Electrocardiogram reconstruction model generating device, electrocardiogram reconstruction model generating method, and electrocardiogram reconstruction method

Abstract

The present invention relates to an electrocardiogram reconstruction model generating device that generates an electrocardiogram reconstruction model based on a generative adversarial network (GAN), including an input unit that receives a single-lead electrocardiogram as input, a generator that reconstructs a 12-lead electrocardiogram based on the single-lead electrocardiogram, and a classifier that uses as an input either a first input value that combines the 12-lead electrocardiogram and the single-lead electrocardiogram or a second input value that combines an actual 12-lead electrocardiogram (ground-truth) and the single-lead electrocardiogram, and determines whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram, and is capable of reconstructing a multi-lead electrocardiogram that more accurately reflects the characteristics of an actual electrocardiogram from a single-lead electrocardiogram.

Description

[Technical Field]

[0001] The disclosed invention relates to an apparatus for generating an electrocardiogram reconstruction model, a method for generating an electrocardiogram reconstruction model, and a method for reconstructing an electrocardiogram based on a generative adversarial neural network, which can reconstruct a multi-lead electrocardiogram from a single-lead electrocardiogram.

[0002] In addition, this invention is the result of research supported by the Korea Research Foundation (Project No. 2022RIS-005, Local Government-University Cooperation-Based Regional Innovation Project, Project ID: 1345370814, Contribution Rate: 50%), funded by the government (Ministry of Education), and the Korea Institute for Information and Communications Technology Planning and Evaluation (No. RS-2022-00155966, Artificial Intelligence Integration Innovation Talent Development (Ewha Womans University), National Project ID: 1711179344, Contribution Rate: 30%), (IITP-2024-RS-2023-00260267, Regional Intelligence Innovation Talent Development (Kangwon National University), Project ID: 1711198468, Contribution Rate: 20%), funded by the government (Ministry of Science and ICT). [Background technology]

[0003] An electrocardiogram is a graphical representation of the heart's electrical activity during each cardiac cycle. Electrocardiography is essential for diagnosing cardiac abnormalities and irregular cardiac arrhythmias, and senses the heart's electrical activity using electrodes attached to the skin. A typical electrocardiogram method is the standard 12-lead electrocardiogram, which uses electrode cables attached to six limb leads (Leads I, II, III, aVR, aVL, aVF) at the extremities and six chest leads (V1, V2, V3, V4, V5, V6) at designated locations on the chest.

[0004] Figure 1 shows a typical electrocardiogram image. As shown in Figure 1, each lead on an electrocardiogram looks similar, but each contains different information about the heart. Generally, to diagnose heart disease, different leads must be focused on. The information about the heart that can be obtained from a single lead is inevitably limited. Therefore, when diagnosing heart disease through an electrocardiogram, various leads must be observed organically to make a comprehensive judgment.

[0005] Recently, electrocardiogram monitoring has become possible through smart devices such as smartwatches like the Galaxy Watch and Apple Watch, and portable electrocardiogram monitors. However, most of these smart devices can only measure single-lead electrocardiograms. While simple atrial fibrillation can be diagnosed with only a single lead, other heart diseases are difficult to measure.

[0006] In recent years, a growing number of studies have been conducted to evaluate the accuracy of cardiac disease diagnosis and the prediction of related diseases using electrocardiograms. Measuring electrocardiograms using smart devices has the advantages of being relatively easy and convenient and capable of long-term recording. However, as mentioned above, most electrocardiograms measured using smart devices are single-lead electrocardiograms. Therefore, they have the disadvantage of being unable to apply diagnostic techniques using 12-lead electrocardiograms, which have been actively researched until now.

[0007] Prior art techniques include restoring missing electrocardiograms using a long-term memory neural network-based artificial intelligence model, as in Korean Patent No. 10-2570598 (published on August 28, 2023), or generating the remaining 12-n electrocardiograms from n electrode information using an adversarial neural network algorithm, as in Korean Patent No. 10-2412974 (published on June 24, 2022). However, these techniques only aim to generate or restore an electrocardiogram well, without performing detailed comparisons on a pixel-by-pixel or time-by-time basis. Furthermore, in order to diagnose diseases using electrocardiograms, it is necessary to verify whether the restored electrocardiogram accurately reflects the characteristics of the actual electrocardiogram. In other words, there is a need for a technology that can accurately reflect the characteristics of existing actual electrocardiograms, rather than simply restoring an electrocardiogram. Summary of the Invention [Problem to be solved by the invention]

[0008] For the above reasons, one aspect of the disclosed invention aims to provide an electrocardiogram reconstruction model generation device, an electrocardiogram reconstruction model generation method, and an electrocardiogram reconstruction method based on a generative adversarial neural network that can reconstruct multiple electrocardiograms from a single electrocardiogram so as to accurately reflect the characteristics of an actual electrocardiogram. [Means for solving the problem]

[0009] According to one aspect of the disclosed invention, an electrocardiogram reconstruction model generating device for generating an electrocardiogram reconstruction model based on a generative adversarial network (GAN) includes an input unit that receives a single-lead electrocardiogram as an input; a generator that reconstructs a 12-lead electrocardiogram based on the single-lead electrocardiogram; and a classifier that uses as an input either a first input value that is a combination of the 12-lead electrocardiogram and the single-lead electrocardiogram, or a second input value that is a combination of an actual 12-lead electrocardiogram (ground-truth) and the single-lead electrocardiogram, and determines whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram.

[0010] The generator may include a label generator that receives the single-lead ECG as input based on a U-Net-based encoder-decoder generative model without skip connections and outputs a second latent vector for outputting a single-lead ECG identical to the single-lead ECG; and an inference generator that performs preprocessing on the single-lead ECG based on a U-Net-based encoder-decoder generative model and reconstructs and outputs the 12-lead ECG based on the first latent vector and the second latent vector of the preprocessed single-lead ECG.

[0011] The inference generator can be trained based on an adversarial loss function for the first input value and the second input value.

[0012] The inference generator can be trained based on an L1 loss function that reflects the Manhattan distance between the 12-lead ECG and the actual 12-lead ECG.

[0013] The inference generator can train based on a latent vector loss function for the first latent vector and the second latent vector.

[0014] The inference generator can perform training based on a global loss function that is a weighted version of the adversarial loss function, the L1 loss function, and the latent vector loss function.

[0015] The classifier can be an encoder-decoder generative model based on 1D U-Net.

[0016] According to one aspect of the disclosed invention, a method for generating an electrocardiogram reconstruction model based on a generative adversarial network (GAN) includes the steps of receiving a single-lead electrocardiogram as an input; reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram; receiving as an input either a first input value that is a combination of the 12-lead electrocardiogram and the single-lead electrocardiogram or a second input value that is a combination of an actual 12-lead electrocardiogram (ground-truth) and the single-lead electrocardiogram; and determining whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram.

[0017] The step of reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram may include the steps of: performing preprocessing on the single-lead electrocardiogram based on a U-Net-based encoder-decoder generative model; receiving the single-lead electrocardiogram as input based on a U-Net-based encoder-decoder generative model without applying skip connections, and outputting a second latent vector for outputting a single-lead electrocardiogram identical to the single-lead electrocardiogram; and reconstructing and outputting the 12-lead electrocardiogram based on the first latent vector and the second latent vector of the preprocessed single-lead electrocardiogram.

[0018] The method for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may further include: performing learning based on an adversarial loss function for the first input value and the second input value.

[0019] The method for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may further include: performing training based on an L1 loss function that reflects the Manhattan distance between the 12-lead electrocardiogram and the actual 12-lead electrocardiogram.

[0020] The method for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may further include the step of performing learning based on a latent vector loss function for the first latent vector and the second latent vector.

[0021] The method for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may further include a step of performing training based on a global loss function obtained by weighting the adversarial loss function, the L1 loss function, and the latent vector loss function, respectively.

[0022] According to one aspect of the disclosed invention, an electrocardiogram reconstruction method using an electrocardiogram reconstruction model based on a generative adversarial neural network may include the steps of receiving a single-lead electrocardiogram as input; preprocessing the single-lead electrocardiogram; inputting the single-lead electrocardiogram into a pre-trained electrocardiogram reconstruction model based on a generative adversarial neural network to reconstruct a 12-lead electrocardiogram; and outputting the reconstructed 12-lead electrocardiogram. [Effects of the Invention]

[0023] According to one aspect of the disclosed invention, it is possible to provide an electrocardiogram reconstruction model generation device, an electrocardiogram reconstruction model generation method, and an electrocardiogram reconstruction method based on a generative adversarial neural network that can reconstruct a multi-lead electrocardiogram that more accurately reflects the characteristics of an actual electrocardiogram from a single-lead electrocardiogram.

[0024] According to one aspect of the disclosed invention, reconstructing a multi-lead electrocardiogram from only a single-lead electrocardiogram can save medical professionals money and time in diagnosing cardiac disease.

[0025] According to one aspect of the disclosed invention, in addition to electrocardiograms, when various specific signals are present, it is possible to restore missing signals or restore signals that are difficult to measure from other signals, and application is possible in various fields such as various electroencephalograms or cerebral blood flow measurements. [Brief explanation of the drawings]

[0026] [Figure 1] 1 is a graph showing a typical 12-lead electrocardiogram image. [Figure 2] 1 is a diagram illustrating a schematic configuration of an electrocardiogram reconstruction model generating device according to an embodiment. [Figure 3] 1 illustrates a schematic diagram of an electrocardiogram reconstruction model according to one embodiment; [Figure 4] 1 is a flowchart illustrating a method for generating an electrocardiogram reconstruction model according to an embodiment. [Figure 5] 10 is a flowchart illustrating in more detail the steps of reconstructing a 12-lead electrocardiogram in a method for generating an electrocardiogram reconstruction model according to an embodiment. [Figure 6] FIG. 10 is a diagram comparing the restoration performance of an electrocardiogram restoration model according to an embodiment with that of an existing model. [Figure 7] FIG. 10 is a diagram illustrating a performance difference between the presence and absence of a classifier and a label generator in an electrocardiogram reconstruction model according to an embodiment. [Figure 8] FIG. 10 is a diagram comparing the prediction performance for three diseases of a 12-lead electrocardiogram restored by an electrocardiogram restoration model according to one embodiment, a 12-lead electrocardiogram restored by an existing model, and an actual 12-lead electrocardiogram. [Figure 9] FIG. 10 is a diagram comparing the restoration performance of an electrocardiogram restoration model according to an embodiment with that of an existing model. DETAILED DESCRIPTION OF THE INVENTION

[0027] The same reference numerals refer to the same elements throughout the specification. This specification does not describe all elements of the embodiments, and content that is common in the technical field to which the disclosed invention belongs or that is duplicated between embodiments will be omitted. The terms "unit, module, component, block" used in the specification can be embodied in software or hardware, and depending on the embodiment, multiple "units, modules, components, blocks" can be embodied as a single component, or one "unit, module, component, block" can include multiple components.

[0028] Throughout this specification, when a part is said to be "connected" to another part, this includes not only a direct connection but also an indirect connection, including a connection via a wireless communication network.

[0029] Furthermore, when a part is said to "comprise" a certain element, this does not mean that it excludes other elements, but that it may further include other elements, unless otherwise specified.

[0030] Throughout this specification, when we say that a member is "on" another member, this includes not only when the member is in contact with the other member, but also when there is another member between the two members.

[0031] The terms "first," "second," etc. are used to distinguish one component from another, and the components are not limited to the terms mentioned above.

[0032] The singular expression includes the plural expression unless the context clearly indicates otherwise.

[0033] In each step, the identification numbers are used for convenience of explanation, and the identification numbers do not describe the order of each step, and each step can be performed in an order different from the order specified unless the context clearly dictates a specific order.

[0034] The principles of operation and embodiments of the disclosed invention will now be described with reference to the accompanying drawings.

[0035] FIG. 2 is a diagram illustrating a schematic configuration of an electrocardiogram reconstruction model generating device 100 according to an embodiment.

[0036] Referring to FIG. 2, an apparatus 100 for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may include an input unit 110, a generator 120, and a classifier .

[0037] The input unit 110 can receive as input a single-lead electrocardiogram of a particular patient. The input unit 110 can preprocess the single-lead electrocardiogram. The input unit 110 can also provide the single-lead electrocardiogram or the preprocessed single-lead electrocardiogram to the generator 120.

[0038] According to one embodiment, the input unit 110 can further receive as input a ground-truth 12-lead electrocardiogram of the particular patient, and can provide the ground-truth 12-lead electrocardiogram to the classifier 130.

[0039] The generator 120 is capable of reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram provided by the input 110 .

[0040] To this end, the generator 120 may include an inference generator 121 and a label generator 122 .

[0041] The inference generator 121 performs preprocessing on the single-lead electrocardiogram based on a U-Net-based encoder-decoder generation model, and can reconstruct and output a 12-lead electrocardiogram based on a first latent vector of the preprocessed single-lead electrocardiogram and a second latent vector provided by the label generator 122 described later.

[0042] The label generator 122 can receive the single-lead electrocardiogram as input and output a second latent vector for outputting a single-lead electrocardiogram identical to the single-lead electrocardiogram based on a U-Net-based encoder-decoder generation model that does not apply skip connections.

[0043] The discriminator 130 can use either a first input value that is a combination of a reconstructed 12-lead electrocardiogram and a single-lead electrocardiogram or a second input value that is a combination of an actual 12-lead electrocardiogram and a single-lead electrocardiogram as an input value to determine whether the input value is a reconstructed 12-lead electrocardiogram or an actual 12-lead electrocardiogram.

[0044] According to one embodiment, the classifier 130 may be an encoder-decoder generative model based on 1D U-Net.

[0045] According to one embodiment, the inference generator 121 may train based on an adversarial loss function for the first input value and the second input value.

[0046] According to one embodiment, the inference generator 121 may be trained based on an L1 loss function that reflects the Manhattan distance between the reconstructed 12-lead ECG and the actual 12-lead ECG.

[0047] According to one embodiment, the inference generator 121 may train based on a latent vector loss function for the first latent vector and the second latent vector.

[0048] According to one embodiment, the inference generator 121 may perform training based on a global loss function that is a weighted combination of an adversarial loss function, an L1 loss function, and a latent vector loss function.

[0049] FIG. 3 is a diagram illustrating a schematic diagram of an electrocardiogram reconstruction model (EKGAN) according to one embodiment.

[0050] The process of restoring a single-lead ECG to a multi-lead ECG can be based on a Generative Adversarial Network (GAN). However, a typical GAN ​​model may have a problem in that it cannot accurately reflect the characteristics of an actual (disease-affected) ECG. To overcome the limitations of using existing GAN models to restore a single-lead ECG to a multi-lead ECG, the present invention discloses an EKGAN (Electrocardiogram Reconstruction Model).

[0051] Referring to Figure 3, the electrocardiogram reconstruction model (EKGAN) is implemented using an inference generator (G I ) 121, Label Generator (G L The electrocardiogram reconstruction model (EKGAN) may further include a generator (Generator) 120 including a classifier (D) 122, and a classifier (D) 130. The electrocardiogram reconstruction model (EKGAN) may further include a preprocessing unit (not shown).

[0052] The generator 120 is responsible for reconstructing a 12-lead electrocardiogram when a single-lead electrocardiogram is input. The generator 120 is composed of two sub-generators: an inference generator (G I ) and label generator (G L )122, each of which plays a different role and allows learning to proceed organically.

[0053] Inference generator (G I ) 121 receives a single-lead electrocardiogram signal as input and is based on an encoder-decoder generation model based on U-Net to restore it to a 12-lead electrocardiogram. At this time, preprocessing of the input signal can be performed at the input end of the encoder to configure an encoder-decoder based model. After that, when the output generated at the output end is transmitted to the input of the discriminator, the discriminator can distinguish between true and false and use it to proceed with learning.

[0054] Inference generator (G IThe detailed structure of 121 may include an input terminal, an encoder-decoder generation model based on U-Net, and an output terminal.

[0055] The input terminal receives a single-lead electrocardiogram as input. Here, the single-lead electrocardiogram has a size of (1, n, 1), where the first 1 represents the number of leads, n represents the length of the electrocardiogram, and the last 1 represents the size of the channel. Here, the length n can be set by the user in various ways, but in the present invention, the value n = 512 is used for explanation.

[0056] To use the encoder-decoder model, the input and output must be the same size, so a preprocessing step is required for the single-lead ECG input. To do this, the single-lead ECG is simply copied 12 times and padded with zeros before and after to change the size to (16, 512, 1). At this time, each convolution filter is configured in the order of 64, 128, 256, 512, 1024.

[0057] The U-Net based encoder-decoder generation model consists of a Contracting Path, which extracts semantic information by viewing a progressively wider range of image pixels in the encoder, and a Latent Vector (Zi) that generates the first latent vector (B) in the section where the contracting path switches to the expanding path. The part corresponding to the ottleneck and decoder can include an Expanding Path that combines the extracted semantic information with pixel position information using skip connections and provides additional information to each pixel.

[0058] U-Net is a model that can extract image features using information that can be confirmed not only in low dimensions but also in high dimensions, and combines this with location information to achieve good performance.

[0059] U-Net is based on providing the decoder with additional information that may be lost when compressing input data into a latent vector.

[0060] However, the latent vectors generated by existing U-Net models are updated using only the loss function between the original data and the generated data, which has the drawback of not checking how well the latent vectors contain the features of the input data. To solve this problem, the present invention uses a label generator (G L ) 122 is introduced to use techniques that help compress the input data without losing information. L ) 122 will be explained later.

[0061] The output terminal has an output size of (16, 512, 1) so that the output has the same size as the input of (16, 512, 1) preprocessed at the input terminal. The output terminal is configured to have a first latent vector (Z i ) to reconstruct a 12-lead ECG. Because it is based on a U-Net, it learns to maximize the information contained in a single-lead ECG through skip connections. The ECG reconstructed at the output terminal is composed of Lead I, II, III, aVR, aVL, aVF, V1-V6 in this order, including zero padding, and each convolution filter can be configured in the order of 512, 256, 128, 64, and 1.

[0062] Label Generator (G L ) 122 is used to minimize information loss in the process of compressing the input data based on U-Net. L ) 122 is an inference generator (G I ) 121 as input and generate the same single-lead ECG output. LThe core role of 122 is to generate a second latent vector (Z) that can best represent the single-lead ECG received as input. l ) that leads to generate the same electrocardiogram as the input. That is, the second latent vector (Z l ) alone can generate a single-lead ECG well, this is because the second latent vector (Z l ) can compress the single-lead ECG signal while retaining much of its input information.

[0063] Label Generator (G L The overall structure of 122 is the inference generator (G I )121, but does not apply skip connections to compress the information as well as possible. This is a general U-Net-based encoder-decoder generative model and a label generator (G L ) 122 are differentiated by the label generator (G L The second latent vector (Z) generated in 122 l ) is the inference generator (G I The first latent vector (Z) generated from the input data of 121 i ) and the loss function, I )121's information compression capability can be improved.

[0064] Since each lead in a 12-lead electrocardiogram has its own unique characteristics, it is very important that the discriminator (D) 130 analyzes each lead signal individually at the pixel level.

[0065] Therefore, when using a classifier based on 2D convolution, which is used in general GAN ​​models for image generation, the unique characteristics of each read may be mixed with other reads, resulting in poor reconstruction quality.

[0066] To solve this problem, the classifier can use an encoder-decoder generative model based on 1D U-Net. I ) 121, along with the single-lead ECG used as input, and trained to distinguish whether the input is a generated ECG or an original ECG.

[0067] That is, the classifier is an inference generator (G I ) 121 and the single-lead electrocardiogram used as input, or a second input value that combines the actual 12-lead electrocardiogram and the single-lead electrocardiogram used as input. [Mode for Carrying Out the Invention]

[0068] FIG. 4 is a flowchart showing a method for generating an electrocardiogram reconstruction model according to an embodiment, and FIG. 5 is a flowchart showing in more detail the steps of reconstructing a 12-lead electrocardiogram in the method for generating an electrocardiogram reconstruction model according to an embodiment.

[0069] Referring to FIG. 4, a method for generating an electrocardiogram reconstruction model according to an embodiment of the present invention may include step 1100 of receiving a single-lead electrocardiogram as input, step 1200 of reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram, step 1300 of receiving as input either a first input value that is a combination of the 12-lead electrocardiogram and the single-lead electrocardiogram or a second input value that is a combination of an actual 12-lead electrocardiogram (ground-truth) and the single-lead electrocardiogram, and step 1400 of determining whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram.

[0070] Referring to FIG. 5 , step 1200 of reconstructing a 12-lead electrocardiogram based on a single-lead electrocardiogram may include step 1210 of performing preprocessing on the single-lead electrocardiogram based on a U-Net-based encoder-decoder generative model, step 1220 of receiving the single-lead electrocardiogram as input based on a U-Net-based encoder-decoder generative model without skip connection and outputting a second latent vector for outputting a single-lead electrocardiogram identical to the single-lead electrocardiogram, and step 1230 of reconstructing and outputting the 12-lead electrocardiogram based on the first latent vector and the second latent vector of the preprocessed single-lead electrocardiogram.

[0071] According to one embodiment, the electrocardiogram reconstruction model generation method may further include a step (not shown) of performing learning based on an adversarial loss function for the first input value and the second input value.

[0072] According to one embodiment, the method for generating an electrocardiogram reconstruction model may further include a step (not shown) of training based on an L1 loss function reflecting the Manhattan distance between the reconstructed 12-lead electrocardiogram and the actual 12-lead electrocardiogram.

[0073] According to one embodiment, the method for generating an electrocardiogram reconstruction model may further include a step (not shown) of performing training based on a latent vector loss function for the first latent vector and the second latent vector.

[0074] According to one embodiment, the method for generating an electrocardiogram reconstruction model may further include a step (not shown) of performing training based on a global loss function in which the adversarial loss function, the L1 loss function, and the latent vector loss function are weighted, respectively.

[0075] An electrocardiogram reconstruction method (not shown) according to one embodiment of the present invention may include the steps of receiving a single-lead electrocardiogram as input, preprocessing the single-lead electrocardiogram, inputting the single-lead electrocardiogram into an electrocardiogram reconstruction model based on a pre-trained generative adversarial neural network to reconstruct a 12-lead electrocardiogram, and outputting the reconstructed 12-lead electrocardiogram.

[0076] The electrocardiogram reconstruction method according to one embodiment is premised on the use of the pre-trained electrocardiogram reconstruction model (EKGAN) disclosed in Figures 1 to 3. Therefore, a 12-lead electrocardiogram can be reconstructed based on an input single-lead electrocardiogram using the electrocardiogram reconstruction model (EKGAN) that already has high performance.

[0077] Meanwhile, in the present invention, the generator 120 and the classifier 130 can use loss functions for training the electrocardiogram reconstruction model (EKGAN). At this time, the loss functions used can be roughly divided into three types, including an adversarial loss function, an L1 loss function, and a latent vector loss function, and an overall loss function obtained by weighting each of these three functions can also be used.

[0078] Before explaining each loss function, the meaning of the symbols used in each loss function is as follows.

[0079] G I :Inference generator 121 G L :Label Generator 122 D: Discriminator 130 e i : Single-lead ECG data replicated 12 times and zero-padded e o :Single-lead ECG vs. actual 12-lead ECG (Ground-truth) z i , z l :G I and G L Latent Vector generated from The adversarial loss function is the inference generator (G I ) 121 is used as a basis for learning the difference between the data reconstructed from the single-lead ECG received as input and the actual 12-lead ECG. The formula is as follows, and the result when the actual data is input to the discriminator (D) 130 and the inference generator (G I The loss function of the Conditional GAN ​​method for the restored 12-lead electrocardiogram data and the input data obtained by inputting input data into the classifier (D) 121. That is, the classifier (D) 130 learns whether the input value is generated data or actual data using an adversarial loss function that uses the electrocardiogram used for restoration and the actual electrocardiogram, and the electrocardiogram used for restoration and the restored electrocardiogram.

[0080]

number

[0081] The L1 loss function can use not only the adversarial loss but also the L1 loss that reflects the Manhattan distance between the actual ECG and the generated ECG. This is because the inference generator (G I )121, as well as the inference generator (G I Label generator (G) for use in 121 L )122 can also be used, and the detailed formula is as follows:

[0082]

number

[0083] The Latent Vector loss function is a label generator (G L The second latent vector (Z) is generated from l ) and the inference generator (G I The first latent vector (Z) generated from i) we can calculate the L1 loss between two latent vectors to check their similarity. The detailed formula is as follows:

[0084]

number

[0085] The total loss function is calculated by applying the loss function to the inference generator (G I ) 121 loss, which is expressed as the following formula: λ and α are variables that control the relative importance and can be set differently depending on the given purpose, and in the embodiment of the present invention, λ=50 and α=1.

[0086]

number

[0087] 6 to 9, the electrocardiogram reconstruction model (EKGAN) according to the embodiment of the present invention receives a single-lead electrocardiogram, Lead 1, as an input and outputs a reconstructed 12-lead electrocardiogram. To verify the reconstruction performance of the electrocardiogram reconstruction model (EKGAN), training and verification were performed using electrocardiograms collected at Ewha Womans University Mokdong Hospital and Ewha Womans University Seoul Hospital, and the model was compared with existing GAN-based models, and other label generators (G L We also performed performance measurements to confirm the usefulness of 122. The performance metrics used were RMSE (Root Mean Square Error), MAE (Mean Absolute Error), PRD (Percentage Root Mean Square Difference), MMD (Maximum Mean Discrepancy), and MAEHR (Mean Absolute Error for Heart Rates).

[0088] FIG. 6 is a diagram comparing the restoration performance of an electrocardiogram restoration model (EKGAN) according to one embodiment with that of an existing model.

[0089] As can be seen from Figure 6, the well-known Pix2Pix, CycleGAN, and C It can be seen that it performs better in all metrics than GAN models such as ardioGAN.

[0090] FIG. 7 shows a classifier (D) 130 and a label generator (G) of an electrocardiogram reconstruction model (EKGAN) according to an embodiment. L 12 is a diagram showing the difference in performance depending on whether or not the .

[0091] Referring to FIG. 7, the EKGAN w / o 1D classifier (D) 130 illustrates a case where the PatchGAN classifier used in Pix2Pix is ​​used instead of the classifier (D) 130 based on the 1D U-Net model in the electrocardiogram reconstruction model (EKGAN) according to an embodiment of the present invention. This results in a decrease in performance.

[0092] Also, EKGAN w / o label generator (G L ) 122 is a label generator (G) using an electrocardiogram reconstruction model (EKGAN) according to an embodiment of the present invention. L ) 122 is not used, and the restoration performance is L It can be seen that the performance is inferior to that of the electrocardiogram reconstruction model (EKGAN) to which 122 is applied.

[0093] FIG. 8 is a graph comparing the prediction performance for three diseases of a 12-lead electrocardiogram restored by an electrocardiogram restoration model (EKGAN) according to one embodiment, a 12-lead electrocardiogram restored by an existing model, and an actual 12-lead electrocardiogram.

[0094] Referring to FIG. 8, a comparison of the prediction results of a conventional electrocardiogram classification model, the prediction results of the original electrocardiogram, and the prediction results of the reconstructed electrocardiogram shows that the disease prediction (classification results) using the electrocardiogram reconstruction model (EKGAN) of the present invention exhibits excellent performance. In this experiment, prediction performance was confirmed for three diseases that can be diagnosed through an electrocardiogram. Atrial fibrillation (AF) can be detected with a single lead, but LBBB and RBBB (Left / Right Bundle Branch Block) require multiple leads to be detected. Thus, the results of FIG. 8 experimentally prove that a 12-lead electrocardiogram reconstructed using the electrocardiogram reconstruction model (EKGAN) according to an embodiment of the present invention can be applied to various downstream tasks.

[0095] 9 is a diagram comparing the restoration performance of an electrocardiogram restoration model (EKGAN) according to an embodiment with that of an existing model. More specifically, FIG. 9 compares a 12-lead electrocardiogram actually restored by the electrocardiogram restoration model (EKGAN) according to an embodiment of the present invention with a model used for the performance comparison, and a portion of the original electrocardiogram (V1, Lead II, V4, V6).

[0096] 9, it can be seen that the electrocardiogram reconstruction model (EKGAN) according to an embodiment of the present invention reconstructs an electrocardiogram with a shape that has a significant similarity to the actual original electrocardiogram compared to other models. This means that the electrocardiogram reconstruction model (EKGAN) according to an embodiment of the present invention can better capture disease characteristics from a single-lead electrocardiogram and can reconstruct a 12-lead electrocardiogram relatively accurately compared to other models.

[0097] The above description merely exemplifies the technical concept of the present invention, and various modifications and variations are possible within the scope of the essential characteristics of the present invention, as long as they are not deviated from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are for illustrative purposes only, and are not intended to limit the technical concept of the present invention. The scope of the present invention should be interpreted by the following claims, and all technical concepts within the scope of the claims should be interpreted as being within the scope of the present invention. [Industrial Applicability]

[0098] The present invention can be used in wearable devices, diagnostic applications, electrocardiogram diagnostic equipment, etc. that use 12-lead electrocardiogram reconstruction technology to enable more accurate and rapid diagnosis.

Claims

1. An electrocardiogram reconstruction model generating device for generating an electrocardiogram reconstruction model based on a generative adversarial neural network (GAN), comprising: an input section that receives as input a single-lead electrocardiogram; a generator that reconstructs a 12-lead electrocardiogram based on the single-lead electrocardiogram; and a discriminator that uses as an input either a first input value that is a combination of the 12-lead electrocardiogram and the single-lead electrocardiogram or a second input value that is a combination of an actual 12-lead electrocardiogram (ground-truth) and the single-lead electrocardiogram, and discriminates whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram; An electrocardiogram reconstruction model generating device comprising:

2. The generator a label generator that receives the single-lead electrocardiogram as an input based on a U-Net-based encoder-decoder generation model without skip connections and outputs a second latent vector for outputting a single-lead electrocardiogram identical to the single-lead electrocardiogram; and an inference generator that performs preprocessing on the single-lead electrocardiogram based on a U-Net-based encoder-decoder generation model, and reconstructs and outputs the 12-lead electrocardiogram based on a first latent vector and the second latent vector of the preprocessed single-lead electrocardiogram; The electrocardiogram reconstruction model generating device according to claim 1 , comprising:

3. The inference generator: based on an adversarial loss function for the first input value and the second input value 3. The electrocardiogram reconstruction model generating device according to claim 2, wherein the electrocardiogram reconstruction model generating device performs learning.

4. The inference generator:

4. The electrocardiogram reconstruction model generating device according to claim 3, wherein the learning is performed based on an L1 loss function that reflects the Manhattan distance between the 12-lead electrocardiogram and the actual 12-lead electrocardiogram.

5. The inference generator: The electrocardiogram reconstruction model generating device according to claim 4 , wherein learning is performed based on a latent vector loss function for the first latent vector and the second latent vector.

6. The inference generator: The electrocardiogram reconstruction model generation device according to claim 5 , wherein learning is performed based on a global loss function obtained by weighting the adversarial loss function, the L1 loss function, and the latent vector loss function.

7. The discriminator The electrocardiogram reconstruction model generating device according to claim 1, wherein the electrocardiogram reconstruction model generating device is an encoder-decoder generated model based on 1D U-Net.

8. A method for generating an electrocardiogram reconstruction model based on a Generative Adversarial Network (GAN), receiving as input a single-lead electrocardiogram; reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram; receiving as input either a first input value that combines the 12-lead electrocardiogram with the single-lead electrocardiogram or a second input value that combines an actual 12-lead electrocardiogram (ground-truth) with the single-lead electrocardiogram; and determining whether the input value is the reconstructed 12-lead electrocardiogram or the actual 12-lead electrocardiogram; The electrocardiogram reconstruction model generating method includes:

9. The step of reconstructing a 12-lead electrocardiogram based on the single-lead electrocardiogram includes: performing pre-processing on the single-lead electrocardiogram based on a U-Net based encoder-decoder generation model; receiving the single-lead electrocardiogram as an input based on a U-Net-based encoder-decoder generation model without skip connections, and outputting a second latent vector for outputting a single-lead electrocardiogram identical to the single-lead electrocardiogram; and reconstructing and outputting the 12-lead electrocardiogram based on the first latent vector and the second latent vector of the preprocessed single-lead electrocardiogram; The method for generating an electrocardiogram reconstruction model according to claim 8 , comprising:

10. performing training based on an adversarial loss function for the first input value and the second input value; The method for generating an electrocardiogram reconstruction model according to claim 9 , further comprising:

11. training based on an L1 loss function reflecting Manhattan distance between the 12-lead electrocardiogram and the actual 12-lead electrocardiogram; The method for generating an electrocardiogram reconstruction model according to claim 10, further comprising:

12. performing training based on a latent vector loss function for the first latent vector and the second latent vector; The method for generating an electrocardiogram reconstruction model according to claim 11 , further comprising:

13. a step of performing learning based on a global loss function obtained by weighting the adversarial loss function, the L1 loss function, and the latent vector loss function; The method for generating an electrocardiogram reconstruction model according to claim 12, further comprising:

14. In an electrocardiogram reconstruction method using an electrocardiogram reconstruction model based on a Generative Adversarial Network (GAN), receiving as input a single-lead electrocardiogram; preprocessing the single-lead electrocardiogram; inputting the single-lead electrocardiogram into a pre-trained generative adversarial neural network-based electrocardiogram reconstruction model to reconstruct a 12-lead electrocardiogram; and outputting the reconstructed 12-lead electrocardiogram; An electrocardiogram reconstruction method comprising:

Images