Coronary heart disease non-invasive early warning system and method based on electrocardiogram and heart sound time sequence parameter fusion

Abstract

The invention discloses a coronary heart disease noninvasive early warning system and method based on electrocardiogram and heart sound time sequence parameter fusion, and relates to the technical field of computer medical assistance. The method is used for solving the problems of inaccurate parameter measurement and insufficient risk assessment precision caused by signal time sequence asynchronization in coronary heart disease screening. The method comprises the following steps: firstly, synchronously collecting electrocardio and heart sound signals in a stepped exercise load process, extracting morphological characteristics of a ventricular depolarization wave group and a first heart sound, and realizing accurate time sequence alignment by constructing a cost matrix and dynamic planning; then, coronary artery function parameters are extracted from the alignment signals, and dynamic trend indexes are calculated; constructing a graph structure by using the dynamic trend indexes, fusing spatial-temporal features by using graph convolution and time convolution, and outputting risk levels of different stenosis degrees; and finally, an early warning index is generated through abnormal index extraction and back propagation optimization, and early monitoring and dynamic evaluation of the coronary heart disease risk are realized.

Description

Technical Field

[0001] This invention relates to the field of computer-aided medical technology, specifically to a non-invasive early warning system and method for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters. Background Technology

[0002] Coronary artery disease (CAD) is the leading cause of cardiovascular disease mortality worldwide, and early detection and intervention are crucial for reducing the incidence of acute myocardial infarction. With the accelerating aging of society and changes in lifestyle, the incidence of CAD continues to rise, creating an urgent need for efficient and accurate non-invasive screening technologies. Traditional CAD diagnostic methods have limitations in identifying early lesions and cannot meet the practical needs of primary healthcare institutions and large-scale population screening.

[0003] Current non-invasive clinical screening methods suffer from several technical limitations: electrocardiograms (ECGs) can only reflect electrophysiological changes following myocardial ischemia; while phonocardiography (PCO) can reveal valvular activity abnormalities, it struggles to differentiate coronary artery disease from other heart conditions; resting state parameter fluctuations fail to reflect potential coronary blood flow abnormalities, making it difficult to capture early lesion characteristics; ECG and phonocardiographic features are often analyzed independently, failing to reflect the temporal coupling relationship of cardiac electromechanical activity; diagnostic results are primarily qualitative, unable to quantify the degree of coronary artery stenosis or changes in cardiac function; and phonocardiographic analysis is highly dependent on physician experience, resulting in poor consistency among operators and hindering standardized assessment. These issues limit the accuracy and reliability of existing methods in early screening and dynamic monitoring. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a non-invasive early warning system and method for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters, thus solving the problems mentioned in the background.

[0005] To achieve the above objectives, this invention provides the following technical solution: a non-invasive early warning system for coronary artery disease based on the fusion of ECG and heart sound temporal parameters, comprising the following modules: a signal synchronous acquisition module, which synchronously acquires ECG and heart sound signals during a stepped exercise load, extracts the local morphological descriptors of the ECG ventricular depolarization group and the first heart sound, constructs a cost matrix between a reference template and the signal segment to be aligned, finds the path with the minimum cumulative cost through a dynamic programming algorithm, and adjusts the relative position of the ECG ventricular depolarization initiation point and the peak point of the first heart sound; and a dynamic trend analysis module, which extracts coronary function parameters from the temporally aligned signals, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion, records parameter values ​​at each stage of the exercise load, and calculates each parameter through linear fitting. The system outputs dynamic trend indicators, including the slope of the change in the number of parameters and the magnitude of the change in statistical parameters from resting to peak load. The graded early warning module uses these dynamic trend indicators as node features, constructs edge weights based on the Pearson correlation coefficient between coronary function parameters, aggregates spatial features through graph convolutional layers, extracts temporal features through temporal convolutional layers, and merges spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesions. The risk level is determined based on this probability distribution. The feedback optimization module extracts abnormal dynamic trend indicators based on the four-category probability distribution, generates early warning indicators by combining core function measurements, constructs a loss function based on the difference between the early warning indicators and the true label of stenosis severity, and optimizes the edge weights and node feature weights of the graph convolutional layers in the graded early warning module through a backpropagation algorithm.

[0006] Furthermore, during the stepped exercise load, ECG and heart sound signals were simultaneously acquired, and local morphological descriptors of the ventricular depolarization group and the first heart sound were extracted. The specific process for constructing the cost matrix between the reference template and the signal segment to be aligned is as follows: A single cardiac cycle with the best signal quality was selected as the reference template. The third-order differential extremum sequence of the ventricular depolarization group and the wavelet packet energy distribution of the first heart sound in that cycle were extracted as morphological descriptors. On the signal segment to be aligned, the same type of morphological descriptor was extracted through a sliding window, with the window length consistent with the reference template. The morphological difference cost between the reference template and the window to be aligned was calculated, including the Euclidean distance of the ventricular depolarization group extremum sequence and the KL divergence of the wavelet packet energy distribution of the first heart sound. The two morphological difference costs were weighted and fused to construct a complete cost matrix, with the ventricular depolarization group morphological cost weight set to 0.6 and the first heart sound morphological cost weight set to 0.4.

[0007] Furthermore, the specific process of adjusting the relative position of the ECG ventricular depolarization start point and the first heart sound peak point by finding the path with the minimum cumulative cost using a dynamic programming algorithm is as follows: Initialize the cumulative cost matrix, set the cumulative cost of the first row and first column to the corresponding morphological difference cost, calculate the minimum cumulative cost element by element using a dynamic programming algorithm, the cumulative cost of each position is equal to the current morphological difference cost plus the minimum cumulative cost among the three adjacent positions to the left, above and upper left; start backtracking from the lower right corner of the cumulative cost matrix to find the path that minimizes the cumulative cost, this path represents the optimal alignment relationship between the reference template and the signal segment to be aligned, and adjust the relative position of the ECG ventricular depolarization start point and the first heart sound peak point of the signal segment to be aligned according to the time offset of the optimal alignment path to achieve sub-sampling precision time alignment.

[0008] Furthermore, the dynamic trend analysis module includes the following steps: Based on the time-aligned signal, four coronary function parameters are calculated respectively: the interval from the onset of the ventricular depolarization wave to the peak value of the aortic valve closure component of the heart sound is measured as the electromechanical delay; the ratio of the time from the first heart sound to the second heart sound to the time from the second heart sound to the next first heart sound is calculated as the systolic-diastolic time ratio; the ratio of the amplitude of the first heart sound to the amplitude of the second heart sound is calculated as the heart sound intensity ratio; and the ratio of the standard deviation to the mean of the systolic time over five consecutive cardiac cycles is calculated as the systolic function dispersion. The numerical sequences of the coronary function parameters are recorded during the resting period, the stable period of each load level, and the recovery period of the exercise load. A linear fit is performed on the numerical sequence of each parameter throughout the exercise load, and the slope of the fitted line is extracted as the trend characteristic of the parameter. The magnitude of the change in the value of each coronary function parameter from the resting period to the peak load period is calculated, and the magnitude and slope of the change are combined to form a dynamic trend index.

[0009] Furthermore, the dynamic trend index is used as a node feature, and the edge weights are constructed using the Pearson correlation coefficient between coronary artery function parameters. Spatial features are aggregated through graph convolutional layers, and temporal features are extracted through temporal convolutional layers. The specific process is as follows: The dynamic trend index is constructed as a four-dimensional feature vector, where each dimension corresponds to the change feature of a coronary artery function parameter. The Pearson correlation coefficient matrix between the four coronary artery function parameters is calculated, and this matrix is ​​used as the adjacency matrix of the graph structure to construct a complete parameter relationship graph. The parameter relationship graph is input into a two-layer graph convolutional network, and spatial features are aggregated through the neighbor node information transmission mechanism. The first graph convolution uses the ReLU activation function, and the second graph convolution uses the linear activation function. At the same time, the time series of the dynamic trend index is input into the temporal convolutional layer, and temporal features are extracted through an expanded causal convolutional architecture. Gradient vanishing is avoided through residual connections.

[0010] Furthermore, the system integrates spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesion. The specific process for determining the risk level based on the probability distribution is as follows: The spatial features output by the graph convolutional layer and the temporal features output by the temporal convolutional layer are concatenated to form a fused feature vector. The fused feature vector is then input into a fully connected layer for dimensionality reduction, and a batch normalization layer is used to accelerate model convergence. The dimensionality-reduced features are then input into a softmax function to calculate the probability distribution of the four categories, where the sum of the probability values ​​is 1. The category corresponding to the maximum value in the probability distribution is selected as the final risk level. When the maximum probability value is lower than a preset threshold, the system is marked as pending re-examination.

[0011] Furthermore, the specific process of extracting abnormal dynamic trend indicators based on the four-category probability distribution and generating early warning indicators by combining them with core function measurement values ​​is as follows: Based on the four-category probability distribution, the dominant risk level is determined, and the abnormal dynamic trend indicators corresponding to the risk level are screened. The core function measurement values ​​corresponding to the abnormal indicators are extracted, including the specific values ​​of electromechanical delay at peak load, the change amplitude relative to the resting value, and the specific values ​​of contraction function dispersion during the recovery period. The abnormal dynamic trend indicators are combined with the core function measurement values ​​to generate a multi-dimensional early warning indicator vector. The early warning indicator vector is normalized to eliminate the dimensional differences between different parameters.

[0012] Furthermore, based on the difference between the warning indicator and the ground truth label of the narrowing degree, a loss function is constructed. The specific process of optimizing the edge weights and node feature weights of the graph convolutional layer in the hierarchical warning module using the backpropagation algorithm is as follows: The difference between the warning indicator and the ground truth label of the narrowing degree is calculated using the cross-entropy loss function; the gradient of the loss function with respect to the graph convolutional layer parameters is calculated using the backpropagation algorithm; the edge weight matrix of the graph convolutional layer is updated using the adaptive moment estimation algorithm, while the node feature weight matrix is ​​optimized; a weight decay strategy is used to prevent overfitting; the change in the loss value for each optimization is recorded, and the optimization process stops when the loss value converges.

[0013] A non-invasive early warning method for coronary artery disease based on the fusion of ECG and heart sound timing parameters includes the following steps: During a stepped exercise load, electrocardiogram (ECG) signals and heart sound signals are acquired simultaneously, and the ventricular depolarization waveforms and heart sounds are extracted. The local morphology descriptor of the peak is used to construct a cost matrix between the reference template and the signal segment to be aligned. A dynamic programming algorithm is then used to find the path with the minimum cumulative cost, and the ECG Q wave initiation and heart sound are adjusted. The relative positions of the peaks; Coronary function parameters, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion, are extracted from the time-aligned signal. Parameter values ​​are recorded at each stage of exercise load. The slope of each parameter change is calculated through linear fitting. The magnitude of parameter change from rest to peak load is statistically analyzed, and dynamic trend indicators are output. Dynamic trend indicators are used as node features, and edge weights are constructed using the Pearson correlation coefficient between coronary function parameters. Spatial features are aggregated through graph convolutional layers, and temporal features are extracted through temporal convolutional layers. Spatial and temporal features are fused to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery disease. The risk level is determined based on the probability distribution. Anomaly dynamic trend indicators are extracted based on the four-class probability distribution, and early warning indicators are generated by combining the core function measurement values. A loss function is constructed based on the difference between the early warning indicators and the true value label of the narrowness. The edge weights and node feature weights of the graph convolutional layer in the hierarchical early warning module are optimized by the backpropagation algorithm.

[0014] The present invention has the following beneficial effects: (1) The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound time parameters has the following beneficial effects in terms of signal synchronization acquisition module and dynamic trend analysis module: By synchronously acquiring ECG and heart sound signals during the step-by-step exercise load, and extracting the local morphological descriptors of the ECG ventricular depolarization wave group and the first heart sound, constructing the cost matrix between the reference template and the signal segment to be aligned, and using dynamic programming algorithm for optimal alignment, high-precision time synchronization of electromechanical events is achieved; further, coronary artery function parameters, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio and systolic function dispersion, are extracted from the time-aligned signals, and parameter values ​​are recorded at different exercise load stages. Combined with linear fitting, the slope and amplitude of each parameter change are calculated to form dynamic trend indicators, thereby continuously and comprehensively reflecting the dynamic evolution of coronary artery function under exercise load, and providing reliable indicators for early detection of functional abnormalities in coronary heart disease.

[0015] (2) The non-invasive early warning method for coronary heart disease based on the fusion of ECG and heart sound temporal parameters uses dynamic trend indicators as node features and constructs a graph convolutional network based on the correlation between coronary function parameters to achieve the fusion of spatial and temporal features. It can output the four-class probability distribution of healthy status, mild stenosis, severe stenosis and left main coronary artery disease, and determine the risk level accordingly. It extracts abnormal dynamic trend indicators by combining core function measurement values, and optimizes the edge weights and node feature weights of the graph convolutional layer by constructing a loss function and using the backpropagation algorithm to achieve dynamic optimization of the early warning indicators, making the risk judgment of the system more accurate and stable under multi-factor conditions.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] Figure 1 This is a flowchart of the non-invasive early warning system for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters of the present invention.

[0018] Figure 2 This is a flowchart of the non-invasive early warning method for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters according to the present invention. Detailed Implementation

[0019] This application presents a non-invasive early warning system and method for coronary artery disease based on the fusion of ECG and heart sound timing parameters. This addresses the problems of low signal alignment accuracy, insufficient dynamic functional assessment, and limited multi-parameter comprehensive analysis capabilities in existing non-invasive screening technologies for early detection of coronary artery functional abnormalities. The system can simultaneously acquire ECG and heart sound signals during exercise, achieving high-precision time synchronization of electromechanical events. Combined with dynamic trend analysis and intelligent risk grading, it provides reliable indicators for the early identification of functional abnormalities in coronary artery disease.

[0020] The overall concept of the solution in this application embodiment is as follows: High-quality ECG and heart sound data are acquired through a signal synchronization acquisition module, and key morphological features are extracted. Dynamic programming algorithm is used to achieve optimal alignment of ECG and heart sound signals. Multiple coronary artery function parameters are calculated in the dynamic trend analysis module to form dynamic trend indicators. Risk grading is achieved through a graded early warning module that combines graph convolution and temporal convolution algorithms, and abnormal indicators are dynamically corrected using a feedback optimization module, thereby providing continuous, comprehensive, and quantifiable coronary artery function assessment and early warning.

[0021] Please see Figure 1This invention provides a technical solution: a non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound temporal parameters, comprising the following modules: a signal synchronous acquisition module, which synchronously acquires ECG and heart sound signals during a stepped exercise load, extracts the local morphological descriptors of the ECG ventricular depolarization wave group and the first heart sound, constructs a cost matrix between a reference template and the signal segment to be aligned, finds the path with the minimum cumulative cost through a dynamic programming algorithm, and adjusts the relative position of the ECG ventricular depolarization initiation point and the peak point of the first heart sound; and a dynamic trend analysis module, which extracts coronary function parameters from the temporally aligned signals, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion, records parameter values ​​at each stage of the exercise load, and calculates the changes of each parameter through linear fitting. The slope, a statistical parameter, represents the magnitude of change from resting to peak load, outputting dynamic trend indicators. The graded early warning module uses these dynamic trend indicators as node features, constructs edge weights based on the Pearson correlation coefficient between coronary function parameters, aggregates spatial features through graph convolutional layers, extracts temporal features through temporal convolutional layers, and merges spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesions. The risk level is determined based on this probability distribution. The feedback optimization module extracts abnormal dynamic trend indicators based on the four-category probability distribution, generates early warning indicators by combining core function measurements, constructs a loss function based on the difference between the early warning indicators and the true label of stenosis severity, and optimizes the edge weights and node feature weights of the graph convolutional layers in the graded early warning module through a backpropagation algorithm.

[0022] In this implementation plan, the signal synchronization acquisition module is used to simultaneously acquire electrocardiogram (ECG) signals and heart sound signals during a stepped exercise load, achieving unified acquisition of multimodal physiological data. The ventricular depolarization complex (QRS complex) in the ECG signal reflects the ventricular depolarization process and is a major indicator of cardiac electrical activity; the first heart sound (… This corresponds to the mechanical vibration signal generated at the end of atrial contraction and the closure of the mitral and tricuspid valves. This module extracts the QRS complex and... The system extracts local morphological descriptors of peaks (such as differential extremum sequence, wavelet packet energy distribution, etc.) and constructs a cost matrix between a reference template and the signal segment to be aligned, quantifying the similarity between signal segments and providing basic data for subsequent electromechanical event alignment. The dynamic trend analysis module extracts coronary functional parameters from the time-aligned signal, including electromechanical delay (from the Q wave origin of the ECG signal to the heart sound). Peak time interval), systolic-diastolic time ratio (heart sounds) Peak Peak time and the next The ratio of peak intervals), the ratio of heart sound intensity ( Peak amplitude and The module analyzes the peak-to-amplitude ratio and systolic function dispersion (the ratio of the standard deviation to the mean of systolic time in a continuous cardiac cycle). By recording parameter values ​​at each stage of exercise load and performing linear fitting on the numerical sequence, the slope and amplitude of each parameter change are extracted to form dynamic trend indicators, reflecting the dynamic evolution of coronary artery function under different exercise loads. The module's innovation lies in coupling electrical and mechanical signals for analysis, capturing early functional abnormalities. The graded early warning module uses dynamic trend indicators as node features and constructs edge weights using the Pearson correlation coefficient between coronary artery function parameters to form a functional parameter relationship graph. Spatial node features are aggregated through a graph convolution layer, and temporal evolution features are extracted through a temporal convolution layer. By fusing spatial and temporal information, a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesion is output, thereby determining the risk level. The module's innovation lies in introducing the concept of a graph neural network, combining the relationships between multiple parameters with time-series changes to achieve comprehensive multi-dimensional feature analysis and improve the accuracy of risk assessment. The feedback optimization module analyzes the four-class probability distribution output by the hierarchical early warning module, extracts abnormal dynamic trend indicators, and generates early warning indicators by combining them with core function measurement values. A loss function is constructed based on the difference between the early warning indicators and the ground truth labels of narrowness. The edge weights and node feature weights of the graph convolutional layer are optimized using the backpropagation algorithm, enabling the system to dynamically and adaptively adjust under complex conditions with multiple factors. The innovation of this module lies in introducing a feedback mechanism into the graph convolutional network, achieving closed-loop optimization between prediction results and actual measurements, and improving the system's adaptability and reliability under different individuals and motion loads.

[0023] Specifically, the process of simultaneously acquiring ECG and heart sound signals during a stepped exercise load, extracting local morphological descriptors of the ventricular depolarization group and the first heart sound, and constructing a cost matrix between the reference template and the signal segment to be aligned is as follows: A single cardiac cycle with the best signal quality is selected as the reference template. The third-order differential extremum sequence of the ventricular depolarization group and the wavelet packet energy distribution of the first heart sound within that cycle are extracted as morphological descriptors. On the signal segment to be aligned, morphological descriptors of the same type are extracted using a sliding window, with the window length consistent with the reference template. The morphological difference cost between the reference template and the window to be aligned is calculated, including the Euclidean distance of the ventricular depolarization group extremum sequence and the KL divergence of the wavelet packet energy distribution of the first heart sound. The two morphological difference costs are weighted and fused to construct a complete cost matrix, with the ventricular depolarization group morphological cost weight set to 0.6 and the first heart sound morphological cost weight set to 0.4.

[0024] In this implementation scheme, during the stepped exercise load, the system simultaneously acquires electrocardiogram (ECG) signals and heart sound signals. First, the single cardiac cycle with the best signal quality is selected as the reference template. From the reference template, the third-order differential extremum sequence of the ventricular depolarization wave group is extracted as a morphological descriptor, reflecting the local morphological changes of ventricular depolarization. Simultaneously, the wavelet packet energy distribution of the first heart sound is extracted as a heart sound morphological descriptor, representing the mechanical vibration energy characteristics when the heart valves close. On the signal segment to be aligned, morphological descriptors of the same type are extracted using a sliding window method, with the window length consistent with the reference template. The morphological difference cost is calculated; the morphological difference cost of the ventricular depolarization wave group is calculated using Euclidean distance. ;in, This represents the cost of the ventricular depolarization wave group morphology difference between the m-th point of the reference template and the n-th point of the window to be aligned. This is the sequence of the p-th extreme points in the reference template; The sequence of extreme points corresponding to the window to be aligned; The total number of extreme point sequences. The cost of the first heart sound morphological difference is calculated using KL divergence: ;in, This represents the cost of the first heart sound morphological difference between the m-th point of the reference template and the n-th point of the window to be aligned. The q-th component of the wavelet packet energy distribution is used as a reference template. The energy distribution components corresponding to the window to be aligned; This represents the total number of wavelet packet decomposition levels. The weighted fusion of the cost differences between the two morphologies forms a complete cost matrix. ;in, This represents the combined cost of point m and point n. The weighting coefficient for the morphological cost of ventricular depolarization waveforms is 0.6; The first heart sound morphology cost weighting coefficient is publicly set to 0.4. This step achieves a quantitative representation of the morphological differences between electrical and mechanical signals, providing a foundation for dynamic programming alignment.

[0025] Specifically, the process of finding the path with the minimum cumulative cost and adjusting the relative position of the ECG ventricular depolarization start point and the first heart sound peak point using a dynamic programming algorithm is as follows: Initialize the cumulative cost matrix, set the cumulative cost of the first row and first column to the corresponding morphological difference cost, and calculate the minimum cumulative cost element by element using a dynamic programming algorithm. The cumulative cost of each position is equal to the current morphological difference cost plus the minimum cumulative cost among the three adjacent positions to the left, above, and upper left. Starting from the lower right corner of the cumulative cost matrix, backtrack to find the path that minimizes the cumulative cost. This path represents the optimal alignment relationship between the reference template and the signal segment to be aligned. Based on the time offset of the optimal alignment path, adjust the relative position of the ECG ventricular depolarization start point and the first heart sound peak point of the signal segment to be aligned to achieve sub-sampling precision time alignment.

[0026] In this implementation scheme, the cumulative cost matrix is ​​first initialized. Set the value of the first row and first column of the matrix to the corresponding cost matrix. The value of . For each element in the matrix, the cumulative cost is calculated using the following formula: ;in, Indicates the reference template number Point and signal segment to be aligned The cumulative cost of points; The cost is for the corresponding morphological differences; , , These are the cumulative costs for the left, top, and adjacent top-left positions, respectively. This calculation ensures that the cumulative cost at each point is the minimum path cost from the starting point to that point. Then, starting from the bottom right corner of the cumulative cost matrix, backtracking is performed to find the path with the minimum cumulative cost. This path represents the optimal alignment between the reference template and the signal segment to be aligned. Based on the time offset of the optimal alignment path, the phase positions of the ventricular depolarization start point and the first heart sound peak of the signal segment to be aligned are adjusted to achieve sub-sampling precision time alignment, thereby ensuring high-precision synchronization of electromechanical events.

[0027] Specifically, the dynamic trend analysis module includes the following steps: Based on the time-aligned signal, four coronary function parameters are calculated respectively: the interval from the onset of the ventricular depolarization wave to the peak value of the aortic valve closure component of the heart sound is measured as the electromechanical delay; the ratio of the time from the first heart sound to the second heart sound to the time from the second heart sound to the next first heart sound is calculated as the systolic-diastolic time ratio; the ratio of the amplitude of the first heart sound to the amplitude of the second heart sound is calculated as the heart sound intensity ratio; and the ratio of the standard deviation to the mean of the systolic time over five consecutive cardiac cycles is calculated as the systolic function dispersion. The numerical sequences of the coronary function parameters are recorded during the resting period, the stable period of each load level, and the recovery period of the exercise load. A linear fit is performed on the numerical sequence of each parameter throughout the exercise load process, and the slope of the fitted line is extracted as the trend characteristic of the parameter. The magnitude of the change in the value of each coronary function parameter from the resting period to the peak load period is calculated, and the magnitude and slope of the change are combined to form a dynamic trend index.

[0028] In this implementation scheme, based on the time-aligned signal, four coronary function parameters are first calculated. Electromechanical delay (EMD) represents the time interval from the onset of the central ventricular depolarization wave in the electrocardiogram signal to the peak value of aortic valve closure in the heart sound, reflecting the temporal relationship between cardiac depolarization and mechanical contraction events; the systolic-diastolic time ratio (CSDR) is determined by the first heart sound (…). ) to the second heart sound ( ) time and To the next The ratio of heart sound intensity to diastolic time reflects the proportion of cardiac systolic and diastolic time within a cardiac cycle; the heart sound intensity ratio (SIR) is... Peak amplitude and The ratio of peak amplitudes represents the change in mechanical energy generated during the closure of heart valves; the systolic dispersion (SDR) is the ratio of the standard deviation to the mean of systolic time over five consecutive cardiac cycles, quantifying the degree of fluctuation in systolic time and reflecting the consistency of ventricular contraction. Numerical sequences of coronary artery function parameters were recorded at each stage of exercise load (resting period, stable period of each load level, and recovery period). Linear fitting was performed on the numerical sequence of each parameter; the fitting formula can be expressed as: ;in, This represents the value of the k-th coronary artery function parameter at time t; To fit the slope of the straight line, representing the trend of this parameter's change; The intercept reflects the initial level of the parameter; These correspond to EMD, CSDR, SIR, and SDR, respectively. Fitting slope This is used to quantify the trend of each parameter changing with exercise load. Further calculations are performed to determine the magnitude of change in each coronary function parameter from the resting period to the peak load period. This represents the absolute magnitude of how the parameter changes with increasing exercise load. Combining the magnitude of change with the slope of the linear fit forms the Dynamic Trend Index (DTI), which can be expressed as: ;in, This is the dynamic trend indicator for the k-th parameter; and The weighting coefficients can be publicly set to 0.5 and 0.5 to ensure a comprehensive reflection of the trend slope and the magnitude of change. The dynamic trend index comprehensively reflects the dynamic evolution of coronary artery function throughout the entire exercise load process, providing a quantitative basis for subsequent graded early warning.

[0029] Specifically, the process of using dynamic trend indicators as node features, constructing edge weights with the Pearson correlation coefficient between coronary artery function parameters, aggregating spatial features through graph convolutional layers, and extracting temporal features through temporal convolutional layers is as follows: Dynamic trend indicators are constructed as four-dimensional feature vectors, where each dimension corresponds to the change feature of a coronary artery function parameter. The Pearson correlation coefficient matrix between the four coronary artery function parameters is calculated, and this matrix is ​​used as the adjacency matrix of the graph structure to construct a complete parameter relationship graph. The parameter relationship graph is input into a two-layer graph convolutional network, where spatial features are aggregated through a neighbor node information transfer mechanism. The first graph convolutional layer uses the ReLU activation function, and the second graph convolutional layer uses a linear activation function. Simultaneously, the time series of dynamic trend indicators is input into the temporal convolutional layer, where temporal features are extracted through an expanded causal convolutional architecture, and gradient vanishing is avoided through residual connections.

[0030] In this implementation scheme, firstly, each Dynamic Trend Indicator (DTI) is constructed as a four-dimensional feature vector. Each dimension corresponds to the variation characteristics of a coronary function parameter (electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion), used to represent the overall dynamic performance of the patient's cardiac function. The Pearson correlation coefficient matrix among the four coronary function parameters is calculated and used as the adjacency matrix of the graph structure to form a parameter relationship graph. This adjacency matrix represents the strength of functional correlation between nodes, where the correlation coefficient ranges from [1,1] and can be transformed into non-negative edge weights through linear mapping. The parameter relationship graph is input into a two-layer graph convolutional network (GCN), aggregating spatial features through a neighbor node information transfer mechanism. The first layer of graph convolution uses the ReLU activation function, and the calculation formula can be expressed as: ;in, This is the node feature matrix output from the first layer; This is the normalized adjacency matrix; This is the node feature matrix (i.e., the DTI vector). This is the convolution weight matrix for the first layer graph; The ReLU activation function is used; the index sign is not repeated from the parameter in the previous step. The second layer of graph convolution uses a linear activation function, calculated as follows: ;in, This is the spatial feature matrix output by the second layer; This is the second-layer weight matrix. This step fully integrates the spatial features of interrelated coronary functional parameters through inter-node information propagation. Simultaneously, the time series of dynamic trend indicators are input into the Temporal Convolutional Network (TCN), and temporal features are extracted through dilated causal convolution. The convolution formula is expressed as: ;in, This represents the output of the l-th convolutional layer at time t. This is the value of the input from the previous layer after a delay of dk; These are the kernel weights; For bias; The kernel length is [length]. This is the expansion factor. Through residual connections, the input of the previous layer is added to the convolutional output, preventing gradient vanishing and improving training stability.

[0031] Specifically, the system integrates spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesion. The specific process for determining the risk level based on the probability distribution is as follows: The spatial features output by the graph convolutional layer and the temporal features output by the temporal convolutional layer are concatenated to form a fused feature vector. The fused feature vector is then input into a fully connected layer for dimensionality reduction, and a batch normalization layer is used to accelerate model convergence. The dimensionality-reduced features are then input into a softmax function to calculate the probability distribution of the four categories, where the sum of the probability values ​​is 1. The category corresponding to the maximum value in the probability distribution is selected as the final risk level. When the maximum probability value is lower than a preset threshold, the system is marked as pending re-examination.

[0032] In this implementation scheme, the spatial features output by the graph convolutional network are... Temporal features output by temporal convolutional networks Perform feature splicing: ;in, This is for fusing feature vectors; [|||] represents the feature concatenation operation. The fused feature vectors are input into a fully connected layer for dimensionality reduction, and batch normalization is used to accelerate model convergence and prevent gradient instability during training. The dimensionality-reduced features are then input into a softmax function to calculate the four-class probability distribution: ;in, For category (Probability values ​​for healthy status, mild stenosis, severe stenosis, and left main coronary artery disease); Output the activation value of the corresponding category for the fully connected layer; ensure The category corresponding to the highest probability value is selected as the final risk level; when the highest probability value is below the threshold... (Can be set to 0.6 publicly), marked as pending re-examination, prompting the physician to conduct further examination.

[0033] Specifically, the process of extracting abnormal dynamic trend indicators based on the four-category probability distribution and generating early warning indicators by combining them with core function measurement values ​​is as follows: Based on the four-category probability distribution, the dominant risk level is determined; the abnormal dynamic trend indicators corresponding to this risk level are screened; the core function measurement values ​​corresponding to the abnormal indicators are extracted, including the specific values ​​of electromechanical delay at peak load, the magnitude of change relative to the resting value, and the specific values ​​of contraction function dispersion during the recovery period; the abnormal dynamic trend indicators are combined with the core function measurement values ​​to generate a multi-dimensional early warning indicator vector; the early warning indicator vector is normalized to eliminate dimensional differences between different parameters.

[0034] In this implementation plan, the process of extracting abnormal indicators and generating early warning indicators is as follows: First, based on the four-category probability distribution... Determine the dominant risk level: ;in, As the primary risk level index; For category The probability value; the category index does not repeat the symbols from the previous steps. Based on Screening for corresponding abnormal dynamic trend indicators (ADT), including electromechanical delay abnormalities: i.e., the time from the onset of ventricular depolarization in the electrocardiogram to the heart sound. The peak time interval shows an abnormal deviation trend during exercise load. For example, excessively long delays or abnormally large fluctuations may indicate abnormal myocardial mechanical response. Abnormal systolic-diastolic time ratio: The ratio of the time from the first heart sound to the second heart sound to the time from the second heart sound to the next first heart sound exceeds the normal range in some stages, reflecting abnormal coordination of cardiac systolic and diastolic function. Abnormal heart sound intensity ratio: An abnormal trend in the ratio of the amplitude of the first heart sound to the amplitude of the second heart sound may indicate changes in cardiac systolic force or abnormal valvular mechanical function. Abnormal systolic function dispersion: An abnormally large increase in the ratio of the standard deviation to the mean of systolic time in consecutive cardiac cycles in some stages indicates unstable systolic function or a trend of rhythm disturbance. Typically, ADT (Anomalous Delay Testing) extracts abnormal deviation indicators by comparing the dynamic trends of various coronary function parameters (such as the slope and amplitude of change from rest to peak load) with the normal reference range. For example, if the trend slope or amplitude of a parameter exceeds a preset threshold or statistical interval, it is marked as an abnormal dynamic trend indicator. The corresponding core function measurements are then extracted, including the specific value of mechanical delay during peak load. ; the range of change relative to the resting period ; numerical value of systolic function dispersion during the recovery period By combining abnormal dynamic trend indicators with core function measurement values, a multi-dimensional early warning indicator vector is formed: ;in, This is the generated multidimensional early warning indicator vector; the meanings of the parameters for each dimension are as shown above. Subsequently, [the following is done / then...] Normalization is performed: ;in, The mean of each dimension of the vector; The standard deviation is used; normalization eliminates the dimensional differences between different indicators, ensuring the stability of the subsequent optimization process.

[0035] Specifically, the process of constructing a loss function based on the difference between the warning indicator and the ground truth label of the narrowing degree, and optimizing the edge weights and node feature weights of the graph convolutional layer in the hierarchical warning module using the backpropagation algorithm is as follows: The difference between the warning indicator and the ground truth label of the narrowing degree is calculated using the cross-entropy loss function; the gradient of the loss function with respect to the graph convolutional layer parameters is calculated using the backpropagation algorithm; the edge weight matrix of the graph convolutional layer is updated using the adaptive moment estimation algorithm, while the node feature weight matrix is ​​optimized; overfitting is prevented using a weight decay strategy; the change in loss value for each optimization is recorded, and the optimization process stops when the loss value converges.

[0036] In this implementation scheme, firstly, the difference between the warning index and the true value label Y of the narrowness is calculated using the cross-entropy loss function: ;in, The value of the loss function; This is the indicator value of the truth label in category c (1 indicates the true category, 0 indicates the non-true category). This represents the predicted probability of the category corresponding to the warning indicator. The loss function is calculated using the backpropagation algorithm on the graph convolutional layer parameters (edge ​​weight matrix). and node feature weight matrix The gradient of ) is calculated, and the parameters are updated using an adaptive moment estimation algorithm (such as Adam): ;in, , These are the edge weight matrix and the node feature weight matrix for the t-th iteration, respectively. The learning rate; , , , Estimate the first and second moments; To prevent small constants from being divided by zero, a weight decay strategy is introduced during the optimization process: ;in, This is the weight decay coefficient, used to prevent overfitting. During the optimization process, the change in the loss value is recorded for each iteration. Convergence or change amplitude below a preset threshold Stop optimization when the time is right.

[0037] Please see Figure 2A non-invasive early warning method for coronary artery disease based on the fusion of electrocardiogram and heart sound timing parameters includes the following steps: During a stepped exercise load, electrocardiogram (ECG) signals and heart sound signals are acquired simultaneously, and the ventricular depolarization waveforms and heart sounds are extracted. The local morphology descriptor of the peak is used to construct a cost matrix between the reference template and the signal segment to be aligned. A dynamic programming algorithm is then used to find the path with the minimum cumulative cost, and the ECG Q wave initiation and heart sound are adjusted. The relative positions of the peaks; Coronary function parameters, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion, are extracted from the time-aligned signal. Parameter values ​​are recorded at each stage of exercise load. The slope of each parameter change is calculated through linear fitting. The magnitude of parameter change from rest to peak load is statistically analyzed, and dynamic trend indicators are output. Dynamic trend indicators are used as node features, and edge weights are constructed using the Pearson correlation coefficient between coronary function parameters. Spatial features are aggregated through graph convolutional layers, and temporal features are extracted through temporal convolutional layers. Spatial and temporal features are fused to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery disease. The risk level is determined based on the probability distribution. Anomaly dynamic trend indicators are extracted based on the four-class probability distribution, and early warning indicators are generated by combining the core function measurement values. A loss function is constructed based on the difference between the early warning indicators and the true value label of the narrowness. The edge weights and node feature weights of the graph convolutional layer in the hierarchical early warning module are optimized by the backpropagation algorithm.

[0038] In this implementation plan, the steps During a stepped exercise load, electrocardiogram (ECG) signals and heart sound signals are simultaneously acquired. The ECG ventricular depolarization waveform and the first heart sound are extracted. A reference template is constructed based on the local morphological features of the heart sound peak, and a morphological difference cost matrix is ​​calculated on the signal segment to be aligned. A dynamic programming algorithm is used to find the path with the minimum cumulative cost, achieving precise temporal alignment between the ECG initiation point and the heart sound peak. This step ensures a high degree of temporal consistency between the ECG and heart sound, providing a reliable foundation for the accurate extraction of subsequent coronary function parameters, and embodies an innovative signal alignment strategy through the morphological difference matrix and dynamic programming algorithm. Four key coronary artery function parameters were extracted from the time-aligned signal: electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion. Parameter values ​​were recorded at each exercise load stage, and the slope of change was calculated using linear fitting. Simultaneously, the magnitude of change from rest to peak load was statistically analyzed to form a dynamic trend indicator. This step, by quantifying changes in cardiac electromechanical coupling, provides continuous functional assessment indicators, enabling the sensitive detection of early coronary artery abnormalities. This method uses dynamic trend indicators as node features, constructs edge weights based on the Pearson correlation coefficient between coronary artery function parameters, forms a graph structure, and aggregates spatial features through graph convolutional layers. Simultaneously, time series data is input into temporal convolutional layers to extract temporal features. Finally, the spatial and temporal features are fused to output a four-class probability distribution, enabling risk assessment of health status and different degrees of coronary artery stenosis. This step integrates the spatial dependence and temporal evolution information of multidimensional parameters to achieve comprehensive risk judgment, and demonstrates an innovative data fusion method through graph convolution and temporal convolution. This process involves identifying abnormal dynamic trend indicators based on a four-category probability distribution, generating multi-dimensional early warning indicators by combining core functional measurements, and optimizing the model parameters through backpropagation by using the difference between the loss function and the ground truth labels. This adjusts the edge weights and node feature weights of the graph convolutional layers. The goal of this step is to improve the model's predictive accuracy and stability, and to demonstrate how this innovative approach supports individualized risk assessment through adaptive optimization based on anomaly indicators.

[0039] In summary, this application has at least the following effects: A non-invasive early warning system and method for coronary artery disease based on the fusion of ECG and heart sound temporal parameters is proposed. By simultaneously acquiring ECG and heart sound signals and using dynamic programming to achieve precise temporal alignment, the system can accurately extract coronary artery function parameters and their dynamic trends, and quantify changes in cardiac electromechanical coupling. By fusing multidimensional features through graph convolution and temporal convolution, the system can predict the graded risks of health status and different degrees of coronary artery stenosis. At the same time, by combining abnormal indicator extraction and backpropagation optimization mechanisms, the system can improve prediction accuracy and model stability, enabling sensitive detection and dynamic assessment of early coronary artery disease risk, and providing reliable non-invasive early warning evidence for clinical practice.

[0040] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0041] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0042] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0043] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0044] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0045] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters, characterized in that, Includes the following modules: The signal synchronous acquisition module synchronously acquires ECG and heart sound signals during the stepped exercise load process, extracts the local morphological descriptors of the ECG ventricular depolarization wave group and the first heart sound, constructs the cost matrix between the reference template and the signal segment to be aligned, finds the path with the minimum cumulative cost through dynamic programming algorithm, and adjusts the relative position of the ECG ventricular depolarization starting point and the peak point of the first heart sound. The dynamic trend analysis module is used to extract coronary function parameters from time-aligned signals, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion. It records parameter values ​​at each stage of exercise load, calculates the slope of change of each parameter through linear fitting, statistically analyzes the change amplitude of parameters from rest to peak load, and outputs dynamic trend indicators. The graded early warning module uses dynamic trend indicators as node features, constructs edge weights with the Pearson correlation coefficient between coronary function parameters, aggregates spatial features through graph convolutional layers, extracts temporal features through temporal convolutional layers, and merges spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery disease, and determines the risk level based on the probability distribution. The feedback optimization module is used to extract abnormal dynamic trend indicators based on the four-class probability distribution, generate early warning indicators by combining the core function measurement values, construct a loss function based on the difference between the early warning indicators and the true value label of the narrowness, and optimize the edge weights and node feature weights of the graph convolutional layer in the hierarchical early warning module through the backpropagation algorithm.

2. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 1, characterized in that: The specific process of simultaneously acquiring electrocardiogram (ECG) and heart sound signals during a stepped exercise load, extracting the local morphological descriptors of the ventricular depolarization group and the first heart sound, and constructing the cost matrix between the reference template and the signal segment to be aligned is as follows: The single cardiac cycle with the best signal quality was selected as the reference template, and the third-order differential extreme point sequence of the ventricular depolarization wave group and the wavelet packet energy distribution of the first heart sound in the intra-electro-optical wave group of that cycle were extracted as morphological descriptors. On the signal segment to be aligned, morphological descriptors of the same type are extracted by a sliding window with the window length consistent with the reference template. The morphological difference cost between the reference template and the window to be aligned is calculated, including the Euclidean distance of the extreme point sequence of the ventricular depolarization wave group and the KL divergence of the energy distribution of the first heart sound wavelet packet. The cost difference between the two morphologies is weighted and fused to construct a complete cost matrix, with the cost weight of the central ventricular depolarization wave group morphology set to 0.6 and the cost weight of the first heart sound morphology set to 0.

4.

3. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 2, characterized in that: The specific process of adjusting the relative positions of the ventricular depolarization initiation point and the first heart sound peak point by finding the path with the minimum cumulative cost using a dynamic programming algorithm is as follows: Initialize the cumulative cost matrix, set the cumulative cost of the first row and first column to the corresponding morphological difference cost, and calculate the minimum cumulative cost element by element using the dynamic programming algorithm. The cumulative cost of each position is equal to the current morphological difference cost plus the minimum cumulative cost among the three adjacent positions to the left, top and top left. Starting from the bottom right corner of the cumulative cost matrix, backtracking is performed to find the path that minimizes the cumulative cost. This path represents the optimal alignment relationship between the reference template and the signal segment to be aligned. Based on the time offset of the optimal alignment path, the relative position of the starting point of the depolarization of the central ventricle of the signal segment to be aligned and the peak point of the first heart sound are adjusted to achieve sub-sampling accuracy time alignment.

4. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 1, characterized in that: The dynamic trend analysis module includes the following steps: Based on the time-aligned signal, four coronary function parameters were calculated: the interval from the onset of the ventricular depolarization wave to the peak value of the aortic valve closure component of the heart sound was measured as the electromechanical delay; the ratio of the time from the first heart sound to the second heart sound to the time from the second heart sound to the next first heart sound was calculated as the systolic-diastolic time ratio; the ratio of the amplitude of the first heart sound to the amplitude of the second heart sound was calculated as the heart sound intensity ratio; and the ratio of the standard deviation to the mean of the systolic time over five consecutive cardiac cycles was calculated as the systolic function dispersion. Numerical sequences of coronary artery function parameters were recorded during the resting period, the stable period of each load level, and the recovery period of the exercise load. Linear fitting was performed on the numerical sequence of each parameter throughout the exercise load process, and the slope of the fitted line was extracted as the trend feature of the parameter. Calculate the magnitude of change of each coronary function parameter from the resting period to the peak load period, and combine the magnitude of change with the slope of change to form a dynamic trend index.

5. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 1, characterized in that: The specific process of using dynamic trend indicators as node features, constructing edge weights with the Pearson correlation coefficient between coronary function parameters, aggregating spatial features through graph convolutional layers, and extracting temporal features through temporal convolutional layers is as follows: The dynamic trend indicators are constructed as a four-dimensional feature vector, where each dimension corresponds to the change feature of a coronary artery function parameter. The Pearson correlation coefficient matrix between the four coronary artery function parameters is calculated, and this matrix is ​​used as the adjacency matrix of the graph structure to construct a complete parameter relationship graph. The parameter relationship graph is input into a two-layer graph convolutional network, and spatial features are aggregated through the neighbor node information transfer mechanism. The first layer of graph convolution uses the ReLU activation function, and the second layer of graph convolution uses the linear activation function. Simultaneously, the time series of dynamic trend indicators are input into the temporal convolutional layer, and temporal features are extracted by expanding the causal convolutional architecture, while gradient vanishing is avoided by using residual connections.

6. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 5, characterized in that: The process of integrating spatial and temporal features to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery lesion, and determining the risk level based on the probability distribution, is as follows: The spatial features output by the graph convolutional layer and the temporal features output by the temporal convolutional layer are concatenated to form a fused feature vector. The fused feature vector is then input into a fully connected layer for dimensionality reduction, and a batch normalization layer is used to accelerate model convergence. The dimensionality-reduced features are input into the softmax function to calculate the probability distribution of the four categories, where the sum of the probability values ​​is 1. The category corresponding to the maximum value in the probability distribution is selected as the final risk level. When the maximum probability value is lower than the preset threshold, it is marked as pending review.

7. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 1, characterized in that: The specific process of extracting abnormal dynamic trend indicators based on the four-category probability distribution and generating early warning indicators by combining them with core function measurement values ​​is as follows: Based on the four-category probability distribution, the dominant risk level is determined, the abnormal dynamic trend indicators corresponding to the risk level are screened, and the core functional measurement values ​​corresponding to the abnormal indicators are extracted, including the specific value of electromechanical delay at peak load, the change range relative to the resting value, and the specific value of contraction function dispersion during the recovery period. By combining abnormal dynamic trend indicators with core function measurement values, a multi-dimensional early warning indicator vector is generated. The early warning indicator vector is then normalized to eliminate the dimensional differences between different parameters.

8. The non-invasive early warning system for coronary heart disease based on the fusion of ECG and heart sound timing parameters according to claim 7, characterized in that: The specific process of constructing a loss function based on the difference between the early warning index and the ground truth label of the narrowness, and optimizing the edge weights and node feature weights of the graph convolutional layer in the hierarchical early warning module through backpropagation is as follows: The difference between the warning index and the true label of narrowness is calculated by using the cross-entropy loss function, and the gradient of the loss function with respect to the parameters of the graph convolutional layer is calculated by using the backpropagation algorithm. The edge weight matrix of the graph convolutional layer is updated by an adaptive moment estimation algorithm, while the node feature weight matrix is ​​optimized. Overfitting is prevented by a weight decay strategy. Record the change in loss value for each optimization, and stop the optimization process when the loss value converges.

9. A non-invasive early warning method for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters, applied to the non-invasive early warning system for coronary heart disease based on the fusion of electrocardiogram and heart sound timing parameters as described in any one of claims 18, characterized in that, Includes the following steps: During a stepped exercise load, electrocardiogram (ECG) signals and heart sound signals are acquired simultaneously, and the ventricular depolarization waveforms and heart sounds are extracted. The local morphology descriptor of the peak is used to construct a cost matrix between the reference template and the signal segment to be aligned. A dynamic programming algorithm is then used to find the path with the minimum cumulative cost, and the ECG Q wave initiation and heart sound are adjusted. The relative positions of the peaks; Coronary function parameters, including electromechanical delay, systolic-diastolic time ratio, heart sound intensity ratio, and systolic function dispersion, are extracted from the time-aligned signal. Parameter values ​​are recorded at each stage of exercise load. The slope of each parameter change is calculated through linear fitting. The magnitude of parameter change from rest to peak load is statistically analyzed, and dynamic trend indicators are output. Dynamic trend indicators are used as node features, and edge weights are constructed using the Pearson correlation coefficient between coronary function parameters. Spatial features are aggregated through graph convolutional layers, and temporal features are extracted through temporal convolutional layers. Spatial and temporal features are fused to output a four-category probability distribution of healthy status, mild stenosis, severe stenosis, and left main coronary artery disease. The risk level is determined based on the probability distribution. Anomaly dynamic trend indicators are extracted based on the four-class probability distribution, and early warning indicators are generated by combining the core function measurement values. A loss function is constructed based on the difference between the early warning indicators and the true value label of the narrowness. The edge weights and node feature weights of the graph convolutional layer in the hierarchical early warning module are optimized by the backpropagation algorithm.

Images