Electrocardiogram and heart sound signal fusion analysis system

Abstract

The invention discloses an electrocardio and heart sound signal fusion analysis system, particularly relates to the technical field of signal analysis, and comprises a preprocessing module, a feature extraction module, a dimension reduction and correlation analysis module, a prediction model construction module, a heart rate influence correction module and a joint feature generation module. Through joint analysis of electrocardio and heart sound signals, the system combines multi-dimensional information of electrophysiological characteristics and mechanical vibration characteristics, the limitation of single signal diagnosis is made up, the splitting degree, intensity and other characteristics of the heart sound signals are quantified through wave packet extraction and information entropy calculation, and subjective dependence of traditional auscultation is reduced; the key waveform features of the electrocardiosignals strengthen the pathological relevance through accurate identification and time sequence analysis, and the sensitivity of key parameters is further optimized through combination of feature generation and dynamic weight distribution, so that the model can capture early pathological changes more accurately, and a reliable basis is provided for quantitative diagnosis of cardiovascular diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of signal analysis, and in particular to a fusion analysis system for electrocardiogram and heart sound signals. Background Art

[0002] Cardiovascular disease is one of the leading causes of death worldwide. Its early diagnosis is crucial to improving patient prognosis. Electrocardiogram (ECG) and cardiac sound (PCG) are core signals for evaluating cardiac function, reflecting cardiac status from two dimensions: electrophysiological activity and mechanical vibration, respectively. In traditional medical practice, ECG and PCG are usually analyzed independently, failing to fully utilize their complementarity, resulting in fragmented diagnostic information and difficulty in comprehensively evaluating complex pathologies.

[0003] In recent years, with the advancement of biomedical signal processing technology, researchers have begun to explore the potential of multimodal signal fusion. However, first, the non-stationarity of PCG signals and environmental noise interference make feature extraction difficult, and traditional wave packet detection methods rely on empirical thresholds and lack robustness. Second, there is a lot of redundancy and noise in the high-dimensional features of ECG and PCG signals, and direct calculation of correlation is susceptible to interference. In addition, the impact of heart rate changes on time-related features (such as ejection time and heart sound duration) is not effectively separated, resulting in the masking of pathological features. In this context, the development of an efficient electrocardiogram and heart sound signal fusion analysis system has become an urgent need. Summary of the Invention

[0004] The main purpose of the present invention is to provide an electrocardiogram and heart sound signal fusion analysis system, which can effectively solve the problems of diagnostic information fragmentation, high-dimensional noise interference and difficulty in nonlinear relationship modeling caused by independent analysis of electrocardiogram and heart sound signals.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A system for fusion analysis of electrocardiogram and heart sound signals, wherein the system is configured to:

[0007] Preprocessing module: used to remove the baseline of the ECG signal and perform bandpass filtering on the heart sound signal;

[0008] Feature extraction module: used to extract the band characteristics of the ECG signal, the wave packet characteristics of the heart sound signal, and auxiliary physiological characteristics, including the wave packet extraction function and the heart rate correction function;

[0009] Dimensionality reduction and correlation analysis module: uses an autoencoder to reduce the dimensionality of the extracted features and calculates the encoded feature correlation matrix;

[0010] Prediction model building module: Screen features based on the correlation matrix and build decision trees and curvilinear regression models to predict heart sound features;

[0011] Heart rate impact correction module: This module removes the heart rate dependency of duration features by adjusting the power parameter of the relationship between features and heart rate.

[0012] Joint feature generation module: performs arithmetic averaging on multiple similar features of heart sound signals to generate joint representation features.

[0013] Preferably, the preprocessing module includes an ECG baseline removal unit, a heart sound filtering unit and a signal interception unit. The ECG baseline removal unit uses a recursive moving window algorithm to detect and eliminate the baseline drift of the ECG signal. The heart sound filtering unit is configured as a bandpass filter from 20Hz to 200Hz to retain the mechanical vibration frequency band of the heart sound signal. The signal interception unit aligns and segments the ECG and heart sound signals according to the cardiac cycle identifier.

[0014] Preferably, the feature extraction module includes a wave packet extraction unit, a frequency domain energy calculation unit, an information entropy calculation unit and an ECG band identification unit, and the wave packet extraction unit is based on a wave packet extraction function:

[0015]

[0016] The range of the cardiac sound wave packet is determined according to the formula, where th = 0.8. The frequency domain energy calculation unit extracts the frequency domain energy of the cardiac sound wave packet in the frequency band of 100 Hz to 200 Hz. The electrocardiogram band identification unit locates the starting and ending points of the P wave, QRS complex and T wave through the gradient window detection method.

[0017] Preferably, the dimensionality reduction and correlation analysis module includes an autoencoder structure unit, a correlation matrix calculation unit and a feature screening unit. The input layer of the autoencoder structure unit is 29 dimensions, the hidden layer is 128-ReLU and 64-ReLU, and the output layer is a 29-dimensional reconstruction layer. The correlation matrix calculation unit calculates the Pearson correlation coefficient matrix based on the encoded low-dimensional features, and the feature screening unit screens significantly correlated features according to the absolute value threshold of the correlation coefficient (≥0.4).

[0018] Preferably, the heart rate influence correction module includes a power parameter adjustment unit and a specific parameter storage unit, and the iterative optimization formula y=x / Rate 8 The power parameter 8 in makes the Pearson coefficient of the corrected feature and the heart rate approach zero, and the specific parameter storage unit stores the optimized parameters 8 for different time-length features.

[0019] Preferably, the prediction model construction module includes a curve regression unit, a fine decision tree unit and a cross-validation unit. The curve regression unit uses a polynomial fitting method to establish a linear mapping relationship between electrocardiographic features and heart sound features. The fine decision tree unit configures the minimum number of leaf node samples to be 4, the input is R_A1, T_A, T_R_D, and the output is S1_pow or st_D. The cross-validation unit performs 5-fold cross-validation and calculates RMSE for model tuning.

[0020] Preferably, the joint feature generation module includes an intensity joint unit, a split degree joint unit and a dynamic weight allocation unit. The intensity joint unit performs arithmetic averaging on s1_A1, s1_A3, and s1_f_E to generate S1_pow. The split degree joint unit performs arithmetic averaging on s1_length, s1_length_D, and s1_S to generate S1_split. The dynamic weight allocation unit dynamically adjusts the weight distribution of the joint features according to the Pearson correlation coefficient between the features.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. Through the joint analysis of ECG and heart sound signals, this system systematically combines multi-dimensional information from electrophysiological and mechanical vibration characteristics, overcoming the limitations of single-signal diagnosis. Characteristics such as the splitting degree and intensity of heart sound signals are quantified through wave packet extraction and information entropy calculation, reducing the subjective reliance of traditional auscultation. Key waveform features of the ECG signal are precisely identified and time-series analyzed to enhance pathological relevance. Joint feature generation and dynamic weight allocation further optimize the sensitivity of key parameters, enabling the model to more accurately capture early pathological changes and providing a reliable basis for the quantitative diagnosis of cardiovascular disease.

[0023] 2. The present invention uses autoencoder dimensionality reduction technology to map high-dimensional features to low-dimensional space, effectively removing noise and redundant information, while capturing nonlinear correlations, significantly improving the efficiency of feature analysis. The heart rate correction function removes the heart rate dependence of duration features through mathematical modeling, allowing the features to more directly reflect pathophysiological changes. In addition, feature screening and weight allocation based on the correlation matrix enhance the model's ability to focus on key pathological features, improve the scientific nature and interpretability of feature engineering, and provide a clear physiological basis for clinical decision-making.

[0024] 3. The present invention achieves wide adaptability to different physiological states and pathological scenarios through signal preprocessing and dynamic feature fusion. Precise signal interception and cycle alignment technology reduces delay interference and ensures data consistency; dynamic weight distribution automatically adjusts feature importance according to disease type, enabling the model to flexibly adapt to the diagnostic needs of various cardiovascular diseases. Experiments have shown that the model performs well in complex pathological analyses such as cardiac ejection function and valvular lesions, providing clinical with an efficient and comprehensive auxiliary diagnostic tool and promoting the practical application of combined electrocardiogram and heart sound diagnosis technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0026] Figure 2 Schematic diagram of the correspondence between ECG and PCG in the cardiac cycle of the present invention;

[0027] Figure 3 This is a schematic diagram of the ECG identification position of the present invention;

[0028] Figure 4 It is a schematic diagram of the correlation matrix of the present invention;

[0029] Figure 5 This is a schematic diagram of the fitting results of 22-5 of the present invention;

[0030] Figure 6 Schematic diagram of the St_D fitting results of the present invention;

[0031] Figure 7 This is a schematic diagram of the S1_pow fitting results of the present invention;

[0032] Figure 8 This is a schematic diagram of the S1_split fitting results of the present invention;

[0033] Figure 9 This is a schematic diagram of the S2_pow fitting results of the present invention. DETAILED DESCRIPTION

[0034] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0035] Example 1, as Figure 1 As shown, a system for fusion analysis of electrocardiogram and heart sound signals is configured as follows:

[0036] Preprocessing module: used to remove the baseline of the ECG signal and perform bandpass filtering on the heart sound signal;

[0037] Feature extraction module: used to extract the band characteristics of the ECG signal, the wave packet characteristics of the heart sound signal, and auxiliary physiological characteristics, including the wave packet extraction function and the heart rate correction function;

[0038] Dimensionality reduction and correlation analysis module: uses an autoencoder to reduce the dimensionality of the extracted features and calculates the encoded feature correlation matrix;

[0039] Prediction model building module: Screen features based on the correlation matrix and build decision trees and curvilinear regression models to predict heart sound features;

[0040] Heart rate impact correction module: This module removes the heart rate dependency of duration features by adjusting the power parameter of the relationship between features and heart rate.

[0041] Joint feature generation module: performs arithmetic averaging on multiple similar features of heart sound signals to generate joint representation features.

[0042] See Figure 2-3 The preprocessing module includes an ECG baseline removal unit, a heart sound filtering unit and a signal interception unit. The ECG baseline removal unit uses a recursive moving window algorithm to detect and eliminate the baseline drift of the ECG signal. The heart sound filtering unit is configured as a bandpass filter from 20Hz to 200Hz to retain the mechanical vibration frequency band of the heart sound signal. The signal interception unit aligns and segments the ECG and heart sound signals according to the cardiac cycle identifier.

[0043] Furthermore, the ECG baseline removal unit mentioned above uses a recursive moving window algorithm (window length 200 ms, corresponding to 1600 sampling points) to calculate the signal mean window by window and generate a baseline curve. The baseline-free ECG signal is obtained by subtracting the baseline from the original signal. The heart sound filtering unit designs a Butterworth second-order bandpass filter (20 Hz-200 Hz) to filter out low-frequency respiratory noise and high-frequency interference in the PCG signal and retain the mechanical vibration frequency band. The signal interception unit aligns and segments the signal according to the cardiac cycle based on the R peak position of the ECG (Feature 13 in Table 2) and the S1 starting point of the PCG (Feature 1 in Table 1).

[0044] See Figure 3 The feature extraction module includes a wave packet extraction unit, a frequency domain energy calculation unit, an information entropy calculation unit, and an ECG band identification unit. The wave packet extraction unit is based on the wave packet extraction function:

[0045]

[0046] The range of the cardiac sound wave packet is determined according to the above formula, where th = 0.8. The frequency domain energy calculation unit extracts the frequency domain energy of the cardiac sound wave packet in the frequency band of 100Hz to 200Hz, and the ECG band identification unit locates the starting and ending points of the P wave, QRS complex and T wave through the gradient window detection method.

[0047] The inner sound of the wave packet is normalized by the following square:

[0048]

[0049] To avoid numerical problems in logarithmic calculations, replace 0 with a small positive number ε=1e-10$, and then calculate the information entropy of Xnormal as follows:

[0050] S=-∑iXnormal(i)·log(Xnormal(i))

[0051] Furthermore, the frequency domain energy calculation unit is used to perform FFT transformation on the signal in the wave packet and calculate the energy in the 100 Hz-200 Hz frequency band (refer to features 3 and 9 in Table 1).

[0052] The ECG band identification unit uses the gradient window detection method to locate the P wave, QRS complex and T wave:

[0053] P wave positioning: Slide the window in the interval before the R peak and calculate the gradient change. The gradient mutation point is marked as P_start. The formula is:

[0054]

[0055] QRS complex and T wave: Use a recursive moving window to detect the gradient extreme point and determine the band range (see Table 2, features 13-23)

[0056] See Table 1-3 below for details:

[0057]

[0058] Table 1: Feature matrix definition

[0059]

[0060]

[0061] Table 2: ECG feature extraction

[0062]

[0063]

[0064] Table 3: Auxiliary feature extraction

[0065] See Figure 4The dimensionality reduction and correlation analysis module includes an autoencoder structure unit, a correlation matrix calculation unit and a feature screening unit. The input layer of the autoencoder structure unit is 29 dimensions, the hidden layer is 128-ReLU and 64-ReLU, and the output layer is a 29-dimensional reconstruction layer. The correlation matrix calculation unit calculates the Pearson correlation coefficient matrix based on the encoded low-dimensional features. The feature screening unit screens significantly correlated features according to the absolute value threshold of the correlation coefficient (≥0.4).

[0066] Furthermore, the encoder structure unit has a 29-dimensional input layer (refer to feature 3 in Table 1), 128-ReLU and 64-ReLU hidden layers, and the output layer reconstructs the input features. The training parameters are Adam optimizer, initial learning rate 0.005, and gradient threshold 1;

[0067] Correlation matrix calculation unit: calculates the Pearson correlation coefficient matrix of the encoded low-dimensional features and selects features ≥ 0.4;

[0068] Feature screening unit: According to the simplified matrix in Table 4, ECG features such as R_A1 and T_A and PCG features such as s1_S and s2_fE are selected for modeling.

[0069]

[0070]

[0071] Table 4: Simplified ECG and PCG feature correlation matrix

[0072] The heart rate impact correction module includes a power parameter adjustment unit and a specific parameter storage unit, and it iteratively optimizes the formula:

[0073] y=x / Rate 8

[0074] The power parameter 8 in the formula is used to make the Pearson coefficient of the corrected feature and the heart rate approach zero, and the specific parameter storage unit stores the optimized parameter 8 for different time-length features.

[0075] Furthermore, the power parameter adjustment unit adopts the formula:

[0076] y=x / Rate s

[0077] The s value is iteratively optimized by this formula. For example, s = 1.213 for s1_length. After 10,000 iterations, the Pearson coefficient with heart rate drops to 1e-4.

[0078] The specific parameter storage unit is used to store the optimized s values of different features (such as s = -0.395 of s2_ength, Table 3) for subsequent analysis and call.

[0079] See Figure 5-9 The prediction model construction module includes a curve regression unit, a fine decision tree unit and a cross-validation unit. The curve regression unit uses a polynomial fitting method to establish a linear mapping relationship between ECG features and heart sound features. The fine decision tree unit configures the minimum number of leaf nodes to be 4, with inputs of R_A1, T_A, and T_R_D, and outputs of S1_pow or st_D. The cross-validation unit performs 5-fold cross-validation and calculates RMSE for model tuning.

[0080] Furthermore, the curve regression unit performs polynomial fitting on the strongly linearly correlated features (such as features 22 and 5 in Table 4):

[0081] f(x)=0.4371x+0.1237

[0082] R 2 =0.253

[0083] Refined decision tree unit: input R_A1, T_A and other features, output S1_poW, the minimum number of leaf node samples is set to 4 to prevent overfitting;

[0084] After 5-fold cross-validation of the cross-validation unit, the RMSE of the ejection duration prediction model was 0.148.

[0085] See Table 5 for details:

[0086] Input x Output y R-squared RMSE R_A1, R_t1_D, T_A, T_R_D, J_T1_k st_D 0.66 0.14854 R_A1, R_A2, T_A S1_pow 0.80 0.11165 R_t1, J_T1, rate_k S1_split 0.4 0.14628 R_A1, R_A2, T_A S2_pow 0.56 0.12313

[0087] Table 5: Fitting results for different inputs and outputs

[0088] The joint feature generation module includes an intensity joint unit, a split degree joint unit and a dynamic weight allocation unit. The intensity joint unit performs arithmetic averaging on s1_A1, s1_A3 and s1_f_E to generate S1_pow. The split degree joint unit performs arithmetic averaging on s1_length, s1_length_D and s1_S to generate S1_split. The dynamic weight allocation unit dynamically adjusts the weight distribution of the joint features according to the Pearson correlation coefficient between the features.

[0089] Furthermore, S1_A1, S1_A3, and S1_fE are input into the strength joint unit, and the arithmetic average of the above three features is performed to generate the joint feature S1_pow:

[0090]

[0091] Multi-dimensional feature fusion reduces the noise interference of single features and enhances the robust representation of heart sound intensity;

[0092] The split degree joint unit inputs the features s1_length, _length_D, and S1_S, performs arithmetic averaging on the above three features, and generates the joint feature S1_split:

[0093]

[0094] The complex characteristics of heart sound splitting are fully characterized by integrating time domain, heart rate correction and information entropy features;

[0095] The dynamic weight allocation unit extracts correlation coefficients, that is, the correlation coefficients between each sub-feature and the ECG features in Table 1-2. For example, the correlation coefficient between s1_S and R_t1_D is -0.519, and the correlation coefficient between s1_length and R_t1_D is -0.519. It then assigns higher weights to sub-features with strong correlations and reduces the weights of sub-features with weak correlations. Finally, through dynamic weight allocation, it highlights heart sound parameters that are strongly correlated with ECG features, thereby improving the model's sensitivity to pathological features.

[0096] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for fusion analysis of electrocardiogram and heart sound signals, characterized in that: the system for fusion analysis of electrocardiogram and heart sound signals is configured as follows: Preprocessing module: used to remove the baseline of the ECG signal and perform bandpass filtering on the heart sound signal; Feature extraction module: used to extract the band characteristics of the ECG signal, the wave packet characteristics of the heart sound signal, and auxiliary physiological characteristics, including the wave packet extraction function and the heart rate correction function; Dimensionality reduction and correlation analysis module: uses an autoencoder to reduce the dimensionality of the extracted features and calculates the encoded feature correlation matrix; Prediction model building module: Screen features based on the correlation matrix and build decision trees and curvilinear regression models to predict heart sound features; Heart rate impact correction module: This module removes the heart rate dependency of duration features by adjusting the power parameter of the relationship between features and heart rate. Joint feature generation module: performs arithmetic averaging on multiple similar features of heart sound signals to generate joint representation features.

2. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The preprocessing module includes an ECG baseline removal unit, a heart sound filtering unit and a signal interception unit. The ECG baseline removal unit uses a recursive moving window algorithm to detect and eliminate the baseline drift of the ECG signal. The heart sound filtering unit is configured as a bandpass filter from 20Hz to 200Hz to retain the mechanical vibration frequency band of the heart sound signal. The signal interception unit aligns and segments the ECG and heart sound signals according to the cardiac cycle identifier.

3. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The feature extraction module includes a wave packet extraction unit, a frequency domain energy calculation unit, an information entropy calculation unit and an ECG band identification unit. The wave packet extraction unit is based on the wave packet extraction function: The range of the cardiac sound wave packet is determined according to the formula, where th = 0.

8. The frequency domain energy calculation unit extracts the frequency domain energy of the cardiac sound wave packet in the frequency band of 100 Hz to 200 Hz. The electrocardiogram band identification unit locates the starting and ending points of the P wave, QRS complex and T wave through the gradient window detection method.

4. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The dimensionality reduction and correlation analysis module includes an autoencoder structure unit, a correlation matrix calculation unit and a feature screening unit. The input layer of the autoencoder structure unit is 29 dimensions, the hidden layer is 128-ReLU and 64-ReLU, and the output layer is a 29-dimensional reconstruction layer. The correlation matrix calculation unit calculates the Pearson correlation coefficient matrix based on the encoded low-dimensional features, and the feature screening unit screens significantly correlated features according to the absolute value threshold of the correlation coefficient (≥0.4).

5. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The heart rate impact correction module includes a power parameter adjustment unit and a specific parameter storage unit, and the iterative optimization formula is: y=x / Rate 8 The power parameter 8 in the formula is used to make the Pearson coefficient of the corrected feature and the heart rate approach zero, and the specific parameter storage unit stores the optimized parameters 8 for features of different durations.

6. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The prediction model construction module includes a curve regression unit, a fine decision tree unit and a cross-validation unit. The curve regression unit uses a polynomial fitting method to establish a linear mapping relationship between electrocardiogram features and heart sound features. The fine decision tree unit configures the minimum number of leaf nodes to be 4, the input is R_A1, T_A, T_R_D, and the output is S1_pow or st_D. The cross-validation unit performs 5-fold cross-validation and calculates RMSE for model tuning.

7. The electrocardiogram and heart sound signal fusion analysis system according to claim 1, characterized in that: The joint feature generation module includes an intensity joint unit, a split degree joint unit and a dynamic weight allocation unit. The intensity joint unit performs arithmetic averaging on s1_A1, s1_A3, and s1_f_E to generate S1_pow. The split degree joint unit performs arithmetic averaging on s1_length, s1_length_D, and s1_S to generate S1_split. The dynamic weight allocation unit dynamically adjusts the weight distribution of the joint features according to the Pearson correlation coefficient between the features.

Images