Healthcare monitoring method for cardiac pacemaker implanter based on data analysis

Abstract

The invention discloses a heart pacemaker implanter health monitoring method based on data analysis, and relates to the technical field of medical health monitoring, and the method comprises the steps: carrying out the biomechanical state mapping of a dynamic impedance sequence and electrocardio phase marking data, generating a myocardial contact stress vector, and carrying out the measurement of the myocardial contact stress vector; performing abnormal mode recognition and probability density modeling on the myocardial contact stress vector to form a contact abnormal event atlas, inputting the contact abnormal event atlas into a multi-mode health monitoring model, performing high-frequency attenuation characteristic decomposition on the heart sound branch layer, performing impedance fluctuation variance extraction and weighted aggregation on the probability inference layer, and outputting a three-color grading early warning instruction. According to the method, dynamic change rate analysis and phase marking are performed on the physiological data set of the implanter, and the multi-modal health monitoring model is constructed synchronously, so that the fusion efficiency of multi-source physiological information is improved, and the overall accuracy and reliability of health monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of medical health monitoring technology, and in particular to a method for health monitoring of pacemaker implantees based on data analysis. Background Technology

[0002] As a crucial medical device for treating arrhythmias, pacemakers have evolved from initially providing only basic cardiac rhythm support to now enabling continuous tracking of patient health status through wireless communication and remote monitoring, playing a significant role in improving patients' quality of life. In recent years, advancements in biomedical engineering technology, particularly data processing and analysis, have made more precise and real-time health monitoring of pacemaker implant recipients possible.

[0003] Nevertheless, existing technologies still have shortcomings. First, most current health monitoring methods focus on processing single types of physiological data, such as using electrocardiogram signals or impedance data alone, failing to fully integrate multi-source information for comprehensive analysis, thus making it difficult to comprehensively reflect cardiac health status. Second, in terms of abnormal event identification and risk prediction, existing models often lack sufficient sensitivity and specificity, failing to accurately distinguish different types of pathological patterns, resulting in a high misjudgment rate and affecting the accuracy and reliability of health monitoring. Summary of the Invention

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

[0005] Therefore, this invention provides a data analysis-based health monitoring method for pacemaker implantees to address the problems of insufficient multidimensional data coupling and high misjudgment rate in existing health monitoring methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for health monitoring of pacemaker implantees based on data analysis, comprising: collecting physiological datasets of implantees, performing dynamic rate of change analysis and phase labeling, and outputting dynamic impedance sequence and electrocardiogram phase labeling data; Biomechanical state mapping is performed on dynamic impedance sequence and electrocardiogram phase label data to generate myocardial contact stress vector. Abnormal pattern recognition and probability density modeling are performed on myocardial contact stress vector to form a contact abnormal event map. The abnormal event spectrum is input into the multimodal health monitoring model. The heart sound branch layer performs high-frequency attenuation feature decomposition, the probability inference layer extracts and weights the impedance fluctuation variance, and outputs a three-color graded early warning command. The execution time decay weighting of the three-color graded early warning instructions is used to obtain the evolution trajectory of risk events; the execution state transition probability analysis of the risk event evolution trajectory is used to generate an implanted health monitoring report.

[0007] As a preferred embodiment of the data analysis-based method for health monitoring of pacemaker implantees described in this invention, the implantee physiological dataset includes pacemaker biphasic impedance data, implantee surface electrocardiogram signals, and time synchronization data.

[0008] As a preferred embodiment of the data analysis-based health monitoring method for pacemaker implantees described in this invention, the output of dynamic impedance sequence and electrocardiogram phase marker data specifically includes the following steps. A sliding window differential is performed on the biphasic impedance data of the cardiac pacemaker to obtain the impedance change vector; the impedance change vector is then integrated over time to generate a dynamic impedance sequence. The implanter performs R-wave reference point division and phase marking on the surface electrocardiogram signal, and outputs electrocardiogram phase marking data.

[0009] As a preferred embodiment of the data analysis-based health monitoring method for pacemaker implantees described in this invention, the generation of the myocardial contact stress vector specifically includes the following steps. Based on time synchronization data, dynamic impedance sequence and ECG phase marker data are dynamically registered, and feature joint encoding is performed synchronously to obtain biomechanical feature tensors. Based on the stress differential equation, the biomechanical characteristic tensor is mapped to a biomechanical state to form the myocardial contact stress vector.

[0010] As a preferred embodiment of the data analysis-based health monitoring method for pacemaker implantees described in this invention, the step of forming a contact abnormality event map specifically includes the following steps. Multi-domain feature decoupling and pathological pattern clustering are performed on the myocardial contact stress vector to obtain a set of abnormal myocardial stress patterns. Parzen window kernel density estimation was applied to optimize the window width and perform nonparametric fitting on the abnormal myocardial stress pattern set to obtain the abnormal probability density distribution. The abnormal probability density distribution is reconstructed in feature space to form a contact anomaly event map.

[0011] As a preferred embodiment of the data analysis-based health monitoring method for pacemaker implantees described in this invention, the step of outputting a three-color graded early warning command specifically includes the following steps. A heart sound branch layer and a probability inference layer were constructed, and multi-scale convolution and layer stacking were performed to build a multimodal health monitoring model; The abnormal event spectrum is input into the multimodal health monitoring model, and the heart sound branch layer uses wavelet packet transform to perform high-frequency attenuation feature decomposition to obtain the normalized spectrum of the second heart sound. The probability inference layer applies an attention mechanism to extract and weighted aggregate impedance fluctuation variance to form the probability of conductor loosening risk. The normalized spectrum of the second heart sound and the probability of wire loosening risk are spatiotemporally aligned and analyzed by joint decision tree, and a three-color graded early warning command is output.

[0012] As a preferred embodiment of the data analysis-based health monitoring method for pacemaker implantees described in this invention, the step of acquiring the evolution trajectory of risk events specifically includes the following steps. The three-color graded early warning instructions are correlated across modal events to form a risk propagation topology; The risk propagation topology is subjected to time decay weighting using an exponentially weighted moving average algorithm to generate a time-varying risk probability distribution. By fitting a random walk trajectory to the time-varying risk probability distribution, the evolution trajectory of risk events can be obtained.

[0013] As a preferred embodiment of the data analysis-based method for health monitoring of pacemaker implantees according to the present invention, the step of generating an implantation health monitoring report specifically includes the following steps. Perform state transition probability analysis on the evolution trajectory of risk events to generate a comprehensive cardiac risk index; Based on comprehensive cardiac risk indicators, a cardiac clinical diagnosis is performed on the implant recipient, and an implantation health monitoring report is generated.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the data analysis-based method for health monitoring of pacemaker implantees as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the data analysis-based method for health monitoring of pacemaker implantees as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By collecting biphasic impedance data from cardiac pacemakers, surface electrocardiogram signals, and time synchronization data, and performing dynamic rate-of-change analysis and phase labeling, the fusion efficiency of multi-source physiological information is improved, enabling a more comprehensive reflection of the real-time health status of the heart. Simultaneously, the multimodal health monitoring model enhances the sensitivity and specificity of abnormal event identification, significantly reducing the false positive rate, thereby improving the overall accuracy and reliability of health monitoring and providing more intelligent and refined early warning and diagnostic support for pacemaker implantees. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of a data analysis-based method for health monitoring of pacemaker implantees.

[0019] Figure 2 This is a schematic diagram of the generation of myocardial contact stress vectors.

[0020] Figure 3 A flowchart for generating three-color graded early warning instructions.

[0021] Figure 4 A flowchart for generating implant health monitoring reports. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] 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.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for health monitoring of pacemaker implantees based on data analysis, including the following steps: S1. Collect physiological data sets of implantees, perform dynamic rate of change analysis and phase labeling, and output dynamic impedance sequence and ECG phase labeling data.

[0026] Specifically, the steps include the following: S1.1 Collect the implantee's physiological dataset, which includes biphasic impedance data of the pacemaker, electrocardiogram signals on the implantee's body surface, and time synchronization data.

[0027] Impedance sensors are installed at the electrode interface of the pacemaker to collect biphasic impedance data; ECG acquisition electrodes are embedded in the outer shell of the pacemaker to collect surface ECG signals from the implanter; and atomic clock devices (such as CSAC chips) are integrated into the clock circuit of the main control chip inside the pacemaker to collect time synchronization data.

[0028] S1.2. Preprocessing of the implantee physiological dataset: Specifically, for the biphasic impedance data of the cardiac pacemaker, a sliding window is used for data smoothing to reduce high-frequency noise interference, and synchronous linear interpolation is used to compensate for missing values, improving the continuity and completeness of the cardiac pacemaker biphasic impedance data; for the implantee's surface electrocardiogram signal, a bandpass filter is used for noise suppression to improve the signal-to-noise ratio, and wavelet transform is used for multi-level decomposition to enhance the R-wave recognition accuracy of the implantee's surface electrocardiogram signal; for the time synchronization data, a high-precision clock source is used for timestamp marking to ensure time synchronization, and Min-Max standardization is used for dimensional unification to eliminate time deviation, outputting the preprocessed implantee physiological dataset.

[0029] S1.3. Perform sliding window differencing and time series integration on the preprocessed biphasic impedance data of the cardiac pacemaker to generate a dynamic impedance sequence. In specific operation, a sliding window with a fixed width and step size (e.g., a width of 500ms and a step size of 250ms) is set to perform a sliding scan on the biphasic impedance data of the cardiac pacemaker to obtain an impedance sliding window dataset. Then, cubic spline interpolation is performed on the impedance sliding window dataset to obtain a continuous impedance sampling sequence. Next, the continuous impedance sampling sequence is sampled at adjacent points to obtain adjacent impedance value pairs. The adjacent impedance value pairs are then subjected to first-order forward differencing to generate an impedance change vector. The impedance change vector represents the local rate of change of impedance amplitude and can reflect the electrode coupling state of the cardiac pacemaker.

[0030] Subsequently, the trapezoidal integral algorithm is used to integrate the impedance change vector over time. Further, the impedance change vector is smoothed and filtered and statistically analyzed using moving averages to obtain the impedance time change. The impedance time change is then numerically integrated to generate an initial impedance time series. The initial impedance time series is then periodically phase aligned to ensure synchronization with the cardiac cycle, and a dynamic impedance sequence is output.

[0031] S1.4. The preprocessed implantable patient surface ECG signal is divided into R-wave reference points and phase-marked, and ECG phase-marked data is output. Specifically, the Pan-Tompkins algorithm is used to divide the preprocessed implantable patient surface ECG signal into R-wave reference points. Further, a differential squared transform is performed on the preprocessed implantable patient surface ECG signal to obtain a feature-enhanced signal. Bandpass filtering is performed on the feature-enhanced signal to eliminate baseline drift and electromyographic interference, obtaining an optimized ECG signal. The optimized ECG signal is then subjected to first-order differentiation to obtain the waveform slope. When the waveform slope exceeds a reference threshold, feature points of the optimized ECG signal at the corresponding waveform slope are extracted as R-wave reference points. It should be noted that the baseline threshold is defined based on the dynamic rate of change of historical optimized ECG signals, with an exemplary range of 0.3-0.7 mV / s.

[0032] During the phase marking stage, the optimized ECG signal is divided into phase intervals based on the R-wave reference point. Furthermore, based on clinical ECG characteristic rules, the waveform range of the optimized ECG signal is set. For example, the range of the PQ segment waveform is defined as 80-120ms before the R-wave reference point, the range of the QRS segment waveform is defined as 20ms before to 40ms after the R-wave reference point, the range of the ST segment waveform is set as 40-120ms after the R-wave reference point, and the range of the T segment waveform is set as 120-300ms after the R-wave reference point. It should be noted that the clinical ECG characteristic rules are based on the individualized ECG parameter definitions of the implantee.

[0033] When the optimized ECG signal is detected within the PQ segment waveform range, it is classified as the atrial depolarization phase; when the optimized ECG signal is within the QRS segment waveform range, it is classified as the ventricular depolarization phase; when the optimized ECG signal is within the ST segment waveform range, it is classified as the early ventricular repolarization phase; and when the optimized ECG signal is within the T segment waveform range, it is classified as the late ventricular repolarization phase. The divided phase intervals are timestamped to obtain time phase marker data. The time phase marker data is then integrated to output ECG phase marker data.

[0034] S2. Biomechanical state mapping is performed on the dynamic impedance sequence and ECG phase label data to generate myocardial contact stress vector. Abnormal pattern recognition and probability density modeling are performed on the myocardial contact stress vector to form a contact abnormal event map.

[0035] Specifically, the steps include the following: S2.1. Based on the time synchronization data, dynamic registration is performed on the dynamic impedance sequence and ECG phase marker data, and joint feature encoding is performed synchronously to obtain biomechanical feature tensors. Specifically, in the dynamic registration stage, the Dynamic Time Warping (DTW) algorithm is used to perform time alignment. Furthermore, non-uniform sampling is performed on the dynamic impedance sequence and ECG phase marker data to ensure consistency in temporal resolution, and multimodal fusion is performed to obtain ECG-mechanical coupling data. Based on the time synchronization data, the ECG-mechanical coupling data is elastically matched on the time axis to eliminate temporal differences caused by heart rate variability, and the registered coupling data is obtained. Feature space mapping is performed on the registered coupling data to output a spatiotemporally aligned feature set. It should be noted that time-axis elastic matching refers to the process of time stretching and phase adjustment of electrocardiogram-mechanical coupling data.

[0036] In the joint feature encoding stage, the spatiotemporally aligned feature set is segmented by a sliding window to extract impedance spatial features. Simultaneously, a Fourier transform is performed on the spatiotemporally aligned feature set to extract phase-time features. 3D convolution kernels are used to jointly encode the impedance spatial features and phase-time features. Furthermore, feature cross-convolution is performed on the impedance spatial features and phase-time features to generate spatiotemporally encoded features. Nonlinear activation transformation is then performed on the spatiotemporally encoded features to obtain myocardial coupling encoded features. Subsequently, tensor dimension reorganization is performed on the myocardial coupling encoded features to form a biomechanical feature tensor. The biomechanical feature tensor not only represents the spatiotemporal characteristics of electrocardiographic-mechanical coupling but also reflects the dynamic changes under different cardiac cycles.

[0037] S2.2. Based on the Navier-Stokes stress differential equation, the biomechanical characteristic tensor is mapped to a biomechanical state to form a myocardial contact stress vector. In specific operations, the biomechanical characteristic tensor is input into a finite element solver to perform mesh discretization and obtain discrete data of myocardial tissue. The stress tensor is solved on the discrete data of myocardial tissue through stress differential equations (such as the Navier-Stokes equation) to obtain a preliminary stress distribution vector. Five layers of Gaussian integration points are set in the direction of ventricular wall thickness, and anisotropic stress calculation is performed on the preliminary stress distribution vector to generate myocardial fiber stress values. ; in, Indicates the stress value of myocardial fibers. This represents the integral range along the thickness of the ventricular wall. This represents the hierarchical index of the Gaussian integral point. Indicates the first The weights of each Gaussian integral point at each level This represents the initial stress distribution vector. A unit vector representing the orientation of myocardial fibers; It should be noted that the ventricular wall thickness direction refers to the normal coordinate axis of the discrete data of myocardial tissue, which is obtained by performing normal fitting on the discrete data of myocardial tissue; the integral range along the ventricular wall thickness direction is determined by the thickness distribution parameters of the discrete data of myocardial tissue; and the unit vector of myocardial fiber orientation is obtained by performing tensor convolution on the discrete data of myocardial tissue.

[0038] Mechanical state mapping is performed on the stress values ​​of myocardial fibers. Furthermore, eigenvalue decomposition is performed on the stress values ​​of myocardial fibers, and the top u eigenvalues ​​are extracted as principal component stresses. Spatial interpolation and stress direction projection are performed on the principal component stresses to form a normalized stress projection vector. Component synthesis and amplitude adjustment are performed on the normalized stress projection vector to output the myocardial contact stress vector.

[0039] S2.3. Perform multi-domain feature decoupling and pathological pattern clustering on the myocardial contact stress vector to obtain a set of abnormal myocardial stress patterns. In specific operations, the myocardial contact stress vector is decoupled from features in multiple dimensions. Furthermore, in the time domain dimension, time-varying principal component analysis (t-PCA) is used to extract time-domain dynamic features: the myocardial contact stress vector is segmented into a time series to obtain time-domain subsequences, and multimodal decomposition and matrix recombination are performed on the time-domain subsequences to obtain the time-domain feature matrix. Simultaneously, the variance of the time-domain feature matrix is ​​normalized to form standardized time-domain features. In the frequency domain dimension, frequency domain energy features are extracted through wavelet packet transform decomposition: the myocardial contact stress vector is decomposed into multiple scales to generate frequency domain coefficients, energy spectrum transformation is performed on the frequency domain coefficients to obtain the frequency domain energy distribution; logarithmic transformation is performed on the frequency domain energy distribution to output normalized frequency domain features. In the spatial dimension, the spatial Laplacian operator is used to extract spatial stress gradient features: the second derivative of the myocardial contact stress vector is taken to form stress divergence parameters, directional filtering and amplitude normalization are performed on the stress divergence parameters to obtain the spatial stress distribution, and Gaussian smoothing is performed on the spatial stress distribution to obtain optimized spatial features; the standardized time-domain features, normalized frequency-domain features and optimized spatial features are integrated into a tensor to output a multi-domain fusion feature set; It should be noted that the spatial Laplacian operator is called directly through the ndimage.laplace function in the SciPy library.

[0040] Fuzzy C-means clustering algorithm is used to cluster pathological patterns in the multi-domain fusion feature set. Furthermore, an initial cluster center with a cluster size of 3 (corresponding to three pathological patterns: normal, local ischemia, and myocardial fibrosis) is set to perform kernel space mapping on the multi-domain fusion feature set to generate a high-dimensional feature projection. The silhouette coefficient is used to optimize the cluster center of the high-dimensional feature projection to generate optimized cluster groups. The similarity measure is performed on the optimized cluster groups to obtain the membership value. It should be noted that the silhouette coefficient is defined based on the intra- and inter-class distance ratio of the initial cluster centers, with an exemplary value range of [-1, 1]. Cluster center optimization refers to the process of iteratively adjusting the centroid and reconstructing the boundaries of the high-dimensional feature projection. Performing similarity measurement on the optimized cluster groups refers to measuring the membership relationship between samples and cluster centers in the optimized cluster groups by performing Gaussian fitting on the optimized cluster groups.

[0041] Based on membership values, pathological patterns are classified using a three-level threshold system. For example, the first-level threshold is set to a range of 0.7-1.0, the second-level threshold to a range of 0.4-0.7, and the third-level threshold to a range of 0-0.4. When the membership value is within the range of the first-level threshold, it is defined as a normal pattern; when the membership value is within the range of the second-level threshold, it is defined as a local ischemia pattern; and when the membership value is within the range of the third-level threshold, it is defined as a myocardial fibrosis pattern. The classified local ischemia and myocardial fibrosis patterns are then hierarchically clustered to output a set of abnormal myocardial stress patterns.

[0042] S2.4. Parzen window kernel density estimation is applied to optimize the window width and perform nonparametric fitting on the abnormal myocardial stress pattern set to obtain the abnormal probability density distribution. Specifically, in the window width optimization stage, a sliding window is applied to monitor the abnormal myocardial stress pattern set, and the parameters of local density, range, and standard deviation are statistically analyzed. The parameters of local density, range, and standard deviation are then integrated to generate the initial window width parameters. The initial window width parameters are then expanded to obtain candidate window width intervals, and the maximum likelihood search is performed on the candidate window width intervals to obtain the optimized window width parameters. It should be noted that the specific process of maximum likelihood search is as follows: Gaussian integral is performed on the candidate window width interval to obtain the likelihood value, and the likelihood values ​​are compared to determine the optimal window width parameter.

[0043] In the nonparametric fitting stage, the optimized window width parameter is projected onto the probability density to form a probability density feature set, and the probability density feature set is uniformly sampled to obtain probability density samples. The probability density samples are then subjected to neighborhood weighted summation and density normalization to obtain a smooth probability distribution. Based on clinical image data, the smooth probability distribution is compared with the feature deformation to ensure that the partition characteristics of the smooth probability distribution are consistent with the partition characteristics of the actual pathological images, and an abnormal probability density distribution is output. It should be noted that clinical imaging data refers to coronary angiography and cardiac MRI, acquired using a DSA angiography machine and a 3.0T magnetic resonance imaging system.

[0044] S2.5. Reconstruct the feature space of the abnormal probability density distribution to form a contact anomaly event map. Specifically, the t-SNE manifold learning algorithm is used to map the abnormal probability density distribution from the three-dimensional feature space to the two-dimensional plane. Furthermore, the abnormal probability density distribution is uniformly sampled to obtain high-dimensional probability feature nodes. Adjacent high-dimensional probability feature nodes are weighted and connected to generate an adjacency matrix. The adjacency matrix is ​​then embedded in a low-dimensional space to obtain the initial projection distribution. The initial projection distribution is then optimized by manifold to generate two-dimensional projection data. It should be noted that the specific process of manifold optimization is as follows: gradient descent is performed on the initial projected distribution to obtain the optimization direction, and the learning rate is adjusted on the optimization direction to minimize the KL divergence of the initial projected distribution.

[0045] The feature space of the two-dimensional projected data is reconstructed using Delaunay triangulation. Further, vertex connections and spatial segmentation are performed on the two-dimensional projected data to obtain a triangulated probability mesh. Gradient tracking is then performed on the triangulated probability mesh to form anomaly boundary lines. Curvature quantization and feature labeling are performed on the anomaly boundary lines to obtain curvature feature sub-blocks. Simultaneously, stress gradient mapping is performed on the curvature feature sub-blocks to reconstruct the stress feature space. Spatial superposition is performed on the reconstructed stress feature space to obtain a superimposed stress field. The stress amplitude in the superimposed stress field is extracted and linearly weighted to generate a contact anomaly feature set. Finally, the contact anomaly feature set is dynamically rendered in three dimensions to form a contact anomaly event map.

[0046] S3. Input the abnormal event spectrum into the multimodal health monitoring model. The heart sound branch layer performs high-frequency attenuation feature decomposition, the feature fusion layer extracts and weights the impedance fluctuation variance, and outputs a three-color graded early warning command.

[0047] Specifically, the steps include the following: S3.1 Construct and train a multimodal health monitoring model. Specifically, in the TensorFlow framework, a convolutional neural network (CNN) is called via Keras API parameters. Wavelet packet transform is embedded into the CNN, and the kernel width is set to 50ms, the stride to 10ms, and the pooling window to 20ms. A gated recurrent unit is then added after the CNN to model temporal features, enhancing heart sound rhythm analysis and completing the construction of the heart sound branch layer. A fully connected network architecture is called via the torch.nn function, with the input dimension set to 256, the hidden layer nodes to 128, and the output dimension to 3. An attention mechanism is then added after the fully connected network architecture for feature weighting to enhance feature representation, and a Dropout layer is used for regularization to prevent overfitting, completing the construction of the probability inference layer. Skip connections are used to perform cross-layer fusion of the heart sound branch layer and the probability inference layer to obtain the electro-acoustic fusion feature tensor. Multi-scale convolution is then performed on the electro-acoustic fusion feature tensor to generate multi-scale heart feature maps. Attention weighting is applied to the multi-scale heart feature maps to obtain weighted heart feature representations. Fully connected transformation is then performed on the weighted heart feature representations to generate hierarchical fusion weights. Based on the hierarchical fusion weights, the heart sound branch layer and the probability inference layer are stacked hierarchically, and gradient optimization is performed using residual connections to complete the construction of the multimodal health monitoring model. Next, the multimodal health monitoring model is trained. Further, tensor decomposition and dimensionality compression are performed on the historical contact anomaly event map to obtain low-dimensional feature representations. These low-dimensional feature representations are then divided into a sample set, a training set, and a validation set. On the sample set, data augmentation is used to expand the samples, and batch normalization is used to standardize the features, forming an augmented training set. On the training set, the Adam optimizer is used to update the parameters of the augmented training set, and a cosine annealing learner is simultaneously applied to adjust the learning rate, obtaining optimized parameters for the multimodal health monitoring model. On the validation set, a smooth L1 loss function is used to measure the error of the optimized multimodal health monitoring model parameters, obtaining the training loss value. When the training loss value exceeds the convergence threshold for several consecutive rounds (e.g., 5 times), training terminates, and the model saver is simultaneously applied to output the trained multimodal health monitoring model. It should be noted that the convergence threshold is defined based on the exponentially weighted moving average of the training loss values, with an exemplary range of 0.01-0.05.

[0048] S3.2 The heart sound branch layer uses wavelet packet transform to decompose high-frequency attenuation features and obtain the normalized spectrum of the second heart sound. In specific operation, the contact abnormal event spectrum is input into the heart sound branch layer of the trained multimodal health monitoring model through a real-time streaming data pipeline (such as Apache Kafka). A three-level convolutional neural network is used to perform heart sound rhythm analysis on the contact abnormal event spectrum. Further, the first level uses a one-dimensional convolutional kernel to perform sliding window scanning on the contact abnormal event spectrum to extract the temporal features of the heart sound, and performs ReLU activation and max pooling on the temporal features of the heart sound to output the primary heart sound features. The second level uses dilated convolution to expand the receptive field of the primary heart sound features and uses batch normalization to stabilize the features to obtain the enhanced heart sound features. The third level captures the rhythmic periodic features of the enhanced heart sound features through a gated recurrent unit, and performs dynamic feature aggregation on the rhythmic periodic features to generate a heart sound rhythm sequence. Then, the heart sound rhythm sequence is compressed to output the heart sound rhythm feature vector.

[0049] Wavelet packet transform is used to decompose the high-frequency attenuation feature vector of the heart sound rhythm. Further, multi-scale decomposition is performed on the heart sound rhythm feature vector to obtain sub-band spectra. Sub-band spectra exceeding the soft threshold are removed, and sub-band spectra below the soft threshold are retained to generate filtered spectra. Feature recombination and energy equalization are performed on the filtered spectra to obtain high-frequency attenuation features. The high-frequency attenuation features are then subjected to frequency band decomposition and amplitude normalization to output the normalized spectrum of the second heart sound. It should be noted that the soft threshold is defined based on the energy distribution of the subband spectrum, with an exemplary value range of 0.15-0.35.

[0050] S3.3, the probabilistic inference layer applies an attention mechanism to extract and weighted aggregate impedance fluctuation variance, forming the probability of conductor loosening risk. Specifically, the attention mechanism is used to analyze the fluctuation variance of the contact anomaly event map. Further, the impedance feature vectors of the contact anomaly event map are extracted, and linear projection is performed on these vectors to obtain key vectors, value vectors, and query vectors. A dot product operation is then performed on the key vectors, value vectors, and query vectors to obtain the attention weight matrix. Within a fixed time step (e.g., 500ms), the variance of the attention weight matrix is ​​quantized to obtain a window variance sequence, and a weighted average is performed on the window variance sequence to generate the fluctuation variance. The specific mathematical formula is as follows. ; in, Indicates the variance of the fluctuation. Indicates the total time step. Indicates the time step index. Indicates the first Query vectors at each time step No. The key vector at each time step This indicates the transpose operation. Scaling factor Indicates the first A vector of values ​​for each time step; It should be noted that the scaling factor is defined based on the feature dimension of the key vector, with exemplary values ​​ranging from 32 to 128.

[0051] A two-layer fully connected network is used to model the risk probability of the volatility variance. The first layer performs a linear transformation and ReLU activation on the volatility variance to obtain nonlinear risk features. The second layer performs weighted aggregation and feature compression on the nonlinear risk features to obtain a risk score vector. The risk score vector is then subjected to Sigmoid projection and probability normalization to output the probability of conductor loosening risk.

[0052] S3.4. Perform spatiotemporal alignment and joint decision tree analysis on the standardized spectrum of the second heart sound and the probability of wire loosening risk to output a three-color graded early warning command. In specific operation, the standardized spectrum of the second heart sound and the probability of wire loosening risk are spatiotemporally aligned using a spatiotemporal coupling registration algorithm (such as the 3D-CPD algorithm). Furthermore, the standardized spectrum of the second heart sound and the probability of wire loosening risk are aligned in execution time to ensure phase synchronization, resulting in an aligned bimodal sequence. The aligned bimodal sequence is then spatially registered to generate a spectrum-probability spatiotemporal coupling matrix. The spectrum-probability spatiotemporal coupling matrix is ​​then subjected to eigenvalue decomposition and dimensionality compression to obtain a spectrum-probability joint feature representation.

[0053] Next, the random forest algorithm is used to perform joint decision tree analysis on the spectrum-probability joint feature representation. Furthermore, the spectrum-probability joint feature representation is sorted according to the Gini coefficient rule to improve feature discrimination and obtain an ordered coupled sequence. The time-domain mean parameter and frequency-domain energy parameter of the ordered coupled sequence are extracted. The time-domain mean parameter is used as a node and the frequency-domain energy parameter is used as a branch to construct a decision tree. The decision tree is then split into nodes to obtain the parent node sample set and the child node sample set. The information entropy difference between the parent node sample set and the child node sample set is calculated to obtain the split gain value. ; in, Indicates the split gain value. Represents the sample set of the parent node. This represents the information entropy of the parent node's sample set. Indicates the total number of child nodes. Indicates the child node index. Indicates the first A sample set of child nodes, Indicates the first Information entropy of the sample set of each child node; It should be noted that the Gini coefficient rule is based on the definition of the sum of squares of sample class impurities in the historical decision tree.

[0054] The split gain value is subjected to a nonlinear transformation to obtain a decision feature vector. Parameter fusion is then performed on the decision feature vector to output joint decision parameters. The joint decision parameters are then classified into risk levels and warnings are issued. Further, a probability transformation is performed on the joint decision parameters to obtain the decision risk value. Risk ranges are defined for the decision risk value; for example, a high-risk range of 0.7-1.0, a medium-risk range of 0.4-0.7, and a safe range of 0-0.4 are defined. When the decision risk value is within the high-risk range, it is determined to be high-risk, and a red warning is triggered. When the decision risk value is within the medium-risk range, it is determined to be medium-risk, and a yellow warning is triggered. When the decision risk value is within the safe range, it is determined to be a safe state, and a green warning is triggered. The risk level and warning color are encapsulated into instructions, and a three-color graded warning instruction is output. It should be noted that the risk range is defined based on the anomaly detection rate of historical joint decision parameters.

[0055] S4. Perform time decay weighted analysis on the execution of the three-color graded early warning instructions to obtain the evolution trajectory of risk events; perform state transition probability analysis on the evolution trajectory of risk events to generate an implanted health monitoring report.

[0056] Specifically, the steps include the following: S4.1. The three-color graded early warning instructions are subjected to cross-modal event association and time decay weighting to generate a time-varying risk probability distribution. In specific operations, the cross-modal event association stage is performed using a Bayesian inference algorithm. Further, the risk parameters of the three-color graded early warning instructions are transformed to generate a risk early warning dataset. Markov chain Monte Carlo sampling is performed on the risk early warning dataset to obtain the initial association distribution. The initial association distribution is subjected to posterior probability inference to obtain the posterior probability distribution. The posterior probability distribution is then subjected to sparsification and matrix filling to output the cross-modal event association matrix. It should be noted that Markov chain Monte Carlo sampling refers to the process of iteratively sampling the risk warning dataset until it approximates the target association distribution (defined based on the convergence characteristics of the historical initial association distribution); posterior probability inference refers to the process of performing Bayesian updates and marginalization on the initial association distribution.

[0057] Extract the diagonal matrix elements of the cross-modal event correlation matrix and define them as topological nodes. Extract the off-diagonal matrix elements of the cross-modal event correlation matrix and define them as topological connecting edges. Combine the topological nodes and topological connecting edges into a graph structure to output the risk propagation topology.

[0058] The risk propagation topology is further spatiotemporally coupled using an exponentially weighted moving average algorithm to generate a risk propagation map. The risk propagation map is then weighted and aggregated using a decay factor to obtain a weighted risk map. Spatial smoothing and noise suppression are then applied to the weighted risk map to obtain a time-varying risk probability distribution. It should be noted that the attenuation factor is defined based on the dynamic stability of historical risk propagation patterns, and the exemplary value range is 0.2-0.5.

[0059] S4.2. Fit the random walk trajectory of the time-varying risk probability distribution to obtain the evolution trajectory of risk events. In specific operations, the time-varying risk probability distribution is sliced ​​over time to extract the time dimension parameters. The time-varying risk probability distribution is decomposed into intensity gradients to extract the risk intensity parameters. At the same time, the time-varying risk probability distribution is spatially mapped to extract the spatial location parameters. The time dimension parameters are used as the horizontal axis, the risk intensity parameters as the vertical axis, and the spatial location parameters as the depth axis to construct a three-dimensional state space (time × risk intensity × spatial location).

[0060] A random walk is performed in the three-dimensional state space using the Metropolis-Hastings algorithm. Furthermore, the spatial extreme points of the time-varying risk probability distribution are used as the starting points of the walk. State transitions and direction guidance are performed on the starting points to obtain the extended trajectory. During the walk, the risk intensity values ​​of the trajectory points in the extended trajectory are recorded. When the risk intensity value of a trajectory point is greater than the intensity threshold, the walk is terminated, and a discrete trajectory point set is output. Spline interpolation is performed on the discrete trajectory point set to generate the risk event evolution trajectory. It should be noted that spatial extreme points are determined by performing local maximum detection on the time-varying risk probability distribution; state transition refers to the process of probabilistically jumping from the starting point of the walk; the intensity threshold is defined based on the statistical distribution of the risk intensity values ​​of historical trajectory points, with an exemplary value range of 0.65-0.75.

[0061] S4.3. Perform state transition probability analysis on the evolution trajectory of risk events to generate a comprehensive cardiac risk index. Based on the comprehensive cardiac risk index, conduct cardiac clinical diagnosis on the implantee and generate an implantation health monitoring report. In specific operations, perform state transition probability analysis on the evolution trajectory of risk events. Further, perform adjacent transition state statistics on the evolution trajectory of risk events to obtain the transition frequency. Normalize the transition frequency to form a state transition probability matrix. Perform eigenvalue decomposition on the state transition probability matrix, extract the first b eigenvalues ​​and concatenate them into tensors to obtain the principal eigenvector. Perform weighted fusion and spatial diffusion quantization on the principal eigenvector to generate the comprehensive cardiac risk index. It should be noted that adjacent transition state statistics refers to the process of counting and accumulating the risk intensity changes in the trajectory of risk events, while spatial diffusion quantification refers to measuring the spatial distance of the main feature vector to determine the spatial propagation range of risk events.

[0062] Based on the cardiac risk comprehensive index, a cardiac clinical diagnosis is performed on implant recipients. Furthermore, a three-tiered diagnostic rule is established based on the cardiac risk comprehensive index: when the cardiac risk comprehensive index is in the first-level warning range (e.g., 0.8-1.0), it is defined as an emergency state requiring immediate clinical intervention; when the cardiac risk comprehensive index is in the second-level warning range (e.g., 0.5-0.8), it is defined as an observation state requiring enhanced monitoring; when the cardiac risk comprehensive index is in the third-level warning range (e.g., 0-0.5), it is defined as a stable state requiring routine follow-up. It should be noted that the warning interval is defined based on the dynamic change rate of historical cardiac risk comprehensive indicators.

[0063] The system integrates cardiac clinical diagnosis results, implantee medication records (collected through electronic medical records), and real-time implantee physiological datasets, transforms them into structured descriptions, and outputs implantation health monitoring reports.

[0064] This embodiment also provides a computer device applicable to the data analysis-based health monitoring method for pacemaker implantees, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the data analysis-based health monitoring method for pacemaker implantees proposed in the above embodiment.

[0065] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0066] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the data analysis-based health monitoring method for pacemaker implantees as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0067] In summary, this invention improves the efficiency of fusing multi-source physiological information by acquiring biphasic impedance data from cardiac pacemakers, surface electrocardiogram signals, and time-synchronized data, and performing dynamic rate-of-change analysis and phase labeling. This allows for a more comprehensive reflection of the real-time health status of the heart. Simultaneously, the multimodal health monitoring model enhances the sensitivity and specificity of abnormal event identification, significantly reducing the false positive rate. This improves the overall accuracy and reliability of health monitoring, providing pacemaker implant recipients with more intelligent and refined early warning and diagnostic support.

[0068] 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 method for health monitoring of pacemaker implantees based on data analysis, characterized in that: include, Collect physiological datasets from implant recipients, perform dynamic rate of change analysis and phase labeling, and output dynamic impedance sequences and ECG phase labeling data; Biomechanical state mapping is performed on dynamic impedance sequence and electrocardiogram phase label data to generate myocardial contact stress vector. Abnormal pattern recognition and probability density modeling are performed on myocardial contact stress vector to form a contact abnormal event map. The abnormal event spectrum is input into the multimodal health monitoring model. The heart sound branch layer performs high-frequency attenuation feature decomposition, the probability inference layer extracts and weights the impedance fluctuation variance, and outputs a three-color graded early warning command. By weighting the execution time decay of the three-color graded early warning instructions, the evolution trajectory of risk events can be obtained; Perform state transition probability analysis on the evolution trajectory of risk events to generate implanted health monitoring reports.

2. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 1, characterized in that: The implantee physiological dataset includes biphasic impedance data from the pacemaker, electrocardiogram signals from the implantee's body surface, and time synchronization data.

3. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 2, characterized in that: The output dynamic impedance sequence and ECG phase marker data specifically include the following steps. A sliding window differential is performed on the biphasic impedance data of the cardiac pacemaker to obtain the impedance change vector; the impedance change vector is then integrated over time to generate a dynamic impedance sequence. The implanter performs R-wave reference point division and phase marking on the surface electrocardiogram signal, and outputs electrocardiogram phase marking data.

4. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 1, characterized in that: The generation of the myocardial contact stress vector specifically includes the following steps. Based on time synchronization data, dynamic impedance sequence and ECG phase marker data are dynamically registered, and feature joint encoding is performed synchronously to obtain biomechanical feature tensors. Based on the stress differential equation, the biomechanical characteristic tensor is mapped to a biomechanical state to form the myocardial contact stress vector.

5. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 4, characterized in that: The formation of the contact anomaly event map specifically includes the following steps. Multi-domain feature decoupling and pathological pattern clustering are performed on the myocardial contact stress vector to obtain a set of abnormal myocardial stress patterns. Parzen window kernel density estimation was applied to optimize the window width and perform nonparametric fitting on the abnormal myocardial stress pattern set to obtain the abnormal probability density distribution. The abnormal probability density distribution is reconstructed in feature space to form a contact anomaly event map.

6. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 1, characterized in that: The output of the three-color graded early warning command specifically includes the following steps. A heart sound branch layer and a probability inference layer were constructed, and multi-scale convolution and layer stacking were performed to build a multimodal health monitoring model; The abnormal event spectrum is input into the multimodal health monitoring model, and the heart sound branch layer uses wavelet packet transform to perform high-frequency attenuation feature decomposition to obtain the normalized spectrum of the second heart sound. The probability inference layer applies an attention mechanism to extract and weighted aggregate impedance fluctuation variance to form the probability of conductor loosening risk. The normalized spectrum of the second heart sound and the probability of wire loosening risk are spatiotemporally aligned and analyzed by joint decision tree, and a three-color graded early warning command is output.

7. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 6, characterized in that: The acquisition of the evolution trajectory of risk events specifically includes the following steps. The three-color graded early warning instructions are correlated across modal events to form a risk propagation topology; The risk propagation topology is subjected to time decay weighting using an exponentially weighted moving average algorithm to generate a time-varying risk probability distribution. By fitting a random walk trajectory to the time-varying risk probability distribution, the evolution trajectory of risk events can be obtained.

8. The method for health monitoring of pacemaker implantees based on data analysis as described in claim 1, characterized in that: The process of generating the implantation health monitoring report specifically includes the following steps. Perform state transition probability analysis on the evolution trajectory of risk events to generate a comprehensive cardiac risk index; Based on comprehensive cardiac risk indicators, a cardiac clinical diagnosis is performed on the implant recipient, and an implantation health monitoring report is generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the data analysis-based cardiac pacemaker implantation health monitoring method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the data analysis-based heart pacemaker implantation health monitoring method according to any one of claims 1 to 8.

Images