and the eigenvector represents the unit vector in that direction. According to the eigenvalue, the eigenvectors are arranged in descending order to form a sorted eigenvector matrix. In the sorted eigenvector matrix, the front eigenvectors correspond to larger eigenvalues, which account for the main variance of the data. The cumulative contribution rate of the sorted eigenvectors is calculated, and the cumulative contribution rate represents the proportion of the total variance of the first few principal components to the total data variance. By calculating the cumulative contribution rate, the amount of data information contained in the first few principal components is determined. In order to determine the appropriate number of eigenvectors k after dimensionality reduction, the first k principal components with a cumulative contribution rate of 95% are selected, and these principal components retain 95% of the original data information. According to the determined k value, the first k columns are selected from the sorted eigenvector matrix to form a dimensionality reduction transformation matrix. The dimensionality reduction transformation matrix is ​​used to project the original data into a low-dimensional space that retains most of the data features. The standardized feature matrix is ​​multiplied by the dimensionality reduction transformation matrix to obtain the initial dimensionality reduction eigenvector. The initial dimensionality reduction eigenvector is the representation of the original eigenvector in the dimensionality reduction space. In order to further optimize the feature representation after dimensionality reduction, the variance of each dimension of the eigenvector after dimensionality reduction is calculated to obtain the importance weight of each feature. The importance weight of the feature reflects the contribution of each feature in the data after dimensionality reduction. The initial dimension reduction feature vector is weighted with the feature importance weight, and the influence of each feature in the final dimension reduction result is adjusted to obtain the dimension reduction feature vector.

[0036] Step S105: Based on the dimension-reduced feature vector, a Gaussian mixture model is used to fit the feature distribution and a dynamic adjustment factor is introduced to obtain an adaptive classification threshold;

[0037] Specifically, cluster analysis is performed on the reduced-dimensional feature vectors to preliminarily identify the potential grouping of the feature vectors and determine the initial clustering center. Based on the initial clustering center, the expectation maximization algorithm is used to estimate the parameters of the Gaussian mixture model to obtain the initial Gaussian mixture model. The model contains several Gaussian components, each of which corresponds to a potential feature distribution. The initial Gaussian mixture model is evaluated by the Bayesian information criterion to determine the optimal number of Gaussian components M. The Bayesian information criterion helps to select the most suitable data description method by balancing the model complexity and data fit. Based on the optimal number of Gaussian components M, the Gaussian mixture model is retrained to obtain the optimized Gaussian mixture model. According to the optimized Gaussian mixture model, the probability density functions of normal heart sounds and abnormal heart sounds are calculated to obtain the probability distribution in the feature space. The probability distribution provides a detailed description of the possibility of each type of heart sound. The probability distribution in the feature space is analyzed by equal probability curves to obtain the initial classification threshold. The initial classification threshold is used to preliminarily divide the boundary between normal and abnormal heart sounds. In order to adapt to the dynamic changes and individual differences of heart sound characteristics, a dynamic adjustment factor α is introduced to adjust the classification threshold. According to the initial classification threshold and the dynamic adjustment factor α, a threshold adjustment formula is constructed. This formula can flexibly adjust the classification boundary to adapt to the changes in real-time data. The classification results of the most recent N samples are statistically analyzed to calculate the ratio P of normal heart sounds to abnormal heart sounds. This ratio reflects the distribution of heart sounds of different categories in the current data set. According to the obtained ratio P of normal heart sounds to abnormal heart sounds, the dynamic adjustment factor α is updated to obtain a new dynamic adjustment factor α'. The factor adjustment process is based on real-time data feedback to ensure that the classification threshold can adapt to the dynamic changes in data distribution. Substitute the new dynamic adjustment factor α' into the threshold adjustment formula to calculate the final adaptive classification threshold. This threshold can effectively distinguish between normal and abnormal heart sounds and improve the accuracy and reliability of heart sound feature recognition.

[0038] Step S106: construct a hidden Markov model by reducing the dimension of the feature vector, and use the forward algorithm to calculate the likelihood probability, compare the likelihood probability with the adaptive classification threshold, and output the heart sound feature recognition result.

[0039] Specifically, the reduced-dimensional feature vectors are sequence analyzed to determine the number of states N of the hidden Markov model and obtain the model state set S. The state set S represents possible hidden states, each of which corresponds to a characteristic state of the heart sound signal. According to the model state set S, the reduced-dimensional feature vectors are clustered to determine the initial state distribution π, and the probability distribution of each state at the beginning is estimated, that is, the initial probability of the system being in a certain state at any time. At the same time, according to the state set S, the state transition probability matrix A is calculated, which describes the probability of transitioning from one state to another. These probabilities reflect the transition law of the heart sound signal between different states and are an important part of the hidden Markov model. Gaussian distribution fitting is performed on the observed values ​​of each state to obtain the observation probability distribution B. The observation probability distribution B describes the probability of observing a certain feature in a given state, which is usually represented by a Gaussian distribution. These probability distributions are combined with the initial state distribution π and the state transition probability matrix A to construct a complete hidden Markov model λ. The model λ can describe the hidden state sequence of the heart sound signal and the statistical characteristics of its observations. The Baum-Welch algorithm is used to optimize the parameters of the complete hidden Markov model λ. The Baum-Welch algorithm is an iterative algorithm that adjusts the model parameters by maximizing the likelihood estimate so that the model fits the training data better. After multiple iterations, the optimized hidden Markov model λ' is obtained, and its parameters more accurately describe the dynamic characteristics of the heart sound signal. After the model optimization is completed, the reduced dimension feature sequence O to be identified is generated according to the reduced dimension feature vector. The sequence O represents the characteristic change pattern of the heart sound signal to be identified. The likelihood probability P(O|λ') of the reduced dimension feature sequence O is calculated using the optimized hidden Markov model λ' and the forward algorithm. The forward algorithm is a dynamic programming algorithm used to calculate the probability of the observation sequence, which can effectively handle the multi-state and complex state transitions in the hidden Markov model. The calculated likelihood probability P(O|λ') is compared with the adaptive classification threshold. If the likelihood probability is higher than the threshold, it is judged as a normal heart sound, otherwise it is an abnormal heart sound. Through this process, the heart sound feature recognition result is output to achieve accurate classification and recognition of the heart sound signal.

[0040] In the embodiment of the present application, the signal preprocessing is performed by combining the adaptive minimum mean square error noise suppression and Hilbert transform, which effectively improves the quality and signal-to-noise ratio of the heart sound signal. The discrete wavelet transform and multi-dimensional statistical feature calculation are introduced to make full use of the time-frequency characteristics of the heart sound signal and improve the distinguishability and characterization ability of the feature. The principal component analysis is used for dimensionality reduction processing to reduce the redundancy of the feature vector, improve the calculation efficiency, and retain the key feature information. The Gaussian mixture model is used to fit the feature distribution and the dynamic adjustment factor is introduced to achieve the dynamic adjustment of the adaptive classification threshold, and enhance the adaptability of the method to different individuals and physiological states. Combined with the hidden Markov model and the forward algorithm, the temporal characteristics of the heart sound signal are fully utilized to improve the accuracy and reliability of feature recognition. By optimizing the selection of wavelet basis functions and the number of decomposition layers, the efficiency of wavelet transform and the pertinence of feature extraction are improved. The energy analysis and information entropy criteria are introduced to achieve the adaptive optimization of the number of wavelet decomposition layers and improve the effectiveness of feature extraction. The method of combining multiple statistical features and scale-related features is used to comprehensively characterize the statistical characteristics and multi-scale features of the heart sound signal and enhance the expression ability of the features. Through the Bayesian information criterion evaluation and expectation maximization algorithm, the parameter estimation of the Gaussian mixture model is optimized, and the accuracy of feature distribution fitting is improved. The dynamic adjustment factor and threshold adjustment formula are introduced to realize the adaptive update of the classification threshold, and improve the generalization ability of the method to different data sets.

[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0042] (1) collecting heart sound signal data, and performing bandpass filtering on the heart sound signal data to obtain a heart sound signal after initial filtering;

[0043] (2) performing adaptive minimum mean square error noise suppression processing on the heart sound signal after the initial filtering to obtain a denoised heart sound signal;

[0044] (3) performing Hilbert transform on the denoised heart sound signal to obtain a signal envelope, and detecting the peak point of the denoised heart sound signal based on the signal envelope to obtain the demarcation point of the heart sound cycle;

[0045] (4) Based on the demarcation point of the heart sound cycle, the heart sound signal after noise reduction is segmented to obtain the heart sound signal of the initial segment, and the short-time energy analysis of the heart sound signal of the initial segment is performed to obtain the energy value of each segment;

[0046] (5) According to the energy value of each segment, the S1 sound and the S2 sound are identified on the initial segmented heart sound signal to obtain a first segmented heart sound signal;

[0047] (6) Performing zero-phase filtering on the first segmented heart sound signal to obtain a second segmented heart sound signal, and resampling the second segmented heart sound signal using a cubic spline interpolation algorithm to obtain a segmented cardiac cycle signal.

[0048] Specifically, heart sound signal data is collected. Heart sound signals are usually obtained through devices such as electronic stethoscopes, which can convert the mechanical vibration of the heart into electrical signals. The heart sound signal data is bandpass filtered to eliminate irrelevant frequency components. The design of the bandpass filter usually selects appropriate passbands and stopbands to retain the frequency range useful for heart sound analysis (such as 20Hz to 400Hz) while filtering out low-frequency and high-frequency noise. The formula for bandpass filtering can be expressed as:

[0049]

[0050] Among them, H(f) represents the frequency response of the filter, f is the signal frequency, and f1 and f2 are the lower and upper limits of the bandpass filter. After bandpass filtering, the heart sound signal after initial filtering is obtained. The adaptive minimum mean square error noise suppression process is performed on the heart sound signal after initial filtering. The adaptive minimum mean square error algorithm improves the signal-to-noise ratio of the signal by minimizing the mean square error of the noise. Its basic principle is to adjust the gain of the signal to suppress the noise. The formula of the adaptive minimum mean square error is:

[0051]

[0052] in, is the signal after denoising, Y(t) is the observed noisy signal, and G(t) is the gain function, whose value depends on the statistical characteristics of the signal and the noise. The denoised heart sound signal is subjected to Hilbert transform. The Hilbert transform is a method of converting a real signal into an analytical signal. The envelope of the signal can be obtained through its transformation. The envelope represents the instantaneous amplitude change of the signal and is used to detect the peak points in the heart sound signal. By detecting these peak points, the dividing points of the heart sound cycle are determined, and the starting and ending points of different heart sound cycles are marked. Based on the dividing points of the heart sound cycle, the denoised heart sound signal is segmented so that each heart sound cycle can be analyzed separately. After segmentation, short-time energy analysis is performed on each segment of the heart sound signal. Short-time energy analysis is used to calculate the energy of the signal in a short time window, and its formula is:

[0053]

[0054] Among them, E(n) is the energy value of the nth segment, x(m+n) is the sample value of the signal, and M is the length of the analysis window. By calculating the energy value of each segment, the energy characteristics of each heart sound cycle are obtained. According to the energy value of each segment, the S1 sound and S2 sound of the initial segmented heart sound signal are identified. S1 sound and S2 sound are the main heart sounds produced during cardiac contraction and diastole, and their identification depends on the timing characteristics of the energy peak. After identifying S1 sound and S2 sound, the first segmented heart sound signal is obtained. The first segmented heart sound signal is zero-phase filtered using a zero-phase filter. The characteristic of zero-phase filtering is that it does not introduce phase distortion and can maintain the original waveform characteristics of the signal. After zero-phase filtering, the second segmented heart sound signal is obtained. In order to ensure the sampling rate consistency of the signal in further analysis, the cubic spline interpolation algorithm is used to resample the second segmented heart sound signal. Cubic spline interpolation smoothes the signal data by fitting a cubic polynomial, and the formula is:

[0055] S(x)=a+b(x i )+c(xx i ) 2 +d(xx i ) 3 ;

[0056] Among them, S(x) is the signal value after interpolation, a, b, c, d are cubic spline coefficients, x i is the location of the data point. After the above processing steps, the segmented cardiac cycle signal is finally obtained. These signals are used for further feature extraction and analysis to achieve recognition and classification of heart sound features. For example, the health status of the heart can be diagnosed by detecting the characteristic frequency and time characteristics of different heart sound components.

[0057] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0058] S1: Select the wavelet basis function for the segmented cardiac cycle signal, and use the orthogonality and compact support evaluation indicators to obtain the optimal Daubechies wavelet basis function;

[0059] S2: According to the optimal Daubechies wavelet basis function, a wavelet filter bank is constructed to obtain a low-pass decomposition filter h and a high-pass decomposition filter g;

[0060] S3: performing convolution operation on the segmented cardiac cycle signal and the low-pass decomposition filter h, and performing 2-fold downsampling to obtain the first-order approximation coefficient A1;

[0061] S4: performing convolution operation on the segmented cardiac cycle signal and the high-pass decomposition filter g, and performing 2-fold downsampling to obtain the first-level detail coefficient D1;

[0062] S5: Repeat the operations of step S3 and step S4 according to the first-level approximation coefficient A1 to obtain the second-level approximation coefficient A2 and the detail coefficient D2;

[0063] S6: cyclically execute the operation of step S5 until the n-th level approximation coefficient An and detail coefficient Dn are obtained;

[0064] S7: Perform energy analysis on the obtained n-level approximate coefficients and detail coefficients, calculate the energy proportion of each level of coefficients, and obtain energy distribution characteristics;

[0065] S8: According to the energy distribution characteristics, the optimal decomposition level n is determined using the information entropy criterion. opt , get the optimized number of wavelet decomposition layers;

[0066] S9: Based on the optimized wavelet decomposition layer number n opt , the segmented cardiac cycle signal is re-transformed by discrete wavelet transform to obtain an approximate coefficient and n detail coefficients.

[0067] Specifically, a suitable wavelet basis function is selected, and orthogonality and compact support are used as evaluation indicators. Orthogonality ensures the non-redundant representation of the wavelet transform, while compact support helps to reduce the amount of calculation and improve the resolution. By comparing the orthogonality and compact support of different Daubechies wavelet basis functions, the optimal wavelet basis function is selected. According to the selected optimal Daubechies wavelet basis function, a wavelet filter group is constructed. The wavelet filter group includes a low-pass decomposition filter port and a high-pass decomposition filter g, which are used to extract the low-frequency and high-frequency components of the signal, respectively. The low-pass filter extracts the general information in the signal, while the high-pass filter captures the detailed changes in the signal. The segmented cardiac cycle signal is compared with the low-pass decomposition filter Perform convolution operation and perform 2 times downsampling to obtain the first-level approximation coefficient A1. The convolution operation is achieved by multiplying the signal with the filter function point by point and summing them. 2 times downsampling is to reduce the amount of data while retaining the main features of the signal. The formula is expressed as:

[0068] A1[k]=∑ n x[n]·h[2k-n];

[0069] Where A1[k] is the first-order approximation coefficient, x[n] is the input signal, and h[n] is the coefficient of the low-pass filter. The same operation is applied to the high-pass decomposition filter g to obtain the first-order detail coefficient D1. The formula is as follows:

[0070] D1[k]=∑ n x[n]·g[2k-n];

[0071] Among them, D1[k] is the first-level detail coefficient, and g[n] is the coefficient of the high-pass filter. After completing the first-level wavelet decomposition, according to the first-level approximate coefficient A1, the convolution and downsampling operations of the low-pass and high-pass filters are repeated to obtain the second-level approximate coefficient A2 and detail coefficient D2 respectively. The decomposition process continues until the nth level is reached, and the nth level approximate coefficient A is obtained. n and detail coefficient D n . By repeatedly performing convolution and downsampling of low-pass and high-pass filters, the frequency components in the signal are gradually separated to achieve multi-scale analysis of the signal. Energy analysis is performed on the obtained n-level approximation coefficients and detail coefficients. Energy analysis aims to calculate the energy proportion of each level of coefficients to understand the contribution of each frequency component to the total signal energy. The energy calculation formula is:

[0072] E i =∑ k |C i [k]| 2 ;

[0073] Among them, E i represents the energy of the i-th level coefficient, C i [k] is the approximation coefficient or detail coefficient of the i-th level. According to the energy distribution characteristics, the information entropy criterion is used to determine the optimal decomposition level n opt The information entropy criterion determines the most suitable number of decomposition levels by measuring the amount of information at each level, ensuring that neither over-decomposition increases the computational complexity nor under-decomposition misses key features. opt , the segmented cardiac cycle signal is re-transformed by discrete wavelet transform to obtain an approximate coefficient and n detail coefficients.

[0074] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0075] (1) Calculate the mean of the approximate coefficient and the n detail coefficients respectively to obtain n+1 mean features, and centralize the approximate coefficient and the n detail coefficients according to the n+1 mean features to obtain a centralized coefficient sequence;

[0076] (2) Calculate the variance of the centering coefficient sequence to obtain n+1 variance features, and calculate the third-order central moment based on the centering coefficient sequence to obtain n+1 skewness features;

[0077] (3) Based on the centralized coefficient sequence, the fourth-order central moment is calculated to obtain n+1 kurtosis features, and the energy of the approximate coefficient and the n detail coefficients is calculated respectively to obtain n+1 energy features;

[0078] (4) According to the n+1 energy features, the relative energy proportion of each coefficient is calculated to obtain n+1 energy ratio features, and the Shannon entropy is calculated for the approximate coefficient and the n detail coefficients respectively to obtain n+1 entropy value features;

[0079] (5) Based on the approximate coefficient and n detail coefficients, the mutual correlation coefficients between adjacent scale coefficients are calculated to obtain n scale-related features;

[0080] (6) The mean feature, variance feature, skewness feature, kurtosis feature, energy feature, energy ratio feature, entropy feature and scale-related feature are combined in sequence to obtain a multidimensional feature vector.

[0081] Specifically, the mean of each approximate coefficient and detail coefficient is calculated respectively. These means reflect the average value of each coefficient on a specific frequency scale and are the basic description of the overall strength of the signal. Through calculation, n+1 mean features are obtained. According to the mean feature, the approximate coefficient and n detail coefficients are centered, that is, each coefficient is subtracted from its mean to obtain a centered coefficient sequence. The purpose of centralization is to eliminate data offset and make subsequent statistical analysis more accurate. The variance of the centered coefficient sequence is calculated to obtain n+1 variance features. The variance describes the degree of dispersion of data points around the mean, and the formula is:

[0082]

[0083] in, is the variance of the i-th level coefficient, reflecting the fluctuation of the signal strength at this scale. The third-order central moment is calculated based on the centralized coefficient sequence to obtain n+1 skewness features. Skewness is a statistic used to describe the symmetry of data distribution. The formula for the third-order central moment is:

[0084]

[0085] Among them, γ i is the skewness characteristic of the i-th level coefficient, which describes the degree of skewness of the coefficient distribution. A positive skewness indicates that the right tail of the distribution is longer, and a negative skewness indicates that the left tail is longer. Similarly, the calculation of the fourth-order central moment gives n+1 kurtosis characteristics. Kurtosis is a statistic that measures the degree of peak distribution, and the formula is:

[0086]

[0087] Among them, κ i is the kurtosis feature of the i-th level coefficient. The kurtosis of the standard normal distribution is 0, and the distribution with large kurtosis is sharper. The energy is calculated for each approximate coefficient and detail coefficient, and n+1 energy features are obtained. The energy feature reflects the strength of the signal at each scale, and the formula is:

[0088]

[0089] Among them, E i is the energy characteristic of the i-th level coefficient. According to the energy characteristics, the relative energy proportion of each coefficient is calculated to obtain n+1 energy ratio characteristics. These characteristics describe the distribution of signal energy at various frequency scales and are important indicators for signal characteristic analysis. In order to further characterize the complexity of the signal, the Shannon entropy is calculated for the approximate coefficient and the detail coefficient respectively, and n+1 entropy value characteristics are obtained.

[0090] Shannon entropy measures the uncertainty of information and its formula is:

[0091]

[0092] Among them, H i is the entropy characteristic of the i-th level coefficient, p ij is the probability distribution of the coefficients. Based on the approximate coefficients and n detail coefficients, the cross-correlation coefficients between adjacent scale coefficients are calculated to obtain n scale-related features. These features reflect the similarity and correlation of signals at different scales. The calculation formula for the cross-correlation coefficient is:

[0093]

[0094] Among them, ρ ij is the mutual correlation coefficient between the coefficients of the i-th level and the j-th level. The mean feature, variance feature, skewness feature, kurtosis feature, energy feature, energy ratio feature, entropy feature and scale-related feature are combined in sequence to form a multidimensional feature vector.

[0095] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0096] (1) Standardizing the multidimensional feature vector to obtain a standardized feature matrix with a mean of 0 and a variance of 1, and calculating the feature covariance matrix based on the standardized feature matrix;

[0097] (2) performing eigenvalue decomposition on the eigencovariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, and arranging the eigenvectors in descending order according to the eigenvalues ​​to obtain a sorted eigenvector matrix;

[0098] (3) Calculate the cumulative contribution rate of the sorted eigenvector matrix to obtain the cumulative contribution rate of each principal component, and select the first k principal components whose cumulative contribution rate reaches 95% based on the cumulative contribution rate to obtain the number k of eigenvectors after dimensionality reduction;

[0099] (4) Based on the number k of eigenvectors after dimensionality reduction, the first k columns are selected from the sorted eigenvector matrix to obtain a dimensionality reduction transformation matrix;

[0100] (5) Perform matrix multiplication on the standardized feature matrix and the dimension reduction transformation matrix to obtain an initial dimension reduction feature vector, and calculate the variance of each dimension based on the initial dimension reduction feature vector to obtain the feature importance weight;

[0101] (6) The initial dimension-reduced feature vector and the feature importance weight are weighted to obtain the dimension-reduced feature vector.

[0102] Specifically, the multidimensional feature vector is standardized so that the mean of each feature is 0 and the variance is 1, eliminating the influence of different feature dimensions on data analysis, and obtaining a standardized feature matrix with a mean of 0 and a variance of 1. Based on the standardized feature matrix, the feature covariance matrix is ​​calculated. The covariance matrix is ​​a symmetric matrix whose elements represent the covariance between different features. The formula is as follows:

[0103]

[0104] Among them, Σ is the feature covariance matrix, N is the number of samples, and Z T is the transposed matrix of the standardized feature matrix. The covariance matrix reflects the linear relationship between features and is the basis of principal component analysis. Perform eigenvalue decomposition on the feature covariance matrix to obtain the eigenvalue and the corresponding eigenvector. The eigenvalue represents the variance of the data in the direction of the corresponding eigenvector, and the eigenvector represents the unit vector in this direction. The eigenvalue decomposition formula is:

[0105] ΣV=VΛ;

[0106] Among them, Σ is the covariance matrix, V is the eigenvector matrix, and Λ is the diagonal eigenvalue matrix. According to the size of the eigenvalue, the eigenvectors are arranged in descending order to form a sorted eigenvector matrix. The column vectors of the sorted eigenvector matrix are arranged from large to small according to the eigenvalue, and the data direction corresponding to these vectors contains the largest variance. Calculate the cumulative contribution rate of the sorted eigenvector matrix. The cumulative contribution rate represents the proportion of the total variance of the data explained after selecting the first K principal components. By calculating the cumulative contribution rate, the amount of original data information retained by the first k principal components is determined. Usually, the first k principal components with a cumulative contribution rate of 95% are selected to ensure that sufficient original data features are retained. Based on the selected number of eigenvectors k after dimensionality reduction, the first k columns are selected from the sorted eigenvector matrix to obtain the dimensionality reduction transformation matrix. The dimensionality reduction transformation matrix can project the original data into a k-dimensional space, simplify the dimension of the data, and retain the main feature information. Perform matrix multiplication on the standardized feature matrix and the dimensionality reduction transformation matrix to obtain the initial dimensionality reduction eigenvector. The formula is:

[0107] Y = ZW;

[0108] Among them, Y is the initial dimension reduction feature vector, Z is the standardized feature matrix, and W is the dimension reduction transformation matrix. This step realizes the projection of data from high-dimensional space to low-dimensional space, making the data more interpretable and processable. In order to optimize the dimension reduction feature vector, the variance of each dimension is calculated to obtain the feature importance weight. The larger the variance, the more information the dimension contains in the data. Therefore, the initial dimension reduction feature vector is weighted according to the variance. Through weighted processing, important features are strengthened and unimportant features are weakened, so that the reduced-dimensional data can better represent the main information of the original data.

[0109] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0110] (1) Performing cluster analysis on the reduced-dimensional feature vectors to obtain initial cluster centers, and using the expectation-maximization algorithm to estimate the parameters of the Gaussian mixture model based on the initial cluster centers to obtain an initial Gaussian mixture model;

[0111] (2) performing a Bayesian information criterion evaluation on the initial Gaussian mixture model to obtain an optimal number of Gaussian components M, and retraining the Gaussian mixture model based on the optimal number of Gaussian components M to obtain an optimized Gaussian mixture model;

[0112] (3) According to the optimized Gaussian mixture model, the probability density functions of normal heart sounds and abnormal heart sounds are calculated to obtain the probability distribution in the feature space;

[0113] (4) Performing equal probability curve analysis on the probability distribution in the feature space to obtain the initial classification threshold, and introducing a dynamic adjustment factor α based on the initial classification threshold to construct a threshold adjustment formula;

[0114] (5) Performing statistical analysis on the classification results of the most recent N samples to obtain the ratio P of normal heart sounds to abnormal heart sounds;

[0115] (6) updating the dynamic adjustment factor α according to the ratio P of normal heart sounds to abnormal heart sounds to obtain a new dynamic adjustment factor α';

[0116] (7) Substitute the new dynamic adjustment factor α' into the threshold adjustment formula to calculate the adaptive classification threshold.

[0117] Specifically, cluster analysis is performed on the feature vector after dimensionality reduction to divide the data points into multiple clusters, each cluster representing a feature pattern. Common clustering algorithms such as K-means are used to determine the initial cluster center. The K-means algorithm determines the center of the cluster by minimizing the sum of squares within the cluster. The formula is as follows:

[0118]

[0119] Among them, J is the objective function, k is the number of clusters, x is the data point, μ i is the center of the ith cluster, C i is the set of data points belonging to the i-th cluster. Through multiple iterations, the K-means algorithm eventually converges to a set of stable cluster centers, which are used as initial cluster centers. Use the initial cluster centers to initialize the parameters of the Gaussian mixture model. The Gaussian mixture model assumes that the data distribution is a mixture of multiple Gaussian distributions, each Gaussian distribution represents a cluster. The expectation maximization (EM) algorithm is used to estimate the parameters of the Gaussian mixture model, including the mean, covariance matrix, and mixing coefficient of each Gaussian distribution. The EM algorithm optimizes the parameters by alternating between the E step and the M step. In the E step, the probability (responsibility) of each data point belonging to each Gaussian distribution is calculated, and in the M step, the parameters are updated according to these probabilities. The goal of the EM algorithm is to maximize the following log-likelihood function:

[0120]

[0121] Among them, X is the data set, θ is the model parameter, and π j is the mixing coefficient of the jth Gaussian distribution, is the probability density function of the Gaussian distribution. After obtaining the initial Gaussian mixture model, the Bayesian Information Criterion (BIC) is used to evaluate the model. BIC is a statistic used to select the best model while considering the complexity of the model. The calculation formula of BIC is:

[0122] BIC = -2logL + plogN;

[0123] Where L is the maximum likelihood estimate of the model, p is the number of parameters of the model, and N is the number of samples. By calculating the BIC values ​​of different numbers of Gaussian components, the number of components M that minimizes the BIC is selected, which corresponds to the optimal number of Gaussian components. Based on the optimal number of Gaussian components M, the Gaussian mixture model is retrained to obtain the optimized model to ensure that the model can best fit the data, thereby improving the classification accuracy. Using the optimized Gaussian mixture model, the probability density functions of normal heart sounds and abnormal heart sounds are calculated, which describe the distribution of different categories of heart sounds in the feature space. The probability distribution in the feature space is analyzed by equal probability curves to determine the initial classification threshold. The equal probability curve refers to a set of points with the same probability density in the feature space, and the initial classification threshold separates different categories of heart sounds. In order to enhance the flexibility of the model, a dynamic adjustment factor α is introduced, which is used to adjust the classification threshold to adapt to the dynamic changes of the data. The classification results of the most recent N samples are statistically analyzed to calculate the ratio P of normal heart sounds to abnormal heart sounds. According to this ratio, the dynamic adjustment factor α is updated to obtain a new adjustment factor α'. The new adjustment factor can reflect the latest data distribution, thereby adjusting the classification threshold more accurately. Substitute the new dynamic adjustment factor α' into the threshold adjustment formula to calculate the adaptive classification threshold. The adaptive threshold can be adjusted according to real-time data to ensure the accuracy and flexibility of heart sound classification. For example, in the heart sound signal, normal heart sounds and abnormal heart sounds may change due to the patient's health status and the detection environment. Through dynamic adjustment, the classification system can better adapt to these changes and improve the detection rate of abnormal heart sounds such as heart murmurs.

[0124] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0125] (1) Perform sequence analysis on the reduced-dimensional feature vector to determine the number of states N of the hidden Markov model and obtain the model state set S;

[0126] (2) Cluster the reduced-dimensional feature vectors according to the model state set S to obtain the initial state distribution π, and calculate the state transition probability matrix A based on the initial state distribution π;

[0127] (3) Fit the observed value of each state with a Gaussian distribution to obtain the observed probability distribution B, and construct a complete hidden Markov model λ based on the initial state distribution π, the state transition probability matrix A and the observed probability distribution B;

[0128] (4) Using the Baum-Welch algorithm to optimize the parameters of the complete hidden Markov model λ, and obtaining the optimized hidden Markov model λ';

[0129] (5) Generate a reduced-dimensionality feature sequence O to be identified based on the reduced-dimensionality feature vector;

[0130] (6) According to the optimized hidden Markov model λ' and the reduced-dimensional feature sequence O to be identified, the forward algorithm is used to calculate the likelihood probability P(O|λ');

[0131] (7) Compare and post-process the likelihood probability P(O|λ') with the adaptive classification threshold, and output the heart sound feature recognition result.

[0132] Specifically, determine the number of states N of the hidden Markov model. The number of states N reflects the model's ability to classify signal features and is usually determined through sequence analysis. Sequence analysis can use BIC (Bayesian Information Criterion) or other model selection criteria to select the optimal number of states to avoid overfitting or underfitting. After determining the number of states N, a model state set S is formed, which represents different potential states, such as different heart sound components in heart sounds. Based on the state set S, cluster analysis is performed on the reduced dimensionality feature vectors. The feature vectors are assigned to different states to estimate the initial state distribution π. The initial state distribution π represents the probability that the system is in each state at the beginning, and the formula is:

[0133] π i =P(Q1=i);

[0134] Among them, π i is the probability of the initial state i, Q1 is the first state of the sequence. Based on the initial state distribution π, the state transition probability matrix A can be further calculated as follows:

[0135] A ij =P(Q t+1 =j|Q t =i);

[0136] Among them, A i j represents the probability of transitioning from state i to state j, Q t and Q t +1 are the states at time t and t+1 respectively. In order to describe the observed value distribution of each state, a Gaussian distribution is fitted to the observed value of each state to obtain the observed probability distribution B. Gaussian distribution is used to estimate the probability density of observed values ​​in the model, and the formula is:

[0137]

[0138] Among them, B j (O t ) is the observed value O in state j t The probability density of j and are the mean and variance of state j respectively. Combining the initial state distribution π, the state transition probability matrix A and the observation probability distribution B, a complete hidden Markov model is constructed, and the formula is:

[0139] λ=(π,A,B);

[0140] In order to optimize the hidden Markov model L, the Baum-Welch algorithm is used for parameter estimation. The Baum-Welch algorithm is an EM (expectation maximization) algorithm that iteratively optimizes the model parameters to maximize the likelihood function of a given observation sequence. The algorithm is divided into E step (expectation step) and M step (maximization step). In the E step, the posterior probability distribution of the observation sequence is calculated, and in the M step, the model parameters are updated. Finally, the optimized hidden Markov model λ' is obtained, and its parameters better fit the observed data. The reduced-dimensional feature sequence O to be identified is generated based on the reduced-dimensional feature vector. These sequences represent the feature changes of the new heart sound data. The forward algorithm is used to calculate the likelihood probability P(O|λ′) of the reduced-dimensional feature sequence O under the optimized hidden Markov model λ'. The forward algorithm is a dynamic programming algorithm that accumulates the probabilities of all possible state paths through recursive calculations. The formula is:

[0141]

[0142] Among them, α t (j) is the probability of being in state j at time t and observing the first part of sequence O. The calculated likelihood probability P(O|λ′) is compared with the adaptive classification threshold. The adaptive classification threshold is dynamically adjusted according to historical data and can flexibly adapt to data changes. If the likelihood probability is greater than the threshold, it is judged as a normal heart sound, otherwise it is an abnormal heart sound. Finally, the recognition result of the heart sound feature is output to help distinguish normal heart sounds from abnormal heart sounds, such as heart murmurs or valve abnormalities.

[0143] The above describes the heart sound feature recognition method based on wavelet spectrum analysis in the embodiment of the present application. The following describes the heart sound feature recognition device based on wavelet spectrum analysis in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a heart sound feature recognition device based on wavelet spectrum analysis includes:

[0144] The acquisition module 201 is used to acquire heart sound signal data, and perform signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals;

[0145] A decomposition module 202 is used to perform n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients;

[0146] A calculation module 203 is used to perform multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector;

[0147] Processing module 204, used for performing principal component analysis and dimensionality reduction processing on the multi-dimensional feature vector to obtain a reduced-dimensional feature vector;

[0148] A fitting module 205 is used to fit the feature distribution using a Gaussian mixture model based on the reduced-dimensional feature vector and introduce a dynamic adjustment factor to obtain an adaptive classification threshold;

[0149] The output module 206 is used to construct a hidden Markov model through the dimension reduction feature vector, calculate the likelihood probability using a forward algorithm, compare the likelihood probability with the adaptive classification threshold, and output the heart sound feature recognition result.

[0150] Through the cooperation of the above components, the signal preprocessing method combining adaptive minimum mean square error noise suppression and Hilbert transform is used to effectively improve the quality and signal-to-noise ratio of heart sound signals. The discrete wavelet transform and multi-dimensional statistical feature calculation are introduced to make full use of the time-frequency characteristics of heart sound signals and improve the distinguishability and characterization ability of features. The dimensionality reduction processing of principal component analysis reduces the redundancy of feature vectors, improves the calculation efficiency, and retains key feature information. The Gaussian mixture model is used to fit the feature distribution and the dynamic adjustment factor is introduced to achieve the dynamic adjustment of the adaptive classification threshold, which enhances the adaptability of the method to different individuals and physiological states. Combining the hidden Markov model and the forward algorithm, the temporal characteristics of heart sound signals are fully utilized to improve the accuracy and reliability of feature recognition. By optimizing the selection of wavelet basis functions and the number of decomposition layers, the efficiency of wavelet transform and the pertinence of feature extraction are improved. The energy analysis and information entropy criteria are introduced to achieve the adaptive optimization of the number of wavelet decomposition layers and improve the effectiveness of feature extraction. The statistical characteristics and multi-scale characteristics of heart sound signals are comprehensively characterized by combining multiple statistical features and scale-related features, which enhances the expression ability of features. Through the Bayesian information criterion evaluation and expectation maximization algorithm, the parameter estimation of the Gaussian mixture model is optimized, and the accuracy of feature distribution fitting is improved. The dynamic adjustment factor and threshold adjustment formula are introduced to realize the adaptive update of the classification threshold, and improve the generalization ability of the method to different data sets.

[0151] The present application also provides a heart sound feature recognition device based on wavelet spectrum analysis, the heart sound feature recognition device based on wavelet spectrum analysis includes a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the heart sound feature recognition method based on wavelet spectrum analysis in the above-mentioned embodiments.

[0152] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the heart sound feature recognition method based on wavelet spectrum analysis.

[0153] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0154] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0155] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application 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 embodiments of the present application.

Claims

1. A heart sound feature recognition method based on wavelet spectrum analysis, characterized in that: The method comprises: Collecting heart sound signal data, and performing signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals; Performing n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients; Performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector; Performing principal component analysis and dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector; Based on the reduced dimension feature vector, a Gaussian mixture model is used to fit the feature distribution and a dynamic adjustment factor is introduced to obtain an adaptive classification threshold; A hidden Markov model is constructed by using the dimension-reduced feature vector, and a forward algorithm is used to calculate the likelihood probability, the likelihood probability is compared with the adaptive classification threshold, and a heart sound feature recognition result is output.

2. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 1 is characterized in that: The collecting of heart sound signal data, and performing signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals, includes: Collecting heart sound signal data, and performing bandpass filtering on the heart sound signal data to obtain a heart sound signal after initial filtering; Performing adaptive minimum mean square error noise suppression processing on the heart sound signal after the initial filtering to obtain a de-noised heart sound signal; Performing Hilbert transform on the denoised heart sound signal to obtain a signal envelope, and detecting the peak point of the denoised heart sound signal according to the signal envelope to obtain a demarcation point of the heart sound cycle; Based on the demarcation point of the heart sound cycle, the heart sound signal after noise reduction is segmented to obtain an initial segmented heart sound signal, and short-time energy analysis is performed on the initial segmented heart sound signal to obtain an energy value of each segment; According to the energy values ​​of each segment, the S1 sound and the S2 sound are identified on the heart sound signal of the initial segment to obtain a first segment heart sound signal; The first segmented heart sound signal is subjected to zero-phase filtering to obtain a second segmented heart sound signal, and the second segmented heart sound signal is subjected to signal resampling using a cubic spline interpolation algorithm to obtain a segmented cardiac cycle signal.

3. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 1 is characterized in that: The segmented cardiac cycle signal is subjected to n-level discrete wavelet transform decomposition to obtain an approximate coefficient and n detail coefficients, including: S1: selecting a wavelet basis function for the segmented cardiac cycle signal, and using orthogonality and compact support evaluation indicators to obtain an optimal Daubechies wavelet basis function; S2: constructing a wavelet filter bank according to the optimal Daubechies wavelet basis function to obtain a low-pass decomposition filter h and a high-pass decomposition filter g; S3: performing a convolution operation on the segmented cardiac cycle signal and the low-pass decomposition filter h, and performing a 2-fold downsampling to obtain a first-order approximation coefficient A1; S4: performing a convolution operation on the segmented cardiac cycle signal and the high-pass decomposition filter g, and performing a 2-fold downsampling to obtain a first-level detail coefficient D1; S5: Repeat the operations of step S3 and step S4 according to the first-level approximation coefficient A1 to obtain the second-level approximation coefficient A2 and the detail coefficient D2; S6: cyclically execute the operation of step S5 until the n-th level approximation coefficient An and detail coefficient Dn are obtained; S7: Perform energy analysis on the obtained n-level approximate coefficients and detail coefficients, calculate the energy proportion of each level of coefficients, and obtain energy distribution characteristics; S8: According to the energy distribution characteristics, the optimal decomposition level n is determined using the information entropy criterion. opt , get the optimized number of wavelet decomposition layers; S9: Based on the optimized wavelet decomposition layer number n opt , re-perform discrete wavelet transform on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients.

4. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 1 is characterized in that: The performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector includes: Calculating means for the approximate coefficient and the n detail coefficients respectively to obtain n+1 mean features, and performing centralization processing on the approximate coefficient and the n detail coefficients according to the n+1 mean features to obtain a centralization coefficient sequence; Calculating the variance of the centering coefficient sequence to obtain n+1 variance features, and calculating the third-order central moment according to the centering coefficient sequence to obtain n+1 skewness features; Based on the centralization coefficient sequence, the fourth-order central moment is calculated to obtain n+1 kurtosis features, and the energy of the approximate coefficient and the n detail coefficients is calculated respectively to obtain n+1 energy features; According to the n+1 energy features, the relative energy proportion of each coefficient is calculated to obtain n+1 energy ratio features, and the Shannon entropy of the approximate coefficient and the n detail coefficients is calculated respectively to obtain n+1 entropy value features; Based on the approximation coefficient and the n detail coefficients, calculating the mutual correlation coefficients between adjacent scale coefficients to obtain n scale-related features; The mean feature, the variance feature, the skewness feature, the kurtosis feature, the energy feature, the energy ratio feature, the entropy feature and the scale-related feature are combined in sequence to obtain a multidimensional feature vector.

5. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 1 is characterized in that: The performing principal component analysis dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector includes: Standardizing the multidimensional feature vector to obtain a standardized feature matrix with a mean of 0 and a variance of 1, and calculating a feature covariance matrix based on the standardized feature matrix; Performing eigenvalue decomposition on the characteristic covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, and arranging the eigenvectors in descending order according to the eigenvalues ​​to obtain a sorted eigenvector matrix; Calculating the cumulative contribution rate of the sorted eigenvector matrix to obtain the cumulative contribution rate of each principal component, and selecting the first k principal components whose cumulative contribution rate reaches 95% according to the cumulative contribution rate to obtain the number k of eigenvectors after dimensionality reduction; Based on the number k of eigenvectors after dimensionality reduction, first k columns are selected from the sorted eigenvector matrix to obtain a dimensionality reduction transformation matrix; Performing a matrix multiplication operation on the standardized feature matrix and the dimension reduction transformation matrix to obtain an initial dimension reduction feature vector, and calculating the variance of each dimension based on the initial dimension reduction feature vector to obtain a feature importance weight; The initial dimension-reduced feature vector and the feature importance weight are weighted to obtain a dimension-reduced feature vector.

6. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 1 is characterized in that: The method of fitting the feature distribution using a Gaussian mixture model and introducing a dynamic adjustment factor based on the reduced dimension feature vector to obtain an adaptive classification threshold comprises: Performing cluster analysis on the dimension-reduced feature vectors to obtain initial cluster centers, and using an expectation-maximization algorithm to perform parameter estimation on a Gaussian mixture model based on the initial cluster centers to obtain an initial Gaussian mixture model; Performing a Bayesian information criterion evaluation on the initial Gaussian mixture model to obtain an optimal number of Gaussian components M, and retraining the Gaussian mixture model based on the optimal number of Gaussian components M to obtain an optimized Gaussian mixture model; Calculating the probability density functions of normal heart sounds and abnormal heart sounds according to the optimized Gaussian mixture model to obtain the probability distribution in the feature space; Performing an equal probability curve analysis on the probability distribution in the feature space to obtain an initial classification threshold, and introducing a dynamic adjustment factor α based on the initial classification threshold to construct a threshold adjustment formula; Perform statistical analysis on the classification results of the most recent N samples to obtain the ratio P of normal heart sounds to abnormal heart sounds; According to the ratio P of the normal heart sounds to the abnormal heart sounds, the dynamic adjustment factor α is updated to obtain a new dynamic adjustment factor α'; The new dynamic adjustment factor α' is substituted into the threshold adjustment formula to calculate the adaptive classification threshold.

7. The heart sound feature recognition method based on wavelet spectrum analysis according to claim 6 is characterized in that: The method of constructing a hidden Markov model by using the dimension-reduced feature vector, calculating the likelihood probability by using a forward algorithm, comparing the likelihood probability with the adaptive classification threshold, and outputting a heart sound feature recognition result includes: Performing sequence analysis on the dimension-reduced feature vector to determine the number of states N of the hidden Markov model and obtain a model state set S; According to the model state set S, the reduced-dimensional feature vectors are clustered to obtain an initial state distribution π, and based on the initial state distribution π, a state transition probability matrix A is calculated; Performing Gaussian distribution fitting on the observed values ​​of each state to obtain the observed probability distribution B, and constructing a complete hidden Markov model λ according to the initial state distribution π, the state transition probability matrix A and the observed probability distribution B; Use the Baum-Welch algorithm to optimize the parameters of the complete hidden Markov model λ to obtain an optimized hidden Markov model λ'; Generate a reduced dimension feature sequence O to be identified according to the reduced dimension feature vector; According to the optimized hidden Markov model λ' and the reduced-dimensional feature sequence O to be identified, a forward algorithm is used to calculate the likelihood probability P(O|λ'); The likelihood probability P(O|λ') is compared with the adaptive classification threshold and post-processed to output a heart sound feature recognition result.

8. A heart sound feature recognition device based on wavelet spectrum analysis, characterized in that: The device for performing the heart sound feature recognition method based on wavelet spectrum analysis according to any one of claims 1 to 7 comprises: An acquisition module is used to acquire heart sound signal data, and perform signal filtering and segmentation processing on the heart sound signal data to obtain segmented cardiac cycle signals; A decomposition module, used for performing n-level discrete wavelet transform decomposition on the segmented cardiac cycle signal to obtain an approximate coefficient and n detail coefficients; A calculation module, used for performing multi-dimensional statistical feature calculation on the approximate coefficient and the n detail coefficients to obtain a multi-dimensional feature vector; A processing module, used for performing principal component analysis and dimensionality reduction processing on the multidimensional feature vector to obtain a reduced-dimensional feature vector; A fitting module, used to fit the feature distribution using a Gaussian mixture model based on the reduced-dimensional feature vector and introduce a dynamic adjustment factor to obtain an adaptive classification threshold; The output module is used to construct a hidden Markov model through the dimension reduction feature vector, calculate the likelihood probability using a forward algorithm, compare the likelihood probability with the adaptive classification threshold, and output the heart sound feature recognition result.

9. A heart sound feature recognition device based on wavelet spectrum analysis, characterized in that: The heart sound feature recognition device based on wavelet spectrum analysis includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the heart sound feature recognition device based on wavelet spectrum analysis to execute the heart sound feature recognition method based on wavelet spectrum analysis according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the heart sound feature recognition method based on wavelet spectrum analysis as described in any one of claims 1 to 7 is implemented.

Images