Heart sound signal classification method and device

Abstract

The invention provides a heart sound signal classification method and device, and belongs to the technical field of heart sound signal processing, and the method comprises the steps: segmenting a to-be-recognized heart sound signal based on a cardiac cycle, and obtaining heart sound signal segments; determining feature information of each heart sound signal segment based on at least one feature extraction algorithm; performing feature fusion and feature selection on the feature information of the heart sound signal fragments, and inputting obtained target feature information into a classification model to obtain a heart sound signal classification result output by the classification model; wherein the classification model is obtained by training a machine learning model through feature information of the heart sound signal fragment samples. The to-be-recognized heart sound signals are segmented based on the cardiac cycle, richer feature information of the to-be-recognized heart sound signals can be obtained, the target feature information is obtained based on feature selection, redundant feature information can be removed, and the accuracy of heart sound signal classification results is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of heart sound signal processing, and in particular to a heart sound signal classification method and device. Background Art

[0002] Cardiovascular disease (CVD) is one of the leading causes of disability and death. Early screening and timely intervention are key to reducing CVD risk. Heart sound signals contain a wealth of physiological and pathological information about the human heart's functional state, closely related to valvular abnormalities, blood flow, and other conditions. Heart sound recognition based on phonocardiogram analysis and processing provides an important medical basis for the early diagnosis and treatment of cardiovascular disease. The richness of heart sound signal features can improve heart sound recognition accuracy. Therefore, it is necessary to propose a heart sound signal classification scheme that enriches the feature information of heart sound signals. Summary of the Invention

[0003] The present invention provides a heart sound signal classification method and device, which enrich the characteristic information of the heart sound signal so as to improve the accuracy of classification and recognition of the heart sound signal.

[0004] The present invention provides a heart sound signal classification method, comprising: Segment the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments; determining feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle; performing feature fusion and feature selection on the feature information of each of the heart sound signal segments, inputting the obtained target feature information into a classification model, and obtaining a heart sound signal classification result output by the classification model; The classification model is obtained by training a machine learning model based on feature information of heart sound signal segment samples.

[0005] As an embodiment, determining the feature information of each of the heart sound signal segments based on at least one feature extraction algorithm includes: Based on the cardiac cycle, at least one feature extraction algorithm among a Mel-frequency cepstral coefficient MFCC feature extraction algorithm, a wavelet transform feature extraction algorithm, a Fourier transform feature extraction algorithm, and a power spectral density feature extraction algorithm is improved to obtain an improved feature extraction algorithm; Feature information of the heart sound signal segment is extracted based on an improved feature extraction algorithm.

[0006] As an embodiment, the Mel-frequency cepstral coefficient (MFCC) feature extraction algorithm is improved based on the cardiac cycle, including: Determining, based on the time frames of each state of the cardiac cycle, the statistical value of the MFCC features, the statistical value of the MFCC first-order difference, and the statistical value of the MFCC second-order difference of each of the heart sound signal segments; Based on the statistical values ​​of the MFCC features corresponding to each state of the cardiac cycle, the statistical values ​​of the MFCC first-order differences, the statistical values ​​of the MFCC second-order differences, and the duration, output feature information of the improved MFCC feature extraction algorithm is determined.

[0007] As an embodiment, the MFCC first-order difference is determined based on a center-weighted difference method.

[0008] As an embodiment, the wavelet transform feature extraction algorithm is improved based on the cardiac cycle, including: determining a first event window of an improved wavelet transform feature extraction algorithm based on the duration of each state of the cardiac cycle so as to extract wavelet feature information of the heart sound signal segment based on the first event window; Based on the wavelet feature information and duration corresponding to each state of the cardiac cycle, output feature information of the improved wavelet transform feature extraction algorithm is determined.

[0009] As an embodiment, the Fourier transform feature extraction algorithm is improved based on the cardiac cycle, including: determining, based on the durations of the various states of the cardiac cycle, a second event window of an improved Fourier transform feature extraction algorithm so as to extract Fourier feature information of the heart sound signal segment based on the second event window; Based on the Fourier feature information and duration corresponding to each state of the cardiac cycle, output feature information of the improved Fourier transform feature extraction algorithm is determined.

[0010] As an embodiment, the power spectrum density feature extraction algorithm is improved based on the cardiac cycle, including: determining Fourier feature information of the heart sound signal segment based on the duration of each state of the cardiac cycle; determining power spectral density characteristic information of each of the heart sound signal segments based on the Fourier characteristic information of each of the heart sound signal segments; Based on the power spectrum density feature information and duration corresponding to different states of the cardiac cycle, output feature information of the improved power spectrum density feature extraction algorithm is determined.

[0011] As an embodiment, the feature fusion and feature selection of the feature information of each of the heart sound signal segments includes: Based on the feature information of all the heart sound signal segments, a feature matrix is ​​constructed to complete feature fusion; The feature information of the feature matrix is ​​sorted by importance, and target feature information is selected based on the sorting result to complete feature selection.

[0012] As an embodiment, the training step of the classification model includes: Segmenting the heart sound signal samples based on the cardiac cycle to obtain heart sound signal segment samples; determining feature information of each of the heart sound signal segment samples based on at least one feature extraction algorithm; Based on an improved five-fold cross-validation algorithm, the characteristic information of all heart sound signal fragment samples is divided into a training set and a test set, the training set is sample-balanced based on a preset sample weight, and the target sample characteristic information is screened from the sample-balanced training set; The machine learning model is trained based on the target sample feature information, and the trained machine learning model is tested based on the test set to obtain the classification model.

[0013] The present invention also provides a heart sound signal classification device, comprising: A signal segmentation module is used to segment the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments; a feature extraction module, configured to determine feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle; a classification module, configured to perform feature fusion and feature selection on the feature information of each of the heart sound signal segments, input the obtained target feature information into a classification model, and obtain a heart sound signal classification result output by the classification model; The classification model is obtained by training a machine learning model based on the feature information of the segmented heart sound signal samples.

[0014] The heart sound signal classification method and device provided by the present invention segment the heart sound signal to be identified based on the cardiac cycle, and extract features based on at least one feature extraction algorithm. The fused features can display rich feature information of the heart sound signal to be identified. Feature selection is conducive to removing redundant feature information and improving the accuracy of the heart sound signal classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 This is one of the flow charts of the heart sound signal classification method provided by the present invention.

[0017] Figure 2 This is the second flow chart of the heart sound signal classification method provided by the present invention.

[0018] Figure 3 It is a schematic diagram of the training process of the classification model provided by the present invention.

[0019] Figure 4 a- Figure 4 d is a schematic diagram of the feature importance ranking of the SVM, random forest, decision tree and XGBoost models provided by the present invention.

[0020] Figure 5 It is a schematic diagram of target sample feature information provided by the present invention.

[0021] Figure 6 a- Figure 6 d is a schematic diagram of the confusion matrix of the SVM, random forest, decision tree and XGBoost models provided by the present invention.

[0022] Figure 7 a- Figure 7 d is a schematic diagram of the ROC curves of the SVM, random forest, decision tree and XGBoost models provided by the present invention.

[0023] Figure 8 It is a structural schematic diagram of the heart sound signal classification device provided by the present invention. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] Figure 1 This is one of the flow charts of the heart sound signal classification method provided by the present invention, such as Figure 1 As shown, the present invention provides a heart sound signal classification method, which includes the following steps.

[0026] Step S100 : segmenting the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments.

[0027] Each heartbeat constitutes a complete cardiac cycle. A cardiac cycle refers to the entire process from the start of one heartbeat to the next, and is primarily divided into two phases: the duration of the heart's contraction and relaxation, known as systole and diastole, respectively. Healthy heart sound signals are generated by the vibrations of the heart valves opening and closing during the cardiac cycle, typically including the first heart sound (S1) and the second heart sound (S2). Accurately locating the timing of S1 and S2 is the primary purpose of segmenting heart sound signals. A phonocardiogram (PCG) characterizes the timing of the mechanical vibrations generated by cardiac activity. A heart murmur, on the other hand, is an abnormal sound produced during the heart's contraction or relaxation, caused by abnormal blood flow near the heart due to valve malfunction.

[0028] Normal heart sound signals are primarily generated by the vibrations of the heart valves as they open and close during each cardiac cycle. A healthy PCG signal consists of two sounds: the first heart sound (S1) and the second heart sound (S2). The first heart sound occurs when the mitral and tricuspid valves close, at the beginning of systole; the second heart sound is produced when the aortic and pulmonary valves close at the beginning of diastole. A normal cardiac cycle consists of four phases: the first heart sound (S1), systole, the second heart sound (S2), and diastole.

[0029] A heart murmur is an abnormal sound produced by the heart during contraction or relaxation. It can be caused by a malfunctioning valve, leading to abnormal blood flow near the heart. By analyzing the characteristics of a heart murmur, doctors can more accurately determine the type and severity of heart disease.

[0030] The present invention utilizes a segmentation algorithm to segment the cardiac sound signal to be identified into signals containing independent cardiac cycles. Each cardiac cycle includes four states: the first cardiac sound S1, systole, the second cardiac sound S2, and diastole. The four states can be represented as S1, sys, S2, and dias. The segmentation algorithm is set according to the contents of the segmentation file to obtain the four states after segmentation: 1 corresponds to S1, 2 corresponds to systole, 3 corresponds to S2, and 4 corresponds to diastole. Preferably, during segmentation, incomplete cardiac cycles and unlabeled partial data are removed, and each time period of each complete cardiac cycle is normalized, that is, the time point of each stage is converted to a relative time position relative to the start time of S1, so that the time characteristics of different heartbeat segments can be compared and analyzed. To reduce the loss of cardiac sound data, the cardiac sound signal to be identified is not denoised.

[0031] Step S200 : determining feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle.

[0032] The feature extraction algorithms include time domain feature extraction algorithms, frequency domain feature extraction algorithms and time-frequency domain feature extraction algorithms. The present invention uses multiple feature extraction algorithms to perform feature extraction on heart sound signal segments respectively, and can effectively extract rich feature information.

[0033] Step S300 , performing feature fusion and feature selection on the feature information of each of the heart sound signal segments, inputting the obtained target feature information into a classification model, and obtaining a heart sound signal classification result output by the classification model.

[0034] The classification model is obtained by training a machine learning model based on feature information of heart sound signal segment samples.

[0035] The feature information of each of the heart sound signal segments is fused, and the fused features can better display the rich feature information of the heart sound signal. Feature selection is then performed to remove redundant features, thereby improving the importance and accuracy of the target feature information, thereby improving the accuracy of the heart sound signal classification results, assisting doctors in diagnosis, improving diagnostic efficiency and accuracy, and reducing medical costs.

[0036] like Figure 2 As shown, based on the above embodiment, as an optional embodiment, the step of determining the feature information of each of the heart sound signal segments based on at least one feature extraction algorithm includes step S210 and step S220.

[0037] Step S210 , based on the cardiac cycle, at least one of the feature extraction algorithms of the Mel-frequency cepstral coefficient MFCC feature extraction algorithm, the wavelet transform feature extraction algorithm, the Fourier transform feature extraction algorithm, and the power spectral density feature extraction algorithm is improved to obtain an improved feature extraction algorithm.

[0038] Step S220 : extracting feature information of the heart sound signal segment based on an improved feature extraction algorithm.

[0039] Based on the cardiac cycle, at least one feature extraction algorithm among the Mel-frequency cepstral coefficient MFCC feature extraction algorithm, the wavelet transform feature extraction algorithm, the Fourier transform feature extraction algorithm and the power spectral density feature extraction algorithm is improved respectively. Specifically, the feature extraction algorithm is improved based on different states of the cardiac cycle to obtain a staged feature extraction algorithm, and the duration of different states of the cardiac cycle is fused with the extracted feature information to construct a multidimensional feature information space and improve the robustness of the classification model.

[0040] It can be understood that the present invention improves the feature extraction algorithm based on the cardiac cycle to obtain a staged feature extraction algorithm, which can characterize the differences in time-frequency characteristics of different cardiac states (such as systole and diastole), improve the discrimination of pathological features, complement the duration of different states of the cardiac cycle with feature information, construct a multi-dimensional feature information space, and improve the robustness of the classification model.

[0041] Based on the above embodiment, as an optional embodiment, the Mel-frequency cepstral coefficient MFCC feature extraction algorithm is improved based on the cardiac cycle, including steps S2111 and S2112.

[0042] Step S2111 : determining the statistical value of the MFCC features, the statistical value of the MFCC first-order difference, and the statistical value of the MFCC second-order difference of each of the heart sound signal segments based on the time frames of each state of the cardiac cycle.

[0043] Optionally, the MFCC first-order difference is determined based on a center-weighted difference method.

[0044] Step S2112: Determine output feature information of an improved MFCC feature extraction algorithm based on the statistical values ​​of the MFCC features, the statistical values ​​of the MFCC first-order differences, the statistical values ​​of the MFCC second-order differences, and the duration corresponding to each state of the cardiac cycle.

[0045] Mel-scale Frequency Cepstral Coefficients (MFCC) are cepstral parameters extracted from audio signals in the Mel-scale frequency domain and are mainly used in the field of audio recognition. f There are different perception abilities. In order to describe this characteristic of the human ear, the Mel frequency is proposed, which is related to the frequency f Into a nonlinear relationship.

[0046] The expression of Mel cepstral coefficient is as follows: in, Represents the Mel-frequency cepstral coefficient.

[0047] In an embodiment of the present invention, the Mel-frequency cepstral coefficient feature extraction algorithm is improved based on the cardiac cycle to obtain a staged dynamic MFCC feature extraction algorithm, which extracts MFCC features and statistical features for the four states of each complete cardiac cycle (assuming that the frame length of the improved MFCC feature extraction algorithm is 20ms and the frame shift is 10ms).

[0048] The steps for extracting MFCC features are as follows: Perform discrete Fourier transform on the heart sound signal segment to obtain the signal amplitude X[K] of the heart sound signal segment in each cardiac cycle state, and calculate the corresponding power spectrum ; Mel frequency conversion is performed on the power spectrum based on the Mel filter. For each Mel filter , the output of the Mel filter is ; m is the mth frame signal of the heart sound signal segment; Perform discrete cosine transform (DCT) on the output of the Mel filter and take the logarithm to get , the calculation formula of MFCC coefficient is as follows: in, Represents the first n coefficients of M Mel filters, and the first 13 coefficients are combined to obtain MFCC features.

[0049] Standard MFCC coefficients only reflect the static characteristics of speech. First-order and second-order differences of these coefficients are used to determine the dynamic characteristics of the signal. First-order differences of MFCC features refer to the relationship between frames (adjacent frames), that is, the difference between adjacent feature frames. Second-order differences are based on first-order differences and calculate the variation between adjacent elements in the first-order difference sequence. Mean and standard deviation processing is then performed to obtain statistical features.

[0050] In a preferred embodiment of the present invention, the MFCC first-order difference is calculated based on the center-weighted difference method using the librosa function to calculate the difference.

[0051] The calculation formula of the MFCC first-order difference is as follows: The calculation formula of the MFCC first-order difference is as follows: in, c (t+ l ) indicates a time point t + l The eigenvalues ​​(MFCC coefficients) at l is the window offset (1≤ l ≤N), and is used as a weight in the calculation; N is the window width for calculating the difference.

[0052] Compared with traditional simple difference In comparison, the present invention improves the robustness of dynamic features through symmetrical window and weight design, uses denominator for normalization to ensure the stability of differential calculation, enhances the classification model's ability to capture timing patterns, captures the dynamic change pattern of heart sounds through multi-order differences, and enhances the sensitivity to transient pathological characteristics.

[0053] The traditional MFCC feature extraction algorithm is based on the MFCC of each frame signal of the global statistics. In this invention, the signal is divided into four states, that is, the feature calculation is limited to a specific time window (S1, sys, S2, dias), and static and dynamic features are calculated; statistics (mean, standard deviation) are extracted based on each physiological event segment (S1, sys, S2, dias), and the event duration, that is, the duration of each state of the cardiac cycle, is explicitly added. , and then fuses multiple features into one feature vector to improve the temporal sensitivity of the features and facilitate processing by machine learning models.

[0054] Statistical values ​​include mean and standard deviation. The formula for calculating the mean is as follows: in, For the i The value of the MFCC coefficient (difference) in the tth frame, where T is the total number of frames.

[0055] The formula for calculating the standard deviation is as follows: The present invention extracts 13-dimensional MFCC features, then obtains 13-dimensional first-order and second-order difference features by difference, and calculates the mean and standard deviation statistical features respectively, plus the duration feature, for a total of 79-dimensional feature vectors. The calculation of each feature and the mean and standard deviation is performed based on the time frame corresponding to each state. Therefore, a complete cardiac cycle will extract a 316-dimensional feature vector, including the mean and standard deviation of the 13-dimensional MFCC features of S1, the mean and standard deviation of the 13-dimensional MFCC first-order differences, the mean and standard deviation of the 13-dimensional MFCC second-order differences, plus the duration of S1, and so on for other states.

[0056] The output feature information of the improved MFCC feature extraction algorithm in any heart sound cycle state can be expressed as , which improves the richness of features.

[0057] It can be understood that the present invention uses staged statistical features to characterize the differences in time-frequency characteristics of different cardiac states (such as systole and diastole), thereby improving the discrimination of pathological characteristics; combining static and dynamic coefficient statistics to enhance the ability to capture transient changes in heart sounds (such as murmurs); and through the complementarity of time domain duration and frequency domain statistics, constructing a multidimensional feature space, thereby improving the robustness of the classification model.

[0058] Based on the above embodiment, as an optional embodiment, the wavelet transform feature extraction algorithm is improved based on the cardiac cycle, including steps S2121 and S2122.

[0059] Step S2121: Based on the duration of each state of the cardiac cycle, a first event window of an improved wavelet transform feature extraction algorithm is determined to extract wavelet feature information of the heart sound signal segment based on the first event window. The first event window is an adaptive variable window that can be dynamically adjusted.

[0060] Step S2122: Determine output feature information of the improved wavelet transform feature extraction algorithm based on the wavelet feature information and duration corresponding to each state of the cardiac cycle.

[0061] The Wavelet Transform captures the local characteristics of a signal by decomposing it into a linear combination of wavelet basis functions at different frequencies and times. Its core idea is to use wavelet decomposition to decompose the signal into different levels, then extract a feature from each level to represent the signal's characteristics within different frequency ranges. The wavelet coefficients reflect the signal's characteristics at different scales and locations, thereby extracting features from the signal and revealing the underlying structural information.

[0062] The calculation formula of the approximate coefficients of wavelet decomposition is as follows: in, is the j-th level approximation coefficient, It is a heart sound signal segment.

[0063] The calculation formula of the detail coefficient of wavelet decomposition is as follows: in, is the j-th level detail coefficient, h is the impulse response of the low-pass filter, g is the impulse response of the high-pass filter.

[0064] This embodiment of the present invention decomposes heart sound signal segments into approximation coefficients (low-frequency components of the signal) and detail coefficients (high-frequency components at different scales) at different levels. Traditional methods typically perform wavelet decomposition on the entire signal and extract global statistics (mean and standard deviation of the entire signal). However, this invention divides the heart sound signal into specific event windows (e.g., S1 corresponds to the first heart sound phase) and performs wavelet decomposition, feature extraction, and statistical analysis on the signal segments within each window. This event window-based approach captures the local characteristics of the signal more precisely, facilitating analysis of the characteristics of different phases within the heart sound signal, and enhancing the expressiveness of the features and task adaptability.

[0065] The formulas for calculating the mean and standard deviation are: ; in, is the number of wavelet coefficients, is the wavelet coefficient.

[0066] The single-segment feature concatenation vector of the wavelet feature of the heart sound signal segment in any heart sound cycle state can be expressed as .

[0067] It can be understood that the present invention independently extracts wavelet feature information in stages, can reduce the dimension of the wavelet features, and use the mean and standard deviation of the wavelet feature information as output feature information to achieve generalization balance, using the mean of the approximate coefficient to reflect the low-frequency energy of the stage and the standard deviation of the detail coefficient to capture high-frequency fluctuations.

[0068] On the basis of the above embodiment, as an optional embodiment, the Fourier transform feature extraction algorithm is improved based on the cardiac cycle, including steps S2131 and S2132.

[0069] Step S2131: Based on the durations of the various states of the cardiac cycle, a second event window is determined for the improved Fourier transform feature extraction algorithm to extract Fourier feature information of the heart sound signal segment based on the second event window. The second event window is a global window. For a short-time Fourier transform, the second event window is a sliding window of fixed length.

[0070] Step S2132: Determine output feature information of the improved Fourier transform feature extraction algorithm based on the Fourier feature information and duration corresponding to each state of the cardiac cycle.

[0071] The Short-Time Fourier Transform (STFT) is a widely used time-frequency analysis method that converts a time-domain signal into a frequency-domain signal. The STFT provides local information about a signal in both time and frequency, allowing the transformation result to reflect changes in the signal's frequency components over different time periods.

[0072] The basic idea of ​​STFT is to divide a long-time signal into multiple short-time windows by windowing, perform Fourier transform in a local time period, and then stack the obtained frequency domain signals, that is, splice the frequency domain representation of each frame output along the time axis to obtain the time-frequency characteristics of the entire signal.

[0073] The STFT divides the signal into multiple short-time windows and then applies the Fourier transform. The Fourier transform uses the discrete Fourier transform DFT. The formula for the short-time Fourier transform (STFT) is as follows: in, It is Frame, The complex STFT coefficients of the frequency bins; is the input signal; is the window function; L is the number of FFT points (n_fft).

[0074] This paper proposes a phased dynamic time-frequency statistical fusion method, namely, a STFT feature extraction method based on phased time-frequency statistical fusion. The FFT is assumed to have 256 points, a window length of 20ms, and a step size of 10ms. The segmented heart sound signal is divided into four states. Each state is framed and windowed to obtain a frame signal. A discrete Fourier transform is then performed to convert the heart sound signal from the time domain to the frequency domain, obtaining a frequency domain representation of the signal.

[0075] Traditional methods typically perform STFT on the entire signal and extract global features (such as spectrogram statistics). However, this invention divides the heart sound signal into specific event windows and extracts features based on the STFT results within each window. This event-window-based approach more precisely captures the local features of the signal at different stages. Furthermore, the duration of each event is included as a feature component, and multi-dimensional feature fusion enhances the expressive power of features.

[0076] For a complete cardiac cycle, the spectrum feature value is first obtained, and then the mean and standard deviation of the spectrum are calculated to obtain statistical features, including the duration of each stage, and these features are spliced ​​into a feature vector.

[0077] After extracting the STFT features of the heart sound signal segment, record it as X, separate the amplitude and phase of the STFT result, and square the amplitude to obtain the power spectrum , reuse The power spectrum is converted to decibel scale (max is the maximum value per frame). The decibel scale can amplify weak signals (such as pathological murmurs), facilitate subsequent analysis and visualization, and improve feature differentiation. Then, features are extracted for each physiological stage (S1, sys, S2, dias) in the heart sound cycle, statistics are calculated, and the time features are fused to form a feature vector. The mean is calculated as follows: , the standard deviation is .

[0078] The single segment splicing feature vector of the heart sound signal segment in any cardiac cycle state can be expressed as ,in =128, number of frequency bins = (1+256 / 2) = 129.

[0079] It can be understood that the present invention extracts features for each stage based on the stage characteristics of heart sound signals, improves the pertinence and effectiveness of features, and solves the problem that global statistics cannot distinguish the abnormal manifestations of pathological features in specific stages; through the multimodal fusion of mean-standard deviation statistics and duration, it breaks through the pure frequency domain limitations of traditional STFT.

[0080] On the basis of the above embodiment, as an optional embodiment, the power spectrum density feature extraction algorithm is improved based on the cardiac cycle, including steps S2141 to S2143.

[0081] Step S2141 : determining Fourier feature information of the heart sound signal segment based on the duration of each state of the cardiac cycle.

[0082] Step S2142: determining the power spectrum density characteristic information of each of the heart sound signal segments based on the Fourier characteristic information of each of the heart sound signal segments.

[0083] Step S2143: Determine output feature information of the improved power spectrum density feature extraction algorithm based on the power spectrum density feature information and duration corresponding to different states of the cardiac cycle.

[0084] Power spectral density (PSD) characterizes the power distribution of a signal per unit frequency. It is a function that reflects the energy distribution of a signal at different frequencies. PSD can be used to describe the time-frequency distribution of heart sound signals by combining features in the time and frequency domains. The PSD is calculated by calculating the square modulus of the DFT and normalizing it. Specifically, the heart sound signal is Fourier transformed, and the squared amplitudes are averaged to obtain an estimated value.

[0085] The power spectral density calculation formula is as follows: in, =( / L)*fs, Indicates the frequency components, fs is the sampling frequency.

[0086] The segmented heart sound signal segments are Fourier transformed from the time domain to the frequency domain. The power spectral density (PSD) is calculated by calculating the square modulus of the STFT result. Each element of the power spectral density matrix represents the signal power at a specific time and frequency. The matrix shape is (129, number of frames). Similarly, statistical features such as mean and standard deviation are calculated. For a complete cardiac cycle, the power spectral density feature vector consists of the mean (average of the power values ​​across all frames within the time window), standard deviation, and duration of each phase of the PSD.

[0087] This paper proposes a multi-dimensional PSD feature fusion extraction method based on STFT. After extracting the STFT feature of the heart sound signal segment through formula (10), it is recorded as X, and the local power spectrum is calculated by the square of the STFT amplitude. , which calculates the power spectral density (PSD) within a specific event window, can capture the local characteristics of heart sound signals at different stages. Traditional PSD analysis (such as the Welch method) typically uses longer segmentation and averaging to reduce noise. However, this method directly utilizes the short-term power spectrum of the STFT, preserving temporal locality and making it more suitable for non-stationary heart sound signals. It combines this with event segment statistics and duration to form a feature vector.

[0088] Mean calculation formula ; The formula for calculating the standard deviation is , Indicates the power spectrum density at the frequency point f , the value at time frame t, is the number of time frames, t0 and t1 are the starting indexes of the time window.

[0089] The power spectral density of the heart sound signal segment in any cardiac cycle state is the single segment feature concatenation vector .

[0090] It can be understood that the present invention makes the features clearer and effectively separates the characteristic distributions of normal and abnormal heart sounds through segmented feature extraction; and combines frequency domain statistics with time domain dynamic features to provide more comprehensive information representation.

[0091] Based on the above embodiment, as an optional embodiment, the feature fusion and feature selection of the feature information of each heart sound signal segment includes steps S310 to S320.

[0092] Step S310: constructing a feature matrix based on the feature information of all the heart sound signal segments to complete feature fusion.

[0093] After the feature extraction algorithm, we obtain 316-dimensional MFCC features, 68-dimensional wavelet transform features, 1036-dimensional STFT features, and 1036-dimensional PSD features. These features are concatenated along the columns of the feature matrix to form a new feature matrix with the shape of (number of samples, 2456).

[0094] Step S320 , sorting the feature information of the feature matrix by importance, and selecting target feature information based on the sorting result to complete feature selection.

[0095] A decision tree algorithm is used for feature selection to select features that are more useful for classification. Therefore, the 2456-dimensional features after feature fusion are input into the decision tree classifier to obtain the feature importance score ranking, and the top 20 features are selected as the input for the subsequent classifier.

[0096] It can be understood that the fused features can show rich feature information of the heart sound signal to be identified, and feature selection is conducive to removing redundant feature information and improving the accuracy of the heart sound signal classification results.

[0097] like Figure 3 As shown, based on the above embodiment, as an optional embodiment, the training step of the classification model includes steps S010 to S040.

[0098] Step S010 : segmenting the heart sound signal samples based on the cardiac cycle to obtain heart sound signal segment samples.

[0099] Step S020 : determining feature information of each of the heart sound signal segment samples based on at least one feature extraction algorithm.

[0100] Step S030 , based on the improved five-fold cross-validation algorithm, the characteristic information of all heart sound signal segment samples is divided into a training set and a test set, the training set is sample balanced based on the preset sample weights, and the target sample characteristic information is screened from the sample-balanced training set.

[0101] Step S040: training the machine learning model based on the target sample feature information, and testing the trained machine learning model based on the test set to obtain the classification model.

[0102] In step S010, the heart sound signal samples are segmented as described in step S100. The public dataset PhysioNet Challenge 2022 is used for heart sound signal segmentation. The training data for the database consists of 3,163 heart sound recordings from 942 patients, 499 of which contain murmurs. Heart sound data were collected and labeled by professional physicians at four primary auscultation sites: the pulmonary valve (PV), aortic valve (AV), mitral valve (MV), and tricuspid valve (TV). Some patients also had recordings from other sites (PhC). The dataset includes waveform files for each heart sound data entry, along with segmentation annotations, including the start and end times of each phase and identifiers for the four states. A file also contains detailed information about each subject. Murmur labels for each subject are categorized into three types: Present (presence of a murmur); Absent (absence of a murmur); and Unknown (unknown). Because heart sound data labeled Unknown are clinically insignificant, only data labeled with murmur and no murmur were used in this experiment. The sampling rate of the heart sound signal is 4000 Hz.

[0103] Table 1. Details of the PhysioNet 2022 Heart Sound Challenge dataset

[0104] In step S020 , the feature information of each of the heart sound signal segment samples may be determined with reference to the solution described in step S200 above.

[0105] When determining the MFCC feature information for a heart sound signal segment, consider the number of beats in the recording. For example, the feature shape of a heart sound recording at a certain location for a patient is (n_beats, 316). Each feature is averaged along the beat axis to obtain (1, 316), where n_beats is the number of beats. This is ultimately saved as an MFCC feature file, which includes the label for each patient and serves as input for the subsequent classifier.

[0106] Table 2 MFCC feature dimensions of a complete cardiac cycle

[0107] To determine the wavelet transform feature information of heart sound signal segments, the wavelet basis function "db10" is used to decompose each segment into seven levels. This generates eight wavelet coefficients, including seven detail coefficients and one approximation coefficient, providing information about the heart sound signal's composition at different frequency bands. The mean and standard deviation of each coefficient are calculated to form statistical features, reflecting the distribution characteristics of the heart sound signal at different scales. Eight mean and eight standard deviation features are calculated for each state, and together with the duration of each state, there are 17 eigenvalues. Therefore, for a complete cardiac cycle, the concatenated feature vector has 68 dimensions. Because the total number of heart beats varies from patient to patient, the feature vector shape is (n_beats, 68). Each feature is then averaged by the number of heart beats, resulting in (1, 68). The feature vector and corresponding labels are saved as a wavelet feature file.

[0108] Table 3 Wavelet feature dimensions of a complete cardiac cycle

[0109] To determine the STFT features of a heart sound signal segment, we set the FFT size to 256, the window length to 20ms, and the step size to 10ms. The segmented heart sound signal is divided into four states. Each state is framed and windowed to obtain each frame. A discrete Fourier transform is then performed to convert the heart sound signal from the time domain to the frequency domain, obtaining a frequency domain representation. This results in a complex matrix with a shape of (129, T), where T is the total number of frames. Each feature is averaged by the number of heart beats, and the feature vector and corresponding label are saved as an STFT feature file.

[0110] Table 4 STFT feature dimensions of a complete cardiac cycle

[0111] When determining the power spectral density features of heart sound signal segments, for a complete heartbeat, the feature vector consists of the mean (average of the power values ​​across all frames within the time window), standard deviation, and duration of each phase of the PSD. Each feature is averaged over the number of heartbeats, and the feature vector and corresponding label are saved as a PSD feature file.

[0112] Table 5. PSD feature dimensions of a complete cardiac cycle

[0113] In step S030, the feature information of all heart sound signal segment samples is divided into five subsets based on an improved five-fold cross-validation algorithm. Four subsets are used as training sets, and one subset is used as a test set. In each fold, the 316-dimensional MFCC features, 68-dimensional wavelet transform features, 1036-dimensional STFT features, and 1036-dimensional PSD features are concatenated along the columns of the feature matrix to form a new feature matrix with the shape of (number of samples, 2456). Preferably, the feature information of all heart sound signal segment samples can be obtained from the MFCC feature file, wavelet feature file, STFT feature file, and PSD feature file.

[0114] Due to the imbalance in the number of positive and negative samples in the dataset, the present invention balances samples based on weighted random sampling. Specifically, by calculating the number of samples in each category, each sample is assigned a weight that is proportional to the inverse of the number of samples in its category, where n(class) is the number of samples. The weights are then normalized so that the sum of the weights is 1. The normalized weights are then used to perform weighted random sampling on the samples. The sampling size is equal to the size of the original dataset, and repeated sampling is allowed. Categories with a smaller number of samples will be sampled multiple times due to a higher probability, thereby increasing their number of appearances in the balanced dataset, while categories with a larger number of samples will be sampled relatively less frequently, ultimately balancing the number of samples in the positive and negative categories in the heart sound dataset.

[0115] The formula for calculating the weight of each sample is as follows: A decision tree algorithm is used for feature selection to select features that are more useful for classification. The 2456-dimensional features after feature fusion are input into a decision tree classifier to obtain a ranking of feature importance scores. The top 20 features are selected as input to the subsequent classifier. The resulting feature matrix after feature selection has a shape of (number of samples, 20).

[0116] In addition, the present invention also adopts a standardization strategy (StandardScaler) based on the cross-validation framework to eliminate the scale differences between features and ensure that the subsequent model can treat each feature fairly during training and prediction. The normalization formula is shown as follows: In the formula is the normalized heart sound characteristic value, is the original eigenvalue, is the mean of the feature in the training set, is the standard deviation of the feature in the training set.

[0117] Preferably, the machine learning model of the present invention includes SVM, random forest, decision tree and XGBoost models. Correspondingly, in step S030, five-fold cross-validation and target sample feature information screening are performed on the four models respectively. Before the machine learning model is trained based on the target sample feature information, the target sample feature information is standardized according to formula (14).

[0118] The top 20 features in terms of importance are the features that appear more frequently using the integration method and are used as the target sample feature information. The feature naming rules are: mfcc(difference)_mean / standard deviation_state_coefficient (1-13), coeffs_coefficient (1-8)_mean / standard deviation_state, stft_state_mean / standard deviation_dimension (1-129), psd_state_mean / standard deviation_dimension (1-129). The features that appear more frequently are: 'stft_sys_mean_10', 'mfcc_delta_meansys_9', 'mfcc_meansys_5', 'coeffs_5_std_sys', 'coeffs_4_std_sys', 'mfcc_delta_meansys_2', 'mfcc_delta_meansys_10', 'mfcc_delta_meansys_8'. Figure 4 a- Figure 4 As shown in Figure d, the first-order difference features of the MFCC account for nearly half of the 20 features, indicating that the dynamic characteristics of the MFCC features are more important in distinguishing normal and abnormal signals. The period sys is more important than other states, while the frequency of occurrence of the period dias is relatively low.

[0119] comprehensive Figure 4 a- Figure 4 The importance of d is ranked, and the final 20 features are as follows Figure 5 As shown, 20 features were input into the SVM, random forest, decision tree, and XGBoost models for classification. Feature fusion, sample balancing, and feature selection were all performed using a five-fold cross-validation approach. Four machine learning methods were used to achieve binary classification of heart sound signals. The SVM used a linear kernel with a random seed of 42 to ensure repeatability. The Decision Tree used a decision tree structure with a random seed of 42. The Random Forest consisted of 100 decision trees with a random seed of 42. The XGBoost classifier used a random seed of 42.

[0120] like Figure 6 a- Figure 6 d and Figure 7 a- Figure 7 As shown in d, Figure 7 a- Figure 7 d represents the receiver operating characteristic curves of the four machine learning models SVM, random forest, decision tree and XGBoost respectively. The horizontal axis is the false positive rate, and the vertical axis is the true positive rate. The present invention uses test data to test the trained classification model and calculates indicators such as accuracy, recall rate, and F1 score. After five-fold cross-validation, the four machine learning models SVM, random forest, decision tree and XGBoost were used respectively, with noise marked as 1 and no noise marked as 0, to achieve the classification of normal heart sounds and abnormal heart sounds in the PhysioNet 2022 Heart Sound Challenge dataset. The classification accuracy of SVM was 89.0%, the classification accuracy of decision tree was 81.0%, the classification accuracy of random forest was 91.0%, and the classification accuracy of XGBoost was 90.0%. By comparison, the classification effect of random forest was the best, with an accuracy rate of 91.0%.

[0121] Table 6 Comparison of classification performance of different heart sound classification models

[0122] In summary, the present invention realizes the normal and abnormal classification of heart sound signals based on a machine learning model, thereby providing help and support for the early diagnosis of cardiovascular diseases. The heart sound signal is segmented according to the segmentation file, so that the heart sound signal is divided into multiple independent cardiac cycles, and four time-frequency features, including dynamic features and statistical features, are extracted for each state of the cardiac cycle to generate a feature vector. The four features are fused to form a complete feature vector matrix. The fused features can better display the rich feature information of the heart sound signal. Feature selection is performed on the fused features to remove redundant features. The filtered features are input into four machine learning models for classification, so as to achieve the purpose of classifying the heart sound signal and improve the accuracy of classification.

[0123] The heart sound signal classification device provided by the present invention is described below. The heart sound signal classification device described below and the heart sound signal classification method described above can be referenced to each other.

[0124] Figure 8 FIG. 1 is a structural diagram of the heart sound signal classification device provided by the present invention. Figure 8 As shown, the present invention also provides a heart sound signal classification device, which includes the following modules.

[0125] A signal segmentation module 810 is configured to segment the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments; a feature extraction module 820 for determining feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle; a classification module 830 for performing feature fusion and feature selection on the feature information of each of the heart sound signal segments, inputting the obtained target feature information into a classification model, and obtaining a heart sound signal classification result output by the classification model; The classification model is obtained by training a machine learning model based on the feature information of the segmented heart sound signal samples.

[0126] As an embodiment, the feature extraction module 820 is further configured to: Based on the cardiac cycle, at least one feature extraction algorithm among a Mel-frequency cepstral coefficient MFCC feature extraction algorithm, a wavelet transform feature extraction algorithm, a Fourier transform feature extraction algorithm, and a power spectral density feature extraction algorithm is improved to obtain an improved feature extraction algorithm; Feature information of the heart sound signal segment is extracted based on an improved feature extraction algorithm.

[0127] As an embodiment, the feature extraction module 820 is further configured to: Determining, based on the time frames of each state of the cardiac cycle, the statistical value of the MFCC features, the statistical value of the MFCC first-order difference, and the statistical value of the MFCC second-order difference of each of the heart sound signal segments; Based on the statistical values ​​of the MFCC features corresponding to each state of the cardiac cycle, the statistical values ​​of the MFCC first-order differences, the statistical values ​​of the MFCC second-order differences, and the duration, output feature information of the improved MFCC feature extraction algorithm is determined.

[0128] As an embodiment, the MFCC first-order difference is determined based on a center-weighted difference method.

[0129] As an embodiment, the feature extraction module 820 is further configured to: determining a first event window of an improved wavelet transform feature extraction algorithm based on the duration of each state of the cardiac cycle so as to extract wavelet feature information of the heart sound signal segment based on the first event window; Determining output feature information of an improved wavelet transform feature extraction algorithm based on the wavelet feature information and duration corresponding to each state of the cardiac cycle; As an embodiment, the feature extraction module 820 is further configured to: determining, based on the durations of the various states of the cardiac cycle, a second event window of an improved Fourier transform feature extraction algorithm so as to extract Fourier feature information of the heart sound signal segment based on the second event window; Based on the Fourier feature information and duration corresponding to each state of the cardiac cycle, output feature information of the improved Fourier transform feature extraction algorithm is determined.

[0130] As an embodiment, the feature extraction module 820 is further configured to: determining Fourier feature information of the heart sound signal segment based on the duration of each state of the cardiac cycle; determining power spectral density characteristic information of each of the heart sound signal segments based on the Fourier characteristic information of each of the heart sound signal segments; Based on the power spectrum density feature information and duration corresponding to different states of the cardiac cycle, output feature information of the improved power spectrum density feature extraction algorithm is determined.

[0131] As an embodiment, the classification module 830 is further configured to: Based on the feature information of all the heart sound signal segments, a feature matrix is ​​constructed to complete feature fusion; The feature information of the feature matrix is ​​sorted by importance, and target feature information is selected based on the sorting result to complete feature selection.

[0132] As an embodiment, a training module is also included; The training module is used to segment the heart sound signal samples based on the cardiac cycle to obtain heart sound signal segment samples; determine the feature information of each of the heart sound signal segment samples based on at least one feature extraction algorithm; divide the feature information of all heart sound signal segment samples into a training set and a test set based on an improved five-fold cross-validation algorithm, balance the training set based on preset sample weights, and screen the target sample feature information from the balanced training set; train the machine learning model based on the target sample feature information, and test the trained machine learning model based on the test set to obtain the classification model.

[0133] The heart sound signal classification device provided by the present invention is used to execute the heart sound signal classification method described in any of the above embodiments, and has technical effects corresponding to the heart sound signal classification method, which will not be described in detail.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A heart sound signal classification method, characterized in that: include: Segment the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments; determining feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle; performing feature fusion and feature selection on the feature information of each of the heart sound signal segments, inputting the obtained target feature information into a classification model, and obtaining a heart sound signal classification result output by the classification model; The classification model is obtained by training a machine learning model based on feature information of heart sound signal segment samples.

2. The heart sound signal classification method according to claim 1, characterized in that: The determining of feature information of each of the heart sound signal segments based on at least one feature extraction algorithm includes: Based on the cardiac cycle, at least one feature extraction algorithm among a Mel-frequency cepstral coefficient MFCC feature extraction algorithm, a wavelet transform feature extraction algorithm, a Fourier transform feature extraction algorithm, and a power spectral density feature extraction algorithm is improved to obtain an improved feature extraction algorithm; Feature information of the heart sound signal segment is extracted based on an improved feature extraction algorithm.

3. The heart sound signal classification method according to claim 2, characterized in that: The Mel-frequency cepstral coefficient (MFCC) feature extraction algorithm is improved based on the cardiac cycle, including: Determining, based on the time frames of each state of the cardiac cycle, the statistical value of the MFCC features, the statistical value of the MFCC first-order difference, and the statistical value of the MFCC second-order difference of each of the heart sound signal segments; Based on the statistical values ​​of the MFCC features corresponding to each state of the cardiac cycle, the statistical values ​​of the MFCC first-order differences, the statistical values ​​of the MFCC second-order differences, and the duration, output feature information of the improved MFCC feature extraction algorithm is determined.

4. The heart sound signal classification method according to claim 3, characterized in that: The first-order difference of MFCC is determined based on the center-weighted difference method.

5. The heart sound signal classification method according to claim 2, characterized in that: The wavelet transform feature extraction algorithm is improved based on the cardiac cycle, including: determining a first event window of an improved wavelet transform feature extraction algorithm based on the duration of each state of the cardiac cycle so as to extract wavelet feature information of the heart sound signal segment based on the first event window; Based on the wavelet feature information and duration corresponding to each state of the cardiac cycle, output feature information of the improved wavelet transform feature extraction algorithm is determined.

6. The heart sound signal classification method according to claim 2, characterized in that: The Fourier transform feature extraction algorithm is improved based on the cardiac cycle, including: determining, based on the durations of the various states of the cardiac cycle, a second event window of an improved Fourier transform feature extraction algorithm so as to extract Fourier feature information of the heart sound signal segment based on the second event window; Based on the Fourier feature information and duration corresponding to each state of the cardiac cycle, output feature information of the improved Fourier transform feature extraction algorithm is determined.

7. The heart sound signal classification method according to claim 2, characterized in that: The power spectrum density feature extraction algorithm is improved based on the cardiac cycle, including: determining Fourier feature information of the heart sound signal segment based on the duration of each state of the cardiac cycle; determining power spectral density characteristic information of each of the heart sound signal segments based on the Fourier characteristic information of each of the heart sound signal segments; Based on the power spectrum density feature information and duration corresponding to different states of the cardiac cycle, output feature information of the improved power spectrum density feature extraction algorithm is determined.

8. The heart sound signal classification method according to any one of claims 1 to 7, characterized in that: The performing feature fusion and feature selection on the feature information of each of the heart sound signal segments includes: Based on the feature information of all the heart sound signal segments, a feature matrix is ​​constructed to complete feature fusion; The feature information of the feature matrix is ​​sorted by importance, and target feature information is selected based on the sorting result to complete feature selection.

9. The heart sound signal classification method according to any one of claims 1 to 7, characterized in that: The training steps of the classification model include: Segmenting the heart sound signal samples based on the cardiac cycle to obtain heart sound signal segment samples; determining feature information of each of the heart sound signal segment samples based on at least one feature extraction algorithm; Based on an improved five-fold cross-validation algorithm, the characteristic information of all heart sound signal fragment samples is divided into a training set and a test set, the training set is sample-balanced based on a preset sample weight, and the target sample characteristic information is screened from the sample-balanced training set; The machine learning model is trained based on the target sample feature information, and the trained machine learning model is tested based on the test set to obtain the classification model.

10. A heart sound signal classification device, characterized in that: include: A signal segmentation module is used to segment the heart sound signal to be identified based on the cardiac cycle to obtain heart sound signal segments; a feature extraction module, configured to determine feature information of each of the heart sound signal segments based on at least one feature extraction algorithm, wherein the feature information of the heart sound signal segment is used to represent feature information of each state of the heart sound signal segment in the cardiac cycle; a classification module, configured to perform feature fusion and feature selection on the feature information of each of the heart sound signal segments, input the obtained target feature information into a classification model, and obtain a heart sound signal classification result output by the classification model; The classification model is obtained by training a machine learning model based on the feature information of the segmented heart sound signal samples.

Images