A heart sound and electrocardio collection and analysis system

Abstract

The application discloses a heart sound and electrocardio collection and analysis system and relates to the technical field of biomedical engineering. The system comprises a time-frequency analysis module, a chaos feature extraction module and the like. The time-frequency analysis module collects heart sound and electrocardio original data sets through an electrocardio and heart sound lead suction ball and an electrocardio and heart sound analog complete machine, and transmits the original data sets to an upper computer to perform wavelet decomposition and short-time Fourier transform, and outputs a heart sound time-frequency spectrum matrix and an electrocardio time-frequency energy matrix. The chaos feature extraction module performs phase space reconstruction on the heart sound time-frequency spectrum matrix and the electrocardio time-frequency energy matrix, forms high-dimensional dynamic trajectories, applies an improved wolf algorithm to calculate dynamic indexes of the high-dimensional dynamic trajectories, and obtains a dynamic parameter set. The improved wolf algorithm and the constructed heart sound space propagation model not only improve the multi-modal fusion capability and analysis discrimination between heart sound and electrocardio signals, but also improve the intelligent level of heart sound and electrocardio collection and analysis.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, and in particular to a heart sound and electrocardiogram acquisition and analysis system. Background Technology

[0002] Heart sounds and electrocardiogram (ECG) signals, as important bioelectrical signals reflecting the physiological and pathological state of the heart, have long been widely used in clinical diagnosis. In recent years, with the deep integration of biomedical engineering and artificial intelligence technologies, heart sound and ECG synchronous acquisition systems based on digital signal processing have gradually replaced traditional auscultation and visual interpretation methods, realizing dynamic monitoring and quantitative analysis of cardiac activity and improving the level of intelligent lesion localization.

[0003] However, existing technologies still have some shortcomings. First, in terms of multimodal signal fusion of heart sounds and electrocardiograms and high-dimensional nonlinear feature modeling, traditional principal component analysis and independent component analysis methods are unable to effectively capture the complex nonlinear correlation between heart sounds and electrocardiogram signals, resulting in insufficient discriminative power and accuracy in heart sound and electrocardiogram analysis. Second, existing graphical models are still weak in mining high-order structural information and modeling pathological propagation patterns, and cannot generate accurate three-dimensional lesion heat maps, thus limiting the accuracy of lesion localization and diagnostic reliability in heart sound and electrocardiogram analysis. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a heart sound and electrocardiogram acquisition and analysis system to solve the problems of insufficient discrimination accuracy and low lesion localization accuracy of existing heart sound and electrocardiogram acquisition and analysis methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a heart sound and electrocardiogram (ECG) acquisition and analysis system, which includes a time-frequency analysis module that acquires raw heart sound and ECG datasets through ECG and heart sound lead suction bulbs and a simulated ECG and heart sound machine, and transmits them to a host computer to perform wavelet decomposition and short-time Fourier transform, outputting a heart sound time-frequency matrix and an ECG time-frequency energy matrix;

[0008] The chaotic feature extraction module performs phase space reconstruction on the heart sound time-frequency spectrum matrix and the electrocardiogram time-frequency energy matrix to form a high-dimensional dynamic trajectory; it applies an improved Wolf algorithm to calculate the dynamic index of the high-dimensional dynamic trajectory to obtain a set of dynamic parameters; and it performs dynamic attention aggregation on the set of dynamic parameters to generate a multi-dimensional chaotic feature vector.

[0009] The lesion localization module inputs multidimensional chaotic feature vectors into the spatial propagation model of heart sounds and electrocardiograms. The biofeature mapping layer performs translead feature fusion and nonlinear spatial embedding to form a set of hypergraph node attributes. The hypergraph propagation layer captures the spatial propagation law of valve vibration and generates pathologically sensitive features. Element-wise multiplication is performed on the set of hypergraph node attributes and pathologically sensitive features to obtain the lesion region feature vector. A softmax classifier is used to calculate the probability distribution of the lesion region feature vector and outputs a three-dimensional pathological heatmap.

[0010] The coupled diagnostic module simultaneously detects the patient's ECG R-wave position sequence, performs pathological-electrophysiological coupling analysis on the ECG R-wave position sequence and three-dimensional pathological thermogram, obtains clinical assessment indicators, and performs joint diagnosis based on the clinical assessment indicators to generate a clinical diagnostic report.

[0011] As a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, wherein:

[0012] As a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the raw heart sound and electrocardiogram dataset includes multi-channel heart sound signal data, multi-lead electrocardiogram signal data, and synchronization timestamp data.

[0013] As a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the output of the heart sound time-frequency spectrum matrix and the electrocardiogram time-frequency energy matrix specifically includes the following steps.

[0014] Multi-level wavelet packet decomposition is performed on the multi-channel heart sound signal data to form effective heart sound segments; time-frequency domain normalization is performed on the effective heart sound segments to output the heart sound time-frequency spectrum matrix;

[0015] The multi-lead ECG signal data is segmented by a sliding window to obtain time-domain signal segments; Hanning windowing and short-time Fourier transform are performed on the time-domain signal segments to generate the ECG time-frequency energy matrix.

[0016] In a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the formation of a high-dimensional dynamic trajectory specifically includes the following steps.

[0017] Cross-modal feature fusion is performed on the time-frequency matrix of heart sounds and the time-frequency energy matrix of electrocardiogram to form a joint feature vector of heart sounds and electrocardiogram.

[0018] Perform cubic spline interpolation on the joint feature vector of heart sound-electrocardiogram to generate a high-dimensional phase space point set;

[0019] The trajectory structure of a high-dimensional phase space point set is fitted using the local weighted projection method to form a high-dimensional dynamic trajectory.

[0020] In a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the generation of the multidimensional chaotic feature vector specifically includes the following steps.

[0021] An improved Wolf algorithm is applied to perform divergent distance quantization on high-dimensional dynamic trajectories, outputting a distance evolution sequence; an exponentially weighted average is then performed on the distance evolution sequence to obtain a set of dynamic parameters.

[0022] Dynamic attention aggregation and tensor recombination are performed on the set of dynamic parameters to generate multidimensional chaotic feature vectors.

[0023] In a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the step of forming a hypergraph node attribute set specifically includes the following steps.

[0024] The biometric mapping layer uses a gated graph attention network to perform cross-lead feature fusion on multi-dimensional chaotic feature vectors to obtain multi-source spatiotemporal features. Then, it performs multi-order tensor decomposition on the multi-source spatiotemporal features to generate heterogeneous feature tensors.

[0025] Nonlinear spatial embedding is performed on heterogeneous feature tensors to form a set of hypergraph node attributes.

[0026] In a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the generation of pathologically sensitive features specifically includes the following steps.

[0027] The hypergraph propagation layer uses a hypergraph neural network to perform high-order neighborhood aggregation on multidimensional chaotic feature vectors to obtain valve vibration features; it then performs Laplace decomposition on the valve vibration features to generate a spatial propagation law matrix.

[0028] The spatial propagation law matrix is ​​subjected to multi-head attention weighting to generate pathologically sensitive features.

[0029] In a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the output of the three-dimensional pathological thermogram specifically includes the following steps.

[0030] Element-wise multiplication is performed on the set of attributes of the hypergraph nodes and the pathologically sensitive features to obtain the associated feature vector; nonlinear transformation is performed on the associated feature vector to generate the lesion region feature vector;

[0031] The softmax classifier is used to calculate the probability distribution of the feature vectors of the lesion area and divide the pathological area, and output a three-dimensional pathological heat map.

[0032] As a preferred embodiment of the cardiac sound and electrocardiogram acquisition and analysis system of the present invention, the acquisition of clinical evaluation indicators refers to performing phase synchronization analysis on the ECG R wave position sequence and the three-dimensional pathological thermogram to form pathological-electrophysiological coupling characteristics; calculating the transfer entropy of the pathological-electrophysiological coupling characteristics to generate clinical evaluation indicators.

[0033] As a preferred embodiment of the heart sound and electrocardiogram acquisition and analysis system of the present invention, the formation of a clinical diagnostic report refers to performing a combined heart sound and electrocardiogram diagnosis on the patient based on clinical assessment indicators to form diagnostic result data; integrating the diagnostic result data with multiple parameters to output a clinical diagnostic report.

[0034] The beneficial effects of this invention are as follows: By using the improved Wolf algorithm to calculate the dynamic exponent of high-dimensional dynamic trajectories and perform dynamic attention aggregation, accurate extraction of multi-dimensional chaotic feature vectors is achieved, significantly improving the multimodal fusion capability and analytical discriminability between heart sounds and electrocardiogram signals. Based on the constructed spatial propagation model of heart sounds, a three-dimensional pathological heatmap is generated, effectively enhancing the spatial positioning accuracy of lesion areas and the modeling ability of pathological propagation patterns, improving the intelligent level of heart sound and electrocardiogram acquisition and analysis, thereby improving the accuracy and reliability of cardiac disease diagnosis. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of a heart sound and electrocardiogram acquisition and analysis system.

[0037] Figure 2 This is a flowchart for generating the spectral matrix of heart sounds.

[0038] Figure 3 This is a flowchart for extracting pathological-electrophysiological coupling features.

[0039] Figure 4 This is a flowchart illustrating the process of constructing a spatial propagation model of heart sounds and electrocardiograms. Detailed Implementation

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0043] Reference Figures 1-4 This is one embodiment of the present invention, which provides a heart sound and electrocardiogram acquisition and analysis system, including the following steps:

[0044] The time-frequency analysis module collects raw data sets of heart sounds and electrocardiograms through the ECG and heart sound lead suction bulb and the ECG and heart sound simulation machine, and transmits them to the host computer to perform wavelet decomposition and short-time Fourier transform, outputting the heart sound time-frequency spectrum matrix and the time-frequency energy distribution of the leads.

[0045] The specific steps are as follows.

[0046] The first step is to collect raw heart sound and electrocardiogram (ECG) datasets, which include multi-channel heart sound signal data, multi-lead ECG signal data, and synchronization timestamp data.

[0047] Six ECG and heart sound lead suction bulbs were attached to the patient's chest measurement sites using conductive gel. Bioelectric and vibration signals were collected and integrated to obtain raw physiological electrical signals. The raw physiological electrical signals were then amplified, filtered, and converted from analog to digital using an ECG and heart sound simulation device to obtain multi-channel heart sound signal data and multi-lead ECG signal data.

[0048] A high-precision clock synchronization device is installed around the patient's ECG and heart sound lead suction bulb. It is started synchronously when the ECG and heart sound lead suction bulb and the ECG and heart sound simulation machine are working, and the synchronization timestamp data is collected.

[0049] The raw heart sound and electrocardiogram datasets not only enable precise monitoring and analysis of heart health status, but also provide high-quality data support for intelligent diagnosis in the biomedical engineering industry.

[0050] The second step involves preprocessing the raw heart sound and electrocardiogram (ECG) dataset. Specifically, for multi-channel heart sound signal data, cubic spline interpolation is used to fill in missing values, improving the continuity and stability of the multi-channel heart sound signal data. Simultaneously, a digital sampler is used for analog-to-digital conversion to achieve digital acquisition of multi-channel heart sound signal data. For multi-lead ECG signal data, a notch filter is used to suppress power frequency interference and eliminate power supply noise. A preamplifier circuit is applied to perform signal conditioning to stabilize the voltage amplitude of the multi-lead ECG signal data. Simultaneously, a synchronous acquisition card is used for continuous sampling to ensure time alignment with the multi-channel heart sound signal data. For synchronous timestamp data, timestamp calibration is performed using the NTP time synchronization protocol to ensure timing consistency. Data is also cached for easy subsequent retrieval and transmission, resulting in the output of the preprocessed raw heart sound and ECG dataset.

[0051] The third step involves performing multi-level wavelet packet decomposition on the preprocessed multi-channel heart sound signal data to form effective heart sound bands. Time-frequency domain normalization is then performed on these effective heart sound bands to output the heart sound time-frequency spectrum matrix. Specifically, the Daubechies 4 wavelet basis function is used to perform 7-level wavelet packet decomposition on the preprocessed multi-channel heart sound signal data, decomposing the heart sound signal data of each channel into sub-signals of different frequency bands. Frequency band recombination and optimization are then performed on the sub-signals of different frequency bands to form a complete wavelet packet tree structure. Feature sub-band separation and noise suppression are then performed on the complete wavelet packet tree structure to obtain the target sub-band set. Important frequency bands (e.g., 20-150Hz) in the target sub-band set are extracted as effective heart sound bands.

[0052] Next, time-frequency domain energy normalization is performed on each effective cardiac sound segment. Further, envelope extraction and transient enhancement are performed on the effective cardiac sound segments to obtain time-varying envelope signals. Logarithmic compression is then performed on the time-varying envelope signals to generate normalized envelope signals. The normalized envelope signals are then time-frequency aligned and energy standardized to output standard time-frequency energy units. These standard time-frequency energy units are arranged according to three dimensions: channel, time, and frequency, constructing a cardiac sound time-frequency spectrum matrix. The channel dimension is determined by the layout of the ECG and cardiac sound lead suction bulbs, the frequency dimension is controlled by the wavelet packet decomposition level, and the frequency dimension is defined by the sampling time step of the multi-channel cardiac sound signal data. The cardiac sound time-frequency spectrum matrix characterizes the time-frequency energy distribution characteristics of the multi-channel cardiac sound signals and can intuitively reflect the spatiotemporal evolution of cardiac mechanical activity.

[0053] The fourth step involves performing sliding window segmentation on the preprocessed multi-lead ECG signal data to obtain time-domain signal segments. Hanning windowing and short-time Fourier transform are then applied to these time-domain signal segments to generate the ECG time-frequency energy matrix. Specifically, the preprocessed multi-lead ECG signal data is segmented using a sliding window. Further, a window with a length of 256 sampling points and a 50% overlap rate is used to segment and divide the multi-lead ECG signal data, generating initial time-domain segments. A waveform integrity check is then performed on the initial time-domain segments, and the time-domain signal segments are output.

[0054] It should be noted that waveform integrity check refers to performing multi-lead feature comparison on the preliminary time domain segment to ensure that the preliminary time domain segment contains at least one complete cardiac cycle and conforms to standard electrocardiogram characteristics.

[0055] Subsequently, the Hanning window function is used to window each time domain signal segment. Furthermore, zero-filling expansion is performed on the time domain signal segment to obtain an optimized time domain signal. Fourier basis transform is performed on the optimized time domain signal to generate the original frequency domain representation. Sidelobe suppression is then applied to the original frequency domain representation to form a clean frequency domain signal.

[0056] It should be noted that sidelobe suppression refers to performing spectral weighted smoothing on the original frequency domain representation in order to suppress energy leakage of non-main lobe frequency components in the original frequency domain representation.

[0057] The clean frequency domain signal is subjected to short-time Fourier transform for spectral energy extraction. Further, power spectral density transformation is performed on the clean frequency domain signal to obtain a preliminary frequency domain energy distribution. Bandwidth limitation is applied to the preliminary frequency domain energy distribution to obtain physiologically relevant frequency domain energy. The physiologically relevant frequency domain energy is then reconstructed into a three-dimensional matrix according to the dimensions of time frame, frequency, and number of leads to generate an ECG time-frequency energy matrix. The time frame dimension is defined based on the sampling time interval of the multi-lead ECG signal data, the frequency dimension is determined by the window length of the sliding window, and the number of leads is determined by the number of leads configured in the multi-lead ECG signal data. The ECG time-frequency energy matrix not only characterizes the time-frequency energy distribution characteristics of each lead ECG signal but also supports the coupled analysis of cardiac electrical and mechanical activities.

[0058] The chaotic feature extraction module performs phase space reconstruction on the heart sound time-frequency spectrum matrix and lead time-frequency energy distribution to form a high-dimensional dynamic trajectory; it applies an improved Wolf algorithm to calculate the dynamic index of the high-dimensional dynamic trajectory to obtain a set of dynamic parameters; and it performs dynamic attention aggregation on the set of dynamic parameters to generate a multi-dimensional chaotic feature vector.

[0059] The specific steps are as follows.

[0060] The first step is to perform phase space reconstruction on the heart sound time-frequency spectrum matrix and the lead time-frequency energy distribution to form a high-dimensional dynamic trajectory. Specifically, the heart sound time-frequency spectrum matrix and the ECG time-frequency energy matrix are fused across modal features. Furthermore, feature alignment and weight fusion are performed on the heart sound time-frequency spectrum matrix and the ECG time-frequency energy matrix to form cross-modal joint features. The cross-modal joint features are then time-normalized and dimension-compressed to obtain standard joint features. At the same time, the standard joint features are subjected to nonlinear transformation to form a heart sound-ECG joint feature vector.

[0061] Cubic spline interpolation is performed on the joint feature vector of heart sound and electrocardiogram. Furthermore, temporal interpolation and spatial expansion are performed on the joint feature vector of heart sound and electrocardiogram to obtain the interpolated feature sequence. Local smoothing and noise suppression are performed on the interpolated feature sequence to generate an optimized feature sequence. Phase space mapping and dimensional adjustment are performed on the optimized feature sequence to form a high-dimensional phase space point set.

[0062] It should be noted that phase space mapping refers to the process of embedding each time-frequency feature point in the optimized feature sequence into a high-dimensional phase space through coordinate transformation.

[0063] A local weighted projection method is used to fit the trajectory structure of a high-dimensional phase space point set. Further, neighborhood features of the high-dimensional phase space point set are extracted and subjected to distance weighting and manifold projection to project these features from the original high-dimensional phase space to a low-dimensional manifold space, generating locally weighted trajectory segments. Dynamic time warping is used to divide these locally weighted trajectory segments into units. Furthermore, temporal alignment of these segments generates aligned trajectory segments. Feature aggregation is performed on these aligned trajectory segments to obtain trajectory feature blocks, and energy normalization is applied to these feature blocks to obtain optimized trajectory units. Finally, segment splicing and Gaussian filtering are applied to these optimized trajectory units to output a high-dimensional dynamic trajectory.

[0064] The second step involves applying the improved Wolf algorithm to quantize the divergence distance of the high-dimensional dynamic trajectory, outputting a distance evolution sequence. This sequence is then subjected to an exponentially weighted average to obtain a set of dynamic parameters. Specifically, the high-dimensional dynamic trajectory is sampled within a fixed radius (e.g., 0.2) to obtain neighborhood point pairs. The divergence distance of these neighborhood point pairs is quantized, and time-cumulative integration is performed to generate a dynamic exponent. The specific mathematical formula is as follows:

[0065] ;

[0066] in, Indicates a time index. Indicates at a point in time The dynamic index, Indicates the time step. Indicates at a point in time The divergence distance of the neighborhood point pairs of the trajectory. Indicates at a point in time The divergence distance of the neighborhood point pairs of the trajectory;

[0067] It should be noted that the divergence distance of the trajectory neighborhood point pairs is obtained by performing Euclidean distance measurement and root mean square integration on the trajectory neighborhood point pairs.

[0068] The dynamic index is piecewise linearly combined and statistical features are extracted to generate piecewise trajectory features. The piecewise trajectory features are then subjected to feature energy balancing and normalization scaling to output the distance evolution sequence.

[0069] An exponentially decaying weighted average is applied to the distance evolution sequence. Further, a temporal weighting is applied to the distance evolution sequence using a decay factor: distance evolution sequences from recent times (e.g., within the last 50ms) are assigned a higher weight (e.g., 0.9), while those from longer times (e.g., 50ms ago) are assigned a lower weight (e.g., 0.6). Based on the assigned weights, a weighted integration is performed on the distance evolution sequence to obtain a weighted evolution sequence. Finally, dimensional expansion and feature recombination are performed on the weighted evolution sequence to obtain a set of dynamic parameters.

[0070] It should be noted that the decay factor is defined based on the time correlation of the distance evolution sequence, and its value ranges from 0.5 to 0.9. The improved Wolf algorithm has made the following improvements on the basis of the existing Wolf algorithm. On the one hand, it improves the accuracy of capturing complex trajectory structures by extracting trajectory neighborhood point pairs through fixed radius neighborhood sampling. On the other hand, it uses time cumulative integral to perform dynamic index calculation, realizing the quantitative evaluation of trajectory divergence characteristics.

[0071] The third step involves performing dynamic attention aggregation and tensor reorganization on the set of dynamic parameters to generate a multidimensional chaotic feature vector. Specifically, during the attention aggregation stage, a multi-head attention mechanism is used to model the feature correlation of the set of dynamic parameters. Furthermore, the set of dynamic parameters is subjected to association interaction and matrix transformation to obtain the parameter correlation matrix. Dynamic aggregation is then performed on the parameter correlation matrix to generate attention-weighted parameters. Linear projection is then performed on the attention-weighted parameters to obtain a dimensionality-reduced feature representation. Min-Max normalization is applied to constrain the amplitude of the dimensionality-reduced feature representation, and the modulated feature vector is output.

[0072] In the tensor reconstruction stage, channel separation and cross-modal linking are performed on the modulated feature vector to form reconstructed feature blocks; convolutional fusion is performed on the reconstructed feature blocks to obtain cross-modal feature integration; feature dimension reconstruction and high-dimensional feature mapping are performed on the cross-modal feature integration to output multi-dimensional chaotic feature vectors.

[0073] The lesion localization module inputs multidimensional chaotic feature vectors into the cardiac sound and electrocardiogram spatial propagation model. The biofeature mapping layer performs translead feature fusion and nonlinear spatial embedding to form a set of hypergraph node attributes. The hypergraph propagation layer captures the spatial propagation law of valve vibration and generates pathologically sensitive features. Element-wise multiplication is performed on the hypergraph node attribute set and pathologically sensitive features to obtain the lesion region feature vector. A softmax classifier is used to calculate the probability distribution of the lesion region feature vector and output a three-dimensional pathological heatmap.

[0074] The specific steps are as follows.

[0075] The first step involves constructing and training a spatial propagation model of heart sounds and electrocardiograms. Specifically, in the PyTorch framework, a gated graph attention network is invoked using the `nn.Module` parameter, with its input dimension set to 128, the number of heads to 8, and the output dimension to 64. Two fully connected layers are then added after the gated graph attention network for non-linear spatial embedding to enhance feature representation. Residual connections are then applied for feature fusion, completing the construction of the biometric mapping layer. Next, a hypergraph neural network is invoked using Keras API functions, with its input dimension set to 64, hidden layer dimension to 128, and output dimension to 256. A multi-head attention mechanism is then applied after the hypergraph neural network for multi-head attention weighting to strengthen key pathological features, completing the construction of the hypergraph propagation layer.

[0076] The nn.Sequential function is used to cascade the biometric mapping layer and the hypergraph propagation layer to obtain a joint feature representation. The parameters of the joint feature representation are standardized to generate an optimized feature representation. A learning rate regulator is used to dynamically adjust the learning rate of the optimized feature representation to generate adaptive learning weights. Based on the adaptive learning weights, the biometric mapping layer and the hypergraph propagation layer are stacked alternately, and gradient clipping is used to optimize stability, thus completing the construction of the cardiac sound and electrocardiogram spatial propagation model.

[0077] Next, the spatial propagation model of heart sounds and electrocardiograms is trained. Further, the historical multidimensional chaotic feature vectors are divided into a sample set, a training set, and a validation set. On the sample set, data augmentation techniques are used for random rotation and scaling, and data normalization is performed through standardization to form augmented training samples. On the training set, the Adam optimizer is used to perform gradient backpropagation on the augmented training samples, and a Dropout layer is applied simultaneously for regularization to obtain optimized parameters for the spatial propagation model of heart sounds and electrocardiograms. On the validation set, the cross-entropy loss function is used to quantify the error of the optimized parameters of the spatial propagation model of heart sounds and electrocardiograms to obtain the validation set accuracy. When the validation set accuracy exceeds the convergence threshold for several consecutive rounds (e.g., 5 times), training terminates, and the trained spatial propagation model of heart sounds and electrocardiograms is output simultaneously.

[0078] It should be noted that the convergence threshold is defined based on the fluctuation range of the validation set accuracy.

[0079] The second step involves the biometric mapping layer performing interlead feature fusion and nonlinear spatial embedding to form a set of hypergraph node attributes. Specifically, the multidimensional chaotic feature vector is input into the trained cardiac sound and electrocardiogram spatial propagation model via the Input interface. The biometric mapping layer applies 1×1 separable convolutions to perform feature alignment and dimension matching on the multidimensional chaotic feature vector to obtain standard-fit features. The standard-fit features are then input into a bidirectional gated graph attention network. The forward gated graph attention network performs multi-head attention weighting on the standard-fit features to generate forward spatiotemporal features. Residual connections are then made on the forward spatiotemporal features to obtain enhanced forward features. The backward gated graph attention network performs reverse attention weighting and gating modulation on the enhanced forward features to generate backward spatiotemporal features. The enhanced forward features and backward spatiotemporal features are then concatenated to form bidirectional gated features. Finally, the bidirectional gated features are subjected to interlead feature fusion and nonlinear activation to generate multi-source spatiotemporal features.

[0080] The Tucker decomposition method is used to perform multi-order tensor decomposition on multi-source spatiotemporal features. Furthermore, mode decomposition is performed on multi-source spatiotemporal features to extract the first 32 core tensors and integrate them to obtain low-rank feature representations. Factor matrix reconstruction is performed on the low-rank feature representations to generate reconstructed feature tensors. At the same time, feature dimensions are reorganized on the reconstructed feature tensors to obtain heterogeneous feature tensors.

[0081] Two fully connected layers are used to perform nonlinear spatial embedding on heterogeneous feature tensors. The first layer expands the heterogeneous feature tensors and performs nonlinear activation to obtain preliminary embedded features. The second layer refines the preliminary embedded features and aligns their dimensions to form the final embedded features. Simultaneously, hypergraph spatial projection is performed on the final embedded features to output a set of hypergraph node attributes. The set of hypergraph node attributes represents the spatiotemporal correlation characteristics of heart sound and electrocardiogram signals, which can support the accurate localization and analysis of pathological features.

[0082] The third step involves the hypergraph propagation layer capturing the spatial propagation pattern of valve vibration and generating pathologically sensitive features. Specifically, in the high-order neighborhood aggregation stage, the hypergraph neural network is used to perform dynamic feature updates and nonlinear transformations on the multidimensional chaotic feature vectors to generate node update features. The node update features are then fused across hyperedges to obtain high-order node features. Neighborhood aggregation is then performed on the high-order node features to output valve vibration features.

[0083] Next, the valve vibration characteristics are subjected to Laplace decomposition. Furthermore, node relationship mapping is performed on the valve vibration characteristics to form a node association structure. Matrix element filling is then performed on the node association structure to construct a Laplace matrix. Eigenvalue decomposition and modal energy sorting are performed on the Laplace matrix to obtain a set of characteristic modes. The set of characteristic modes is then subjected to matrix recombination and weight adjustment to generate a spatial propagation law matrix. The row vectors of the spatial propagation law matrix reflect the participation intensity of different nodes in the signal propagation process, and the column vectors reflect the contribution of different frequency components in spatial propagation.

[0084] A multi-head attention mechanism is used to weight the spatial propagation law matrix. Furthermore, by having eight attention heads work in parallel, the spatial propagation law matrix is ​​decomposed into key vectors, value vectors, and query vectors. Dot product operations and softmax normalization are then performed on the key vectors, value vectors, and query vectors to obtain the attention weight matrix. Weighted summation is then performed on the attention weight matrix to generate eight single-head attention features. These eight single-head attention features are then concatenated and linearly transformed to form pathologically sensitive features. These pathologically sensitive features characterize the spatiotemporal correlation pattern between heart valve vibration and electrocardiogram signals, laying the foundation for subsequent precise localization of lesion areas.

[0085] The fourth step involves performing element-wise multiplication on the hypergraph node attribute set and the pathologically sensitive features to obtain the lesion region feature vector. Specifically, the hypergraph node attribute set and the pathologically sensitive features are dimensionally aligned to ensure that their feature dimensions are completely consistent. Then, the Hadamard product is performed to generate preliminary product features. The preliminary product features are then filtered using an intensity threshold, removing those that exceed the threshold and retaining significant pathological features. The filtered features are then output. Finally, the filtered features are multiplied element-wise with the hypergraph node attribute set at corresponding positions to generate enhanced product features.

[0086] It should be noted that the intensity threshold is defined based on the magnitude concentration of the preliminary product features, and its value ranges from 0.05 to 0.2.

[0087] The enhanced product features are augmented using an LRN (Local Response Normalization) operation with a neighborhood window size of 5, and the magnitude is adjusted using a learnable scaling factor to optimize the feature representation and generate a lesion region feature vector.

[0088] It should be noted that the learnable scaling factor is defined based on the dynamic energy distribution range of the enhanced product feature, and its value ranges from 1.0 to 1.5.

[0089] The fifth step involves using a softmax classifier to calculate the probability distribution of the lesion region feature vectors and outputting a 3D pathological heatmap. Specifically, the softmax classifier performs a linear transformation on the lesion region feature vectors, mapping them to the dimension of the number of pathological categories (e.g., 5 categories), generating a category space vector. This category space vector is then smoothed with a Gaussian kernel to obtain a category transformation score. An exponential moving average is applied to the category transformation score to output an exponentially weighted value. This exponentially weighted value is then logarithmically transformed and linearly fitted to form the original probability distribution. Finally, probability normalization is performed on the original probability distribution to generate probability distribution values ​​for different pathological categories. The specific mathematical formulas are as follows.

[0090] ;

[0091] in, Indicates the pathology category index. Indicates the first The probability distribution values ​​of each pathological category. Indicates the first Category transformation score for each pathology category Indicates the first The exponential weighted value of the category transformation score for each pathology category; Indicates the total number of pathology categories. This represents the index for summing pathology categories. Indicates the preceding Category transformation score for each pathology category Indicates the preceding The exponential weighted value of the category transformation score for each pathology category;

[0092] According to the anatomical location mapping rules, the probability distribution values ​​of different pathological categories are projected into three-dimensional space, and linear interpolation is performed simultaneously to ensure spatial continuity, outputting a three-dimensional feature distribution field. Pathological regions are divided into three-dimensional feature distribution fields using probability thresholds: regions exceeding the probability threshold are classified as high-probability pathological regions, and regions below the probability threshold are classified as low-probability background regions. The transparency of the divided regions is adjusted and color rendered to generate a three-dimensional pathological heatmap. In the three-dimensional pathological heatmap, red areas represent high-probability pathological regions, and blue areas represent low-probability background regions.

[0093] It should be noted that the anatomical position mapping rule refers to the mapping relationship between the physical coordinates of the ECG and cardiac sound lead suction bulb and the three-dimensional model of the heart, based on the clinical anatomical standard definition; the probability threshold is defined based on the energy concentration of the three-dimensional feature distribution field, with a value range of 0.3 to 0.7.

[0094] The coupled diagnostic module simultaneously detects the patient's ECG R-wave position sequence, performs pathological-electrophysiological coupling analysis on the ECG R-wave position sequence and three-dimensional pathological thermogram, obtains clinical assessment indicators, and performs joint diagnosis based on the clinical assessment indicators to generate a clinical diagnostic report.

[0095] The specific steps are as follows.

[0096] The first step involves performing phase synchronization analysis on the ECG R-wave position sequence and the 3D pathological heatmap to form pathological-electrophysiological coupling characteristics. Specifically, an ECG signal processor is used to perform R-wave detection and noise filtering on the patient to obtain the ECG R-wave position sequence. The ECG R-wave position sequence and the 3D pathological heatmap are then time-aligned using a dynamic time warping algorithm. Furthermore, nonlinear time calibration is performed on the ECG R-wave position sequence and the 3D pathological heatmap to eliminate the time difference between them, ensuring that each ECG R-wave position sequence is precisely matched with the timestamp of cardiac motion, and outputting synchronized spatiotemporal data.

[0097] Next, cross-wavelet transform is used to analyze the coupling relationship of synchronous spatiotemporal data in the time-frequency domain. Furthermore, time-frequency decomposition is performed on the synchronous spatiotemporal data to generate multi-scale time-frequency bands. Coherence quantization is performed on the multi-scale time-frequency bands to obtain coherence coefficients, and phase extraction is performed on the coherence coefficients to obtain phase differences. Based on the coherence coefficients and phase differences, significant region labeling is performed on the multi-scale time-frequency bands. For example, multi-scale time-frequency bands with coherence coefficients greater than the coherence threshold (e.g., 0.7) and phase differences less than the phase threshold (e.g., π / 4) are identified as significant coupling regions. Feature extraction and dimensionality compression are performed on the significant coupling regions to output pathological-electrophysiological coupling features.

[0098] It should be noted that the coherence threshold is defined based on the percentile distribution of historical coherence coefficients; the phase threshold is defined based on the fluctuation range of historical phase differences.

[0099] The second step involves calculating the transfer entropy of the pathological-electrophysiological coupling features to generate clinical assessment indicators. Specifically, the pathological-electrophysiological coupling features are analyzed for correlation using a time-delay embedding method. Furthermore, the pathological-electrophysiological coupling features are expanded according to a time series to obtain a time-delay feature matrix. A recombination parameter with an embedding dimension of 5 and a time delay of 10ms is set to perform feature sequence recombination on the time-delay feature matrix, outputting the delay correlation features.

[0100] Next, a Gaussian kernel function is used to perform joint probability estimation and entropy calculation on the delay correlation features using a sliding window with a fixed step size (e.g., 50ms), and the propagation entropy value is output. The specific mathematical formula is as follows.

[0101] ;

[0102] in, Indicates the propagation entropy value. The total number of dimensions representing the delayed correlation features. Indicates a dimension index. Indicates the first The probability density values ​​of each dimension;

[0103] The transfer entropy values ​​are subjected to moving average filtering to obtain a smoothed entropy value sequence. The smoothed entropy value sequence is then integrated over a time window to obtain the cumulative transfer entropy. The cumulative transfer entropy is normalized by z-score standardization to eliminate individual differences in baseline fluctuations and output a standardized transfer entropy index. Finally, the standardized transfer entropy index is mapped to clinical indicators according to clinical grading standards (such as NYHA functional classification) to generate clinical assessment indicators.

[0104] The third step is to perform a combined diagnosis based on clinical assessment indicators to generate a clinical diagnostic report. In this process, key pathological parameters (such as transmission entropy value and pathological probability) are extracted from the clinical assessment indicators. Based on these key pathological parameters, a combined diagnosis of heart sounds and electrocardiograms is performed on the patient. For example, when the transmission entropy value is >1.2 and the pathological probability is >0.8, key lesions are marked. When the transmission entropy value is <0.5 and the pathological probability is <0.3, normal areas are excluded. The results of the combined diagnosis of heart sounds and electrocardiograms are integrated, and the diagnostic results data are output.

[0105] The diagnostic results data are integrated with multiple parameters to obtain a structured treatment decision set; the structured treatment decision set is then converted into natural language to generate a sequence of medical semantic units; the sequence of medical semantic units is then converted into a text description that conforms to clinical reporting standards using OpenAI Clinician software, and a clinical diagnostic report is output.

[0106] In summary, this invention achieves accurate extraction of multidimensional chaotic feature vectors by using an improved Wolf algorithm for dynamic exponent calculation and dynamic attention aggregation of high-dimensional dynamic trajectories. This significantly enhances the multimodal fusion capability and analytical discriminative ability between heart sounds and electrocardiogram signals. Furthermore, by generating a three-dimensional pathological heatmap based on the constructed heart sound spatial propagation model, it effectively enhances the spatial localization accuracy of lesion areas and the modeling ability of pathological propagation patterns, thereby improving the intelligence level of heart sound and electrocardiogram acquisition and analysis, and ultimately increasing the accuracy and reliability of cardiac disease diagnosis.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A heart sound and electrocardiogram acquisition and analysis system, characterized in that: include, The time-frequency analysis module collects raw data sets of heart sounds and electrocardiograms through the ECG and heart sound lead suction bulb and the ECG and heart sound simulation machine, and transmits them to the host computer to perform wavelet decomposition and short-time Fourier transform, outputting the heart sound time-frequency matrix and the ECG time-frequency energy matrix; The chaotic feature extraction module performs phase space reconstruction on the heart sound time-frequency spectrum matrix and the electrocardiogram time-frequency energy matrix to form a high-dimensional dynamic trajectory. An improved Wolf algorithm is then applied to calculate the dynamic exponent of the high-dimensional dynamic trajectory, specifically including the following steps. A fixed-radius neighborhood sampling method is used to sample the high-dimensional dynamic trajectory, forming trajectory neighborhood point pairs. The divergence distance of the trajectory neighborhood point pairs is quantized, and time-cumulative integration is performed to generate a dynamic exponent. The specific mathematical formula is as follows. ； in, Indicates a time index. Indicates at a point in time The dynamic index, Indicates the time step. Indicates at a point in time The divergence distance of the neighborhood point pairs of the trajectory. Indicates at a point in time The divergence distance of the neighborhood point pairs of the trajectory; The dynamic index is piecewise linearly combined and statistical features are extracted to generate piecewise trajectory features. The piecewise trajectory features are then subjected to feature energy balancing and normalized scaling to output the distance evolution sequence. Temporal weighting is performed on the distance evolution sequences using a decay factor: higher weights are assigned to recent distance evolution sequences, and lower weights are assigned to distant distance evolution sequences; weighted integration is performed on the distance evolution sequences according to the assigned weights to obtain a weighted evolution sequence; dimensional expansion and feature recombination are performed on the weighted evolution sequence to obtain a set of dynamic parameters. Dynamic attention aggregation is performed on the set of dynamic parameters to generate a multidimensional chaotic feature vector; The lesion localization module inputs multidimensional chaotic feature vectors into the spatial propagation model of heart sounds and electrocardiograms. The biofeature mapping layer performs translead feature fusion and nonlinear spatial embedding to form a set of hypergraph node attributes. The hypergraph propagation layer captures the spatial propagation law of valve vibration and generates pathologically sensitive features. Element-wise multiplication is performed on the set of hypergraph node attributes and pathologically sensitive features to obtain the lesion region feature vector. A softmax classifier is used to calculate the probability distribution of the lesion region feature vector and outputs a three-dimensional pathological heatmap. The construction of the spatial propagation model of heart sounds and electrocardiograms includes the following steps: In the PyTorch framework, a gated graph attention network is invoked through the nn.Module parameter, and two fully connected layers are added after the gated graph attention network for nonlinear spatial embedding to enhance feature representation capabilities. Residual connections are then applied for feature fusion to complete the construction of the biometric feature mapping layer. Keras API functions are used to invoke the hypergraph neural network, and a multi-head attention mechanism is added after the hypergraph neural network for multi-head attention weighting to strengthen key pathological features, thus completing the construction of the hypergraph propagation layer. The nn.Sequential function is used to cascade the biometric mapping layer and the hypergraph propagation layer to obtain a joint feature representation. The parameters of the joint feature representation are standardized to generate an optimized feature representation. A learning rate regulator is used to dynamically adjust the learning rate of the optimized feature representation to generate adaptive learning weights. Based on the adaptive learning weights, the biometric mapping layer and the hypergraph propagation layer are stacked alternately, and gradient clipping is used to optimize stability, thus completing the construction of the cardiac sound and electrocardiogram spatial propagation model. The coupled diagnostic module simultaneously detects the patient's ECG R-wave position sequence, performs pathological-electrophysiological coupling analysis on the ECG R-wave position sequence and three-dimensional pathological thermogram, obtains clinical assessment indicators, and performs joint diagnosis based on the clinical assessment indicators to generate a clinical diagnostic report.

2. The heart sound and electrocardiogram acquisition and analysis system as described in claim 1, characterized in that: The raw heart sound and electrocardiogram dataset includes multi-channel heart sound signal data, multi-lead electrocardiogram signal data, and synchronous timestamp data.

3. The heart sound and electrocardiogram acquisition and analysis system as described in claim 2, characterized in that: The output heart sound time-frequency spectrum matrix and electrocardiogram time-frequency energy matrix specifically include the following steps. Multi-level wavelet packet decomposition is performed on the multi-channel heart sound signal data to form effective heart sound segments; time-frequency domain normalization is performed on the effective heart sound segments to output the heart sound time-frequency spectrum matrix; The multi-lead ECG signal data is segmented by a sliding window to obtain time-domain signal segments; Hanning windowing and short-time Fourier transform are performed on the time-domain signal segments to generate the ECG time-frequency energy matrix.

4. The heart sound and electrocardiogram acquisition and analysis system as described in claim 1, characterized in that: The formation of the high-dimensional dynamic trajectory specifically includes the following steps. Cross-modal feature fusion is performed on the time-frequency matrix of heart sounds and the time-frequency energy matrix of electrocardiogram to form a joint feature vector of heart sounds and electrocardiogram. Perform cubic spline interpolation on the joint feature vector of heart sound-electrocardiogram to generate a high-dimensional phase space point set; The trajectory structure of a high-dimensional phase space point set is fitted using the local weighted projection method to form a high-dimensional dynamic trajectory.

5. The heart sound and electrocardiogram acquisition and analysis system as described in claim 4, characterized in that: The generation of multidimensional chaotic feature vectors specifically includes the following steps. An improved Wolf algorithm is applied to perform divergent distance quantization on high-dimensional dynamic trajectories, outputting a distance evolution sequence; an exponentially weighted average is then performed on the distance evolution sequence to obtain a set of dynamic parameters. Dynamic attention aggregation and tensor recombination are performed on the set of dynamic parameters to generate multidimensional chaotic feature vectors.

6. The heart sound and electrocardiogram acquisition and analysis system as described in claim 1, characterized in that: The process of forming a set of hypergraph node attributes specifically includes the following steps. The biometric mapping layer uses a gated graph attention network to perform cross-lead feature fusion on multi-dimensional chaotic feature vectors to obtain multi-source spatiotemporal features. Then, it performs multi-order tensor decomposition on the multi-source spatiotemporal features to generate heterogeneous feature tensors. Nonlinear spatial embedding is performed on heterogeneous feature tensors to form a set of hypergraph node attributes.

7. The heart sound and electrocardiogram acquisition and analysis system as described in claim 1, characterized in that: The generation of pathologically sensitive features specifically includes the following steps. The hypergraph propagation layer uses a hypergraph neural network to perform high-order neighborhood aggregation on multidimensional chaotic feature vectors to obtain valve vibration features; it then performs Laplace decomposition on the valve vibration features to generate a spatial propagation law matrix. The spatial propagation law matrix is ​​subjected to multi-head attention weighting to generate pathologically sensitive features.

8. The heart sound and electrocardiogram acquisition and analysis system as described in claim 7, characterized in that: The output of the three-dimensional pathological thermogram specifically includes the following steps. Element-wise multiplication is performed on the set of attributes of the hypergraph nodes and the pathologically sensitive features to obtain the associated feature vector; nonlinear transformation is performed on the associated feature vector to generate the lesion region feature vector; The softmax classifier is used to calculate the probability distribution of the feature vectors of the lesion area and divide the pathological area, and output a three-dimensional pathological heat map.

9. The heart sound and electrocardiogram acquisition and analysis system as described in claim 1, characterized in that: The acquisition of clinical assessment indicators refers to performing phase synchronization analysis on the ECG R-wave position sequence and the three-dimensional pathological thermogram to form pathological-electrophysiological coupling characteristics. The transfer entropy of the pathophysiological coupling characteristics is calculated to generate clinical assessment indicators.

10. The heart sound and electrocardiogram acquisition and analysis system as described in claim 9, characterized in that: The formation of a clinical diagnostic report refers to the process of performing a combined diagnosis of heart sounds and electrocardiograms on patients based on clinical assessment indicators to generate diagnostic result data; integrating the diagnostic result data with multiple parameters to output a clinical diagnostic report.

Images