Method, device and system for reconstructing twelve-lead signal

Abstract

The invention discloses a 12-lead signal reconstruction method, device and system, and belongs to the technical field of biomedical signal detection. According to the method, a single-lead electrocardiosignal is input into a trained twelve-lead signal reconstruction model, and then a clinically available twelve-lead electrocardiogram can be synthesized. Specifically, the 12-lead signal reconstruction model comprises a pre-coding module, a coding module, a migration module and a reconstruction module which are connected in sequence; global features are extracted through collaborative coding, target lead features are mapped through a domain self-adaption module, and finally 12-lead signals conforming to the electrophysiological law are generated through denoising in combination with a conditional diffusion model. Compared with a traditional method, the method has the advantages that the signal fidelity, the heart beat retentivity and the diagnosis consistency are remarkably improved on multiple data sets, the generalization ability in an open scene is considered, and the application value of wearable equipment in remote monitoring and early screening of cardiovascular diseases is effectively enhanced.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal detection technology, and more specifically, relates to a method, apparatus and system for reconstructing a twelve-lead signal. Background Technology

[0002] Cardiovascular disease is a leading cause of death worldwide, projected to cause more than 20 million deaths annually by 2030. Therefore, early detection and intervention of cardiovascular disease are of significant public health importance. A 12-lead electrocardiogram (ECG) is the gold standard for clinical assessment of cardiac electrical activity and diagnosis of complex arrhythmias, myocardial ischemia, and other diseases. It simultaneously measures ECG signals from multiple locations on the limbs and chest, providing a multi-dimensional depiction of the heart's electrical activity.

[0003] However, traditional 12-lead electrocardiogram (ECG) acquisition relies on specialized equipment and operational experience, making long-term continuous monitoring difficult. While wearable devices are portable, they typically only acquire single-lead signals, limiting their diagnostic capabilities in complex cardiac pathologies. To address this contradiction, researchers have attempted to reconstruct 12-lead signals from single leads using linear transformations or deep learning, but the reconstruction accuracy has been poor in practical applications.

[0004] Therefore, there is an urgent need for a method that can fully utilize the implicit correlation between the twelve leads under single-lead input conditions to meet the needs of remote continuous monitoring and early screening. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a method, apparatus and system for reconstructing a twelve-lead signal, the purpose of which is to solve the technical problem that existing wearable devices cannot reconstruct a twelve-lead signal from a single lead with high precision.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for reconstructing a twelve-lead signal is provided, comprising: The single-lead electrocardiogram signal of the subject being tested Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. ; The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them; The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve diffusion-based joint conditions, eleven-lead reconstruction results were obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

[0007] Furthermore, the precoding module includes: a main UNet network. With the UNet network When the input is a real 12-lead ECG signal When using a single training sample, The main UNet network Using a single-lead ECG signal from a single training sample as input, and outputting the first type of twelve-lead ECG signal... The training objective is to minimize the difference between the training sample and the single training sample. The sub-UNet network The input is any ECG signal from any lead other than the single-lead ECG signal in the single training, and the output is a second type of twelve-lead ECG signal. The training objective is to minimize the difference between the training sample and the single training sample.

[0008] Furthermore, during training, the joint loss of the precoding modules for: ;in, Loss due to the reconstruction of the main UNet network. , For the loss of UNet network reconstruction, , These are the weight coefficients for the feature alignment loss. For feature alignment loss, , , These are the main UNet network and the intermediate features of the main UNet network, respectively. This represents the Euclidean norm.

[0009] Furthermore, the encoding module includes an encoder and a decoder; during training, when the input is a real twelve-lead electrocardiogram signal... When training on a single training sample, the training objective of the encoding module is: the encoder encodes the masked single training sample to obtain global features. The decoder for the global features The difference between the third type of twelve-lead ECG signal obtained by decoding and the single training sample is minimized.

[0010] Furthermore, the total loss of the encoding module is: ; in, for A single heartbeat segment, This refers to a single reconstructed segment from a third type of twelve-lead ECG signal, where n represents the segment. and Their respective totals For the i-th mask, For Hadamard's element-wise product, Denotes the Euclidean norm. For frequency domain loss weights, For spectral alignment loss, , This is the amplitude spectrum obtained from the short-time Fourier transform.

[0011] Furthermore, the migration module includes a domain migration network. and rhythm classifier After the encoding module is trained, during the training of the transfer module, the actual 12-lead ECG signal will be used. The single training sample is input into the encoder in the encoding module. Make it output global features , Then utilize the domain migration network Map it to the target lead domain to obtain And using a rhythm classifier right Perform category discrimination.

[0012] Furthermore, the total loss of the migration module is: ;in, For domain alignment loss, , For the target lead domain reference features, , For rhythm classification loss, The weights are used for classification loss.

[0013] According to another aspect of the present invention, a twelve-lead signal reconstruction device is provided for reconstructing single-lead electrocardiogram signals of a subject to be tested. Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them; The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve diffusion-based joint conditions, eleven-lead reconstruction results were obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

[0014] According to another aspect of the present invention, a system for reconstructing a 12-lead signal is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for reconstructing the 12-lead signal.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for reconstructing the twelve-lead signal.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention relies solely on single-lead ECG signals easily acquired by wearable devices. By inputting these signals into a pre-trained twelve-lead signal reconstruction model, a clinically usable twelve-lead ECG can be synthesized, providing a highly feasible solution for remote cardiovascular disease screening in home and community settings. Specifically, the twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module. Global features are extracted through co-coding, target lead features are mapped using a domain adaptation module, and finally, a conditional diffusion model is used to denoise and generate a twelve-lead signal that conforms to electrophysiological laws. Compared with traditional methods, this invention significantly improves signal fidelity, heart rate maintenance, and diagnostic consistency across multiple datasets, while also considering generalization capabilities in open scenarios, effectively enhancing the application value of wearable devices in remote monitoring and early screening of cardiovascular diseases.

[0017] (2) In a preferred embodiment, the precoding module adopts a dual UNet structure (main UNet and secondary UNet). The main UNet takes a single-lead signal as input, and the secondary UNet takes other randomly selected leads as input. During training, by minimizing the reconstruction loss of both and the alignment loss of intermediate features at the encoding end, the model is forced to learn shared global latent features that are independent of specific leads, thereby improving the ability to represent the spatial structure of multiple leads.

[0018] (3) In a preferred embodiment, the encoding module is trained by a heartbeat-level masking autoencoder and a spectrum alignment task. Specifically, the module extracts a joint feature representation that is both temporally and spatially consistent by calculating the reconstruction loss in the masking time domain and the spectrum alignment loss in the frequency domain.

[0019] (4) In a preferred embodiment, the transfer module utilizes a Transformer encoder and a domain transfer network to map the global features derived from a single lead to the target lead domain. The training objectives include minimizing the domain alignment loss, i.e., the mapped features being close to the true target lead features, and the rhythm classification loss, thereby enhancing the model's generalization ability to unknown individuals and rare pathological patterns in open scenarios.

[0020] (5) In a preferred embodiment, the reconstruction module employs a conditional denoising diffusion probability model. Noise is gradually added to the real signal during the forward process; during the backward reconstruction process, single-lead signal and domain adaptive features are introduced as dual conditions. To balance signal fidelity and the influence of domain priors, the module adopts a dual-path noise fusion strategy, which combines the noise prediction results of "single-lead only condition" and "single-lead + domain feature condition" by weighting with adjustable fusion coefficients. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a twelve-lead signal reconstruction model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of another twelve-lead signal reconstruction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-lead electrocardiogram signal reconstruction results for atrial fibrillation (AF) patients provided in an embodiment of the present invention; Figure 4 Comparison of the stepwise reconstruction results of twelve-lead ECG signals at different stages provided in the embodiments of the present invention; wherein, (a) is the reconstruction result using only the global features obtained in the precoding stage, (b) is the reconstruction result obtained after introducing domain adaptive enhancement features, and (c) is the final reconstruction result obtained under the full domain enhancement diffusion model; Figure 5 This is a comparison diagram of the reconstruction results of a long-term 12-lead electrocardiogram signal and the actual signal provided in an embodiment of the present invention; Figure 6 Distribution of (a) MSE and (b) Pearson correlation coefficients of signals reconstructed by different methods over a longer span. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 This embodiment provides a method for reconstructing a twelve-lead signal, including: a single-lead electrocardiogram signal of the subject to be tested. Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them.

[0024] The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve diffusion-based joint conditions, eleven-lead reconstruction results were obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

[0025] like Figure 1 and Figure 2 As shown, this model aims to address the problem of missing information in single-lead observations. Instead of directly employing end-to-end mapping, the model decouples representation learning into two stages: first, global features are extracted through bidirectional collaborative precoding; then, a domain adaptation module maps these features to the target lead domain; and finally, a conditional diffusion model is used to generate a high-fidelity signal. This design ensures both the electrophysiological consistency of the signal and enhances its robustness to pathological patterns.

[0026] Specifically, the twelve-lead electrocardiogram reconstruction model provided by the present invention includes: an input and preprocessing module, a precoding module, an encoding module, a migration module, a reconstruction module, and an output module. Figure 2 The detailed structure of the LSDA-Diff model proposed in this invention is given, including a dual UNet precoding network for initial mapping and latent feature alignment, a Transformer-based encoding module, a transfer module for lead domain adaptation and rhythm information injection, and a conditional diffusion reconstruction module that incorporates domain priors and cross-lead related information during the diffusion process.

[0027] The input and preprocessing module is used to acquire the single-lead ECG signal of the subject being tested and perform preprocessing operations such as filtering, normalization, and heartbeat segmentation on the raw signal. Preferably, the single-lead ECG signal is acquired in real time by a wearable ECG device, stored in the form of a time-amplitude binary, and after removing baseline drift and high-frequency noise through bandpass filtering, the amplitude is normalized to a specific range to reduce amplitude shift caused by individual differences.

[0028] The precoding module is used to jointly model the input lead and other randomly selected leads using a bidirectional collaborative coding structure, align the latent features of different leads to a shared latent space, and extract global ECG features that are independent of specific leads.

[0029] The encoding module is used to perform heart rate-level masking autoencoding and spectral alignment modeling on the twelve-lead ECG signal output by the precoding module, and learn a joint feature representation that combines time-domain waveform and frequency-domain energy distribution.

[0030] The transfer module maps global features derived from input leads to the target lead domain, fuses inter-lead correlations and rhythm diagnostic information to obtain domain-adaptive features, and alleviates the problem of lead-specific domain shift.

[0031] The reconstruction module is used to perform multi-step denoising and reconstruction of the noisy twelve-lead ECG signal under the combined constraints of single-lead ECG signal and domain adaptive features, generating a twelve-lead ECG signal that conforms to physiological constraints.

[0032] The output module is used to output the reconstructed twelve-lead ECG signal and can further calculate indicators such as mean square error, Pearson correlation coefficient, R-R interval error and diagnostic consistency with the real signal to evaluate the reconstruction quality.

[0033] In one optional implementation, the optimization objective of the present invention for the twelve-lead electrocardiogram reconstruction task is defined as follows: Let the single-lead input be... The actual twelve-lead signal is ,in Indicates the number of signal segments. Let represent the sampling length of each segment. The model optimization objective is defined as minimizing the mean squared error of the joint channel-time dimension, as shown in the following formula:

[0034] in, For lead index, These are the time sampling points.

[0035] Furthermore, the precoding module includes: a main UNet network. With the UNet network When the input is a real 12-lead ECG signal When training a single sample, the main UNet network and the aforementioned UNet network The training process is as follows.

[0036] The main UNet network Using a single-lead ECG signal from a single training sample as input, and outputting the first type of twelve-lead ECG signal... The training objective is to minimize the difference between the training sample and the single training sample.

[0037] The sub-UNet network The input is any ECG signal from any lead other than the single-lead ECG signal in the single training, and the output is a second type of twelve-lead ECG signal. The training objective is to minimize the difference between the training sample and the single training sample.

[0038] Main UNet network The input is a single-lead ECG signal. The output is a preliminary reconstruction of the twelve-lead signal. The reconstruction losses are:

[0039] Sub-UNet network The input is a signal from a lead randomly selected from other leads. The output is The reconstruction losses are:

[0040] in, and The network structures are preferably identical but do not share parameters, which are used to learn alignable latent features from different input leads.

[0041] In the precoding module, the intermediate features of the main UNet and the secondary UNet at the encoding end are denoted as follows: and To align two paths in a shared latent space, this invention introduces a feature alignment loss:

[0042] The total loss of co-training in the precoding stage is:

[0043] in, Let be the weight hyperparameters of the feature alignment loss. Minimize This can improve the consistency of latent features learned from different lead inputs, enabling them to better characterize the global structural features of ECG signals.

[0044] Furthermore, the encoding module includes an encoder and a decoder; during training, when the input is a real twelve-lead electrocardiogram signal... When training on a single training sample, the training objective of the encoding module is: the encoder encodes the masked single training sample to obtain global features. The decoder for the global features The difference between the third type of twelve-lead ECG signal obtained by decoding and the single training sample is minimized.

[0045] In one alternative implementation, the encoding module employs a Transformer-based masked autoencoder structure. The twelve-lead signal obtained via the precoding module is divided into multiple cardiac segment segments. Construct masking markers for each fragment ,in Indicates the masked segment. This represents the preserved segment. Using an encoder-decoder structure, the masked segment is reconstructed within the context provided by the unmasked segment. The temporal reconstruction loss is defined as:

[0046] in, To reconstruct the fragment, Represents element-wise product. The loss weights are used for spectrum alignment.

[0047] To ensure that the reconstructed signal is consistent with the original signal in terms of frequency components, this invention performs a short-time Fourier transform on each segment to obtain the amplitude spectrum. and Define the spectral alignment loss as:

[0048] By minimizing simultaneously and The coding module is able to learn a joint representation that is consistent in both the time and frequency domains.

[0049] Furthermore, the migration module includes a domain migration network. and rhythm classifier After the encoding module is trained, during the training of the transfer module, the actual 12-lead ECG signal will be used. The single training sample is input into the encoder in the encoding module. Make it output global features , Then utilize the domain migration network Map it to the target lead domain to obtain And using a rhythm classifier right Perform category discrimination.

[0050] Furthermore, the preliminary reconstruction results of the twelve leads obtained from the main UNet will be used... Feed into Transformer encoder The global features are obtained as follows:

[0051] Real twelve-lead electrocardiogram signal The data is fed into the same encoder to obtain the target lead domain reference features:

[0052] Domain migration network Used to transfer global features Mapped to the target lead domain, the output domain adaptive features Rhythm classifier right It can be used to classify rhythms into normal and abnormal categories.

[0053] Domain alignment loss is defined as:

[0054] The preferred loss for rhythm classification is binary cross-entropy loss, defined as:

[0055] The total loss of the migration module is:

[0056] in, The weights are used for classification loss. This is achieved by simultaneously minimizing... and This invention can align the distribution of input leads and target leads in the feature space and incorporate prior information related to arrhythmia diagnosis, thereby enhancing the generalization ability to unknown individuals and rare pathological patterns in open scenarios.

[0057] In one alternative implementation, the reconstruction module employs a conditional denoising diffusion model. Let the actual twelve-lead ECG signal be... During the forward noise addition process, a preset noise scheduling sequence is used. Gaussian noise is gradually injected into the signal, defined as:

[0058]

[0059] The conditional probability distribution of the forward process is:

[0060] This can be written in explicit sampling form:

[0061] in, .

[0062] In the reverse denoising process, the conditions are extended to a single-lead input. Adaptive features of the domain Construction conditions for the reverse process distribution:

[0063] To fully utilize lead-domain prior information while ensuring signal fidelity, this invention designs a dual-path fusion strategy in the noise prediction network: one path uses only single-lead conditions. Noise prediction is obtained. Another approach uses a joint conditional approach combining single-lead and domain-adaptive features. Noise prediction is obtained. The reasoning phase adopts the following fusion form:

[0064] in, , which is the fusion coefficient used to adjust the strength of the influence of domain priors on the generation process.

[0065] By adjusting It can strike a balance between the overall fidelity of the reconstructed signal and the sensitivity to rare pathological patterns.

[0066] In one optional implementation, in the above-mentioned input and preprocessing module, the length of the heartbeat segment obtained by uniformly dividing the continuous electrocardiogram signal according to the heartbeat is preferably 3 to 10 heartbeats, so as to reduce the computational burden of the model while ensuring that a single input contains the complete heartbeat structure.

[0067] Example 2 This embodiment provides a twelve-lead signal reconstruction device for reconstructing single-lead electrocardiogram signals of the tested subject. Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them; The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve diffusion-based joint conditions, eleven-lead reconstruction results were obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

[0068] For example, using lead II as the input lead, a specific twelve-lead ECG reconstruction model uses the single-lead ECG signal of lead II as the detection signal, and completes the conversion from single lead to twelve lead through bidirectional collaborative precoding, feature coding, domain adaptive feature transfer, and conditional diffusion reconstruction. Figure 3 The reconstruction results of multi-lead ECG signal segments, including leads II, aVF, and V1, are presented using the method of this invention with only a single-lead ECG signal as input. This demonstrates the ability of this invention to preserve pathological features such as the disappearance of P waves and the appearance of f waves in abnormal cardiac rhythm scenarios and the reliability of reconstruction. Figure 4 By comparing the three, we can illustrate the gradual improvement in reconstruction quality at each stage of this invention. Figure 5 The ECG signals from top to bottom are those of leads I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6.

[0069] Specifically, the following processes are involved: Signal Acquisition and Preprocessing: The ECG signal from lead II of the user's wearable device is acquired and stored in time-amplitude format. The raw signal is bandpass filtered and normalized, and the R-peak is located using the QRS detection algorithm. After removing obvious artifacts and extremely short invalid segments, the remaining valid segments longer than several seconds are divided into fixed-length ECG segments required for training and inference according to a fixed number of heartbeats.

[0070] Precoding and Feature Learning: Precoding each single-lead segment Corresponding to the actual twelve-lead segment The input precoding module learns global ECG features through collaborative training of the main UNet and the secondary UNet, and uses a loss function... Constrain the coding alignment under different lead perspectives. Subsequently, through Transformer-based masking autoencoders and spectral alignment tasks, further improve the consistency of latent features in the time and frequency dimensions, providing high-quality priors for subsequent domain adaptive migration and diffusion reconstruction.

[0071] Domain Adaptive Transfer and Conditional Diffusion Reconstruction: Utilizing a transfer module, global features corresponding to the input leads are reconstructed. Mapped to target lead domain features and through loss The discriminative power of features is enhanced by incorporating rhythmic diagnostic information. Finally, in the diffusion reconstruction phase, single-lead input is used... Domain Adaptive Features To obtain the reconstructed twelve-lead configuration, the data was obtained by stepwise sampling from Gaussian noise, based on the combined conditions. And through multi-step denoising iterations, It closely resembles a real twelve-lead ECG signal in terms of waveform morphology, heart rate index, and diagnostic consistency.

[0072] In summary, this embodiment expands the easily accessible single-lead ECG signal from wearable devices into a clinically usable twelve-lead ECG through two-stage feature representation learning and domain-enhanced diffusion reconstruction, significantly improving the accuracy and practicality of cardiovascular disease screening and monitoring based on single-lead ECG signals. Figure 6 The distribution of (a) MSE and (b) Pearson correlation coefficients of signals reconstructed by different methods over a longer span.

[0073] Example 3 This embodiment provides a system for reconstructing a 12-lead signal, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the 12-lead signal reconstruction method.

[0074] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the twelve-lead signal reconstruction method.

[0075] Example 5 This invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this invention.

[0076] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.

[0077] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for reconstructing a twelve-lead signal, characterized in that, include: The single-lead electrocardiogram signal of the subject being tested Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. ; The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them; The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve the diffusion-linked conditions, the eleven-lead reconstruction result was obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

2. The method for reconstructing a twelve-lead signal as described in claim 1, characterized in that, The precoding module includes: a main UNet network. With the UNet network When the input is a real 12-lead ECG signal When using a single training sample, The main UNet network Using a single-lead ECG signal from a single training sample as input, and outputting the first type of twelve-lead ECG signal... The training objective is to minimize the difference between the training sample and the single training sample. The sub-UNet network The input is any ECG signal from any lead other than a single-lead ECG signal in the single training sample, and the output is a second type of twelve-lead ECG signal. The training objective is to minimize the difference between the training sample and the single training sample.

3. The method for reconstructing a twelve-lead signal as described in claim 2, characterized in that, During training, the joint loss of the precoding modules for: ; in, Loss due to the reconstruction of the main UNet network. , For the loss of UNet network reconstruction, , These are the weight coefficients for the feature alignment loss. For feature alignment loss, , , These are the main UNet network and the intermediate features of the main UNet network, respectively. This represents the Euclidean norm.

4. The method for reconstructing a twelve-lead signal as described in claim 1, characterized in that, The encoding module includes an encoder and a decoder; during training, when the input is a real twelve-lead electrocardiogram signal... When training on a single training sample, the training objective of the encoding module is: the encoder encodes the masked single training sample to obtain global features. The decoder for the global features The difference between the third type of twelve-lead ECG signal obtained by decoding and the single training sample is minimized.

5. The method for reconstructing a twelve-lead signal as described in claim 4, characterized in that, The total loss of the encoding module is: ； in, for A single heartbeat segment, This refers to a single reconstructed segment from a third type of twelve-lead ECG signal, where n represents the segment. and Their respective totals For the i-th mask, For Hadamard's element-wise product, Describes the Euclidean norm. For frequency domain loss weights, For spectral alignment loss, , This is the amplitude spectrum obtained from the short-time Fourier transform.

6. The method for reconstructing a twelve-lead signal as described in claim 4, characterized in that, The migration module includes a domain migration network. and rhythm classifier ; After the encoding module is trained, the transfer module training process will use real 12-lead ECG signals. The single training sample is input into the encoder in the encoding module. Make it output global features , Then utilize the domain migration network Map it to the target lead domain to obtain And using a rhythm classifier right Perform category discrimination.

7. The method for reconstructing a twelve-lead signal as described in claim 6, characterized in that, The total loss of the migration module is: ; in, For domain alignment loss, , For the target lead domain reference features, , For rhythm classification loss, The weights are used for classification loss.

8. A device for reconstructing a twelve-lead signal, characterized in that, Single-lead electrocardiogram signal used for the subject being tested Input the trained 12-lead signal reconstruction model to obtain the 12-lead ECG target signal. The twelve-lead signal reconstruction model comprises, in sequence, a precoding module, an encoding module, a transfer module, and a reconstruction module; during training, the twelve-lead signal reconstruction model minimizes the twelve-lead ECG target signal output during training. Compared with the real twelve-lead ECG signal input during training The target is the error between them; The precoding module is used to process the single-lead electrocardiogram signal. Preliminary reconstruction yielded a twelve-lead electrocardiogram signal. The encoding module is used to process the twelve-lead electrocardiogram signal. Spatiotemporal joint coding is performed to obtain global latent features. The migration module is used to transfer the global latent features. Mapping to the target lead domain yields domain-adaptive features. The reconstruction module is used to convert the single-lead electrocardiogram signal into a single-lead ECG signal. and the domain adaptive features To achieve diffusion-based joint conditions, eleven-lead reconstruction results were obtained by stepwise sampling from Gaussian noise, and the single-lead ECG signal was then used. The 12-lead ECG target signal is obtained by fusing the reconstruction results of the 11-lead ECG with the 11-lead ECG signal. .

9. A system for reconstructing a twelve-lead signal, 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 method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images