Method and system for screening and selecting cardiac murmur frequency observation points

Abstract

The invention relates to the technical field of heart sound signal processing, and discloses a heart noise frequency observation point screening and selecting method and system. The method comprises the following steps: carrying out preprocessing and time-frequency analysis on heart sound signal data to obtain a target time-frequency spectrogram and a corresponding energy distribution diagram; constructing a multi-layer spectrum network to obtain a three-layer spectrum network; performing propagation model analysis to obtain an energy-frequency propagation model, an amplitude-duration propagation model and a frequency-phase propagation model; carrying out network propagation characteristic calculation to obtain a propagation characteristic evaluation index; screening a corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and performing adaptive density peak clustering to obtain a first heart sound characteristic observation point subset; and performing optimization processing on the first heart sound feature observation point subset through a PSO-Means algorithm to obtain a second heart sound feature observation point subset, thereby improving the screening and selection accuracy of the heart noise frequency observation points.

Description

Technical Field

[0001] The present application relates to the technical field of cardiac sound signal processing, and in particular to a method and system for screening and selecting cardiac murmur frequency observation points. Background Art

[0002] Heart murmur is one of the important indicators in the diagnosis of cardiovascular diseases. Its accurate identification and analysis are crucial for the early detection of heart problems. Traditional heart murmur analysis methods mainly rely on the doctor's auscultation experience, which is highly subjective and has limited accuracy. With the development of digital signal processing technology, automatic analysis methods based on heart sound signals have gradually become a research hotspot. However, the non-stationary and complex characteristics of heart sound signals make it still challenging to accurately extract and analyze murmur features.

[0003] Most existing heart sound signal analysis methods focus on a single dimension of the time domain or frequency domain, making it difficult to fully capture the time-frequency characteristics of the murmur. In addition, these methods often ignore the interaction and propagation characteristics between different frequency components in the heart sound signal, resulting in an inaccurate and incomplete description of the murmur. In terms of the selection of frequency observation points, existing methods mostly use fixed thresholds or simple peak detection, lacking consideration of the overall structure and dynamic characteristics of the signal. Summary of the invention

[0004] The present application provides a method and system for screening and selecting heart murmur frequency observation points, which are used to improve the accuracy of screening and selecting heart murmur frequency observation points.

[0005] In a first aspect, the present application provides a method for screening and selecting observation points of heart murmur frequency, the method comprising:

[0006] Preprocess and perform time-frequency analysis on the heart sound signal data to obtain the target time-frequency spectrum and the corresponding energy distribution diagram;

[0007] Performing a multi-layer spectrum network construction on the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer;

[0008] Performing a propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model, and a frequency-phase propagation model;

[0009] Performing network propagation characteristic calculation on the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain propagation characteristic evaluation indicators;

[0010] According to the propagation characteristic evaluation index, a corresponding frequency observation point candidate set is selected from the heart sound signal data, and adaptive density peak clustering is performed on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset;

[0011] The first heart sound feature observation point subset is optimized by using the PSO-Means algorithm to obtain a second heart sound feature observation point subset.

[0012] In a second aspect, the present application provides a heart murmur frequency observation point screening and selection system, the heart murmur frequency observation point screening and selection system comprising:

[0013] A preprocessing module is used to preprocess and perform time-frequency analysis on the heart sound signal data to obtain a target time-frequency spectrum and a corresponding energy distribution diagram;

[0014] A construction module, used for constructing a multi-layer spectrum network for the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer;

[0015] An analysis module, used to perform a propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model and a frequency-phase propagation model;

[0016] A calculation module, used to calculate network propagation characteristics of the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain propagation characteristic evaluation indicators;

[0017] A screening module, used for screening a corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and performing adaptive density peak clustering on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset;

[0018] The optimization module is used to optimize the first heart sound feature observation point subset by using the PSO-Means algorithm to obtain the second heart sound feature observation point subset.

[0019] A third aspect of the present application provides a heart murmur frequency observation point screening and selection device, comprising: 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 murmur frequency observation point screening and selection device to perform the above-mentioned heart murmur frequency observation point screening and selection method.

[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 murmur frequency observation point screening and selection method.

[0021] In the technical solution provided by the present application, by constructing a three-layer spectrum network including the bottom layer, the middle layer and the top layer, the multi-scale frequency characteristics of the heart sound signal are fully captured, which helps to more accurately characterize the complex structure of the murmur. The three propagation models of energy-frequency, amplitude-duration and frequency-phase are adopted to deeply analyze the interaction between frequency components and reveal the dynamic propagation characteristics of the murmur. By calculating indicators such as node energy importance, amplitude stability and frequency synergy, the importance of frequency observation points is comprehensively evaluated to improve the accuracy of screening. Combining spectral clustering and density peak detection, adaptive clustering of frequency observation points is achieved, overcoming the limitations of the fixed threshold method. The particle swarm optimization algorithm is introduced in combination with K-means clustering to further optimize the selection of observation points and improve the stability and representativeness of the screening results. Taking into account multiple evaluation indicators, a comprehensive score is obtained through a multi-criteria decision-making method to ensure the comprehensiveness and objectivity of the screening process. In the signal preprocessing stage, time-frequency analysis is integrated to extract Mel frequency features, which enhances the characterization ability of murmur features. From time domain features to frequency domain features, and then to network topology features, signal features are gradually extracted in depth, thereby improving the accuracy of screening and selection of heart murmur frequency observation points. 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 A schematic diagram of an embodiment of a method for screening and selecting observation points of heart murmur frequency in an embodiment of the present application;

[0024] Figure 2 This is a schematic diagram of an embodiment of a system for screening and selecting heart murmur frequency observation points in an embodiment of the present application. DETAILED DESCRIPTION

[0025] Embodiments of the present application provide a method and system for selecting observation points for heart murmur frequency. 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 interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. 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 In one embodiment of the present application, a method for selecting and screening heart murmur frequency observation points includes:

[0027] Step S101, preprocessing and time-frequency analysis are performed on the heart sound signal data to obtain a target time-frequency spectrum and a corresponding energy distribution diagram;

[0028] It is understandable that the execution subject of the present application may be a heart murmur frequency observation point screening and selection system, 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 original heart sound signal data is subjected to bandpass filtering to remove low-frequency noise and high-frequency interference in the signal, retain the main frequency components of the heart sound signal, and obtain the heart sound signal after initial filtering. The heart sound signal after initial filtering is subjected to wavelet threshold denoising, and the random noise and murmur in the signal are effectively removed by wavelet transform and threshold processing to obtain the heart sound signal after denoising. The heart sound signal after denoising is subjected to heart sound cycle segmentation. By detecting the periodic components in the heart sound signal, the signal is divided into multiple heart sound cycle segments. These segments represent different heart sound cycles and contain various characteristic information of the heart sound signal. Time domain feature extraction is performed on the heart sound cycle segments, and a series of characteristic values ​​including the mean value, standard deviation, kurtosis, skewness, amplitude, etc. of the signal are extracted. These characteristic values ​​can effectively describe the time domain characteristics of the heart sound signal. Based on the time domain feature set, the heart sound cycle segments are normalized. The characteristic values ​​of different heart sound segments are adjusted to a standard range, the characteristic deviation caused by the amplitude difference is eliminated, and the standardized heart sound cycle segments are obtained. Perform short-time Fourier transform on the standardized heart sound cycle segment, convert the time domain signal into the frequency domain signal, and generate the initial time-frequency spectrum. Short-time Fourier transform can provide balanced resolution between the time domain and the frequency domain, so that the spectral characteristics of the heart sound signal can be observed over time. Calculate the Mel-frequency cepstral coefficients based on the initial time-frequency spectrum, and analyze and extract the Mel-frequency characteristics of the signal. Mel-frequency cepstral coefficients are a characteristic parameter commonly used in speech and audio signal processing. Based on the human ear's perception of different frequencies, they can more closely reflect the actual characteristics of the heart sound signal. By calculating these coefficients, a Mel-frequency feature matrix is ​​constructed. Based on the Mel-frequency feature matrix, a target time-frequency spectrum is constructed, and the energy density of the spectrum is calculated to obtain the corresponding energy distribution diagram. The energy distribution diagram reflects the energy distribution of the signal at different frequencies and is used for frequency observation point screening and analysis.

[0030] Step S102: construct a multi-layer spectrum network for the target time-frequency spectrum graph and the energy distribution graph to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer;

[0031] Specifically, the frequency peak detection is performed on the target time-frequency spectrum to identify a set of significant frequency points, which represent important frequency features in the heart sound signal. According to the corresponding positions of the significant frequency point set and the energy distribution map, each frequency point is assigned an energy weight to form a weighted frequency point set. The weighted frequency point set is subjected to frequency correlation analysis to calculate the correlation coefficient matrix between these frequency points. The correlation coefficient matrix reflects the similarity and dependency between different frequency points. On this basis, the weighted frequency point set is subjected to hierarchical analysis to classify the frequency points with higher correlation into the same category to obtain the underlying frequency node set. The underlying frequency node set is subjected to neighborhood analysis to determine the local connection relationship of each node. By analyzing the interaction of each node in the frequency space around it, an underlying spectrum network is constructed, in which each node represents a frequency point and the edge represents the connection relationship and energy weight between nodes. In the underlying spectrum network, the frequency pattern community set is identified by the community detection algorithm to find groups with similar frequency features in the spectrum, and the related communities are merged into middle-level nodes according to the internal structure of these groups to form a middle-level spectrum network. Based on the middle-level spectrum network, a global topological analysis is performed to calculate the centrality indicators of the network, such as betweenness centrality and degree centrality. These indicators help identify the key nodes in the network, that is, those nodes that have a significant impact on the network structure. According to the centrality indicators, the target middle-level nodes are selected from the middle-level nodes and used as the basis for building the top-level spectrum network. In the top-level spectrum network, the nodes represent the important frequency pattern communities in the middle-level network, while the edges represent the connection relationship and interaction between these communities. Based on the structure of the bottom-level spectrum network, the middle-level spectrum network, and the top-level spectrum network, a three-layer spectrum network is generated. The three-layer network structure can provide a multi-level and all-round analysis of the frequency characteristics of heart sound signals, which helps to identify the key frequency observation points in heart murmurs.

[0032] Step S103: Perform propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model, and a frequency-phase propagation model;

[0033] Specifically, the energy distribution of each node in the three-layer spectrum network is analyzed to obtain the node energy distribution matrix, which describes the energy distribution of each node at different frequencies. According to the node energy distribution matrix, the energy transfer coefficient between adjacent nodes in the network is calculated to obtain the energy transfer coefficient matrix. The energy transfer coefficient represents the efficiency and probability of energy propagation from one node to adjacent nodes. The energy transfer coefficient matrix is ​​normalized to eliminate the influence of energy differences between different nodes, and a standardized energy-frequency propagation probability matrix is ​​obtained. The standardized value of this matrix can be regarded as a state transition probability in a Markov chain. Based on this matrix, a Markov chain model is constructed to obtain an energy-frequency propagation model. This model is used to simulate and predict the propagation behavior of energy between different nodes in the spectrum network and reveal the energy transfer path of the frequency characteristics in the heart sound signal. The amplitude-time series of each node in the three-layer spectrum network is extracted to obtain the node amplitude-time feature matrix, which records the change of the amplitude of each node over time, and can reflect the change characteristics of the heart sound signal in the time domain. By analyzing the node amplitude-time feature matrix, the amplitude attenuation coefficient and duration of each node are calculated to obtain the amplitude-duration feature vector. The calculation of amplitude attenuation coefficient and duration helps to understand the stability and persistence of the signal. Exponential fitting is performed on the eigenvector to obtain the amplitude-duration propagation model. This model is used to describe the attenuation law of signal amplitude over time and help understand the dynamic characteristics of heart sound signals. Phase spectrum analysis is performed on each node in the three-layer spectrum network to obtain the node phase-frequency relationship matrix, which describes the phase change of each node at different frequencies. Based on the node phase-frequency relationship matrix, the phase difference between adjacent frequency components is calculated to obtain the phase difference distribution vector. Phase difference is an important parameter between signal frequency characteristics, which can reveal the coherence and waveform characteristics of the signal. By performing statistical analysis and periodicity detection on the phase difference distribution vector, a frequency-phase propagation model is constructed. The model helps to understand the phase change law of heart sound signals between different frequency components and reveals the phase characteristics of heart sound signals and their potential periodic changes.

[0034] Step S104, calculating network propagation characteristics of the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain propagation characteristic evaluation indicators;

[0035] Specifically, the steady-state distribution of the energy-frequency propagation model is calculated to obtain the energy propagation balance vector. The energy propagation balance vector reflects the proportion of energy distribution occupied by each node in the steady state. The energy centrality of each node is calculated based on the energy propagation balance vector to obtain the energy importance ranking of the node, and identify the nodes that play a key role in the energy propagation process. The amplitude-duration propagation model is decomposed into time series, and the time series data is decomposed into trend terms, periodic terms and random terms. These decomposition terms help understand the long-term trend, periodic changes and random fluctuation characteristics of the signal. By analyzing the trend term and the periodic term, the amplitude persistence index is calculated, and the random term is used to calculate the fluctuation coefficient to obtain the amplitude stability index. The amplitude stability index is a key parameter to measure the stability and consistency of the signal in time, reflecting the persistence and stability of the signal. At the same time, the phase synchronization analysis of the frequency-phase propagation model is performed to calculate the phase coupling strength matrix between frequency pairs. This matrix represents the degree of phase coupling between different frequency components and is an indicator to measure the synchronization between the frequency components of the signal. By analyzing the phase coupling strength matrix, the global phase synchronization index is calculated to obtain the frequency synergy index. The frequency synergy index is used to evaluate the coordination degree of heart sound signals at different frequencies and reveal the synchronization and coherence of signals in the frequency domain. A multi-criteria decision analysis is performed on the node energy importance ranking, amplitude stability index and frequency synergy index to obtain a comprehensive scoring vector. The vector combines the scores of various indicators and is a comprehensive indicator to measure the importance of each node in the multidimensional propagation feature space. Based on the comprehensive scoring vector, the importance of network nodes is ranked, and those nodes that perform outstandingly in energy propagation, amplitude stability and frequency synergy are identified to form a target frequency node set. The network topology analysis is performed on the target frequency node set to calculate the topological characteristics of the nodes, such as degree centrality, betweenness centrality and closeness centrality. These centrality indicators are used to evaluate the position and role of each node in the network structure and reveal the key position of the node in the propagation process. According to the topological feature matrix and the comprehensive scoring vector, a multidimensional propagation feature space is constructed, and the propagation feature evaluation index is calculated on this basis.

[0036] Step S105: selecting a corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and performing adaptive density peak clustering on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset;

[0037] Specifically, the frequency points in the heart sound signal data are ranked by importance according to the propagation characteristic evaluation index to form a frequency point importance sequence. The sequence shows the relative importance of each frequency point in the propagation characteristic, which helps to identify which frequency points have a higher priority in the heart sound signal analysis. The frequency point importance sequence is truncated, and the top N frequency points are selected to form a candidate set of frequency observation points. The frequency observation point candidate set is subjected to time-frequency correlation analysis, and the correlation coefficient between each pair of frequency points is calculated to obtain the correlation coefficient matrix between the observation points. The matrix shows the similarity and correlation between the frequency points. Based on the correlation coefficient matrix, a weighted undirected graph structure is constructed, in which the nodes represent the frequency observation points and the edge weights represent the similarity between the frequency points. The weighted undirected graph structure is clustered by the spectral clustering method to preliminarily obtain the cluster distribution of the frequency points, that is, the initial clustering result. According to the initial clustering result, the local density and relative distance of each observation point are calculated to generate a density-distance distribution map. The local density reflects the density of neighboring points around an observation point, while the relative distance represents the distance between the observation point and other high-density points. By performing adaptive threshold analysis on the density-distance distribution graph, the density peak points are determined. These points usually represent the center or core area of ​​the cluster. The set of density peak points is the first cluster center set. Based on the first cluster center set, the remaining observation points are assigned to the nearest neighbors, and each unclassified observation point is assigned to the nearest cluster center to obtain the optimized clustering result. During the optimization process, each observation point is assigned to the cluster corresponding to its closest density peak point to ensure the rationality and accuracy of the clustering results. The most representative observation point is selected in each cluster to form a subset of the first heart sound feature observation points.

[0038] Step S106: Optimize the first heart sound feature observation point subset using the PSO-Means algorithm to obtain a second heart sound feature observation point subset.

[0039] Specifically, feature extraction is performed on the first heart sound feature observation point subset to generate a multidimensional feature vector for each observation point, which includes multiple feature parameters of the observation point in the heart sound signal. According to the multidimensional feature vector, the position and speed of each particle in the particle swarm are initialized to obtain an initial particle swarm. Each particle in the initial particle swarm represents a possible solution, that is, a set of cluster centers. Fitness evaluation is performed on each particle in the initial particle swarm. The fitness value reflects the clustering effect of each particle, which is usually measured by calculating the ratio of the internal distance of the cluster to the distance between clusters. Through fitness evaluation, an initial fitness value set is obtained, and the global optimal particle and the individual optimal particle, that is, the initial optimal solution, are selected from them. The global optimal particle represents the optimal solution in the current particle swarm, and the individual optimal particle is the optimal solution reached by each particle in its historical trajectory. The initial particle swarm is updated in speed and position. Through the update formula, the speed and position of the particle will be gradually adjusted to move toward the global optimal solution and the individual optimal solution. The updated particle swarm is the new generation particle swarm. K-means clustering is performed on the new generation particle swarm to obtain the second cluster center set. The clustering ability of the K-means algorithm is used to help the particle swarm further optimize the quality of its solution. According to the second cluster center set, the distance ratio between the internal and inter-cluster distances is calculated to obtain the cluster evaluation index. The internal distance of a cluster refers to the distance between points in the same cluster, while the inter-cluster distance refers to the distance between different cluster centers. By comparing the cluster evaluation index, the global optimal solution and the individual optimal solution are updated. This process is repeated, and the clustering results of each particle are continuously optimized through multiple iterations until the fitness value reaches the predetermined convergence condition, and finally the target optimization result is obtained. According to the target optimization result, the point closest to the cluster center is selected from the first heart sound feature observation point subset to form the second heart sound feature observation point subset. These observation points are representative points after optimization, which can more accurately reflect the characteristics of the heart sound signal. Through the optimization process of the PSO-Means algorithm, the accuracy and stability of the clustering results are effectively improved, making the final selected observation points more representative.

[0040] In the embodiment of the present application, by constructing a three-layer spectrum network including the bottom layer, the middle layer and the top layer, the multi-scale frequency characteristics of the heart sound signal are fully captured, which helps to more accurately characterize the complex structure of the murmur. The three propagation models of energy-frequency, amplitude-duration and frequency-phase are adopted to deeply analyze the interaction between frequency components and reveal the dynamic propagation characteristics of the murmur. By calculating indicators such as node energy importance, amplitude stability and frequency synergy, the importance of frequency observation points is comprehensively evaluated to improve the accuracy of screening. Combining spectral clustering and density peak detection, adaptive clustering of frequency observation points is achieved, overcoming the limitations of the fixed threshold method. The particle swarm optimization algorithm is introduced in combination with K-means clustering to further optimize the selection of observation points and improve the stability and representativeness of the screening results. Taking into account multiple evaluation indicators, a comprehensive score is obtained through a multi-criteria decision-making method to ensure the comprehensiveness and objectivity of the screening process. Time-frequency analysis is integrated into the signal preprocessing stage to extract Mel frequency features, which enhances the characterization ability of murmur features. From time domain features to frequency domain features, and then to network topology features, signal features are gradually extracted in depth, thereby improving the accuracy of screening and selection of heart murmur frequency observation points.

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

[0042] (1) performing bandpass filtering on the heart sound signal data to obtain an initially filtered heart sound signal, and performing wavelet threshold denoising on the initially filtered heart sound signal to obtain a denoised heart sound signal;

[0043] (2) performing heart sound cycle segmentation on the denoised heart sound signal to obtain a plurality of heart sound cycle segments, and performing time domain feature extraction on the plurality of heart sound cycle segments to obtain a time domain feature set;

[0044] (3) normalizing the plurality of heart sound cycle segments according to the time domain feature set to obtain standardized heart sound cycle segments, and performing short-time Fourier transform on the standardized heart sound cycle segments to obtain an initial time-frequency spectrum;

[0045] (4) Calculating the Mel-frequency cepstrum coefficients according to the initial time-frequency spectrum to obtain a Mel-frequency feature matrix, and constructing the target time-frequency spectrum according to the Mel-frequency feature matrix;

[0046] (5) Calculate the energy density of the target time-frequency spectrum to obtain the corresponding energy distribution diagram.

[0047] Specifically, the heart sound signal data is band-pass filtered to remove the low-frequency and high-frequency noise in the heart sound signal, and only the effective signal components in the target frequency band are retained. Commonly used band-pass filters can be implemented by selecting appropriate cutoff frequencies. Assuming that the main frequency components of the heart sound signal are concentrated between 20Hz and 200Hz, a band-pass filter can be designed, and its lower cutoff frequency f is set to L is 20Hz, the upper cutoff frequency f H After being processed by a bandpass filter, the initial filtered heart sound signal S is obtained. f (t). Perform wavelet threshold denoising on the heart sound signal after initial filtering. Wavelet threshold denoising decomposes the signal into multiple scales by selecting appropriate wavelet basis and decomposition level, and applies thresholds at different scales to remove noise components. Select the soft threshold method and appropriate wavelet basis (such as Daubechies wavelet), decompose the signal into several levels, and then apply the threshold function to the wavelet coefficients of each level to remove noise, and finally reconstruct the signal to obtain the denoised heart sound signal S d (t). Perform heart sound cycle segmentation on the denoised heart sound signal. Utilize the periodic characteristics of the heart sound signal, for example, by detecting the S1 and S2 sounds (i.e., the sound of the heart valve closing) in the heart sound to mark the beginning and end of each heart sound cycle. Through these marks, the heart sound signal is segmented into multiple heart sound cycle segments P i (t), where i represents the i-th heart sound cycle segment. These segments are analyzed to extract time domain features, such as peak value, mean value, standard deviation, energy, etc. These features form the time domain feature set {F j}, where F j Represents the jth time domain feature. According to the time domain feature set, multiple heart sound cycle segments are normalized. The eigenvalues ​​of different segments are scaled to the same range to eliminate the eigenvalue deviation caused by amplitude differences. The standardized heart sound cycle segments are short-time Fourier transformed to obtain the time-frequency characteristics of the signal. The short-time Fourier transform performs a Fourier transform on the signal in each window by sliding the time window, generates a spectrum diagram that changes with time, and obtains the initial time-frequency spectrum diagram. The mathematical expression of the short-time Fourier transform is:

[0048]

[0049] Among them, S(t,f) is the initial time-frequency spectrum, w(τ-t) is the time window function, f is the frequency variable, and τ is the time variable. According to the initial time-frequency spectrum, the Mel-frequency cepstral coefficients are calculated. The Mel-frequency cepstral coefficients simulate the human ear's perception characteristics of different frequencies through a set of filter banks, which are usually triangular filters on the Mel scale. The time-frequency spectrum is filtered by these filters, and then the filtered energy spectrum is logarithmically operated and discrete cosine transformed to obtain the Mel-frequency feature matrix C. Each row of the Mel-frequency feature matrix represents a frequency feature, and each column corresponds to a time window. The target time-frequency spectrum is constructed according to the Mel-frequency feature matrix, and the energy density of the target time-frequency spectrum is calculated. The energy density can be calculated by squaring the signal amplitude and then integrating it, which represents the energy distribution of the signal at different frequencies.

[0050] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0051] (1) Perform frequency peak detection on the target time-frequency spectrum to obtain a set of significant frequency points, and assign energy weights to each frequency point based on the set of significant frequency points and the energy distribution diagram to obtain a weighted frequency point set;

[0052] (2) Performing frequency correlation analysis on the weighted frequency point set to obtain a correlation coefficient matrix between the frequency points, and performing hierarchical analysis on the weighted frequency point set based on the correlation coefficient matrix to obtain a bottom-level frequency node set;

[0053] (3) Perform neighborhood analysis on the underlying frequency node set to obtain the local connection relationship of each node, and construct the underlying spectrum network based on the local connection relationship and node energy weight;

[0054] (4) Perform community detection on the underlying spectrum network to obtain a frequency pattern community set, and merge related communities into middle-layer nodes based on the frequency pattern community set to obtain a middle-layer spectrum network;

[0055] (5) Perform global topological analysis on the middle-layer spectrum network to obtain the network centrality index, and select the target middle-layer nodes to construct the top-layer spectrum network based on the network centrality index;

[0056] (6) Generate a corresponding three-layer spectrum network according to the bottom-layer spectrum network, the middle-layer spectrum network and the top-layer spectrum network.

[0057] Specifically, frequency peak detection is performed on the target time-frequency spectrum to identify significant frequency points in the heart sound signal, that is, those frequencies that appear as energy peaks on the time-frequency spectrum. By setting a threshold, frequency points whose energy exceeds the threshold are screened out to form a set of significant frequency points. These significant frequency points usually correspond to important features in the heart sound signal, such as specific frequency components of S1, S2 sounds or murmurs. After obtaining the set of significant frequency points, an energy weight is assigned to each frequency point based on the frequency point and its corresponding energy distribution diagram by calculating the energy proportion of each frequency point in the energy distribution diagram. For each frequency point f i , whose energy weight w i It can be expressed as:

[0058]

[0059] Among them, E(f i ) represents the frequency point f i The energy of j E(f j ) is the total energy of all significant frequency points. Through this formula, a weighted frequency point set is obtained, in which each frequency point includes not only its frequency value but also its weight in the overall energy distribution. Frequency correlation analysis is performed on the weighted frequency point set to calculate the correlation coefficient matrix between frequency points. Each element R in the matrix ij Reflects the frequency point f i and f jThe correlation between the points is usually achieved by calculating the mutual information or other correlation measures between these points. In the hierarchical analysis, the frequency points with higher correlation are aggregated together according to the correlation coefficient matrix to form a set of underlying frequency nodes. The aggregation process usually uses a clustering algorithm, such as hierarchical clustering, which can effectively group nodes with similar frequency characteristics. The neighborhood analysis is performed on the set of underlying frequency nodes to understand the local connection relationship of each node. By constructing an adjacency matrix, in which the matrix elements represent the connection strength or distance between nodes. Based on the adjacency matrix and the energy weights of the nodes, an underlying spectrum network is constructed. In this network, the nodes represent frequency points, and the weights of the edges reflect the similarity and correlation between the frequency points. Community detection is performed on the underlying spectrum network to identify the frequency pattern community set in it. The purpose of community detection is to discover group structures in the network, which represent clusters of frequency points with similar spectral characteristics. These communities are identified through algorithms such as the Louvain method or the Girvan-Newman algorithm, and the interconnected communities are merged into mid-level nodes to form a mid-level spectrum network. A global topological analysis is performed on the middle-layer spectral network to calculate network centrality indicators such as degree centrality, betweenness centrality, and closeness centrality. These indicators help identify key nodes in the network, that is, those nodes that have an important impact on the network structure and information propagation. Based on these centrality indicators, target middle-layer nodes are selected, and these nodes are used to construct the top-layer spectral network. Based on the structure of the bottom-layer spectral network, the middle-layer spectral network, and the top-layer spectral network, a three-layer spectral network is generated. For example, if certain frequency points are found to occupy a central position in the top-layer network during the analysis, this may mean that these frequency points are associated with important physiological or pathological features in the heart sound signal.

[0060] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0061] (1) Perform energy distribution analysis on each node in the three-layer spectrum network to obtain a node energy distribution matrix, and calculate the energy transfer coefficient between adjacent nodes in the network based on the node energy distribution matrix to obtain an energy transfer coefficient matrix;

[0062] (2) normalizing the energy transfer coefficient matrix to obtain a standardized energy-frequency propagation probability matrix, and constructing a Markov chain model based on the standardized energy-frequency propagation probability matrix to obtain an energy-frequency propagation model;

[0063] (3) Extract the amplitude-time series of each node in the three-layer spectrum network to obtain the node amplitude-time feature matrix, and calculate the amplitude attenuation coefficient and duration of each node based on the node amplitude-time feature matrix to obtain the amplitude-duration feature vector;

[0064] (4) Perform exponential fitting on the amplitude-duration characteristic vector to obtain the amplitude-duration propagation model;

[0065] (5) Perform phase spectrum analysis on each node in the three-layer spectrum network to obtain the node phase-frequency relationship matrix, and calculate the phase difference between adjacent frequency components based on the node phase-frequency relationship matrix to obtain the phase difference distribution vector;

[0066] (6) Perform statistical analysis and periodic detection on the phase difference distribution vector to obtain the frequency-phase propagation model.

[0067] Specifically, the energy information of the node is obtained. By calculating the energy of the node at different time-frequency positions, a node energy distribution matrix is ​​formed. Assume that the node set is {N i}, each node N i At different frequencies f j The energy on i j, then the node energy distribution matrix can be expressed as E = [E ij ], where i represents the node and j represents the frequency. The energy transfer coefficient between adjacent nodes in the network is calculated according to the matrix to obtain the energy transfer coefficient matrix. The element T of the energy transfer coefficient matrix i k represents the node N i To node N k The ratio of energy transferred can be calculated using the following formula:

[0068]

[0069] Among them, E ik Represents slave node N i Pass to node N k The energy of j E ij is node N iThe total energy output. This formula shows that the energy transfer coefficient is the relative proportion of energy transfer between nodes. The energy transfer coefficient matrix is ​​normalized to obtain the standardized energy-frequency propagation probability matrix, which adjusts these coefficients to a uniform range for subsequent analysis. The standardized matrix is ​​the propagation probability matrix, which represents the probability of energy transfer from one node to another in the network. Based on the standardized matrix, a Markov chain model is constructed to obtain the energy-frequency propagation model. In this model, each node represents a state, the edges between nodes represent possible paths for energy transfer, and the elements of the propagation probability matrix represent the transition probability from one state (node) to another state (node). The steady-state distribution of the Markov chain model can reveal the long-term distribution characteristics of energy in the network. The amplitude-time series is extracted for each node in the three-layer spectrum network to obtain the amplitude-time feature matrix of the node. The amplitude-time feature matrix records the amplitude changes of the node at different time points, expressed as A=[A ij ], where A ij Represents node N i At time t j The amplitude on the matrix. According to this matrix, the amplitude attenuation coefficient and duration of each node are calculated. The amplitude attenuation coefficient describes the rate at which the signal amplitude decays over time, while the duration indicates the length of time the signal maintains a certain amplitude. These two indicators can be obtained by fitting the amplitude-time data. The amplitude-duration feature vector is exponentially fitted to obtain the amplitude-duration propagation model. This model is used to describe the change law of the signal amplitude, and is usually fitted using an exponential decay model, which is in the form of:

[0070] A(t)=A0e -λt ;

[0071] Among them, A(t) represents the signal amplitude at time t, A0 is the initial amplitude, and λ is the attenuation coefficient. By fitting the actual data, the most suitable attenuation coefficient and initial amplitude are obtained, and then an accurate propagation model is constructed. Phase spectrum analysis is performed on each node in the three-layer spectrum network to obtain the node's phase-frequency relationship matrix, and the phase information of each node at different frequencies is recorded, which is expressed as Φ=[Φ ij ], where Φ ij Represents node N i At frequency f jThe phase on the frequency components. By calculating the phase difference between adjacent frequency components, the phase difference distribution vector is obtained. The phase difference represents the phase change between frequency components and is an important parameter for understanding the signal in the frequency domain. Statistical analysis and periodicity detection are performed on the phase difference distribution vector to obtain a frequency-phase propagation model and identify the coherence and synchronization characteristics between frequency components. For example, if the phase difference distribution shows obvious periodic changes, this may indicate the presence of periodic physiological or pathological characteristics in the heart sound signal.

[0072] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0073] (1) The steady-state distribution of the energy-frequency propagation model is calculated to obtain the energy propagation balance vector. Based on the energy propagation balance vector, the energy centrality of each node is calculated to obtain the node energy importance ranking.

[0074] (2) The amplitude-duration propagation model is decomposed into time series to obtain trend terms, period terms, and random terms. Based on the trend terms, period terms, and random terms, the amplitude persistence index and fluctuation coefficient are calculated to obtain the amplitude stability index.

[0075] (3) Perform phase synchronization analysis on the frequency-phase propagation model to obtain the phase coupling strength matrix between frequency pairs, and calculate the global phase synchronization index based on the phase coupling strength matrix to obtain the frequency synergy index;

[0076] (4) Perform multi-criteria decision analysis on node energy importance ranking, amplitude stability index, and frequency coordination index to obtain a comprehensive scoring vector;

[0077] (5) According to the comprehensive score vector, the network nodes are ranked by importance to obtain the target frequency node set;

[0078] (6) Performing network topology analysis on the target frequency node set and calculating the node degree centrality, betweenness centrality, and closeness centrality to obtain a topological feature matrix;

[0079] (7) Based on the topological feature matrix and the comprehensive scoring vector, a multidimensional propagation feature space is constructed and the corresponding propagation feature evaluation index is calculated.

[0080] Specifically, the steady-state distribution vector of the Markov chain is calculated. Assume that the transition probability matrix in the energy-frequency propagation model is P, which describes the probability of energy transfer from one frequency node to another. The steady-state distribution vector π satisfies πP=π, and ∑ i π i =1, where π iRepresents the energy proportion of frequency node i in steady state. These steady-state probabilities reflect the long-term distribution of energy on each frequency node. The steady-state distribution vector can be calculated by iteration or by solving the eigenvector. The energy centrality of each node is calculated according to π. The energy centrality is defined as the value of the steady-state distribution vector, that is, the relative importance of the energy distribution of each node in the entire network. This indicator is used to rank the importance of nodes, thereby identifying which nodes play a key role in the energy propagation of the network. Time series decomposition of the amplitude-duration propagation model. The amplitude-time series data of each node is decomposed into three parts: trend term, periodic term and random term. The trend term represents the long-term trend of the data, the periodic term represents the periodic fluctuation, and the random term captures the random fluctuation in the data. Time series decomposition can be implemented by using an additive model or a multiplicative model. The additive model is expressed as Y(t)=T(t)+C(t)+R(t), where Y(t) is the original data, T(t) is the trend term, C(t) is the periodic term, and R(t) is the random term. Based on these decomposition terms, the amplitude persistence index and the fluctuation coefficient are calculated. The amplitude persistence index can be expressed as the ratio of the standard deviation of the trend term to the total amplitude, while the fluctuation coefficient can be expressed as the ratio of the standard deviation of the random term to the total amplitude. The amplitude persistence index measures the stability of the signal over a long time scale, while the fluctuation coefficient measures the short-term fluctuation characteristics of the signal. Together, they constitute the amplitude stability index. Phase synchronization analysis is performed on the frequency-phase propagation model. A phase coupling strength matrix is ​​constructed, in which each element represents the degree of phase synchronization between two frequency components. Phase information is extracted using methods such as Hilbert transform or wavelet transform, and then the phase difference and phase synchronization between the frequency components are calculated. The global phase synchronization index is a summary of the phase coupling of the entire network, which can usually be expressed by the mean or maximum value of the phase coupling strength matrix. The global phase synchronization index provides a quantitative indicator to measure the frequency coordination of the entire signal and reveals the degree of synchronization of each frequency component in the signal. Multi-criteria decision analysis is performed on the node energy importance ranking, amplitude stability index, and frequency coordination index. A weighted comprehensive scoring method is used, in which the calculation formula of the comprehensive scoring vector is:

[0081] S i =w1·E i +w2·F i +w3·P i ;

[0082] Among them, S i is the comprehensive score of node i, E i 、F i , P iare the energy centrality, amplitude stability and frequency coordination index of the node, and w1, w2 and w3 are the corresponding weights. By adjusting the weights, the importance of different indicators is highlighted. After obtaining the comprehensive scoring vector, the importance of the network nodes is sorted according to the comprehensive score, and the target frequency node set is determined. These nodes are the most representative nodes after comprehensive evaluation of multiple indicators and have high analysis value. The network topology analysis of the target frequency node set is performed to calculate the degree centrality, betweenness centrality and closeness centrality of the nodes. Degree centrality D i Refers to the number of connections of a node, that is, the number of nodes it is directly connected to. i It measures the frequency of a node being an intermediate node in the shortest path in the network, reflecting the importance of the node to the network information flow. i It is the reciprocal of the average shortest path length from a node to all other nodes, reflecting the efficiency of information dissemination of nodes in the network. These indicators form the topological feature matrix T =

[0083] [D i ,B i ,C i ], which helps analyze the characteristics of network structure. According to the topological feature matrix and the comprehensive scoring vector, a multi-dimensional propagation feature space is constructed, and the corresponding propagation feature evaluation index is calculated.

[0084] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0085] (1) According to the propagation characteristic evaluation index, the frequency points in the heart sound signal data are sorted by importance to obtain a frequency point importance sequence;

[0086] (2) The frequency point importance sequence is truncated and the top N frequency points are selected to obtain a candidate set of frequency observation points. The candidate set of frequency observation points is subjected to time-frequency correlation analysis to obtain a correlation coefficient matrix between the observation points.

[0087] (3) Based on the correlation coefficient matrix, a similarity graph of the observation points is constructed to obtain a weighted undirected graph structure, and spectral clustering is performed on the weighted undirected graph structure to obtain an initial clustering result;

[0088] (4) According to the initial clustering results, the local density and relative distance of each observation point are calculated to obtain a density-distance distribution map, and an adaptive threshold analysis is performed on the density-distance distribution map to determine the density peak point and obtain the first cluster center set;

[0089] (5) According to the first cluster center set, the remaining observation points are assigned to the nearest neighbors to obtain an optimized clustering result, and a representative observation point is selected for each cluster in the optimized clustering result to obtain a first heart sound feature observation point subset.

[0090] Specifically, the importance of each frequency point is evaluated according to the propagation characteristic evaluation index. These indicators include energy centrality, amplitude stability, and frequency synergy, which together reflect the importance of each frequency point in the heart sound signal analysis. These indicators are combined into a comprehensive score vector, expressed as:

[0091] S i =αE i +βF i +γP i ;

[0092] Among them, S i is the comprehensive score of frequency point i, E i 、F i , P iare the energy centrality, amplitude stability and frequency synergy index of the frequency point, and α, β and γ are the corresponding weight coefficients. By adjusting the weights, the influence of certain features can be highlighted according to different analysis objectives. After obtaining the comprehensive score, all frequency points are sorted to form a frequency point importance sequence. The frequency point importance sequence is truncated, and the top N frequency points are selected to form a candidate set of frequency observation points. The size of N can be determined according to the actual analysis needs and the amount of data. Selecting a suitable N value can ensure that enough frequency points are included for analysis while avoiding information overload. The candidate set is subjected to time-frequency correlation analysis, and the correlation coefficient matrix between observation points is obtained by calculating the correlation between each pair of frequency points. According to the correlation coefficient matrix of the observation points, a similarity graph is constructed. This graph can be represented as a weighted undirected graph structure, in which the nodes represent frequency observation points and the edge weights are the correlation coefficients between frequency points. In this graph, higher edge weights indicate stronger similarity. This graph structure helps to understand the relationship between frequency points. The weighted undirected graph is clustered using the spectral clustering method to obtain the initial clustering results. Spectral clustering uses the Laplacian matrix of the graph to discover the natural clustering structure of the data. Its principle is to minimize the cutting cost between clusters while maximizing the similarity within the cluster. Based on the initial clustering results, the local density and relative distance of each observation point are calculated to generate a density-distance distribution map. Local density refers to how many neighboring points there are around an observation point, while relative distance indicates the distance between a point and its higher density point. These two parameters help identify the core points and edge points of different clusters in the data set. Through the density-distance distribution map, adaptive threshold analysis is performed to determine the density peak point. The density peak point represents the core point or center point of the cluster, which is the point with the highest density in the cluster and farthest from other density centers. After determining these points, the first cluster center set is obtained. According to the first cluster center set, the nearest neighbor assignment is performed on the remaining observation points, and each unclassified observation point is assigned to its nearest cluster center to ensure that each observation point is reasonably assigned to the most suitable cluster. After obtaining the optimized clustering results, the observation points in each cluster are screened to select the most representative observation points. Representative observation points are usually those points that best reflect clustering characteristics and are of great significance in analysis and interpretation. These representative observation points constitute the first heart sound feature observation point subset.

[0093] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0094] (1) extracting features from the first heart sound feature observation point subset to obtain a multidimensional feature vector for each observation point, and initializing the position and velocity of each particle in the particle swarm based on the multidimensional feature vector to obtain an initial particle swarm;

[0095] (2) Perform fitness evaluation on each particle in the initial particle swarm to obtain an initial fitness value set, and select the global optimal particle and the individual optimal particle based on the initial fitness value set to obtain the initial optimal solution;

[0096] (3) The velocity and position of the initial particle swarm are updated to obtain a new generation of particle swarm, and K-means clustering is performed on the new generation of particle swarm to obtain the second cluster center set;

[0097] (4) According to the second cluster center set, the distance ratio within the cluster and between the clusters is calculated to obtain the cluster evaluation index, and the cluster evaluation index is compared to update the global optimal solution and the individual optimal solution to obtain the updated optimal solution;

[0098] (5) Iteratively optimize the updated optimal solution until the convergence condition is reached to obtain the target optimization result, and select the point closest to the cluster center from the first heart sound feature observation point subset according to the target optimization result to obtain the second heart sound feature observation point subset.

[0099] Specifically, a multidimensional feature vector is extracted from each observation point. These feature vectors may include features such as frequency, energy, and phase. For example, suppose there is a heart sound observation point with a frequency of f i , the corresponding energy is E i , the phase is θ i , then its eigenvector can be expressed as v i =[f i ,E i ,θ i ]. By performing similar feature extraction on all observation points, a set of multidimensional feature vectors is obtained. Based on the multidimensional feature vectors, the position and velocity of each particle in the particle swarm are initialized. Particle swarm optimization is an optimization algorithm based on swarm intelligence, in which each particle represents a possible solution. The position of a particle represents the feature vector of a solution, and the velocity represents the rate of change of the solution. During the initialization process, the position and velocity of the particle can usually be randomly generated or set based on specific rules. Suppose there is a particle whose position vector is x i =[x i1 ,x i2 ,x i3 ], the velocity vector is v i =[v i1 ,v i2 ,v i3], then the initial particle group is a collection of multiple such particles. Perform fitness evaluation on each particle in the initial particle group and calculate the fitness value of each particle. The fitness function is usually defined according to the specific requirements of the problem, which can be to minimize a certain loss function or maximize a certain evaluation index. For example, if the goal is to maximize the compactness and separation of clusters, the ratio of the internal distance of the cluster to the distance between clusters can be selected as the fitness function. For each particle, calculate its fitness value f i =f(x i ), then select the particle with the highest fitness value as the global optimal particle, and retain the best position reached by each particle in its historical trajectory as the individual optimal particle to obtain the initial optimal solution. Update the speed and position of the initial particle group. The update formula is:

[0100]

[0101] Among them, v id and x id are the velocity and position of particle i in dimension d, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0, 1], and p id is the individual optimal position of particle i, g d is the global optimal position. Through the updating process, the position and speed of the particles are constantly adjusted and gradually tend to the global optimal solution. The updated particle swarm forms a new generation of particle swarm, which is subjected to K-means clustering to obtain the second cluster center set. K-means clustering is a commonly used clustering algorithm that divides data points into different clusters by minimizing the variance within the cluster. Each particle is assigned to its closest cluster center, and the updated cluster center set reflects the clustering situation under the current optimal solution. In order to evaluate the clustering effect, the ratio of the distance within the cluster and between clusters is calculated to obtain the clustering evaluation index. The internal distance of the cluster refers to the average distance between points in the same cluster, while the distance between clusters is the average distance between different cluster centers. The clustering evaluation index usually selects the ratio of maximizing the distance between clusters to minimizing the internal distance of the cluster, which helps to ensure the compactness and separation of the clusters. By comparing these indicators, the global optimal solution and the individual optimal solution are updated to obtain the updated optimal solution. This process is iterated continuously until the convergence condition is reached. The convergence condition may be that the global optimal solution does not change significantly within several generations, or reaches a predetermined number of iterations. The target optimization result is obtained. According to the target optimization result, the point closest to the cluster center is selected from the first heart sound feature observation point subset to form a second heart sound feature observation point subset.

[0102] The above describes the heart murmur frequency observation point screening and selection method in the embodiment of the present application. The following describes the heart murmur frequency observation point screening and selection system in the embodiment of the present application. Figure 2 In one embodiment of the present application, a heart murmur frequency observation point screening and selection system includes:

[0103] The preprocessing module 201 is used to preprocess and perform time-frequency analysis on the heart sound signal data to obtain a target time-frequency spectrum and a corresponding energy distribution diagram;

[0104] A construction module 202 is used to construct a multi-layer spectrum network for the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer;

[0105] The analysis module 203 is used to perform propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model and a frequency-phase propagation model;

[0106] A calculation module 204 is used to calculate the network propagation characteristics of the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain a propagation characteristic evaluation index;

[0107] A screening module 205 is used to screen the corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and perform adaptive density peak clustering on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset;

[0108] The optimization module 206 is used to optimize the first heart sound feature observation point subset by using the PSO-Means algorithm to obtain the second heart sound feature observation point subset.

[0109] Through the collaboration of the above components, a three-layer spectrum network including the bottom layer, middle layer and top layer is constructed to fully capture the multi-scale frequency characteristics of the heart sound signal, which helps to more accurately characterize the complex structure of the murmur. Three propagation models, namely energy-frequency, amplitude-duration and frequency-phase, are used to deeply analyze the interaction between frequency components and reveal the dynamic propagation characteristics of the murmur. By calculating indicators such as node energy importance, amplitude stability and frequency synergy, the importance of frequency observation points is comprehensively evaluated to improve the accuracy of screening. Combining spectral clustering and density peak detection, adaptive clustering of frequency observation points is achieved, overcoming the limitations of the fixed threshold method. The particle swarm optimization algorithm is introduced in combination with K-means clustering to further optimize the selection of observation points and improve the stability and representativeness of the screening results. Taking into account multiple evaluation indicators, a comprehensive score is obtained through a multi-criteria decision-making method, which ensures the comprehensiveness and objectivity of the screening process. Time-frequency analysis is incorporated into the signal preprocessing stage to extract Mel frequency features, which enhances the characterization ability of murmur features. From time domain features to frequency domain features, and then to network topology features, signal features are gradually extracted in depth, thereby improving the accuracy of screening and selection of heart murmur frequency observation points.

[0110] The present application also provides a device for screening and selecting heart murmur frequency observation points, the device comprising a memory and a processor, the memory storing computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the heart murmur frequency observation point screening and selection method in the above-mentioned embodiments.

[0111] 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, wherein instructions are stored in the computer, and when the instructions are executed on a computer, the computer executes the steps of the heart murmur frequency observation point screening and selection method.

[0112] 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.

[0113] 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.

[0114] 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 method for selecting heart murmur frequency observation points, characterized in that: The method comprises: Preprocess and perform time-frequency analysis on the heart sound signal data to obtain the target time-frequency spectrum and the corresponding energy distribution diagram; Performing a multi-layer spectrum network construction on the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer; Performing a propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model, and a frequency-phase propagation model; Performing network propagation characteristic calculation on the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain propagation characteristic evaluation indicators; According to the propagation characteristic evaluation index, a corresponding frequency observation point candidate set is selected from the heart sound signal data, and adaptive density peak clustering is performed on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset; The first heart sound feature observation point subset is optimized by using the PSO-Means algorithm to obtain a second heart sound feature observation point subset.

2. The heart murmur frequency observation point screening and selection method according to claim 1, characterized in that: The preprocessing and time-frequency analysis of the heart sound signal data to obtain a target time-frequency spectrum diagram and a corresponding energy distribution diagram includes: Performing bandpass filtering on the heart sound signal data to obtain an initially filtered heart sound signal, and performing wavelet threshold denoising on the initially filtered heart sound signal to obtain a denoised heart sound signal; Performing heart sound cycle segmentation on the denoised heart sound signal to obtain a plurality of heart sound cycle segments, and performing time domain feature extraction on the plurality of heart sound cycle segments to obtain a time domain feature set; Normalizing the multiple heart sound cycle segments according to the time domain feature set to obtain standardized heart sound cycle segments, and performing short-time Fourier transform on the standardized heart sound cycle segments to obtain an initial time-frequency spectrum; Calculating the Mel-frequency cepstrum coefficients according to the initial time-frequency spectrogram to obtain a Mel-frequency feature matrix, and constructing a target time-frequency spectrogram according to the Mel-frequency feature matrix; The energy density of the target time-frequency spectrum is calculated to obtain a corresponding energy distribution diagram.

3. The heart murmur frequency observation point screening and selection method according to claim 1, characterized in that: The multi-layer spectrum network is constructed for the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer, including: Performing frequency peak detection on the target time-frequency spectrum to obtain a significant frequency point set, and assigning an energy weight to each frequency point according to the significant frequency point set and the energy distribution diagram to obtain a weighted frequency point set; Performing frequency correlation analysis on the weighted frequency point set to obtain a correlation coefficient matrix between frequency points, and performing hierarchical analysis on the weighted frequency point set according to the correlation coefficient matrix to obtain a bottom-level frequency node set; Performing neighborhood analysis on the underlying frequency node set to obtain a local connection relationship of each node, and constructing an underlying spectrum network according to the local connection relationship and node energy weight; Performing community detection on the underlying spectrum network to obtain a frequency mode community set, and merging related communities into middle-layer nodes according to the frequency mode community set to obtain a middle-layer spectrum network; Performing a global topological analysis on the middle-layer spectrum network to obtain a network centrality index, and selecting target middle-layer nodes to construct a top-layer spectrum network based on the network centrality index; A corresponding three-layer spectrum network is generated according to the bottom-layer spectrum network, the middle-layer spectrum network and the top-layer spectrum network.

4. The heart murmur frequency observation point screening and selection method according to claim 3, characterized in that: The propagation model analysis of the three-layer spectrum network is performed to obtain an energy-frequency propagation model, an amplitude-duration propagation model and a frequency-phase propagation model, including: Performing energy distribution analysis on each node in the three-layer spectrum network to obtain a node energy distribution matrix, and calculating energy transfer coefficients between adjacent nodes in the network based on the node energy distribution matrix to obtain an energy transfer coefficient matrix; Normalizing the energy transfer coefficient matrix to obtain a standardized energy-frequency propagation probability matrix, and constructing a Markov chain model based on the standardized energy-frequency propagation probability matrix to obtain an energy-frequency propagation model; Extracting amplitude-time series from each node in the three-layer spectrum network to obtain a node amplitude-time feature matrix, and calculating amplitude attenuation coefficient and duration for each node based on the node amplitude-time feature matrix to obtain an amplitude-duration feature vector; Performing exponential fitting on the amplitude-duration characteristic vector to obtain an amplitude-duration propagation model; Performing phase spectrum analysis on each node in the three-layer spectrum network to obtain a node phase-frequency relationship matrix, and calculating the phase difference between adjacent frequency components according to the node phase-frequency relationship matrix to obtain a phase difference distribution vector; Statistical analysis and periodic detection are performed on the phase difference distribution vector to obtain a frequency-phase propagation model.

5. The heart murmur frequency observation point screening and selection method according to claim 4, characterized in that: The network propagation characteristic calculation is performed on the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain a propagation characteristic evaluation index, including: Performing steady-state distribution calculation on the energy-frequency propagation model to obtain an energy propagation balance vector, and calculating the energy centrality of each node based on the energy propagation balance vector to obtain a node energy importance ranking; Decomposing the amplitude-duration propagation model in time series to obtain trend terms, periodic terms and random terms, and calculating the amplitude persistence index and the fluctuation coefficient based on the trend terms, the periodic terms and the random terms to obtain an amplitude stability index; Performing phase synchronization analysis on the frequency-phase propagation model to obtain a phase coupling strength matrix between frequency pairs, and calculating a global phase synchronization index based on the phase coupling strength matrix to obtain a frequency synergy index; Performing a multi-criteria decision analysis on the node energy importance ranking, the amplitude stability index, and the frequency coordination index to obtain a comprehensive scoring vector; According to the comprehensive scoring vector, the network nodes are sorted by importance to obtain a target frequency node set; Performing network topology analysis on the target frequency node set and calculating node degree centrality, betweenness centrality and closeness centrality to obtain a topological feature matrix; According to the topological feature matrix and the comprehensive scoring vector, a multi-dimensional propagation feature space is constructed and the corresponding propagation feature evaluation index is calculated.

6. The heart murmur frequency observation point screening and selection method according to claim 1, characterized in that: The step of selecting a corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and performing adaptive density peak clustering on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset includes: According to the propagation characteristic evaluation index, the frequency points in the heart sound signal data are sorted by importance to obtain a frequency point importance sequence; The frequency point importance sequence is truncated, and the top N frequency points are selected to obtain a candidate set of frequency observation points, and a time-frequency correlation analysis is performed on the candidate set of frequency observation points to obtain a correlation coefficient matrix between observation points; According to the correlation coefficient matrix, a similarity graph of the observation points is constructed to obtain a weighted undirected graph structure, and spectral clustering is performed on the weighted undirected graph structure to obtain an initial clustering result; According to the initial clustering result, the local density and relative distance of each observation point are calculated to obtain a density-distance distribution map, and an adaptive threshold analysis is performed on the density-distance distribution map to determine the density peak point to obtain a first cluster center set; According to the first cluster center set, nearest neighbor assignment is performed on the remaining observation points to obtain an optimized clustering result, and representative observation points are selected for each cluster in the optimized clustering result to obtain a first heart sound feature observation point subset.

7. The heart murmur frequency observation point screening and selection method according to claim 6, characterized in that: The optimizing process of the first heart sound feature observation point subset by the PSO-Means algorithm to obtain the second heart sound feature observation point subset includes: Performing feature extraction on the first heart sound feature observation point subset to obtain a multidimensional feature vector of each observation point, and initializing the position and velocity of each particle in the particle swarm according to the multidimensional feature vector to obtain an initial particle swarm; Performing fitness evaluation on each particle in the initial particle group to obtain an initial fitness value set, and selecting a global optimal particle and an individual optimal particle according to the initial fitness value set to obtain an initial optimal solution; updating the velocity and position of the initial particle swarm to obtain a new generation of particle swarm, and performing K-means clustering on the new generation of particle swarm to obtain a second cluster center set; According to the second cluster center set, the distance ratio within the cluster and between the clusters is calculated to obtain a cluster evaluation index, and the cluster evaluation index is compared to update the global optimal solution and the individual optimal solution to obtain an updated optimal solution; The updated optimal solution is iteratively optimized until a convergence condition is reached to obtain a target optimization result, and based on the target optimization result, a point closest to the cluster center is selected from the first heart sound feature observation point subset to obtain a second heart sound feature observation point subset.

8. A heart murmur frequency observation point screening and selection system, characterized in that: The system for executing the heart murmur frequency observation point screening and selection method according to any one of claims 1 to 7 comprises: A preprocessing module is used to preprocess and perform time-frequency analysis on the heart sound signal data to obtain a target time-frequency spectrum and a corresponding energy distribution diagram; A construction module, used for constructing a multi-layer spectrum network for the target time-frequency spectrum diagram and the energy distribution diagram to obtain a three-layer spectrum network including a bottom layer, a middle layer and a top layer; An analysis module, used to perform a propagation model analysis on the three-layer spectrum network to obtain an energy-frequency propagation model, an amplitude-duration propagation model and a frequency-phase propagation model; A calculation module, used to calculate network propagation characteristics of the energy-frequency propagation model, the amplitude-duration propagation model and the frequency-phase propagation model to obtain propagation characteristic evaluation indicators; A screening module, used for screening a corresponding frequency observation point candidate set from the heart sound signal data according to the propagation characteristic evaluation index, and performing adaptive density peak clustering on the frequency observation point candidate set to obtain a first heart sound characteristic observation point subset; The optimization module is used to optimize the first heart sound feature observation point subset by using the PSO-Means algorithm to obtain the second heart sound feature observation point subset.

9. A device for selecting and screening heart murmur frequency observation points, characterized in that: The heart murmur frequency observation point screening and selection device comprises: 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 murmur frequency observation point screening and selection device to perform the heart murmur frequency observation point screening and selection method according to any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the heart murmur frequency observation point screening and selection method according to any one of claims 1 to 7 is implemented.

Images