Heart sound-based coronary heart disease detection method and device, storage medium and electronic equipment
Through the extraction and fusion of multi-domain and medical multi-domain feature, combined with one-dimensional heart sound deep learning features, the information loss problem caused by two-dimensional feature conversion in the existing technology is solved, the accuracy and efficiency of coronary heart disease detection is improved, and a Non-invasive detection methods.
Patent Information
- Application Number
- CN202411093477.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-06
AI Technical Summary
现有基于深度学习的心音特征提取和冠心病检测方法需要二维心音数据特征转换,导致一维心音数据特征捕获不稳定且可能丢失关键信息,影响检测精度和一致性。
Using multi-domain feature extraction model and medical multi-domain feature extraction model, combining one-dimensional heart sound deep learning features, denoising and cardiac cycle cleavage through wavelet transformation and logistic regression, fusing multi-domain and medical multi-domain features Perform detection and directly use original heart sound data to avoid information loss and enhance the capture of coronary heart disease characteristics.
It improves the accuracy and efficiency of coronary heart disease detection, enhances insight into coronary heart disease-related characteristics, and provides a non-invasive and effective detection method.
Smart Images

Figure CN119092096A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and is a method, device, storage medium and electronic equipment for detecting coronary heart disease based on heart sounds. Background Art
[0002] Auscultation of heart sounds depends on the subjective experience of cardiologists, and the test results of different experts may vary significantly. Therefore, building a coronary heart disease detection model based on heart sounds and realizing rapid screening of coronary heart disease can effectively reduce the reliance on subjective judgment of experts, improve the accuracy and consistency of detection, make early detection and treatment of coronary heart disease possible, and thus significantly improve the prognosis and quality of life of patients.
[0003] Current automatic heart sound signal analysis can be summarized into two types: one is to manually extract heart sound features and classify them based on machine learning; the other is to automatically complete heart sound feature extraction and classification through deep learning.
[0004] Manually extracted heart sound features are usually extracted from multiple domains such as time domain, frequency domain, time-frequency domain, wavelet, entropy, etc., and then the heart sound features are input into the classifier using machine learning for detection. Common classifiers include support vector machine, naive Bayes, k nearest neighbor and random forest. The manual feature extraction method relies on the combination of features from different domains, which may reveal disease-related features. However, this manual feature extraction method has the following limitations, including: (1) The manual feature extraction process is usually time-consuming and costly; (2) Feature selection depends largely on the researcher's experience and intuition, resulting in differences in feature extraction results between different researchers, affecting the consistency and repeatability of the final results; (3) Manually selected features may not fully capture all the key information of heart sounds, thereby limiting the performance of the coronary heart disease detection model.
[0005] In terms of automatic heart sound feature extraction and classification based on deep learning, the current mainstream method is to convert one-dimensional heart sound signals into two-dimensional feature maps as input to deep learning models to detect coronary heart disease. For example, studies have converted one-dimensional heart sound signals into two-dimensional (Mel-Frequency Cepstral Coefficients, MFCC) MFCC feature maps for classification, achieving 94.1% sensitivity and 88.7% specificity respectively; by converting heart sound signals into two-dimensional feature maps Mel spectrum and Log-Mel spectrum, 71.4% sensitivity and 95.1% specificity were achieved. This method has the following limitations: (1) Different two-dimensional feature maps have different abilities to capture the features of one-dimensional heart sound signals, which may lead to the loss of key information during the conversion process; (2) Since the selection of two-dimensional images is often based on empirical judgment, it affects the performance of the model. Summary of the invention
[0006] The present invention provides a coronary heart disease detection method, device, storage medium and electronic equipment based on heart sound, which are used for detecting coronary heart disease through heart sound data.
[0007] In view of this, the first aspect of the present application provides a method for detecting coronary heart disease based on heart sounds, comprising: Obtain the denoised heart sound data of the person to be predicted; The heart sound data is input into the coronary heart disease detection model to obtain the corresponding detection classification results, wherein the coronary heart disease detection model is obtained by deep learning training using a number of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection classification result, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model that fuses and classifies heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
[0008] Preferably, the process of constructing a coronary heart disease detection model includes: Obtaining a training set, a test set, and a validation set, wherein each sample data in the training set, the test set, and the validation set includes heart sound data and a label identifying a corresponding coronary heart disease detection classification result; The number of training times is set, and the preset coronary heart disease detection model is trained multiple times using the training set according to the number of training times to obtain the corresponding coronary heart disease detection model, wherein the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively; The validation set is used to adjust the model parameters of the coronary heart disease detection model, and the test set is used to test the coronary heart disease detection model in combination with the evaluation conditions, and the coronary heart disease detection model that meets the evaluation conditions is output.
[0009] Preferably, obtaining a training set, a test set, and a validation set includes: Acquire a number of sample data, wherein each sample data includes heart sound data and a label identifying a corresponding coronary heart disease detection classification result; Preprocess the heart sound data in the sample data; Perform cardiac cycle cutting on the heart sound data in the sample data; The sample data after cutting several cardiac cycles are divided into training set, test set and validation set according to the set ratio.
[0010] Preferably, the one-dimensional heart sound deep learning feature extraction model includes multiple Inception modules, multiple residual addition modules and a global pooling module; Multiple Inception modules are arranged in a sorted manner, and the output of the Inception module with a sequence number of 3 and its multiples is added to a residual value through a residual addition module, and the addition result is the input data of the next Inception module, wherein the residual value of the first residual addition module is the original heart sound data, and the residual values of all residual addition modules except the first residual addition module are the output data of the first residual addition module, wherein each Inception module includes a multi-scale convolutional layer, an inverted residual attention layer, a global pooling layer, and a discard layer; The output data of the last Inception module is input into the global pooling module for processing.
[0011] Preferably, the multi-domain features extracted by the multi-domain feature extraction model include: 12 time domain features, 8 cardiac cycle spectrum features, 6 entropy features, 10 energy ratio features, 8 statistical features, 8 spectrum features of 5-second heart sound data and 22 wavelet features.
[0012] Preferably, the medical multi-domain features extracted by the medical multi-domain feature extraction model include: 6 time domain features, 4 cardiac cycle spectrum features, 6 entropy features, 6 energy ratio features, 4 statistical features and 22 wavelet features of 5 seconds of heart sound data.
[0013] Preferably, the coronary heart disease detection method based on heart sound is characterized in that the step of obtaining the denoised heart sound data of the person to be predicted comprises: Acquire the heart sound data of the person to be predicted; The Daubechies 8 wavelet function in wavelet transform is used to denoise the heart sound data; A hidden semi-Markov model based on logistic regression is used to segment the cardiac cycle of the denoised heart sound data.
[0014] In view of this, the second aspect of the present application provides a coronary heart disease detection device based on heart sounds, characterized by comprising: A data acquisition unit, which acquires the denoised heart sound data of the person to be predicted; The detection and classification unit inputs the heart sound data into the coronary heart disease detection model to obtain the corresponding detection and classification results, wherein the coronary heart disease detection model is obtained by deep learning training using a plurality of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection and classification results, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
[0015] In view of this, the third aspect of the present application provides a storage medium, on which is stored a computer program that can be read by a computer, and the computer program is configured to execute a coronary heart disease detection method based on heart sounds when running.
[0016] In view of this, the fourth aspect of the present application provides an electronic device, including a processor, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the program includes instructions for executing steps in a coronary heart disease detection method based on heart sounds.
[0017] The present invention can extract features from various angles of the heart sound data to be detected through the model, that is, extract heart sound deep learning features, multi-domain features and medical multi-domain features, and then fuse the heart sound deep learning features with the multi-domain features and medical multi-domain features respectively, and then complete the detection and classification based on the fused features. This process directly uses the original heart sounds, retains the complete feature information in the heart sound data, and fuses different domain features (multi-domain features and medical multi-domain features) on the basis of the heart sound deep learning features, effectively avoiding the information loss caused by conversion to a two-dimensional image and enhancing the insight into the subtle changes in heart sounds caused by coronary heart disease, thereby more comprehensively capturing the features closely related to coronary heart disease, improving the accuracy and efficiency of coronary heart disease detection, and providing an effective and non-invasive detection method for coronary heart disease detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Attached Figure 1 The present invention is a flow chart of a detection method according to an embodiment of the present invention.
[0019] Attached Figure 2 The figure is a flow chart of a method for acquiring heart sound data in one embodiment of the present invention.
[0020] Attached Figure 3 The figure is a flow chart of a method for constructing a coronary heart disease detection model in one embodiment of the present invention.
[0021] Attached Figure 4Schematic diagram of the structure of a coronary heart disease detection model in one embodiment of the present invention.
[0022] Attached Figure 5 Schematic diagram of the structure of a one-dimensional heart sound deep learning feature extraction model in one embodiment of the present invention.
[0023] Attached Figure 6 FIG. 4 is a schematic diagram of the structure of ECA in one embodiment of the present invention.
[0024] Attached Figure 7 Schematic diagram of a time domain waveform of heart sound data in one embodiment of the present invention.
[0025] Attached Figure 8 It is a schematic diagram of a time domain waveform of heart sound data after noise reduction in one embodiment of the present invention.
[0026] Attached Fig. 9 It is a schematic diagram of a time domain waveform diagram of heart sound data after cardiac cycle segmentation processing in one embodiment of the present invention.
[0027] Attached Fig.10 It is a schematic diagram of the structure of the fusion classification model in one embodiment of the present invention.
[0028] Attached Fig.11 The figure is a schematic diagram of the device structure of an embodiment of the present invention. DETAILED DESCRIPTION
[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0030] The present application provides a method, device, storage medium and electronic device for detecting coronary heart disease based on heart sounds, which are used to solve the technical problem that the existing heart sound feature extraction and coronary heart disease classification detection methods based on deep learning require the conversion of two-dimensional heart sound data features and one-dimensional heart sound data features, thereby causing unstable capture of one-dimensional heart sound data features and loss of key information.
[0031] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0032] The embodiments of the present application are described in detail below.
[0033] Embodiment 1: As attached Figure 1 As shown, this embodiment discloses a coronary heart disease detection method based on heart sounds, comprising: Step S110, obtaining the denoised heart sound data of the person to be predicted.
[0034] The above steps are as follows Figure 2 As shown, specifically including: Step S111, obtaining the heart sound data of the person to be predicted; When implementing this step, 30 seconds of heart sound audio can be recorded at a sampling rate of 4000 Hz using a 3M Littmann 3200 electronic stethoscope. 30 seconds of heart sound audio can be recorded at a sampling rate of 4000 Hz using a 3M Littmann 3200 electronic stethoscope.
[0035] Step S112, using Daubechies 8 wavelet function in wavelet transform to perform denoising on the heart sound data; Denoising can improve data quality and provide a clearer and more accurate data basis for subsequent heart sound analysis. Therefore, this example requires denoising the heart sound data of the person being predicted.
[0036] Wavelet transform has advantages in time-frequency analysis, and can perform localized analysis of signals at different scales, thereby effectively capturing subtle features in heart sound data, and the multi-resolution analysis capability of wavelet transform enables it to adapt to mutation points in the signal, thereby efficiently suppressing noise. Therefore, this embodiment uses the Daubechies 8 wavelet function in wavelet transform to perform noise reduction processing on heart sound data, which can provide a moderate smoothing effect while retaining key features in the signal. Further, by setting the number of decomposition layers to 10, an in-depth analysis of the heart sound data at multiple levels is achieved, thereby dividing the data more finely, so as to more effectively weaken or eliminate noise components at each level and retain important information in the heart sound data.
[0037] Step S113, using a hidden semi-Markov model based on logistic regression to perform cardiac cycle segmentation on the denoised heart sound data.
[0038] Cardiac cycle segmentation is an important part of clinical heart disease diagnosis. It can accurately identify and isolate the first heart sound (S1) and the second heart sound (S2) in the cardiac cycle. S1 is usually caused by the vibration generated when the mitral valve and tricuspid valve are closed, marking the beginning of the heart's systolic period. S2 is caused by the vibration generated when the aortic valve and pulmonary valve are closed, marking the beginning of the heart's diastolic period. The interval between the two heart sounds is the systolic period, and the interval from S2 to the next cardiac cycle S1 is the diastolic period. These periodic events of heart sounds are the key to understanding the functional state of the heart, but their differences in time and characteristics increase the difficulty of analyzing heart sound signals. This embodiment uses a hidden semi-Markov model based on logistic regression to segment the heart sound signal within the cardiac cycle, and accurately locates and segmented the heart sound by analyzing the hidden state sequence and time characteristics of the heart sound signal.
[0039] Step S120, input the heart sound data into the coronary heart disease detection model to obtain the corresponding detection classification result, wherein the coronary heart disease detection model is obtained by deep learning training using a number of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection classification result, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
[0040] The present invention discloses a coronary heart disease detection method based on heart sounds, which can extract features from various angles of heart sound data to be detected through a model, that is, extract heart sound deep learning features, multi-domain features and medical multi-domain features, and then fuse the heart sound deep learning features and the multi-domain features and medical multi-domain features respectively, and then complete detection and classification based on the fused features. In this process, the original heart sounds are directly used to extract heart sound deep learning, which retains the complete feature information in the heart sound data, and different domain features (multi-domain features and medical multi-domain features) are respectively fused on the basis of the heart sound deep learning features, which can effectively avoid the information loss caused by conversion to a two-dimensional image, thereby more comprehensively capturing the features closely related to coronary heart disease and improving the accuracy and efficiency of coronary heart disease detection.
[0041] The multi-domain features extracted by the multi-domain feature extraction model include: 12 time domain features, 8 cardiac cycle spectrum features, 6 entropy features, 10 energy ratio features, 8 statistical features, 8 spectrum features of 5-second heart sound data and 22 wavelet features.
[0042] The medical multi-domain features extracted by the above medical multi-domain feature extraction model include: 6 time domain features, 4 cardiac cycle spectrum features, 6 entropy features, 6 energy ratio features, 4 statistical features and 22 wavelet features of 5-second heart sound data.
[0043] Among them, the time domain features (12×4, 6×4) are as follows: The time domain features of normal and abnormal heart sounds may contain features that are beneficial for the detection of coronary heart disease. The multi-domain features calculate the duration of the four states of S1, systole, S2, and diastole in the cardiac cycle, as well as the time ratio and amplitude ratio within the cardiac cycle. Considering that the blood flow velocity of the left coronary artery is the largest during diastole and the blood flow velocity of the right coronary artery is approximately equal during systole and diastole, the medical multi-domain features calculate some time domain features, as shown in Table 1.
[0044] Table 1 Extraction of time domain features for each cardiac cycle Feature Abbreviation describe Medical multi-domain features DurCc The sum of the durations of the four phases of the cardiac cycle √ DurS1 The interval from S1 to the next adjacent contraction DurSys The interval from systole to the next adjacent S2 √ DurS2 The interval from S2 to the next adjacent diastole DurDia The interval from diastole to the next adjacent S1 √ SysToCc Ratio of systolic duration to cardiac cycle duration √ DiaToCc Ratio of diastolic duration to cardiac cycle duration √ SysToDia Systolic to diastolic ratio √ SysToS1_Amp Ratio of mean systolic amplitude to S1 DiaToS2_Amp Ratio of mean diastolic amplitude to S2 S1ToCc_Amp Ratio of the mean amplitude of S1 to the cardiac cycle S2ToCc_Amp The ratio of the average amplitude of S2 to the cardiac cycle The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0045] Among them, the cardiac cycle spectrum characteristics (8×4, 4×4) specifically include: The multi-domain feature analyzes the spectrum characteristics of the four states of the cardiac cycle, setting 50 Hz as the low-frequency threshold and 200 Hz as the high-frequency threshold. Considering that the main frequency and center frequency of the heart sound signal after fast Fourier transform are usually 20-40 Hz, the medical multi-domain feature sets the low-frequency threshold to 20-150 Hz and the high-frequency threshold to 150-1000 Hz, and calculates the systolic and diastolic characteristics, as shown in Table 2.
[0046] Table 2 Extract all spectral features of each cardiac cycle Feature Abbreviation describe Medical multi-domain features HF_S1 The proportion of high frequency components in S1 LF_S1 The proportion of low-frequency components in S1 HF_S2 The proportion of high-frequency components in S2 LF_S2 The proportion of low-frequency components in S2 HF_Sys The proportion of high-frequency components during systole √ LF_Sys The proportion of low-frequency components during systole √ HF_Dia The proportion of high-frequency components in diastole √ LF_Dia The proportion of low-frequency components in diastole √ The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0047] Among them, the entropy features (6×4, 6×4) specifically include: The nonlinear and non-stationary characteristics of heart sound signals make it difficult for traditional linear analysis methods to accurately capture their dynamic characteristics. Entropy features can provide deeper insights into the intrinsic dynamics of heart sound signals by measuring the uncertainty and complexity of the system. Multi-domain features and medical multi-domain features jointly analyze the three entropy features of systole and diastole. The detailed information is shown in Table 3.
[0048] Table 3 Extract all entropy features for each cardiac cycle Feature Abbreviation describe Medical multi-domain features Samp_Sys Systolic sample entropy √ Samp_Dia Diastolic sample entropy √ Fuzzy_Sys Contraction cycle fuzzy entropy √ Fuzzy_Dia Diastolic fuzzy entropy √ Dist_Sys Systolic distribution entropy √ Dist_Dia Diastolic distribution entropy √ The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0049] (1) Sample entropy is a statistical measure used to quantify the regularity and complexity of time series. It evaluates the degree of synchronization of time series by analyzing the similarity patterns of time series data. It is calculated as follows: (1) Where: is the length of the heart sound signal, is the embedding dimension, is the threshold parameter, is the probability that any two fixed time windows match each other.
[0050] (2) Fuzzy entropy is similar to sample entropy, but a certain degree of fuzziness is taken into account when calculating similarity, making it more robust to small sample data or noisy data. and Fuzzy similarity between fuzzy entropy The membership function of is usually chosen to be an exponential decay function, calculated as follows: (2) Where: yes and The distance between them, the pattern length of SampEn and FuzzyEn is set to 2, and the matching tolerance Set to 0.2 times the input time series.
[0051] (3) Distribution entropy quantifies the uncertainty and complexity of time series data by analyzing its probability distribution. Distribution entropy uses empirical probability distribution functions (ePDF) to achieve global measurement of the distance matrix, avoiding the parameter dependence caused by local evaluation.
[0052] (3) Where: Represents a random variable The probability distribution function of Represents the number of bins of the histogram, the present invention Set to 2*8.
[0053] Among them, the energy ratio features (10×4, 6×4) specifically include: The energy ratio feature is used to evaluate the energy level of the heart sound signal. The multi-domain feature analyzes the energy ratio of different cardiac cycle states to the entire cardiac cycle, as well as the energy ratio between two states. Considering the possible presence of murmurs caused by coronary heart disease in the diastolic period, and the possible decrease in the amplitude of S1 and S2, the medical multi-domain feature calculates the characteristics of S1, S2, and diastolic period, as shown in Table 4. The energy calculation method is as follows: (4) Where: represents discrete time points, Indicates signal Length.
[0054] Table 4. Extraction of energy ratio features for each cardiac cycle Feature Abbreviation describe Medical multi-domain features En_S1Sys The ratio of S1 energy to systolic period energy En_S1Dia The ratio of S1 energy to diastole period energy √ En_S2Sys The ratio of S2 energy to systolic period energy En_S2Dia The ratio of S2 energy to diastole period energy En_DiaSys The ratio of diastole period energy to systolic period energy √ En_S1Cc The ratio of S1 energy to cardiac cycle energy √ En_SysCc The ratio of systolic period energy to cardiac cycle energy En_S2Cc The ratio of S2 energy to cardiac cycle energy √ En_DiaCc The ratio of diastole period energy to cardiac cycle energy √ En_S1S2Cc The ratio of the sum of S1 and S2 energies to cardiac cycle energy √ The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0055] Among them, the statistical features (8×4, 4×4) specifically include: In the analysis of heart sound signals, skewness and kurtosis are used to evaluate the asymmetry and peak distribution characteristics of heart sound signals at different stages of the cardiac cycle. Multi-domain features calculate the skewness and kurtosis of the four states of each cardiac cycle. Medical multi-domain features calculate the statistical features of S1 and S2. The skewness and kurtosis are calculated as follows: (5) (6) Where: and Represents time series The mean and standard deviation of .
[0056] Among them, the heart sound spectrum domain features (8×1, 0×0) specifically include: By analyzing the frequency components, energy distribution and other spectrum information in the heart sound signal, the spectrum features related to coronary heart disease can be extracted, which is beneficial for detecting coronary heart disease. The multi-domain feature sets 50Hz and 200Hz as the low-frequency threshold and high-frequency threshold respectively to calculate the spectrum of the entire heart sound signal. The medical multi-domain feature selects not to calculate this feature, as shown in Table 5. The entropy is calculated as follows: (7) Where: Represents the sequence length.
[0057] Table 5 Extract all spectral features of each 5-s heart sound signal Feature Abbreviation describe Medical multi-domain features Mean_spec The average spectrum value of the entire heart sound signal Sd_spec The standard deviation of the spectrum value of the entire heart sound signal Ske_spec The skewness of the spectrum value of the entire heart sound signal Kur_spect The kurtosis of the spectrum value of the entire heart sound signal HighF_ratio The proportion of high-frequency components in the spectrum value of the entire heart sound signal LowF_ratio The proportion of low-frequency components in the spectrum value of the entire heart sound signal HighToLow _ratio The ratio of high-frequency components to low-frequency components in the entire heart sound signal spectrum Ep_spec The entropy of the spectrum value of the entire heart sound signal The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0058] Among them, the wavelet features (22×1, 22×1) specifically include: Wavelet transform is an effective time-frequency analysis tool. Multi-domain features and medical multi-domain features When the sampling rate of the heart sound signal is 2000Hz, Daubechies 6 wavelets are used for three-layer decomposition, and the approximate coefficients and detail coefficients are calculated to obtain 9 wavelet features, including wavelet energy in different frequency bands and total energy. In order to deeply analyze the high-frequency and low-frequency parts of the 5-second heart sound signal, eight reconstruction coefficient vectors are calculated by three-level wavelet packet decomposition, and 13 wavelet packet features are extracted based on them, including total energy, energy in different frequency bands (Formula 1) and energy entropy (Formula 7), providing an in-depth perspective on the high-frequency analysis of heart sound signals. Table 6 gives the specific wavelet feature information.
[0059] Table 6 Wavelet feature extraction for each 5-s heart sound signal Feature Abbreviation describe Medical multi-domain features Wdec_ratio1 The wavelet approximation coefficient energy is applicable to the entire heart sound band in the range of 0 ~ 62.5 Hz √ Wdec_ratio2 The wavelet detail coefficient energy of the entire heart sound band is between 62.5 and 125 Hz √ Wdec_ratio3 The energy range of wavelet detail coefficient is 125 ~ 250hz, which is applicable to the entire heart sound band √ Wdec_ratio4 The wavelet detail coefficient energy range is 250 ~ 500hz, which is applicable to the entire heart sound band √ Wdec_ratio5 The wavelet detail coefficient energy is in the range of 500 to 1000 Hz, applicable to the entire heart sound band √ Wdec_total The total wavelet coefficient energy of the entire heart sound band is in the range of 0 to 1000 Hz √ Wdec_l_ratio The total energy of the wavelet coefficients of the entire heart sound band is less than 125 Hz √ Wdec_m_ratio The total energy of the wavelet coefficients of the entire heart sound band is between 125 and 500 Hz √ Wdec_h_ratio The total energy characteristics of the wavelet packet decomposition of the entire heart sound band are above 500 Hz √ Wpac_ratio1 Reconstruction coefficient energy across heart sounds in the range 0 to 125 Hz √ Wpac_ratio2 Energy reconstruction coefficient spans heart sounds in the range of 125 ~ 250hz √ Wpac_ratio3 Energy reconstruction coefficient spans heart sounds in the range of 250 ~ 375 Hz √ Wpac_ratio4 Energy reconstruction coefficient spans heart sounds in the range of 375 ~ 500 Hz √ Wpac_ratio5 Energy reconstruction coefficient spans heart sounds in the range of 500 ~ 625hz √ Wpac_ratio6 Energy reconstruction coefficient spans heart sounds in the range of 625 ~ 750hz √ Wpac_ratio7 Energy reconstruction coefficient spans heart sounds in the range of 750 ~ 875 Hz √ Wpac_ratio8 Energy reconstruction coefficient spans heart sounds in the range of 875 ~ 1000hz √ Wpec_total The total energy of the reconstruction coefficient of the entire heart sound in the range of 0 ~ 1000hz √ Wpac_l_ratio Total energy of the reconstruction coefficient of the whole heart sound segment below 125 Hz √ Wpac_m_ratio Total energy of reconstruction coefficients in the range of 125 ~ 500 Hz √ Wpac_h_ratio The reconstruction coefficient of the entire heart sound band is greater than the total energy of 500hz √ Wpac_entropy Energy Energy entropy of reconstruction coefficients of eight frequency bands in the whole heart sound band √ The multi-domain features are shown in the table. The third column "√" indicates medical multi-domain features.
[0060] Example 2: Figure 3 As shown, this embodiment discloses a coronary heart disease detection method based on heart sound, wherein the coronary heart disease detection model construction process includes: Step S210, obtaining a training set, a test set and a validation set, wherein each sample data in the training set, the test set and the validation set includes heart sound data and a label identifying the corresponding coronary heart disease detection classification result.
[0061] In the above steps, the training set, test set, and validation set are obtained, including: (1) obtaining a number of sample data, each of which includes heart sound data and a label identifying a corresponding coronary heart disease detection classification result; (2) Preprocessing the heart sound data in the sample data; (3) Perform cardiac cycle segmentation on the heart sound data in the sample data; (4) Divide the sample data after cutting several cardiac cycles into training set, test set and validation set according to the set ratio.
[0062] The specific methods of the above steps (2) and (3) are the same as those in Example 1 and will not be repeated here. The ratio in step (4) is set according to actual conditions.
[0063] Step S220, setting the number of training times, and using the training set to train the preset coronary heart disease detection model multiple times according to the number of training times to obtain the corresponding coronary heart disease detection model, wherein the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
[0064] The structure of the above coronary heart disease detection model is shown in the attached Figure 4 As shown, the details are as follows: Among them, the one-dimensional heart sound deep learning feature extraction model is shown in the attached Figure 5 As shown in the figure, it is a one-dimensional heart sound deep learning feature extraction model based on one-dimensional convolution, which combines the residual design concept of the Inception module and ResNet.
[0065] Specifically, it includes multiple Inception modules, multiple residual addition modules and global pooling modules; each Inception module includes a multi-scale convolution layer, an inverted residual attention layer, a global pooling layer, and a discard layer. The inverted residual attention layer is the Channel Attention Inversion Module (CAIM). CAIM is specially designed to enhance the processing effect after each convolution and pooling, aiming to ensure that the model fully utilizes the heart sound features extracted from different levels, thereby improving the efficiency of extracting favorable features for detecting coronary heart disease.
[0066] Among them, multiple Inception modules are arranged in a sorted manner, and the output of the Inception module with serial number 3 and its multiples is added to a residual value through a residual addition module, and the addition result is the input data of the next Inception module, wherein the residual value of the first residual addition module is the original heart sound data, and the residual values of all residual addition modules except the first residual addition module are the output data of the first residual addition module; the output data of the last Inception module is input to the global pooling module for processing.
[0067] The discard layer is Dropout, which is an effective regularization strategy suitable for processing heart sound signals containing a lot of redundant information. Therefore, Dropout is placed after multi-scale convolution to improve the generalization ability of the model and reduce the risk of overfitting.
[0068] Among them, ECA (Efficient Channel Attention, ECA) channel attention is used in the one-dimensional heart sound deep learning feature extraction model, and its structure is shown in the attached figure. Figure 6 shown.
[0069] In the above inverted residual attention layer, first, each input channel is independently weighted multiplied according to Formula 8. Input tensor ( represents the length of a one-dimensional feature, represents the number of channels) in the input channel (1 to ) with a one-dimensional convolution kernel (1 to ) to perform product, and sum all products to get the output tensor , where the size of the convolution kernel is 1 and the number of channels is Became , achieving dimensionality upgrade.
[0070] Next, apply batch normalization (BN) (Formula 9) and ReLU6 activation function (Formula 10); (8) Where: are input features, is the weight of the convolution kernel, is the output feature after the convolution operation, is the index of the feature vector, ranging from 1 to ,and and are the indices of the input and output channels respectively, and are followed by Represents a convolution operation.
[0071] (9) Where: is the input tensor, yes The mean of the channel, It is The variance of the channels, is a small constant to prevent division by zero errors, and are learnable parameters for scaling and translation, which are used later Represents a ReLU6 operation.
[0072] (10) Where: follow-up Represents a BN operation.
[0073] Subsequently, the tensor is slid across each channel according to Formula 11, taking into account the current position and surrounding context information. Use a one-dimensional convolution of size 3 for feature extraction and get the output tensor , this step is also followed by batch normalization and ReLU6 activation function: (11) Where: is the index along the spatial dimension, from -1 to 1 represents the convolution kernel size of 3; It is the index along the channel dimension of the input feature map, corresponding to all channels at each position; is the input feature map at position and Channel The tensor on , Is the convolution kernel from the input channel To output channel The weight of From -1 to 1 represents the current position and the previous and next positions. is the spatial position of the currently processed feature map, is the index of the input channel, is the index of the output channel.
[0074] Next, the tensor Enter the ECA channel attention module. Unlike other channel attention mechanisms, the ECA module believes that convolution has good cross-channel information acquisition capabilities, so capturing the dependencies of all channels is inefficient and unnecessary. All ECA only focuses on the input tensor Information interaction with k adjacent neighbors (channels).
[0075] ECA first performs a global average pooling operation on the input feature map, then performs a 1D convolution operation with a convolution kernel size of K, and obtains the weights of each channel through a Sigmoid activation function. , as shown in Formula 12: (12) Where: Represents the convolution kernel size is K ( ), Stands for Sigmoid.
[0076] Convolution kernel size It can be determined adaptively by formula 13.
[0077] (13) Where: Represents the odd number closest to the inner number.
[0078] Obtained in ECA The channel attention weight of , according to formula 14, the input tensor Multiply it with the attention weight to get the output tensor ,This process aims to assign higher weights to the more critical features in the current ,coronary heart disease detection process, thereby building a detection model with better ,performance.
[0079] (14) After the ECA module, the tensor Use a one-dimensional convolution with a size of 1 for dimensionality reduction and batch normalization. According to Formula 15, the output tensor is .
[0080] (15) Where: Represents a one-dimensional convolution with a size of 1.
[0081] Finally, residual connections are made according to A and B of CAIM, and the specific implementation strategy depends on the relationship between the step value and the number of input and output feature channels. When the step is 1 and the number of input and output feature channels is the same, the network directly adds the output of the current ECA module to the initial input of the network, using direct addition to achieve residual connections, maintaining the integrity of information and the simplicity of the model. When the step is 2 or the number of feature channels does not match, the network first performs one-dimensional convolution and batch normalization on the input to adjust its size and number of channels to match the output of the ECA module, and then performs addition operations to complete the residual connection.
[0082] Among them, the multi-domain feature extraction model and the medical multi-domain feature extraction model extract multi-domain features and medical multi-domain features from heart sound data respectively. Their structures are the same and are both based on a one-dimensional CNN network structure.
[0083] Among them, the fusion classification model is as follows Fig.10 As shown in the figure, the deep learning features of heart sounds are combined with multi-domain features and medical multi-domain features and input into MLP (multi-layer perceptron) for further processing. Subsequently, through a specific fusion strategy, the same-dimensional features are merged to form a comprehensive feature representation. In order to avoid overfitting and optimize model performance, the Dropou method is applied to the fused features of the two models. Finally, the Sigmoid activation function is used to map the features into two-class probabilities to achieve coronary heart disease detection.
[0084] Step S230, using the validation set to adjust the model parameters of the coronary heart disease detection model, using the test set to test the coronary heart disease detection model in combination with the evaluation conditions, and outputting the coronary heart disease detection model that meets the evaluation conditions.
[0085] The above evaluation conditions can be set using four different evaluation indicators, including accuracy, sensitivity, specificity, and area under the curve (AUC).
[0086] Example 3: The coronary heart disease detection method based on heart sounds disclosed in the embodiment of the present invention is used for the following applications, specifically including: (1) Collect the heart sound data of patients with coronary heart disease and display them in time domain waveform diagrams, as shown in the attached figure. Figure 7 As shown in the figure, the horizontal axis is time in seconds, and the vertical axis is amplitude in millivolts.
[0087] (2) De-noise the heart sound data. The results are shown in the attached figure. Figure 8 shown.
[0088] (3) Cardiac cycle segmentation and processing. The results are shown in the attached figure. Fig. 9 shown.
[0089] (4) As attached Fig. 9 The heart sound data shown are respectively input into various detection models integrating CAIM module, multi-domain feature extraction, and medical multi-domain feature extraction, and the accuracy, sensitivity, specificity and AUC value are introduced as evaluation indicators to obtain the evaluation analysis table shown in the following table.
[0090] Table 7 Evaluation analysis table CAIM Multi-domain features Medical multi-domain features Accuracy(%) Sensitivity(%) Specificity(%) AUC(%) × × × 78.667 77.659 80.706 88.338 √ × × 83.453 85.628 79.049 90.727 × √ × 86.506 89.985 79.719 93.203 × × √ 84.745 91.227 72.065 89.866 √ √ × 87.861 90.668 82.383 94.704 √ × √ 85.599 88.036 80.832 92.741 It can be seen from the above table that various detection models integrated with CAIM module, multi-domain feature extraction, and medical multi-domain feature extraction have improved accuracy and AUC. Therefore, the coronary heart disease detection method based on heart sound disclosed in the present invention can effectively increase the accuracy of detection.
[0091] Among them, the following is a partial display of multi-domain features and medical multi-domain features extracted from 5-second heart sounds.
[0092]
[0093] (a) Multi-domain features
[0094] (b) Medical multi-domain features Embodiment 4: As attached Fig.11 As shown, this embodiment discloses a coronary heart disease detection device based on heart sound, comprising: A data acquisition unit, which acquires the denoised heart sound data of the person to be predicted; The detection and classification unit inputs the heart sound data into the coronary heart disease detection model to obtain the corresponding detection and classification results, wherein the coronary heart disease detection model is obtained by deep learning training using a plurality of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection and classification results, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
[0095] Embodiment 5: This embodiment discloses a storage medium, on which a computer program readable by a computer is stored, and the computer program is configured to execute a coronary heart disease detection method based on heart sounds when running.
[0096] The memory here may include but is not limited to: a USB flash drive, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and other media that can store computer programs.
[0097] Example 6: This embodiment discloses an electronic device, including a processor, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the program includes instructions for executing steps in a coronary heart disease detection method based on heart sounds.
[0098] The processor may be a central processing unit (CPU), a general purpose processor, a digital signal processor (DSP), an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. It may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like. The memory may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk.
[0099] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The schemes in the embodiments of the present application may be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.
[0100] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0101] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0102] Although the present invention has been described in detail above with general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements may be made thereto based on the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.
Claims
1. A method for detecting coronary heart disease based on heart sound, It is characterized in that include: Obtain the denoised heart sound data of the person to be predicted; The heart sound data is input into the coronary heart disease detection model to obtain the corresponding detection classification results, wherein the coronary heart disease detection model is obtained by deep learning training using a number of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection classification result, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model that fuses and classifies heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
2. The method for detecting coronary heart disease based on heart sound according to claim 1, It is characterized in that The coronary heart disease detection model construction process includes: Obtaining a training set, a test set, and a validation set, wherein each sample data in the training set, the test set, and the validation set includes heart sound data and a label identifying a corresponding coronary heart disease detection classification result; The number of training times is set, and the preset coronary heart disease detection model is trained multiple times using the training set according to the number of training times to obtain the corresponding coronary heart disease detection model, wherein the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively; The validation set is used to adjust the model parameters of the coronary heart disease detection model, and the test set is used to test the coronary heart disease detection model in combination with the evaluation conditions, and the coronary heart disease detection model that meets the evaluation conditions is output.
3. The method for detecting coronary heart disease based on heart sounds according to claim 2, It is characterized in that The obtaining of the training set, the test set and the validation set comprises: Acquire a number of sample data, wherein each sample data includes heart sound data and a label identifying a corresponding coronary heart disease detection classification result; Preprocess the heart sound data in the sample data; Perform cardiac cycle cutting on the heart sound data in the sample data; The sample data after cutting several cardiac cycles are divided into training set, test set and validation set according to the set ratio.
4. The method for detecting coronary heart disease based on heart sounds according to claim 1, 2 or 3, It is characterized in that The crown one-dimensional heart sound deep learning feature extraction model includes multiple Inception modules, multiple residual addition modules and a global pooling module; Multiple Inception modules are arranged in a sorted manner, and the output of the Inception module with a sequence number of 3 and its multiples is added to a residual value through a residual addition module, and the addition result is the input data of the next Inception module, wherein the residual value of the first residual addition module is the original heart sound data, and the residual values of all residual addition modules except the first residual addition module are the output data of the first residual addition module, wherein each Inception module includes a multi-scale convolutional layer, an inverted residual attention layer, a global pooling layer, and a discard layer; The output data of the last Inception module is input into the global pooling module for processing.
5. The method for detecting coronary heart disease based on heart sound according to claim 1, 2 or 3, It is characterized in that The multi-domain features extracted by the multi-domain feature extraction model include: 12 time domain features, 8 cardiac cycle spectrum features, 6 entropy features, 10 energy ratio features, 8 statistical features, 8 spectrum features of 5-second heart sound data and 22 wavelet features.
6. The method for detecting coronary heart disease based on heart sound according to claim 1, 2 or 3, It is characterized in that The medical multi-domain features extracted by the medical multi-domain feature extraction model include: 6 time domain features, 4 cardiac cycle spectrum features, 6 entropy features, 6 energy ratio features, 4 statistical features and 22 wavelet features of 5-second heart sound data.
7. The method for detecting coronary heart disease based on heart sounds according to any one of claims 1 to 6, It is characterized in that The step of obtaining the denoised heart sound data of the person to be predicted includes: Acquire the heart sound data of the person to be predicted; The Daubechies 8 wavelet function in wavelet transform is used to denoise the heart sound data; A hidden semi-Markov model based on logistic regression is used to segment the cardiac cycle of the denoised heart sound data.
8. A coronary heart disease detection device based on heart sound using the method according to any one of claims 1 to 7, It is characterized in that include: A data acquisition unit, which acquires the denoised heart sound data of the person to be predicted; The detection and classification unit inputs the heart sound data into the coronary heart disease detection model to obtain the corresponding detection and classification results, wherein the coronary heart disease detection model is obtained by deep learning training using a plurality of sample data, each group of sample data in the multiple groups of sample data includes heart sound data and a label identifying the corresponding coronary heart disease detection and classification results, and the coronary heart disease detection model includes a one-dimensional heart sound deep learning feature extraction model for extracting heart sound deep learning features, a multi-domain feature extraction model for extracting multi-domain features, a medical multi-domain feature extraction model for extracting medical multi-domain features, and a fusion classification model for fusing and classifying heart sound deep learning features with multi-domain features and medical multi-domain features respectively.
9. A storage medium, It is characterized in that The storage medium stores a computer program that can be read by a computer, and the computer program is configured to execute the coronary heart disease detection method based on heart sounds as described in any one of claims 1 to 7 when running.
10. An electronic device, It is characterized in that It comprises a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the coronary heart disease detection method based on heart sounds as described in any one of claims 1 to 7.
