Explainable heart sound anomaly recognition method and system based on fractional fourier transform

Abstract

The application discloses an interpretable heart sound anomaly recognition method and system based on a fractional domain Fourier transform, which comprises preprocessing, feature extraction, model establishment and model interpretation. The preprocessing comprises shearing, downsampling, filtering, amplitude normalization, heart cycle segmentation, frame segmentation and windowing in sequence; the feature extraction is configured to firstly perform fractional domain Fourier transform on the preprocessed heart sound, then extract frame-level Shannon entropy features of one-dimensional fractional domain heart sound signals, and calculate 13 statistical functions on the frame-level features as final features; the model establishment selects an XGBoost classifier; and the model interpretation selects a SHAP (SHapley Additive exPlanation) interpretation model. The application is easy to realize, simple in method, low in feature dimension, fast in model fitting, and has model prediction interpretability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent biomedical signal processing, and particularly relates to an interpretable method and system for identifying abnormal heart sounds based on fractional-domain Fourier transform. Background Technology

[0002] Currently, researchers both domestically and internationally have proposed numerous methods and applications for identifying cardiac abnormalities based on heart sounds, some of which have achieved good recognition results. The implementation process of most of these methods and applications involves first extracting time-domain, frequency-domain, or time-frequency-domain features, and then using traditional machine learning or deep learning models to model and classify these features to identify normal and abnormal heart sounds. While they have achieved good results, they have overlooked the importance of the interpretability of the proposed models. SHAP-based model interpretation methods can calculate global and local feature contributions, which is essential for verifying whether the behavior of the heart sound abnormality recognition model meets expectations and can better establish a trusting relationship with doctors and patients.

[0003] Based on traditional machine learning models, related invention patents (CN110610204A) mostly use Mel frequency cepstral coefficients as features, or extract multiple low-level audio descriptors and fuse them as features, which has a high dimensionality and results in long model fitting time.

[0004] Existing patents related to heart sound recognition (CN113345471A) focus more on the model's recognition performance, neglecting the importance of model interpretability and failing to establish a trust relationship between the model and doctors and patients. For example, the paper: TanY, Wang Z, Qian K, et al. Heart Sound Classification based on Fractional Fourier Transformation Entropy[C] / / 2022 IEEE 4th GlobalConference on LifeSciences and Technologies (LifeTech). IEEE, 2022: 588-589. discloses: This paper proposes a heart sound classification method. The entire process involves heart sound clipping, feature extraction, and classification, using fractional Fourier transform entropy (FRFE) as features and support vector machine (SVM) as the classification model. This paper does not perform sufficient preprocessing of the heart sound signal, uses low-level features with large dimensions, has a slow model fitting speed, and lacks model interpretability. Summary of the Invention

[0005] In view of this, the present invention aims to propose an interpretable method and system for identifying abnormal heart sounds based on fractional-domain Fourier transform. Specifically, it is an algorithm for extracting the entropy features of one-dimensional fractional-domain heart sound signals.

[0006] The first aspect of this invention discloses an interpretable method for identifying abnormal heart sounds based on fractional-domain Fourier transform. The method includes:

[0007] Step S100: Preprocess the heart sound signal;

[0008] Step S200: Extract the fractional-domain Fourier transform entropy features of the preprocessed heart sound signal;

[0009] Step S300, Model building, i.e., training the XGBoost classifier to build the model, and evaluating the results by unweighted average recall: select multiple parameters that need to be optimized, set the range of grid search for each parameter, set random seed values ​​for the parameters, and keep the default values ​​for the remaining parameters;

[0010] A grid search is performed on each parameter of XGBoost, and the model that maximizes the unweighted average recall of the test set is found, using the unweighted average recall as the performance metric of the model.

[0011] Step S400, Model Interpretation: Calculate SHAP values ​​for statistical features to obtain the SHAP value for each feature.

[0012] Furthermore, step S100 includes:

[0013] Step S110: Cut a portion of the original heart sound signal as a heart sound sample for subsequent processing;

[0014] Step S120: Downsample the sampling frequency of the heart sound signal to a certain number of values;

[0015] Step S130: The heart sound signal is filtered using a filter;

[0016] Step S140: Normalize the amplitude of all heart sound signals using the following formula:

[0017]

[0018] in, It is the heart sound signal after amplitude normalization. It is the original heart sound signal;

[0019] Step S150: Segment the second heart sound component of the heart sound, and splice the multiple second heart sound components of each example heart sound signal together;

[0020] Step S160: Perform frame segmentation on the heart sound signal processed in step S150.

[0021] Step S170: Add a Hamming window to each frame of heart sound signal:

[0022]

[0023] in, For Hamming window functions, These are the sampling points for the signal;

[0024] Furthermore, step S200 includes:

[0025] Step S210: Perform a fractional-domain Fourier transform on the heart sound signal to obtain the time-frequency domain signal, also known as the fractional-domain signal;

[0026] Step S220: Take the modulus of the fractional-domain spectrum of the heart sound signal;

[0027] Step S230: Extract the Shannon entropy features of the one-dimensional fractional domain heart sound signal, that is, extract the Shannon entropy features of each frame of fractional domain heart sound signal.

[0028] Step S240: Apply multiple statistical functions to the frame-level fractional domain Fourier transform entropy features as the final features extracted for each heart sound signal.

[0029] Furthermore, step S230 includes:

[0030] Step S231, let a frame of fractional-domain heart sound signal be... The total number of sampling points is ;

[0031] Step S232, the data sequence According to the following formula, it is divided into Duan got , :

[0032] ;

[0033] Step S233: Compare and determine the maximum value of the heart sound signal in this frame. minimum value ;

[0034] Step S234: Establish a continuous data segment The range is ( , Similarly, divided into Duan got , ;

[0035] Step S235: Calculate the segmented heart sound signal sequence Located in the corresponding Number of , then the first The formula for calculating the Shannon entropy of a segment is:

[0036]

[0037] Among them, let ;

[0038] Step S236, then the fractional-domain Fourier transform entropy of this frame of heart sound signal is:

[0039]

[0040] Or it can be expressed as:

[0041] .

[0042] A second aspect of the present invention discloses an interpretable heart sound abnormality recognition system based on fractional-domain Fourier transform. The system includes:

[0043] The preprocessing module is used to preprocess the heart sound signals;

[0044] The feature extraction module is used to extract the fractional-domain Fourier transform entropy features of the preprocessed heart sound signal;

[0045] The model building module is used to train the XGBoost classifier and build the model. The results are evaluated by unweighted average recall: select multiple parameters to be optimized, set the range of grid search for each parameter, set random seed values ​​for the parameters, and keep the default values ​​for the remaining parameters.

[0046] A grid search is performed on each parameter of XGBoost, and the model that maximizes the unweighted average recall of the test set is found, using the unweighted average recall as the performance metric of the model.

[0047] The model interpretation module is used to calculate the SHAP value of statistical features and obtain the SHAP value of each feature.

[0048] Furthermore, the preprocessing module includes:

[0049] A portion of the original heart sound signal is cut out and used as a heart sound sample for subsequent processing;

[0050] The sampling frequency of the heart sound signal is downsampled to a certain number of values;

[0051] The heart sound signal is filtered using a filter;

[0052] The amplitude of all heart sound signals is normalized using the following formula:

[0053]

[0054] in, It is the heart sound signal after amplitude normalization. It is the original heart sound signal;

[0055] The second heart sound component is segmented, and multiple second heart sound components of each example heart sound signal are spliced ​​together.

[0056] The processed heart sound signal is then subjected to frame segmentation.

[0057] Add a Hamming window to each frame of heart sound signal:

[0058]

[0059] in, For Hamming window functions, These are the sampling points for the signal;

[0060] Furthermore, the feature extraction module includes:

[0061] The fractional-domain Fourier transform of the heart sound signal yields the time-frequency domain signal, also known as the fractional-domain signal.

[0062] Fractional-domain spectral modulus of heart sound signals;

[0063] Extract the Shannon entropy features of the one-dimensional fractional domain heart sound signal, that is, extract the Shannon entropy features of each frame of the fractional domain heart sound signal;

[0064] Multiple statistical functions are applied to the frame-level fractional-domain Fourier transform entropy features as the final features extracted for each heart sound signal.

[0065] Furthermore, the feature extraction module includes:

[0066] Let a fractional-domain heart sound signal be... The total number of data points is ;

[0067] Data sequence According to the following formula, it is divided into Duan got , :

[0068] ;

[0069] The maximum value of the heart sound signal in this frame was determined by comparison. minimum value ;

[0070] Establish a continuous data The range is ( , Similarly, divided into Duan got , ;

[0071] Calculate the segmented heart sound signal sequence Located in the corresponding Number of , then the first The formula for calculating the Shannon entropy of a segment is:

[0072]

[0073] Among them, let ;

[0074] Then the fractional-domain Fourier transform entropy of this frame of heart sound signal is:

[0075]

[0076] Or it can be expressed as:

[0077] .

[0078] The present invention also discloses an application of the above-mentioned interpretable heart sound abnormality recognition method and system based on fractional domain Fourier transform to heart sound abnormality recognition.

[0079] The technical solution of the present invention can achieve the following beneficial technical effects:

[0080] 1. The method of this invention transforms the heart sound signal to the time-frequency domain (also known as the fractional domain) using a fractional-domain Fourier transform, extracting the entropy features of the one-dimensional fractional-domain heart sound signal. This easily reflects the time-frequency domain information features of the heart sound signal. The method of this invention develops an algorithm for extracting the entropy features of the one-dimensional fractional-domain heart sound signal. It transforms the heart sound signal to the time-frequency domain (also known as the fractional domain) using a fractional-domain Fourier transform with an order of 0.9, and then calculates the entropy features of the one-dimensional fractional-domain heart sound signal using the algorithm proposed in this invention, easily reflecting the time-frequency domain information of the heart sound signal. (Note: When the order is 1.0, the fractional-domain Fourier transform transforms the signal from the time domain to the frequency domain.)

[0081] 2. Based on the extraction of low-level audio descriptors, this invention further extracts 13 statistical functions as input to the classification model. Compared with the high-dimensional features of existing technologies, this invention has low feature dimensionality and fast model fitting. The 13 statistical functions are the minimum value, maximum value, range, mean, variance, standard deviation, skewness, peak value, first percentile, second percentile, third percentile, slope and offset of the first linear regression fitting function. Therefore, the final feature dimension of the heart sound signal in this invention is 13-dimensional.

[0082] 3. This invention achieves model interpretability by selecting the SHAP (SHapley Additive exPlanation) method to calculate the local and global contributions of features. Based on SHAP theory, SHAP values ​​are calculated for each of the 13 statistical features. The SHAP value characterizes the degree of contribution of a feature to the model fit, and the sign and magnitude of the SHAP value can explain the positive or negative impact of the feature on the model fit and the importance of that impact.

[0083] 4. This invention is easy to implement and the method is simple. Attached Figure Description

[0084] Figure 1 This is a system block diagram of an interpretable heart sound abnormality recognition method based on fractional domain Fourier transform according to the present invention.

[0085] Figure 2 This is a flowchart of an interpretable heart sound abnormality identification method based on fractional domain Fourier transform according to the present invention.

[0086] Figure 3 This is a flowchart for extracting features from heart sound signals. Detailed Implementation

[0087] The following is combined Figure 1 and Figure 2 This paper describes the specific implementation method and system of an interpretable heart sound abnormality identification method based on fractional domain Fourier transform of the present invention.

[0088] The first aspect of this invention discloses an interpretable method for identifying abnormal heart sounds based on fractional-domain Fourier transform, the specific steps of which are as follows:

[0089] Step S100 involves preprocessing the heart sound signal. Specifically, step S100 includes:

[0090] Step S110: Cut the original heart sound signal for 0.5-5.5 seconds as the heart sound sample for subsequent processing;

[0091] Step S120: Downsample the sampling frequency of the heart sound signal to 1000Hz;

[0092] The heart sound segmentation method in step S150 below is for signals with a sampling rate of 1000Hz; in addition, reducing the sampling frequency can effectively reduce the computing cost of the computer.

[0093] Step S130: The heart sound signal is filtered using a digital bandpass sixth-order Butterworth filter with zero phase and a bandwidth of 25-400Hz.

[0094] Specifically, the audio signals that are usually collected contain not only noise from the environment and the heart sound acquisition equipment, but also interference from the subject's breathing vibrations, bowel sounds, skin friction, and heart and lung murmurs from other organs around the heart; in addition, useful heart sound signals are mainly concentrated between 25-400Hz.

[0095] Step S140: Normalize the amplitude of all heart sound signals using the following formula:

[0096]

[0097] in, It is the heart sound signal after amplitude normalization. This is the original heart sound signal. This is because amplitude normalization can transform each data point of the heart sound signal into the interval [-1, 1].

[0098] Step S150: Perform the segmentation of the second heart sound component based on the logistic regression-HSMM-based heart sound segmentation (Springer DB, Tarassenko L, Clifford G D. Logisticregression-HSMM-based heart sound segmentation[J]. IEEE Transactions on Biomedical Engineering, 2015, 63(4): 822-832.), and splice the multiple second heart sound components of each example heart sound signal.

[0099] In step S160, the heart sound signal processed in step S150 is subjected to frame segmentation with a frame length of 25ms and an overlap rate of 50%.

[0100] Specifically, heart sound signals are non-stationary signals, while a long audio segment divided into many short frames can be considered stationary.

[0101] Step S170: Add a Hamming window to each frame of heart sound signal:

[0102]

[0103] This is because windowing can enhance the continuity between the left and right ends of frames. Specifically, For Hamming window functions, For the sampling points of the signal, The total number of sampling points in one frame of signal;

[0104] Step S200, feature extraction, which involves extracting the fractional-domain Fourier transform entropy features of the preprocessed heart sound signal from step S100. The feature extraction process is as follows: Figure 3 As shown, specifically:

[0105] Step S210: Perform a fractional Fourier transform on the heart sound signal with an order of 0.9 to obtain the time-frequency domain signal, also known as the fractional domain signal;

[0106] Step S220: Take the modulus of the fractional-domain spectrum of the heart sound signal;

[0107] Unless otherwise specified, all calculations in this invention refer to the analog calculations of the fractional-domain spectrum of the heart sound signal.

[0108] Step S230: Extract the Shannon entropy features of the one-dimensional fractional-domain heart sound signal, that is, extract the Shannon entropy features for each frame of the fractional-domain heart sound signal, specifically as follows:

[0109] Step S231, let a frame of fractional-domain heart sound signal be... The total number of sampling points is ;

[0110] Step S232, the data sequence According to the following formula, it is divided into Duan got , :

[0111] ;

[0112] Step S233: Compare and determine the maximum value of the heart sound signal in this frame. minimum value ;

[0113] Step S234: Establish a continuous data segment The range is ( , Similarly, divided into Duan got , ;

[0114] Step S235: Calculate the segmented heart sound signal sequence Located in the corresponding Number of , then the first The formula for calculating the Shannon entropy of a segment is:

[0115]

[0116] Among them, let ;

[0117] Step S236, then the fractional-domain Fourier transform entropy of this frame of heart sound signal is:

[0118]

[0119] Or it can be expressed as:

[0120]

[0121] Step S240: Thirteen statistical functions are applied to the frame-level fractional-domain Fourier transform entropy features as the final features extracted from each heart sound signal. Therefore, the feature dimension of this invention is 13-dimensional. The 13 statistical functions are: minimum value, maximum value, range, mean, variance, standard deviation, skewness, peak value, first percentile, second percentile, third percentile, slope and offset of the first linear regression fitting function.

[0122] Step S300, model building, that is, training the XGBoost classifier to build the model, and evaluating the results by unweighted average recall.

[0123] Specifically, step S300 includes:

[0124] Step S310, the classifier parameter settings and adjustments are as follows:

[0125] Step S311: Select the six parameters to be optimized: learning rate, maximum tree depth, minimum leaf node sample weight and min_child_weight, minimum loss function descent value gamma required for node splitting, proportion of random sampling for each tree (subsample) and proportion of columns of random sampling for each tree (each column is a feature) colsample_bytree, and set the range of grid search for each parameter.

[0126] Step S312: Set the random seed parameter random_state to 42, and keep the other parameters at their default values;

[0127] Step S320: Perform a grid search on each parameter of XGBoost, using unweighted average recall as the model's performance metric, and find the model that maximizes the unweighted average recall of the test set.

[0128] Step S400, Model Interpretation.

[0129] Specifically, step S400, model interpretation, includes:

[0130] The SHAP (SHapley Additive exPlanations) theory is used to calculate the SHAP values ​​for 13 statistical features. This is because the model will output different predictions when a feature is present and when it is absent; comparing the difference between the two predictions yields the contribution of that feature. After performing this operation on each subset of the 13 statistical features, the SHAP value for each feature is obtained by averaging these contributions.

[0131] A second aspect of the present invention discloses an interpretable heart sound abnormality recognition system based on fractional-domain Fourier transform. The system includes:

[0132] The preprocessing module is used to preprocess the heart sound signals;

[0133] The feature extraction module is used to extract the fractional-domain Fourier transform entropy features of the preprocessed heart sound signal;

[0134] The model building module is used to train the XGBoost classifier and build the model. The results are evaluated by unweighted average recall: select multiple parameters to be optimized, set the range of grid search for each parameter, set random seed values ​​for the parameters, and keep the default values ​​for the remaining parameters.

[0135] A grid search is performed on each parameter of XGBoost, and the model that maximizes the unweighted average recall of the test set is found, using the unweighted average recall as the performance metric of the model.

[0136] The model interpretation module is used to calculate the SHAP value of statistical features and obtain the SHAP value of each feature.

[0137] The preprocessing module includes:

[0138] Cut the original heart sound signal for 0.5-5.5 seconds as the heart sound sample for subsequent processing;

[0139] The sampling frequency of the heart sound signal was downsampled to 1000Hz.

[0140] The heart sound segmentation method described below is for signals with a sampling rate of 1000Hz; in addition, reducing the sampling frequency can effectively reduce the computing cost of the computer.

[0141] The heart sound signal was filtered using a digital bandpass sixth-order Butterworth filter with zero phase and a bandwidth of 25-400Hz.

[0142] Specifically, the audio signals that are usually collected contain not only noise from the environment and the heart sound acquisition equipment, but also interference from the subject's breathing vibrations, bowel sounds, skin friction, and heart and lung murmurs from other organs around the heart; in addition, useful heart sound signals are mainly concentrated between 25-400Hz.

[0143] The amplitude of all heart sound signals is normalized using the following formula:

[0144]

[0145] in, It is the heart sound signal after amplitude normalization. This is the original heart sound signal. This is because amplitude normalization can transform each data point of the heart sound signal into the interval [-1, 1].

[0146] The second heart sound component was segmented using a hidden Markov model based on logistic regression (Springer DB, Tarassenko L, Clifford G D. Logistic regression-HSMM-based heart soundsegmentation[J]. IEEE Transactions on Biomedical Engineering, 2015, 63(4): 822-832.), and the multiple second heart sound components of each example heart sound signal were spliced ​​together.

[0147] The heart sound signal after the above processing is processed by framing with a frame length of 25ms and an overlap rate of 50%.

[0148] Specifically, heart sound signals are non-stationary signals, while a long audio segment divided into many short frames can be considered stationary.

[0149] Add a Hamming window to each frame of heart sound signal:

[0150]

[0151] This is because windowing can enhance the continuity between the left and right ends of frames. Specifically, For Hamming window functions, For the sampling points of the signal, The total number of sampling points in one frame of signal;

[0152] The feature extraction module includes:

[0153] The time-frequency domain signal, also known as the fractional domain signal, is obtained by performing a fractional Fourier transform of the heart sound signal with an order of 0.9.

[0154] Modulo the fractional-domain spectrum of the heart sound signal;

[0155] Extract the Shannon entropy features of the one-dimensional fractional domain heart sound signal, that is, extract the Shannon entropy features of each frame of the fractional domain heart sound signal;

[0156] Multiple statistical functions are applied to the frame-level fractional-domain Fourier transform entropy features as the final features extracted for each heart sound signal.

[0157] Furthermore, the feature extraction module includes:

[0158] Let a fractional-domain heart sound signal be... The total number of sampling points is ;

[0159] Data sequence According to the following formula, it is divided into Duan got , :

[0160] ;

[0161] The maximum value of the heart sound signal in this frame was determined by comparison. minimum value ;

[0162] Establish a continuous data The range is ( , Similarly, divided into Duan got , ;

[0163] Calculate the segmented heart sound signal sequence Located in the corresponding Number of , then the first The formula for calculating the Shannon entropy of a segment is:

[0164]

[0165] Among them, let ;

[0166] Then the fractional-domain Fourier transform entropy of this frame of heart sound signal is:

[0167]

[0168] Or it can be expressed as:

[0169]

[0170] Thirteen statistical functions are applied to the frame-level fractional-domain Fourier transform entropy features as the final features extracted from each heart sound signal; therefore, the feature dimension of this invention is 13-dimensional. The 13 statistical functions are: minimum value, maximum value, range, mean, variance, standard deviation, skewness, peak value, first percentile, second percentile, third percentile, slope of the first linear regression fitting function, and offset.

[0171] The model building module trains the XGBoost classifier to build the model and evaluates the results using unweighted average recall.

[0172] The specific settings and adjustments for classifier parameters are as follows:

[0173] Select six parameters to be optimized: learning rate, maximum tree depth, minimum leaf node sample weight and min_child_weight, minimum loss function descent value gamma required for node splitting, proportion of random sampling per tree (subsample) and proportion of random sampling columns per tree (each column is a feature) (colsample_bytree), and set the range of grid search for each parameter;

[0174] Set the random seed parameter `random_state` to 42, and keep the other parameters at their default values.

[0175] Furthermore, a grid search is performed on each parameter of XGBoost, using unweighted average recall as the model's performance metric, to find the model that maximizes the unweighted average recall on the test set.

[0176] It also includes a model interpretation module.

[0177] The SHAP (SHapley Additive exPlanations) theory is used to calculate the SHAP values ​​for 13 statistical features. This is because the model will output different predictions when a feature is present and when it is absent; comparing the difference between the two predictions yields the contribution of that feature. After performing this operation on each subset of the 13 statistical features, the SHAP value for each feature is obtained by averaging these contributions.

[0178] The present invention also discloses an application of the above-mentioned interpretable heart sound abnormality recognition method and system based on fractional domain Fourier transform to heart sound abnormality recognition.

[0179] In summary, this invention proposes an interpretable method and system for identifying abnormal heart sounds based on fractional-domain Fourier transform, including preprocessing, feature extraction, model building, and model interpretation. Preprocessing sequentially includes: cropping, downsampling, filtering, amplitude normalization, cardiac cycle segmentation, framing, and windowing. Feature extraction is configured to first perform a fractional-domain Fourier transform on the preprocessed heart sounds, then extract frame-level Shannon entropy features of the one-dimensional fractional-domain heart sound signal, and calculate 13 statistical functions on the frame-level features as the final features. Model building uses an XGBoost classifier. Model interpretation uses SHAP (SHapley Additive exPlanation) to interpret the model. This invention is easy to implement, simple in method, has low feature dimensionality, fast model fitting, and interpretable model predictions.

[0180] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An interpretable heart sound anomaly recognition method based on fractional domain Fourier transform, characterized in that, The method comprises: Step S100, preprocessing the heart sound signal; Step S200, extracting the fractional domain Fourier transform entropy feature of the preprocessed heart sound signal; Step S210, performing fractional domain Fourier transform on the heart sound signal to obtain a time-frequency domain signal, also referred to as a fractional domain signal; Step S220, taking the fractional domain spectrum of the heart sound signal as a modulus; Step S230, extracting the Shannon entropy feature of the one-dimensional fractional domain heart sound signal, that is, extracting the Shannon entropy feature of each frame of the fractional domain heart sound signal respectively; Step S240, applying multiple statistical functions to the frame-level fractional domain Fourier transform entropy feature as the final feature extracted for each heart sound signal; Step S300, model establishment, that is, training an XGBoost classifier to establish a model, and evaluating the result by using the unweighted average recall: selecting multiple parameters that need to be optimized, setting the range of grid search for each parameter, setting the random seed value of the parameter, and keeping the remaining parameters at the default value; Performing grid search on the parameters of the XGBoost one by one, taking the unweighted average recall as the performance indicator of the model, and finding the model that makes the unweighted average recall of the test set maximum; Step S400, model interpretation, calculating the SHAP value of the statistical feature to obtain the SHAP value of each feature. 2.The interpretable heart sound anomaly recognition method based on fractional Fourier transform according to claim 1, wherein, The step S100 comprises: Step S110, cutting a part of the original heart sound signal as a heart sound sample for subsequent processing; Step S120, down-sampling the sampling frequency of the heart sound signal to a certain number of values; Step S130, filtering the heart sound signal by using a filter; Step S140, performing amplitude normalization on all heart sound signals; Step S150, segmenting the second heart sound component of the heart sound, and splicing multiple second heart sound components of each example heart sound signal respectively; Step S160, performing frame segmentation on the heart sound signal processed in step S150; Step S170, adding a Hamming window to each frame of the heart sound signal. 3.The fractionally Fourier transform-based interpretable heart sound anomaly recognition method of claim 1, wherein, The step S230 comprises: Step S231, set a frame score domain heart sound signal as , the total number of sampling points is ; Step S232, the data sequence is divided according to the following formula , :​ ； Step S233, comparing to obtain the maximum value of the heart sound signal of this frame , minimum value ; Step S234, a continuous data is established , ranging from , ), and similarly, divided into segments to obtain , ; Step S235, calculating the segmented heart sound signal sequence located in the corresponding number of , then the segment of the Shannon entropy formula is: ； wherein let ; Step S236, then the fractional domain Fourier transform entropy of this frame of heart sound signal is: ； Or expressed as: 。 4. An interpretable heart sound anomaly recognition system based on fractional Fourier transform, characterized by, The system comprises: A preprocessing module for preprocessing the heart sound signal; A feature extraction module for extracting the fractional domain Fourier transform entropy feature of the preprocessed heart sound signal; performing fractional domain Fourier transform on the heart sound signal to obtain a time-frequency domain signal, also referred to as a fractional domain signal; Taking the fractional domain spectrum of the heart sound signal as a modulus; Extracting the Shannon entropy feature of the one-dimensional fractional domain heart sound signal, that is, extracting the Shannon entropy feature of each frame of the fractional domain heart sound signal respectively; Applying multiple statistical functions to the frame-level fractional domain Fourier transform entropy feature as the final feature extracted for each heart sound signal; A model establishment module for training an XGBoost classifier to establish a model, and evaluating the result by using the unweighted average recall: selecting multiple parameters that need to be optimized, setting the range of grid search for each parameter, setting the random seed value of the parameter, and keeping the remaining parameters at the default value; Performing grid search on the parameters of the XGBoost one by one, taking the unweighted average recall as the performance indicator of the model, and finding the model that makes the unweighted average recall of the test set maximum; The model explanation module is configured to calculate SHAP values of the statistical features, and obtain SHAP values of each feature.

5. The interpretable heart sound anomaly recognition system based on fractional Fourier transform according to claim 4, characterized in that, The preprocessing module is configured to: cut a part of the original heart sound signal as a heart sound sample for subsequent processing; downsample the sampling frequency of the heart sound signal to a certain number of values; filter the heart sound signal by using a filter; perform amplitude normalization on all the heart sound signals; segment the second heart sound components of the heart sounds, and splice the second heart sound components of each example heart sound signal respectively; perform frame processing on the processed heart sound signals; add a Hamming window to each frame of heart sound signal.

6. The interpretable heart sound anomaly recognition system based on fractional Fourier transform according to claim 5, characterized in that, The feature extraction module is configured to: Let a frame of fractional domain heart sound signal be , and the total number of data points be ; The data sequence is divided into segments according to the following formula , : ； The maximum value of the heart sound signal of this frame is compared The minimum value ; establishing a continuous data segment , ranging from , , and similarly, dividing into segments results in , ; Computing a sequence of segmented heart sound signals located at the corresponding number of , then the shannon entropy calculation formula of the first segment is: ； wherein let ; The score domain Fourier transform entropy of this frame of heart sound signal is: ； or expressed as: 。

Images