A multi-fidelity electrocardiogram fusion method and device based on lesion semantic space

Abstract

The application provides a multi-fidelity electrocardio fusion method and device based on lesion semantic space, which comprises: acquiring electrical physiological feature representations corresponding to simulation electrocardio data and real electrocardio data respectively, and clustering and grouping the electrical physiological feature representations according to lesion semantic similarity to obtain a lesion cluster; determining physiological detail residuals of real electrocardio data in the lesion cluster and simulation electrocardio data in the lesion cluster; correcting the simulation electrocardio data in the lesion cluster according to the physiological detail residuals to obtain a fused ventricular electrical physiological feature representation; and constructing a multi-fidelity fusion training set for training an artificial intelligence model according to the electrical physiological feature representation of the real electrocardio data and the fused ventricular electrical physiological feature representation. Through lesion-level multi-fidelity residual fusion, the electrical physiological features in the real data are migrated to the lesions of a large number of simulation samples, so that the model has good sample coverage capability and the ability to depict real physiological details, and the stability and accuracy in real clinical scenarios are improved.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram imaging, and specifically to a multi-preservation electrocardiogram fusion method and apparatus based on lesion semantic space. Background Technology

[0002] In electrocardiogram imaging (ECGI) tasks, real-world data with paired annotations of "12-lead ECG + LGE-CMR infarction image" are extremely scarce. However, "simulated ECG + transmembrane potential distribution on the heart surface" generated based on bioelectric models and numerical simulations can be obtained on a large scale, but there is a "low fidelity" problem, that is, it is difficult to fully depict the complex anatomical variations, fiber orientation, membrane electrical properties and conduction velocity of real patients.

[0003] Existing multi-fidelity modeling methods integrate low-fidelity and high-fidelity data through methods such as sharing latent spaces, transfer learning, or weighted combination. However, in the specific scenario of electrocardiogram imaging: (1) High-fidelity data is scarce and the distribution of cases is uneven. When directly projected onto a unified potential space, it is easy to be biased towards a small amount of real data, making it impossible for the model to make full use of the rich simulation data.

[0004] (2) To improve diversity, simulation data often generates a large number of widely distributed samples by changing parameters such as ventricular conduction velocity, resulting in a significant deviation between the characteristic distribution and the real cases.

[0005] In summary, existing methods address the significant discrepancies between simulated and real ECGs in terms of lesion distribution and electrophysiological details. Summary of the Invention

[0006] In view of the aforementioned problems, this application is proposed to provide a multi-sensor electrical fusion method and apparatus based on lesion semantic space to overcome or at least partially solve the aforementioned problems, comprising: A multi-sensor ECG fusion method based on lesion semantic space is used to construct a training set that fuses simulated and real ECG data, including the following steps: Electrophysiological feature representations corresponding to simulated ECG data and real ECG data are obtained respectively, and the electrophysiological feature representations are clustered and grouped according to the semantic similarity of lesions to obtain lesion clusters; Determine the physiological detail residuals between the actual electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster; Based on the physiological detail residuals, the simulated electrocardiogram data within the lesion clusters are corrected to obtain a fused ventricular electrophysiological feature representation; A multi-fidelity fusion training set for training an artificial intelligence model is constructed based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

[0007] Further, the steps of obtaining the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively include: The simulated ECG data and the real ECG data are mapped to the grid node space of the preset ventricular geometric model to obtain the corresponding node-level initial features; Based on the ventricular topology of the ventricular geometric model, graph convolution operation is performed on the node-level initial features to obtain the corresponding node-level temporal features; The node-level temporal features are subjected to temporal modeling to generate the corresponding electrophysiological feature representations.

[0008] Furthermore, the step of clustering the electrophysiological feature representations based on lesion semantic similarity to obtain lesion clusters includes: The electrophysiological features are input into a preset lesion embedding network to generate a corresponding low-dimensional lesion semantic embedding vector. Clustering is performed based on the distance between the low-dimensional lesion semantic embedding vectors to obtain lesion clusters.

[0009] Furthermore, the step of clustering and grouping lesion clusters based on the distance between the low-dimensional lesion semantic embedding vectors includes: Determine the initial centroid vector in the low-dimensional lesion semantic embedding vector; Determine the Euclidean distance between the low-dimensional lesion semantic embedding vector and the initial centroid vector, and assign the low-dimensional lesion semantic embedding vector to the cluster corresponding to the nearest centroid; Determine the mean of all low-dimensional lesion semantic embedding vectors within the corresponding cluster, and update the mean vector to the new centroid vector of the corresponding cluster; After the iteration is completed, the set of all original ECG data samples of the corresponding cluster is defined as a lesion cluster.

[0010] Further, the step of determining the physiological detail residuals between the actual electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster includes: Determine the first statistical center representing the electrophysiological characteristics of the actual electrocardiogram data within the lesion cluster; Determine the second statistical center for the electrophysiological characteristics of the simulated electrocardiogram data within the lesion cluster; The physiological detail residuals are determined based on the first statistical center and the second statistical center.

[0011] Furthermore, the formula for calculating the physiological detail residual is as follows:

[0012] In the formula, For the physiological detail residuals; This refers to the first statistical center; It is the second statistical center; The calculation formula for the first statistical center is as follows:

[0013] In the formula, A collection of high-fidelity samples within a lesion cluster; For lesion embedding; The calculation formula for the second statistical center is as follows:

[0014] In the formula, This is a collection of low-fidelity samples within a lesion cluster.

[0015] Further, the step of correcting the simulated electrocardiogram data within the lesion cluster based on the physiological detail residuals to obtain a fused representation of ventricular electrophysiological characteristics includes: Based on the physiological detail residuals, the electrophysiological feature representation of the simulated electrocardiogram data within the lesion cluster is modified by residual enhancement through a learnable mapping network to obtain the fused ventricular electrophysiological feature representation.

[0016] A multi-sensor ECG fusion device based on lesion semantic space, used to construct a training set for fusing simulation and real ECG data, includes: The clustering and grouping module is used to obtain the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively, and to cluster and group the electrophysiological feature representations according to the semantic similarity of the lesions to obtain lesion clusters; The residual calculation module is used to determine the physiological detail residuals between the actual ECG data within the lesion cluster and the simulated ECG data within the lesion cluster. The correction and fusion module is used to correct the simulated electrocardiogram data within the lesion cluster based on the physiological detail residuals, so as to obtain a fused ventricular electrophysiological feature representation; The training set construction module is used to construct a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

[0017] A computer electronic device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When executed by the processor, the computer program implements the steps of the multi-security true-to-false electrical fusion method based on lesion semantic space as described above.

[0018] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multi-security true-to-false electrical fusion method based on lesion semantic space as described above.

[0019] This application has the following advantages: In the embodiments of this application, addressing the significant discrepancies between simulated and real ECGs in terms of lesion region distribution and electrophysiological details in existing technologies, this application provides a multi-fidelity ECG fusion solution based on lesion regions as the basic unit. Specifically, this involves: acquiring the electrophysiological feature representations corresponding to simulated and real ECG data respectively; clustering these electrophysiological feature representations based on lesion semantic similarity to obtain lesion clusters; determining the physiological detail residuals between real and simulated ECG data within the lesion clusters; correcting the simulated ECG data within the lesion clusters based on these physiological detail residuals to obtain a fused ventricular electrophysiological feature representation; and constructing a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representations of the real ECG data and the fused ventricular electrophysiological feature representations. Through lesion-level multi-fidelity residual fusion, key electrophysiological features from real data are selectively transferred to the lesion representations of a large number of simulated samples, enabling the model to simultaneously possess good sample coverage and the ability to depict real physiological details, thereby improving stability and accuracy in real clinical scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solution of this application, the drawings used in the description of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the steps of a multi-factorial electronic fusion method based on lesion semantic space provided in an embodiment of this application; Figure 2 This is a diagram showing the physiological detail differences in a multi-sensor electrocardiogram provided in an embodiment of this application; Figure 3 This is a schematic diagram of a heart-shaped cutting procedure provided in an embodiment of this application; Figure 4 This is a graph showing the experimental results of different simulation sample sizes provided in one embodiment of this application; Figure 5 These are various types of myocardial infarction detection results provided in one embodiment of this application; Figure 6This is a structural block diagram of a multi-factorial electro-pharmacy fusion device based on lesion semantic space provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer electronic device provided in an embodiment of the present invention; 1. Computer electronic device; 2. External device; 3. Processing unit; 4. Bus; 5. Network adapter; 6. I / O interface; 7. Display; 8. Memory; 9. Random access memory; 10. Cache memory; 11. Storage system; 12. Program / utility; 13. Program module. Detailed Implementation

[0022] To make the objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] The inventors, through analysis of existing technologies, discovered that there is currently no mature, specialized modeling and reconstruction method for multi-fidelity ECG data in the field of electrocardiogram (ECG) imaging. Existing research on multi-fidelity data fusion and multi-fidelity learning mainly focuses on computer science and engineering simulation, such as multi-fidelity fluid dynamics calculations, structural mechanics simulations, and multi-fidelity prior construction in Bayesian optimization. These methods typically target approximate calculations of scalar or low-dimensional physical quantities, lacking specific design for high-dimensional spatiotemporal signals and anatomically constrained scenarios like ECG-electrophysiology. Therefore, directly transferring these general multi-fidelity fusion ideas to the ECG-GI problem reveals a series of incompatibilities, mainly in the following aspects: (1) The prior of “lesion-electrophysiological consistency” is not explicitly utilized. Common multifidelity methods usually only align data of different fidelity at the sample level or function value level, without considering the medical prior of “samples in the same lesion area should be similar in potential representation” in disease scenarios such as myocardial infarction, and also lack the mechanism to organize samples according to clinical semantics such as coronary blood supply area and infarction combination; (2) It is difficult to handle the strong distribution shift between simulated ECG and real ECG. Existing multifidelity methods often assume that high-fidelity and low-fidelity data only have "precision differences" on the same physical quantity. However, simulated ECG often generates a large number of samples by changing parameters such as conduction velocity to increase diversity, making its feature distribution more dispersed and its morphological differences greater than those of real ECG. Simple kernel regression, linear conformal transformation or shared latent space methods are difficult to effectively align this shift. (3) Lack of fusion strategy that deeply integrates with ventricular geometry and inverse problem structure. Most existing multifidelity fusion is carried out on numerical scalar or low-dimensional vector spaces. In ECGI, a highly ill-posed problem that is strongly constrained by ventricular geometry and forward operators, there is no systematic solution that integrates multifidelity fusion with unified ventricular grid, electrophysiological time series and lesion spatial distribution modeling.

[0024] To address the aforementioned shortcomings, this invention proposes a multi-fidelity real ECG fusion method based on detailed patient physiological information. Within a unified ventricular geometric grid and lesion semantic space, it collaboratively models and performs residual fusion of high-fidelity real ECG and low-fidelity simulated ECG, effectively injecting the electrophysiological details contained in limited real data into large-scale simulation samples, thereby obtaining a fusion representation that has both broad coverage and preserves the physiological characteristics of real patients.

[0025] It should be noted that, in any embodiment of the present invention, the direct purpose of the multi-fidelity ECG fusion method based on lesion semantic space provided by the present invention is to construct a high-quality, high-generalization-capability multi-fidelity fusion training set. This training set aims to integrate the breadth and scale of "simulated ECG data" with the physiological detail and realism of "real ECG data" to address the fundamental constraint on the performance of data-driven models caused by the scarcity of real clinical labeled data. Ultimately, the model trained using this fusion training set can significantly improve its accuracy and robustness in locating and classifying myocardial infarction or reconstructing the cardiac electrophysiological field in real, complex clinical scenarios.

[0026] Reference Figure 1 This application illustrates a multi-factor authentication method for electronic communication fusion based on lesion semantic space, comprising the following steps: S110. Obtain the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively, and cluster the electrophysiological feature representations according to the semantic similarity of the lesions to obtain lesion clusters; S120. Determine the physiological detail residuals between the real electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster; S130. Based on the physiological detail residuals, the simulated electrocardiogram data within the lesion cluster is corrected to obtain a fused ventricular electrophysiological feature representation. S140. Construct a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

[0027] In the embodiments of this application, addressing the significant discrepancies between simulated and real ECGs in terms of lesion region distribution and electrophysiological details in existing technologies, this application provides a multi-fidelity ECG fusion solution using lesion regions as the basic unit. Through lesion-level multi-fidelity residual fusion, key electrophysiological features from real data are selectively transferred to the lesion representation of a large number of simulated samples, enabling the model to simultaneously possess good sample coverage and the ability to depict real physiological details, thereby improving its stability and accuracy in real clinical scenarios.

[0028] The following will further explain a multi-factor electronic fusion method based on lesion semantic space in this exemplary embodiment.

[0029] In one embodiment of the present invention, the specific process of step S110, which involves "obtaining the electrophysiological feature representations corresponding to simulated ECG data and real ECG data respectively, and clustering the electrophysiological feature representations according to the semantic similarity of lesions to obtain lesion clusters", can be further explained in conjunction with the following description.

[0030] In one embodiment of the present invention, the specific process of "obtaining the electrophysiological feature representations corresponding to simulated ECG data and real ECG data respectively" can be further explained in conjunction with the following description.

[0031] The simulated ECG data and real ECG data are mapped to the grid node space of the preset ventricular geometric model to obtain the corresponding node-level initial features; Based on the ventricular topology of the ventricular geometric model, graph convolution operation is performed on the node-level initial features to obtain the corresponding node-level temporal features; The node-level temporal features are subjected to temporal modeling to generate the corresponding electrophysiological feature representations.

[0032] It should be noted that the simulated (low-fidelity) electrocardiogram (ECG) data described in this invention refers to data generated through numerical simulation using a computer-aided cardiac electrophysiological model. During its generation, key physiological parameters are allowed to be perturbed within a reasonable range to increase data diversity. This type of data typically includes simulated surface ECGs, a strictly corresponding spatiotemporal distribution field of transmembrane potentials on the cardiac surface, and idealized myocardial infarction region labels directly determined by simulation parameters. Its advantage is that it can generate tens of thousands of data points, covering various situations; its disadvantage is that it is too idealized and does not consider the complexities of real-life situations, such as the different orientations of blood vessels in each person's heart, differences in myocardial conduction velocity, and slight deviations in electrode placement.

[0033] The real (high-fidelity) electrocardiogram (ECG) data described in this invention refers to data collected from real patients in a clinical setting. It includes at least a simultaneously recorded standard 12-lead ECG, and spatial annotation information of the corresponding myocardial infarction area is obtained through LGE-CMR (delayed gadolinium-enhanced cardiac magnetic resonance imaging) images or clinical expert interpretation. This type of data reflects real physiological details and pathological conditions. Its advantages are realistic detail and perfect matching of clinical conditions; its disadvantages are a small sample size and uneven case distribution.

[0034] like Figure 2 As shown, a diagram illustrating the physiological differences between a simulated electrocardiogram (ECG) and a real ECG is presented.

[0035]

[0036] Table 1. Composition of low-fidelity and high-fidelity datasets As shown in Table 1, the case composition distributions of the two types of data differ significantly. If they are directly mixed, the model will be heavily biased towards the distribution with the larger sample size during learning, thus failing to learn the important but rare case types in the real world.

[0037] Both real and simulated ECG data are standard 12-lead ECGs in form, but they differ in their generation methods and the richness of physiological details: For simulated data, in addition to the 12-lead ECG, a complete ventricular transmembrane potential field and an idealized label of the myocardial infarction area can usually be obtained; for real data, in addition to ECG, spatial distribution information of the infarct area can be obtained through LGE-CMR or expert annotation in most cases, but it is difficult to directly obtain an accurate TMP distribution field (transmembrane potential distribution field).

[0038] As an example, 12-lead ECG signals from simulation and clinical settings underwent preprocessing including noise suppression, baseline drift correction, cardiac cycle clipping, and amplitude normalization to standardize them into cardiac cycle sequences of consistent length and dimensions. Corresponding infarct region annotations were then compiled onto a unified ventricular geometry model, constructing a multi-fidelity dataset composed of standardized ECG and ventricular infarction spatial information. Furthermore, surface ECG data were uniformly mapped to an initial electrophysiological field defined on a ventricular grid through ECG inverse problem modeling. The unified ventricular geometry model is a discretized three-dimensional mesh structure G=(V, E). All ECG data were mapped to this same geometry model for subsequent calculations to ensure consistency of spatial coordinates.

[0039] In a specific implementation, to facilitate multi-fidelity fusion, this example logically organizes all data into binary tuples. :in This is a 12-lead ECG sequence. This provides the corresponding myocardial infarction location information on a unified cardiac grid.

[0040] For simulated electrocardiogram data The infarct area set during simulation can be mapped onto a uniform grid. For real ECG data, The scar areas identified through LGE-CMR images are then registered and mapped onto the same uniform grid after image processing.

[0041] To address the discrepancies in physiological details between simulated and real electrocardiograms (ECGs), particularly the potential for myocardial infarction localization errors due to a lack of patient-specific details in the underlying representations, a mapping from the body surface ECG space to the ventricular electrophysiological space is first constructed on a unified ventricular geometric model. The core idea of ​​this mapping is to treat the 12-lead body surface ECG signal as an observation of the body surface domain. By applying linear dimensionality increase and ventricular topological constraints, the ventricular transmembrane potential is projected onto the ventricular surface grid nodes, resulting in a representation of the ventricular transmembrane potential that reflects the local lesion electrophysiological pattern. This serves as the basis for a unified electrophysiological representation for subsequent multi-fidelity fusion and detailed modeling of lesions.

[0042] Specifically, a unified ventricular geometric model is selected, and the ventricular surface is discretized into... A fixed set of ventricular nodes and their adjacency relationships are obtained from a grid of nodes. A standardized 12-lead ECG matrix is ​​generated for each sample. (in m =12, t (for time length), through a linear dimension-upgrading operator This elevates the signal from the lead dimension to the ventricular node dimension, expanding the "12-lead signal" to... n "Initial features of each ventricular node", resulting in a size of The intermediate representation. This step can be understood as rearranging and projecting the surface electrical information onto various nodes on the ventricular surface based on the forward conduction relationship and network parameters.

[0043]

[0044] After obtaining the initial node-level features, the adjacency matrix of the ventricular topology is further utilized. and the corresponding degree matrix Apply graph convolution or graph attention operations to the ventricular grid. By performing feature aggregation between each node and its neighbors. , Transforming the spatial correlations implicit in surface electrocardiograms into electrophysiological distributions on the ventricular surface allows nodes within the same coronary artery supply region to exhibit greater consistency in the feature space, while nodes crossing anatomical boundaries maintain appropriate differences. This process is equivalent to embedding surface electrocardiogram features into the ventricular geometry, reorganizing and rearranging the information originally "mixed in 12 leads" according to the true ventricular topology.

[0045] To better characterize the temporal evolution of depolarization-repolarization, this invention, after embedding in the graph structure space, can further refine the process using a thinning operator in the temporal direction. Recursively updating node-level temporal features, for example using one-dimensional convolution, recurrent neural networks, or self-attention structures, allows for dynamic modeling of the time series corresponding to each node. , and Combination operations can convert surface electrocardiogram data. Mapped to transmembrane potentials defined on ventricular node × time. This representation not only preserves the basic morphology and propagation direction of myocardial action potentials in terms of numerical values, but also aggregates the electrophysiological characteristics of the same lesion area under the constraint of ventricular geometry.

[0046]

[0047] Through the above mapping steps, multi-source, multi-fidelity 12-lead ECG data are uniformly transformed into the same ventricular electrophysiological space: on the one hand, each sample obtains an electrophysiological field with clear anatomical meaning on the ventricular grid. This facilitates the extraction of electrophysiological features within the lesion area and the construction of lesion embeddings. On the other hand, samples within the same blood supply area (such as LAD, RCA, LCX and their combinations) naturally exhibit similar latent patterns in this space, enabling the electrophysiological prior that "the same coronary blood supply area has similar latent electrophysiological patterns" to be explicitly reflected in the representation.

[0048] In one embodiment of the present invention, the specific process of "clustering the electrophysiological feature representations according to the semantic similarity of lesions to obtain lesion clusters" can be further explained in conjunction with the following description.

[0049] The electrophysiological features are input into a preset lesion embedding network to generate a corresponding low-dimensional lesion semantic embedding vector. Clustering is performed based on the distance between the low-dimensional lesion semantic embedding vectors to obtain lesion clusters. Specifically, the initial centroid vector in the low-dimensional lesion semantic embedding vectors is determined; the Euclidean distance between the low-dimensional lesion semantic embedding vectors and the initial centroid vector is determined, and the low-dimensional lesion semantic embedding vectors are assigned to the cluster corresponding to the nearest centroid; the mean of all low-dimensional lesion semantic embedding vectors in the corresponding cluster is determined, and the mean vector is updated to the new centroid vector of the corresponding cluster; after iteration, the set of all original ECG data samples of the corresponding cluster is defined as a lesion cluster.

[0050] It should be noted that after completing the unified preprocessing of multi-source ECG data and the ventricular electrophysiological spatial mapping, an electrophysiological representation defined on the ventricular grid is obtained. Subsequently, a lesion-oriented embedding and clustering mechanism was introduced to organize multi-fidelity samples in the "lesion semantic space," alleviating the mismatch in feature distribution between simulated and real ECGs. Simulated data typically compensates for the lack of real physiological details by adjusting key electrophysiological parameters such as conduction velocity, resulting in a large number of samples with diverse conditions, and their distribution in the feature space is often quite dispersed. In contrast, real-world ECG samples are limited in number, but their feature distribution within the same lesion type or blood supply area is more concentrated. If the two are simply mixed without distinction, the overall feature distribution will more closely resemble the statistical characteristics of the simulated data, and the key patient physiological details carried in the real data are easily weakened during training, making effective multi-fidelity fusion difficult to achieve.

[0051] Lesion semantics refers to the structured and characterizable clinicopathological attributes used to describe and differentiate different myocardial infarction lesions. Its core dimensions include, but are not limited to: the coronary artery supply area involved, the degree of infarction transmural damage, and the spatial distribution pattern of the infarction. The coronary artery supply area includes, for example, the area supplied by the left anterior descending artery (LAD), right coronary artery (RCA), left circumflex artery (LCX), and their combinations; the degree of infarction transmural damage includes, for example, transmural or non-transmural infarction; and the spatial distribution pattern of the infarction includes, for example, the specific location and three-dimensional morphology of the apex, lateral wall, and inferior wall.

[0052] As an example, based on a unified ventricular electrophysiological field, combined with LGE-CMR and expert annotation, infarct areas and their coronary blood supply areas are marked on the ventricular grid. Low-dimensional lesion vectors are extracted using a lesion embedding network, and the simulation data and real data are clustered in the embedding space according to lesion type, so that high-fidelity and low-fidelity samples of the same blood supply area or the same lesion pattern form a stable cluster structure in the feature space.

[0053] By completing registration, region division, and lesion labeling on a unified ventricular geometric model, samples from different centers, data sources, and simulation conditions are uniformly represented as a set of lesion-level vectors. This allows multi-fidelity fusion to shift from the traditional overall sample or electrophysiological field level to a finer-grained lesion level for information alignment and fusion.

[0054] In one specific implementation, a lesion-guided embedding mechanism is designed based on a core electrophysiological assumption: samples from the same coronary artery supply region (such as LAD, RCA, LCX, and their pairwise combinations) should have similar underlying electrophysiological patterns, regardless of whether they originate from simulation or real data. Let the representation of the i-th sample in the ventricular electrophysiological space be denoted as . This invention constructs a discriminative feature embedding network. , .express Subsequently, this invention treats each sample as a node in the embedding space, constructs affinity relationships based on the similarity between embedding vectors, and performs lesion-guided clustering on this basis. During clustering, using the embedding vector as input, starting from several randomly initialized centroids, through iterative processes of sample assignment and centroid updates, samples with similar embeddings are automatically grouped into the same cluster, ensuring that each cluster primarily contains samples from the same coronary artery supply region or similar lesion patterns. Since real samples are more concentrated under the same lesion type, while simulated samples are relatively dispersed after changes in parameters such as conduction velocity, this lesion-guided clustering strategy can reduce the distance between the simulated sample and its corresponding real sample representation in the embedding space, making the distribution of simulated features more concentrated due to parameter diversification, thereby mitigating the mismatch in feature distribution of multi-fidelity data. The ventricular potential distribution of each sample is then analyzed. The mapping is represented as a low-dimensional embedding vector. The embedding network consists of multiple nonlinear transformations, and its training objective is to minimize the distance between samples of the same coronary artery supply area or the same lesion pattern in the embedding space, while increasing the interval between samples of different supply areas or lesion types, thereby obtaining a discriminative representation space centered on lesion distribution and supply area semantics. Using this embedding representation, this invention treats each sample as a node in the embedding space, constructs affinity relationships based on feature similarity, and implements lesion-guided clustering on this basis: firstly, random initialization is performed in the embedding space. Centroid The process then iterates between the sample allocation and centroid update steps until convergence. In the allocation step, for the first... Each sample, its cluster index Defined as the centroid that is closest to its embedding vector:

[0055] In the formula, The total number of samples, This represents the Euclidean distance. In the update step, the centroid of each cluster is recalculated based on the current assignment results. There are clusters, including:

[0056] In the formula, This is an indicator function; the value is set to true when the condition within the parentheses is true. Otherwise, the value is 0. Through the allocation step, each sample embedding vector... Each centroid is assigned to the nearest centroid; through the update step, each centroid... The update is performed using the mean of the in-cluster sample embeddings. After iterative convergence, we can obtain... A set of samples:

[0057] Each collection Including the same centroid in the embedded space The samples typically correspond to a certain type of lesion pattern or blood supply area. Within this cluster, the potential representation of simulated data is more consistent with that of real data, thereby alleviating the distribution mismatch caused by factors such as changes in conduction velocity.

[0058] In one embodiment of the present invention, the specific process of "determining the physiological detail residuals between the real electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster" in step S120 can be further explained in conjunction with the following description.

[0059] Determine the first statistical center representing the electrophysiological characteristics of the actual electrocardiogram data within the lesion cluster; Determine the second statistical center for the electrophysiological characteristics of the simulated electrocardiogram data within the lesion cluster; The physiological detail residuals are determined based on the first statistical center and the second statistical center.

[0060] It should be noted that after clustering, physiological details from real-world electrocardiograms are fused within each cluster and injected into the corresponding multi-fidelity sample representation. On the other hand, because clustering shifts the feature space from the original comprehensive feature space to a clustered feature space that focuses more on lesion differentiation, some comprehensive information helpful for fine localization of myocardial infarction may be lost. To simultaneously preserve both the comprehensive features extracted from the electrocardiogram and the lesion-specific features, a residual fusion layer is further introduced after embedding and clustering to fuse the ventricular electrophysiological field representation.

[0061] As an example, before fusion, the embedding centers of high-fidelity and low-fidelity lesions are calculated separately within each lesion cluster, and their differences are used to construct the physiological detail residual for that lesion type. The physiological detail residual is a quantified value of the difference in true detail between real ECG data, real clinical data, and simulated ECG data in the electrophysiological feature representation within the same lesion cluster. Specifically, it is reflected as the difference between the center (mean) of the electrophysiological features of high-fidelity samples and the center (mean) of the electrophysiological features of low-fidelity samples within the cluster. This residual can capture the physiological detail information of real patients.

[0062] In a specific implementation, residual fusion can be expressed as:

[0063] In the formula, For fusion networks, used to synthesize original features lesion-specific embedding characteristics Perform joint modeling, This is the fused ventricular electrophysiological representation. While retaining the original comprehensive representation, the residual structure introduces discriminative embedding information obtained through lesion-guided clustering, achieving lesion-level fusion of multifidelity electrocardiograms. This allows for more efficient utilization of the limited but high-value patient physiological details in real-world electrocardiograms, providing a more stable and clinically interpretable input representation for subsequent lesion-based multifidelity fusion and myocardial infarction localization.

[0064] In each cluster Further distinguish between high-fidelity and low-fidelity sample sets, denoted as...

[0065]

[0066] In the formula, For lesion embedding, calculate the center of high-fidelity and low-fidelity embeddings within the cluster separately:

[0067]

[0068] Difference between high-fidelity centers and low-fidelity centers

[0069] Used to characterize the physiological detail residuals of the lesion cluster between real and simulated data.

[0070] In one embodiment of the present invention, the specific process of "correcting the simulated electrocardiogram data in the lesion cluster based on the physiological detail residuals to obtain a fused ventricular electrophysiological feature representation" in step S130 can be further explained in conjunction with the following description.

[0071] Based on the physiological detail residuals, the electrophysiological feature representation of the simulated electrocardiogram data within the lesion cluster is modified by residual enhancement through a learnable mapping network to obtain the fused ventricular electrophysiological feature representation.

[0072] It should be noted that the correction does not target all simulated ECG data, but rather simulated ECG data that has the same lesion semantics as real ECG data within the same lesion cluster. That is, the sample and the high-fidelity sample within the cluster belong to the same coronary artery supply area or the same lesion pattern, ensuring the semantic matching of the residual correction. The correction is not a replacement or rewriting of the low-fidelity data, but rather injects physiological detail residuals as supplementary signals to the electrophysiological features of the low-fidelity data, so that the corrected features possess both the parameter coverage capability of the simulated data and the clinical adaptability of the real data.

[0073] As an example, this can be corrected by linearly weighting the physiological detail residuals with learnable weights and combining them with the electrophysiological feature representations of low-fidelity samples. The result of the linear weighted combination can then be further input into a learnable residual fusion network for nonlinear transformation to obtain the fused lesion representation.

[0074] By constructing a difference vector between high-fidelity and low-fidelity lesion embeddings, this vector is modeled as a physiological detail residual signal and injected into the low-fidelity embedding using a learnable mapping function. This allows the low-fidelity samples to retain the original lesion location, extent, and parameter diversity information while making their lesion-level feature distribution closer to the electrophysiological pattern of real patients. The fused lesion embeddings are then mapped back to a unified ventricular grid to update the electrophysiological field of the corresponding region, providing more reliable input for downstream tasks such as myocardial infarction localization and blood supply area discrimination.

[0075] In a specific implementation, for clusters For any low-fidelity sample, its embedding representation The multifidelity residual fusion form can be written as

[0076] in, For residual fusion networks, For learnable fusion strength coefficients, This is the fused lesion-level embedding representation. Through the above operations, the average physiological characteristics of high-fidelity real data within each lesion cluster are injected as residuals into the corresponding low-fidelity simulation sample representation. This maintains the coverage and diversity of the simulation data while making its lesion-level feature distribution closer to the electrophysiological patterns of real patients. Based on the correspondence between the lesion embedding and the ventricular grid, the fused embedding... It can be further mapped back to the ventricular grid or used to update the ventricular electrophysiological field representation, providing multifidelity fusion input for subsequent myocardial infarction localization and functional assessment.

[0077] In one embodiment of the present invention, the specific process of "constructing a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation" in step S140 can be further described in conjunction with the following description.

[0078] It should be noted that the samples in the multi-fidelity fusion training set are all associated with the semantic information of lesions on the unified ventricular geometric model. The sample labels are consistent with the lesion annotations in the original multi-fidelity ECG dataset. They are used to train ECG analysis-related artificial intelligence models, including but not limited to myocardial infarction localization models, coronary artery blood supply area classification models, ventricular transmembrane potential field reconstruction models, etc.

[0079] As an example, the fused lesions are embedded and mapped back to the ventricular grid, the transmembrane potential field is updated, and then input into the inverse electrocardiogram problem solving network or infarction discrimination network to achieve downstream tasks such as ventricular TMP reconstruction, infarct region localization, and blood supply area classification.

[0080] The following are the experimental data of this invention. 1. Heartbeat Classification Results like Figure 3 As shown, for a real dataset with continuous heartbeats, this study uses a QRS initiation point detection algorithm to periodically divide the heartbeats. By identifying the initiation point of the QRS wave as the periodic start point of the heartbeat and truncating it with a period end of 600ms, the continuous ECG signal was successfully decomposed into independent cardiac cycle segments.

[0081] 2. Purely simulated sample size experiment This experiment aims to investigate the impact of increasing or decreasing the number of low-fidelity simulation samples in the training dataset on the model's final performance in accurately locating myocardial infarction. The experiment was conducted for training and testing on heartbeats with multiple infarct regions (anterior and inferior walls). For the same heartbeat, this study analyzes the results obtained by training with different amounts of simulation samples. Figure 4As shown, the experimental results clearly demonstrate a positive correlation between the model's localization results and the training sample size. With the gradual increase in the sample size for this myocardial infarction location (from 285 to 655), the model's predictions became increasingly closer to the actual myocardial infarction location. However, with a small sample size, the model's learning of the complex nonlinear mapping between surface signals and the myocardial infarction region is incomplete and inaccurate. The predicted myocardial infarction region often has blurred boundaries and is scattered, making it difficult to accurately cover the actual lesion area. This indicates that while low-fidelity data provides broad background information, it lacks sufficient high-precision anchors to calibrate the model's output. Furthermore, even with 655 samples, the model results consistently showed some error compared to the actual myocardial infarction location (the anterior wall prediction always had a certain offset). This offset suggests that a model trained solely on simulation data cannot completely eliminate the gap between the simulation and reality domains. While simulation data can provide a wealth of electrophysiological prior information and broad parameter space coverage, they fail to fully capture the complexity of the real human conduction media, subtle differences in electrode placement, and other physiological details.

[0082] 3. Verification of multiple types of myocardial infarction The model was validated on various types of myocardial infarction to assess its generalization ability. Clinical cases with varying degrees of transmural penetration (transmural and non-transmural) and location were used. Figure 5 As shown, we performed a qualitative analysis on three representative patients and compared our predictions with the actual infarct areas identified on the patients' LGE-CMR images. For each case, we present the long-axis and corresponding short-axis images side-by-side. In all three cases, the present invention showed a high degree of spatial consistency with the LGE-CMR images. In the first case, the predicted infarct area showed spatial consistency with the reference image across multiple sections and correctly reflected the three-dimensional transmural extent. Similar consistency was observed for both large and small infarcts, indicating that the present invention can reliably locate myocardial infarctions of different sizes and types.

[0083] This invention utilizes both simulated ECG (low fidelity) and clinical ECG (high fidelity) on a unified ventricular geometry model. Through lesion-guided embedding, clustering, and residual fusion, it transfers the physiological details of real patients to a large number of simulated samples, thereby improving the accuracy and robustness of downstream tasks such as myocardial infarction localization.

[0084] (1) Compared with ECGI or myocardial infarction localization methods that use only a single data source (simulated ECG or real ECG), the advantage of this invention is that it can systematically integrate the advantages of both types of data. Although methods that rely solely on simulated data can easily generate large-scale samples and facilitate the training of complex models, simulated ECGs differ significantly from real patients in terms of tissue heterogeneity, surface noise, and electrode placement errors, which limits their generalization performance on clinical data. Methods that rely solely on real data are closer to the actual situation in terms of physiological details and noise distribution, but they are limited by the limited number of samples, uneven case distribution, and high collection costs, making it difficult to cover different lesion locations, ranges, and combinations of various physiological parameters. This invention uses lesion-level multi-fidelity residual fusion to selectively transfer key electrophysiological features from real data to the lesion representation of a large number of simulated samples, enabling the model to have both good sample coverage and the ability to depict real physiological details, thereby improving its stability and accuracy in real clinical scenarios. (2) Compared with existing general multifidelity modeling and multifidelity learning methods in the field of computer science, this invention is more suitable for ECG and ventricular electrophysiological scenarios in terms of modeling objects and structural design. Existing multifidelity methods mostly focus on scalar or low-dimensional physical quantities, paying attention to the deviation correction between high-fidelity and low-fidelity values. They lack targeted support for high-dimensional spatiotemporal fields with anatomical structural constraints, such as surface ECG-transmembrane potential. Their multifidelity fusion usually focuses on the overall sample level and rarely combines the specific lesion location and the semantics of the coronary artery supply area. This invention combines clinically commonly used region divisions on a unified ventricular geometric model, refines electrophysiological information to the lesion level, and completes multifidelity alignment and residual fusion in the lesion embedding space, making the fusion result more consistent with the spatial distribution characteristics of myocardial infarction lesions and the needs of clinical interpretation.

[0085] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0086] Reference Figure 6 This application illustrates an embodiment of a multi-sensor ECG fusion device based on lesion semantic space, used to construct a training set that fuses simulated and real ECG data. Specifically, it includes the following modules: Specifically, it includes: The clustering and grouping module 610 is used to obtain the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively, and to cluster and group the electrophysiological feature representations according to the semantic similarity of the lesions to obtain lesion clusters; The residual calculation module 620 is used to determine the physiological detail residuals between the real electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster. The correction and fusion module 630 is used to correct the simulated electrocardiogram data in the lesion cluster based on the physiological detail residuals to obtain a fused ventricular electrophysiological feature representation. The training set construction module 640 is used to construct a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

[0087] In one embodiment of the present invention, the clustering and grouping module 610 includes: The initial feature generation submodule is used to map the simulated ECG data and the real ECG data to the grid node space of the preset ventricular geometry model to obtain the corresponding node-level initial features; The temporal feature generation submodule is used to perform graph convolution operation on the node-level initial features based on the ventricular topology of the ventricular geometric model to obtain the corresponding node-level temporal features. The temporal modeling submodule is used to perform temporal modeling on the node-level temporal features and generate the corresponding electrophysiological feature representations.

[0088] In one embodiment of the present invention, the clustering and grouping module 610 includes: The embedding vector generation submodule is used to input the electrophysiological feature representation into a preset lesion embedding network to generate a corresponding low-dimensional lesion semantic embedding vector. The grouping submodule is used to perform clustering based on the distance between the low-dimensional lesion semantic embedding vectors to obtain lesion clusters.

[0089] In one embodiment of the present invention, the grouping submodule includes: An initial centroid unit is used to determine the initial centroid vector in the low-dimensional lesion semantic embedding vector; The distance calculation unit is used to determine the Euclidean distance between the low-dimensional lesion semantic embedding vector and the initial centroid vector, and to assign the low-dimensional lesion semantic embedding vector to the cluster corresponding to the nearest centroid; The vector update unit is used to determine the mean of all low-dimensional lesion semantic embedding vectors within the corresponding cluster, and update the mean vector to the new centroid vector of the corresponding cluster. The iteration completion unit is used to define the set of all original electrocardiogram data samples of the corresponding cluster as a lesion cluster after the iteration is completed.

[0090] In one embodiment of the present invention, the residual calculation module 620 includes: The first calculation submodule is used to determine the first statistical center of the electrophysiological feature representation of the real electrocardiogram data within the lesion cluster; The second calculation submodule is used to determine the second statistical center of the electrophysiological characteristics representation of the simulated electrocardiogram data within the lesion cluster; The center subtraction submodule is used to determine the physiological detail residuals based on the first statistical center and the second statistical center.

[0091] In one embodiment of the present invention, the central subtraction submodule includes:

[0092] In the formula, For the physiological detail residuals; This refers to the first statistical center; It is the second statistical center; The calculation formula for the first statistical center is as follows:

[0093] In the formula, A collection of high-fidelity samples within a lesion cluster; For lesion embedding; The calculation formula for the second statistical center is as follows:

[0094] In the formula, This is a collection of low-fidelity samples within a lesion cluster.

[0095] In one embodiment of the present invention, the correction fusion module 630 includes: The residual enhancement correction submodule is used to perform residual enhancement correction on the electrophysiological feature representation of the simulated electrocardiogram data within the lesion cluster through a learnable mapping network based on the physiological detail residuals, so as to obtain the fused ventricular electrophysiological feature representation.

[0096] Reference Figure 7 The illustration shows a computer electronic device for implementing a multi-security real-time fusion method based on lesion semantic space according to the present invention, which may specifically include the following: The aforementioned computer electronic device 1 is manifested in the form of a general-purpose computing device. The components of the computer electronic device 1 may include, but are not limited to: one or more processors or processing units 3, memory 8, and a bus 4 connecting different system components (including memory 8 and processing unit 3).

[0097] Bus 4 represents one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Audio / Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0098] Computer electronic device 1 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer electronic device 1, including volatile and non-volatile media, removable and non-removable media.

[0099] Memory 8 may include computer system readable media in the form of volatile memory, such as random access memory 9 and / or cache memory 10. Computer electronic device 1 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 11 may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). Although Figure 7 As not shown, a disk drive for reading and writing to a removable non-volatile disk (such as a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (such as a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 4 via one or more data media interfaces. The memory may include at least one program product having a set (e.g., at least one) of program modules 13 configured to perform the functions of the embodiments of this application.

[0100] A program / utility 12 having a set (at least one) of program modules 13 may be stored, for example, in memory. Such program modules 13 include—but are not limited to—an operating system, one or more application programs, other program modules 13, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 13 typically perform the functions and / or methods described in the embodiments of this application.

[0101] The computer electronic device 1 can also communicate with one or more external devices 2 (e.g., keyboard, pointing device, display 7, camera, etc.), and with one or more devices that enable an operator to interact with the computer electronic device 1, and / or with any device that enables the computer electronic device 1 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through the I / O interface 6. Furthermore, the computer electronic device 1 can also communicate with one or more networks (e.g., local area network (LAN)), wide area network (WAN), and / or public networks (e.g., the Internet) through the network adapter 5. Figure 7 As shown, network adapter 5 communicates with other modules of computer electronic device 1 via bus 4. It should be understood that, although... Figure 7 Not shown, it may be combined with other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing unit 3, external disk drive array, RAID system, tape drive and data backup storage system 11, etc.

[0102] The processing unit 3 executes various functional applications and data processing by running programs stored in memory 8, such as implementing a multi-security electronic fusion method based on lesion semantic space provided in the embodiments of this application.

[0103] That is, when the above-mentioned processing unit 3 executes the above-mentioned program, it performs the following: acquiring the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively, and clustering the electrophysiological feature representations according to the semantic similarity of the lesions to obtain lesion clusters; determining the physiological detail residuals between the real ECG data and the simulated ECG data in the lesion cluster; correcting the simulated ECG data in the lesion cluster according to the physiological detail residuals to obtain the fused ventricular electrophysiological feature representation; and constructing a multi-fidelity fusion training set for training the artificial intelligence model based on the electrophysiological feature representation of the real ECG data and the fused ventricular electrophysiological feature representation.

[0104] In the embodiments of this application, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a multi-security true-to-false electrical fusion method based on lesion semantic space as provided in all embodiments of this application.

[0105] That is, when the program is executed by the processor, it performs the following: acquiring the electrophysiological feature representations corresponding to simulated ECG data and real ECG data respectively, and clustering the electrophysiological feature representations according to the semantic similarity of lesions to obtain lesion clusters; determining the physiological detail residuals between the real ECG data and the simulated ECG data within the lesion clusters; correcting the simulated ECG data within the lesion clusters according to the physiological detail residuals to obtain a fused ventricular electrophysiological feature representation; and constructing a multi-fidelity fusion training set for training the artificial intelligence model based on the electrophysiological feature representations of the real ECG data and the fused ventricular electrophysiological feature representations.

[0106] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0107] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0108] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the operator's computer, partially on the operator's computer, as a standalone software package, partially on the operator's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the operator's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

[0109] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0110] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0111] The above provides a detailed description of the multi-faceted electronic fusion method and apparatus based on lesion semantic space provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-sensor ECG fusion method based on lesion semantic space, used to construct a training set that fuses simulated and real ECG data, characterized in that, Including the following steps: Electrophysiological feature representations corresponding to simulated ECG data and real ECG data are obtained respectively, and the electrophysiological feature representations are clustered and grouped according to the semantic similarity of lesions to obtain lesion clusters; Determine the physiological detail residuals between the actual electrocardiogram data within the lesion cluster and the simulated electrocardiogram data within the lesion cluster; Based on the physiological detail residuals, the simulated electrocardiogram data within the lesion clusters are corrected to obtain a fused ventricular electrophysiological feature representation; A multi-fidelity fusion training set for training an artificial intelligence model is constructed based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

2. The method according to claim 1, characterized in that, The steps for obtaining the electrophysiological feature representations corresponding to simulated ECG data and real ECG data respectively include: The simulated ECG data and the real ECG data are mapped to the grid node space of the preset ventricular geometric model to obtain the corresponding node-level initial features; Based on the ventricular topology of the ventricular geometric model, graph convolution operation is performed on the node-level initial features to obtain the corresponding node-level temporal features; The node-level temporal features are subjected to temporal modeling to generate the corresponding electrophysiological feature representations.

3. The method according to claim 1, characterized in that, The step of clustering the electrophysiological features based on semantic similarity of lesions to obtain lesion clusters includes: The electrophysiological features are input into a preset lesion embedding network to generate a corresponding low-dimensional lesion semantic embedding vector. Clustering is performed based on the distance between the low-dimensional lesion semantic embedding vectors to obtain lesion clusters.

4. The method according to claim 3, characterized in that, The step of clustering and grouping lesion clusters based on the distance between the low-dimensional lesion semantic embedding vectors includes: Determine the initial centroid vector in the low-dimensional lesion semantic embedding vector; Determine the Euclidean distance between the low-dimensional lesion semantic embedding vector and the initial centroid vector, and assign the low-dimensional lesion semantic embedding vector to the cluster corresponding to the nearest centroid; Determine the mean of all low-dimensional lesion semantic embedding vectors within the corresponding cluster, and update the mean vector to the new centroid vector of the corresponding cluster; After the iteration is completed, the set of all original ECG data samples of the corresponding cluster is defined as a lesion cluster.

5. The method according to claim 1, characterized in that, The steps for determining the physiological detail residuals between the actual ECG data within the lesion cluster and the simulated ECG data within the lesion cluster include: Determine the first statistical center representing the electrophysiological characteristics of the actual electrocardiogram data within the lesion cluster; Determine the second statistical center for the electrophysiological characteristics of the simulated electrocardiogram data within the lesion cluster; The physiological detail residuals are determined based on the first statistical center and the second statistical center.

6. The method according to claim 5, characterized in that, The formula for calculating the physiological detail residuals is as follows: In the formula, For the physiological detail residuals; This refers to the first statistical center; It is the second statistical center; The calculation formula for the first statistical center is as follows: In the formula, A collection of high-fidelity samples within a lesion cluster; For lesion embedding; The calculation formula for the second statistical center is as follows: In the formula, This is a collection of low-fidelity samples within a lesion cluster.

7. The method according to claim 1, characterized in that, The step of correcting the simulated electrocardiogram data within the lesion cluster based on the physiological detail residuals to obtain a fused representation of ventricular electrophysiological characteristics includes: Based on the physiological detail residuals, the electrophysiological feature representation of the simulated electrocardiogram data within the lesion cluster is modified by residual enhancement through a learnable mapping network to obtain the fused ventricular electrophysiological feature representation.

8. A multi-sensor electrocardiogram (ECG) fusion device based on lesion semantic space, used to construct a training set that fuses simulated and real ECG data, characterized in that, include: The clustering and grouping module is used to obtain the electrophysiological feature representations corresponding to the simulated ECG data and the real ECG data respectively, and to cluster and group the electrophysiological feature representations according to the semantic similarity of the lesions to obtain lesion clusters; The residual calculation module is used to determine the physiological detail residuals between the actual ECG data within the lesion cluster and the simulated ECG data within the lesion cluster. The correction and fusion module is used to correct the simulated electrocardiogram data within the lesion cluster based on the physiological detail residuals, so as to obtain a fused ventricular electrophysiological feature representation; The training set construction module is used to construct a multi-fidelity fusion training set for training an artificial intelligence model based on the electrophysiological feature representation of the real electrocardiogram data and the fused ventricular electrophysiological feature representation.

9. A computer electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the steps of the multi-preservation true-to-false electrical fusion method based on the lesion semantic space as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the multi-security true-to-false electrical fusion method based on the lesion semantic space as described in any one of claims 1 to 7.

Images