Deep learning-based electrocardiogram data denoising system

Abstract

The present invention discloses a deep learning-based noise removal system for electrocardiogram data, which includes an electrocardiogram measurement unit (110) that measures electrocardiogram data for each lead from the body of a medical examinee, and a style-based electrocardiogram generation unit (120) that extracts an inherent style from the electrocardiogram data measured by the electrocardiogram measurement unit (110) by reflecting the characteristics of the medical examinee and the characteristics of the measurement method through an electrocardiogram generation deep learning algorithm (121) constructed by learning in advance a set of learning data of electrocardiogram data for each lead and an inherent style of the electrocardiogram data for each lead with low noise based on a large number of electrocardiogram data, and converts and generates electrocardiogram data of a specific lead style that does not include noise through the extracted inherent style, thereby generating electrocardiogram data from which noise has been removed, thereby improving the accuracy of disease diagnosis and prediction.

Description

[Technical Field]

[0001] The present invention relates to a deep learning-based electrocardiogram data noise removal system that can maintain the cardiac-specific electrical signals from the measured electrocardiogram data, reflect the characteristics of the subject and the characteristics of the measurement method, and effectively remove noise caused by external electrical signals.

[0002] [Background technology]

[0003] As is well known, after the development of the electrocardiogram, knowledge related to electrocardiograms has expanded exponentially, and electrocardiograms can be used to obtain information about the electrical function of the heart and diagnose various heart diseases such as arrhythmias, coronary artery disease, and myocardial disease.

[0004] Recently, active research has been conducted on AI algorithms for electrocardiograms, and AI algorithms can be used to detect heart failure, predict atrial fibrillation from arrhythmia rhythms, or even determine gender.

[0005] In this way, by overcoming human limitations, AI algorithms can detect subtle changes in ECG waveforms, thereby improving ECG interpretation.

[0006] For example, the electrocardiogram used in the medical field is a 12-lead electrocardiogram, which is measured by attaching 10 electrodes: three limb electrodes, six chest electrodes, and one ground electrode, and the measured electrocardiogram data can be transmitted remotely.

[0007] On the other hand, an electrocardiogram is data output by measuring electrical signals from the heart on the body surface, and thus is subject to considerable noise. Here, electrocardiograms may contain noise generated by electromyograms due to contractions of muscles in the chest and limbs, and noise generated at the contact surface between the skin and the electrodes. Breathing during an electrocardiogram may cause the chest to rise and fall, resulting in chest movement being included in the electrocardiogram. Noise may also be generated in the wires connecting the electrodes in contact with the body to the electrocardiogram device. Noise may also be generated by the alternating current in the electrocardiogram device itself. In the case of exercise electrocardiograms and activity electrocardiograms, which are not measured in a stationary state but involve movement after wearing the device, noise caused by walking or other movements may also be included in the electrocardiogram.

[0008] Conventionally, noise has been removed by collectively removing noise below and above a certain frequency, i.e., signals below 0.05 Hz or above 150 Hz, which are frequency bands not primarily generated by the heart. However, this method causes distortion of the ECG signal itself, which reduces the accuracy of disease diagnosis.

[0009] As described above, it is not easy to distinguish and remove noises generated by the characteristics of the examinee, the measurement environment, and the measurement equipment from the electrical signals of the heart itself, and a method of removing noises collectively based on a certain frequency distorts the signal specific to the electrocardiogram, thereby reducing the accuracy of diagnosing or predicting the occurrence of diseases based on the electrocardiogram.

[0010] Therefore, there is a need for a technique for removing noise from electrocardiogram data to generate high-quality electrocardiograms and ensure high diagnostic accuracy. Summary of the Invention [Problem to be solved by the invention]

[0011] The technical problem that the concept of the present invention aims to achieve is to provide a deep learning-based noise removal system for electrocardiogram data that can effectively remove noise caused by external electrical signals by maintaining the electrical signals specific to the heart from the measured electrocardiogram data and reflecting the characteristics of the subject and the characteristics of the measurement method.

[0012] [Means for solving the problem]

[0013] To achieve the above-mentioned object, an embodiment of the present invention provides a deep learning-based noise removal system for electrocardiogram data, including: an electrocardiogram measurement unit that measures electrocardiogram data for each lead from the body of a medical subject; and a style-based electrocardiogram generation unit that extracts an eigenstyle from the electrocardiogram data measured by the electrocardiogram measurement unit, reflecting the characteristics of the medical subject and the characteristics of the measurement method, through an electrocardiogram generation deep learning algorithm constructed by pre-learning a set of learning data for each lead of electrocardiogram data and eigenstyles of each lead of electrocardiogram data with low noise based on a large number of electrocardiogram data, and converts and generates electrocardiogram data of a specific lead style that is noise-free using the extracted eigenstyle.

[0014]

[0015] Here, the electrocardiogram generation deep learning algorithm may generate a plurality of electrocardiogram data for each lead, and may construct the set of learning data by learning in advance.

[0016]

[0017] In addition, the deep learning algorithm for generating an electrocardiogram for each lead can learn electrocardiogram data having one or more eigenstyles.

[0018]

[0019] In addition, the deep learning algorithm for generating an electrocardiogram may be implemented using an autoencoder or a generative adversarial network, either singly or in combination, to extract the inherent style of the electrocardiogram data for each lead.

[0020]

[0021] In addition, the autoencoder is composed of an encoder that expresses the eigenstyle of electrocardiogram data and a decoder that restores the eigenstyle to the original electrocardiogram data, and is capable of learning the eigenstyle of the electrocardiogram data.

[0022]

[0023] In addition, the generative adversarial network is composed of a generative network that generates synthetic electrocardiogram data using randomly generated variables as input, and a discriminative network that classifies whether the synthetic electrocardiogram data is similar to actual electrocardiogram data, and the generative network can learn the inherent style of the electrocardiogram data.

[0024]

[0025] In addition, the generative network can be converted into the form of the autoencoder, and the autoencoder and the generative adversarial network can be merged.

[0026] [Effects of the Invention]

[0027] According to the present invention, the electrical signals specific to the heart can be maintained from the measured electrocardiogram data, and noise due to external electrical signals can be effectively removed by reflecting the characteristics of the subject and the characteristics of the measurement method. This enables diseases to be more accurately predicted and diagnosed using the noise-removed electrocardiogram data, and the degree of noise can be quantified by comparing and analyzing the noise-removed electrocardiogram data with the original electrocardiogram data.

[0028] [Brief explanation of the drawings]

[0029] [Figure 1] FIG. 1 is a schematic configuration diagram of an electrocardiogram data denoising system based on deep learning according to an embodiment of the present invention.

[0030] [Figure 2] FIG. 2 illustrates an autoencoder of the deep learning-based electrocardiogram data denoising system of FIG. 1.

[0031] [Figure 3] FIG. 2 illustrates an example of a generative adversarial network for the deep learning-based electrocardiogram data denoising system of FIG. 1.

[0032] [Figure 4] 2 is a diagram illustrating noise removal by the deep learning-based electrocardiogram data noise removal system of FIG. 1. FIG.

[0033] DETAILED DESCRIPTION OF THE INVENTION

[0034] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the preferred embodiments of the present invention having the above-mentioned features will be described in more detail with reference to the accompanying drawings.

[0035]

[0036] The electrocardiogram data noise removal system based on deep learning according to an embodiment of the present invention includes an electrocardiogram measurement unit 110 that measures electrocardiogram data for each lead from the body of a medical examinee, and a style-based electrocardiogram generation unit 120 that extracts an eigenstyle from the electrocardiogram data measured by the electrocardiogram measurement unit 110 by reflecting the characteristics of the medical examinee and the characteristics of the measurement method through an electrocardiogram generation deep learning algorithm 121 that is constructed by pre-learning a set of learning data of electrocardiogram data for each lead and eigenstyles for each lead with low noise based on a large number of electrocardiogram data, and converts the extracted eigenstyle into electrocardiogram data of a specific lead style that does not include noise, thereby generating electrocardiogram data from which noise has been removed, thereby improving the accuracy of disease diagnosis and prediction.

[0037]

[0038] Hereinafter, the electrocardiogram data noise removal system based on deep learning having the above-described configuration will be specifically described in detail with reference to FIGS.

[0039]

[0040] First, the electrocardiogram measurement unit 110 measures one or more lead-specific electrocardiogram data from the examinee's body and transmits the data to the style-based electrocardiogram generation unit 120 for provision.

[0041] For example, the electrocardiogram measurement unit 110 may include a wearable electrocardiogram patch 111 capable of contact or non-contact electrocardiogram measurement in daily life, a smart watch 112, a 6-lead electrocardiogram bar for short-term measurement, or a medical institution electrocardiogram device capable of one or more standard lead electrocardiogram measurements, and may measure asynchronous or synchronous electrocardiograms.

[0042] Here, the electrocardiogram measurement unit 110 can measure the subject's continuous electrocardiogram and transmit it to the style-based electrocardiogram generation unit 120, or can measure two electrocardiograms at a time interval and transmit them to the style-based electrocardiogram generation unit 120.

[0043]

[0044] Next, the style-based electrocardiogram generation unit 120 converts the electrocardiogram data for each lead from the electrocardiogram measurement unit 110 into electrocardiogram data with noise removed, reflecting the characteristics of the subject and the characteristics of the measurement method, through the electrocardiogram generation deep learning algorithm 121.

[0045] Specifically, the style-based electrocardiogram generation unit 120 extracts the corresponding eigenstyle from the electrocardiogram data measured by the electrocardiogram measurement unit 110, reflecting the characteristics of the subject and the characteristics of the measurement method, through an electrocardiogram generation deep learning algorithm 121 constructed by pre-learning a set of learning data of electrocardiogram data for each lead with low noise and eigenstyles for each lead based on a large number of electrocardiogram data, for example, standard 12-lead electrocardiogram data accumulated in a medical institution, and converts the extracted eigenstyle into electrocardiogram data of a specific lead style that does not contain noise, thereby generating electrocardiogram data.

[0046] That is, each lead electrocardiogram has its own unique style due to the three-dimensional electrical flow measured in a direction (electrical potential vector) when looking at the heart from a specific direction. For example, just as a V6 lead electrocardiogram always has a style drawn by quantities, each lead electrocardiogram has its own unique style and may include not only the characteristics of P, Q, R, S, and T waveforms, but also curve inflection or noise signals that cannot be identified with existing medical knowledge.

[0047]

[0048] Here, the electrocardiogram generation deep learning algorithm 121 extracts unique styles, which are characteristics of each lead electrocardiogram data, from large-scale electrocardiogram data accumulated in medical institutions and uses them as a set of learning data, and can generate multiple sets of learning data for each lead of the electrocardiogram data and build them in advance by learning.The electrocardiogram generation deep learning algorithm for each lead learns electrocardiogram data having one or more unique styles, and one electrocardiogram generation deep learning algorithm 121 can learn based on not only one electrocardiogram data but also electrocardiogram data having two or more multiple unique styles.

[0049] For example, in order to generate V1 lead electrocardiogram data, the electrocardiogram generation deep learning algorithm 121 can learn the eigenstyle of V1 lead electrocardiogram data based on large-scale V1 lead electrocardiogram data, and convert the V1 lead electrocardiogram data from the electrocardiogram measurement unit 110 into the form of that eigenstyle to generate the V1 lead electrocardiogram data.

[0050]

[0051] Meanwhile, the electrocardiogram generation deep learning algorithm 121 can accurately grasp the lead-specific style of each electrocardiogram data by reflecting characteristics such as the age, sex, and disease of the examinee, as well as characteristics of the measurement method such as the electrode attachment position and the electrocardiogram device, thereby more accurately converting and generating electrocardiogram data for each lead.

[0052]

[0053] In addition, the ECG generation deep learning algorithm 121 may extract an eigenstyle, which is a latent characteristic of given ECG data, through learning using a self-supervised learning method, and generate ECG data by converting ECG data input from the ECG measurement unit 110 into ECG data of each eigenstyle. For example, an autoencoder or a generative adversarial network may be implemented independently or in combination to extract the eigenstyle of ECG data for each lead and generate ECG data of the corresponding eigenstyle.

[0054] For example, referring to FIG. 2, the autoencoder is composed of an encoder 121a that expresses the eigenstyle of electrocardiogram data and a decoder 121b that restores the eigenstyle to the original electrocardiogram data. The electrocardiogram data is input to the encoder 121a, which includes one or more hidden layers Hidden1 and Hidden2, and the decoder 121b, which includes one or more hidden layers Hidden2 and Hidden3, adds Gaussian noise to the extracted eigenstyle to grasp the structure of the original electrocardiogram data, and restores the electrocardiogram data used as input. This process allows the autoencoder to learn the eigenstyle of the electrocardiogram data in order to restore the original electrocardiogram data.

[0055]

[0056] Alternatively, referring to FIG. 3, the generative adversarial network is composed of a generative network 121c that generates synthetic electrocardiogram data using randomly generated variables as input, and a discriminative network 121d that classifies whether the synthetic electrocardiogram data is similar to actual electrocardiogram data. The generative network 121c attempts to generate synthetic electrocardiogram data such that the discriminative network 121d cannot clearly distinguish between the synthetic electrocardiogram data and the actual electrocardiogram data, and the discriminative network 121d attempts to clearly distinguish between the synthetic electrocardiogram data and the actual electrocardiogram data, thereby enabling the generative network 121c to learn the inherent style of the electrocardiogram data.

[0057]

[0058] Meanwhile, the generative network 121c can be changed to the form of an autoencoder, merging an autoencoder with a generative adversarial network, and can be implemented by conditionally inputting an index for desired electrocardiogram data along with a random variable.

[0059]

[0060] FIG. 4 is a diagram illustrating noise removal by the deep learning-based electrocardiogram data noise removal system of FIG. 1. Referring to this figure, the noise removal will be described in detail as follows.

[0061] The electrocardiogram data stored in medical institutions is an electrocardiogram with noise reduced to the maximum extent possible. In actual medical practice, electrocardiogram electrodes are attached to the patient's body, and the patient is observed for a certain period of time, and the electrocardiogram is taken after confirming that the noise has been reduced by, for example, holding one's breath or relaxing the entire body. If the taken electrocardiogram has too much noise and is not valuable as a medical record, the electrocardiogram can be deleted and taken again, so that the electrocardiogram data for learning the electrocardiogram generation deep learning algorithm 121 can be an electrocardiogram with little noise.

[0062] Therefore, when noisy electrocardiogram data measured by the electrocardiogram measurement unit 110 is input to the electrocardiogram generation deep learning algorithm 121, it can be converted into electrocardiogram data with less noise and eigenstyle and output.

[0063] In addition, noise filtering tailored to each individual patient can be applied, reflecting the characteristics of the patient and the characteristics of the measurement method, to remove distortion from the ECG signal, leaving the inherent cardiac vibration intact and removing only the noise.

[0064] Figure 4 is a graph illustrating the experimental results using the electrocardiogram generation deep learning algorithm 121, where the thick line is the actual electrocardiogram generated by the electrocardiogram measurement unit 110 and the thin line is the result of the electrocardiogram generated by the electrocardiogram generation deep learning algorithm 121.

[0065] Looking at the example of lead V6 at the bottom right, it can be seen that the baseline is actually moving due to noise, and that when this is input into the ECG generation deep learning algorithm 121 to generate lead V6, the noise is removed.

[0066] In this way, by grasping the characteristics of each electrocardiogram and determining which lead the electrocardiogram information input into the electrocardiogram generation deep learning algorithm 121 is, this can be used to generate new lead electrocardiogram information.

[0067] It is also possible to generate synchronized electrocardiogram information, but if the model used for disease diagnosis uses asynchronous electrocardiograms or is a model that does not use synchronized time-series information, it is also possible to generate asynchronous electrocardiogram information or to generate electrocardiogram information without considering synchronization.

[0068] More specifically, when trying to predict disease using electrocardiogram data, it is important to measure the noise contained in the input electrocardiogram data, which allows one to decide whether to use the electrocardiogram for disease prediction or to re-measure it, and also allows one to confirm in advance the reliability of the results output after inputting the electrocardiogram data.

[0069] Therefore, by using a method for measuring noise from electrocardiogram data, the initially input electrocardiogram data can be compared with the electrocardiogram from which noise has been removed to measure the level of noise in the electrocardiogram.

[0070] In addition, if some leads or sections of the electrocardiogram data contain a lot of noise or cannot be measured due to reasons such as electrode contact being separated, electrocardiogram data from noise-free leads can be generated to fill in the missing electrocardiogram data, thereby enabling more accurate disease diagnosis, prediction, and screening.

[0071]

[0072] Furthermore, the health condition prediction unit 130 can use the electrocardiogram data generated by the style-based electrocardiogram generation unit 120 and from which noise has been removed as input to more accurately predict diseases that can be predicted from electrocardiogram data.

[0073] For example, the health condition prediction unit 130 can diagnose and predict circulatory system diseases, endocrine, nutritional and metabolic diseases, neoplastic diseases, mental and behavioral disorders, nervous system diseases, eye and appendicular diseases, ear and mastoid diseases, respiratory system diseases, digestive system diseases, skin and skin tissue diseases, musculoskeletal and connective tissue diseases, genitourinary system diseases, pregnancy, childbirth and postpartum diseases, and congenital malformations, deformities and chromosomal abnormalities.

[0074] In addition, the health condition prediction unit 130 can check damage caused by physical trauma, check prognosis, measure pain, predict the risk of death or worsening risk due to trauma, capture or predict complications, and identify specific pathologies that appear before or after birth.In the healthcare field, the health condition prediction unit 130 can measure, diagnose, examine, and predict the health condition of the examinee, which can lead to services such as aging, sleep, weight, blood pressure, blood sugar, oxygen saturation, metabolism, stress, tension, fear, drinking, smoking, problem behaviors, lung capacity, exercise amount, pain management, obesity, body mass, body composition, menus, exercise types, lifestyle pattern recommendations, emergency situation management, late-onset disease management, drug prescriptions, test recommendations, checkup recommendations, nursing care, remote health management, remote medical treatment, vaccination and post-vaccination management.

[0075]

[0076] Meanwhile, the health condition prediction unit 130 generates a generalized reference electrocardiogram of the subject's normal, healthy state, and then compares and analyzes it with electrocardiogram data provided in real time from the electrocardiogram measurement unit 110 to monitor whether the electrocardiogram is an electrocardiogram measured without error and whether there are any abnormalities in the subject's health condition.When an error or abnormality is predicted, the warning unit 140 generates warning information and transmits the warning information together with a beep via the single-lead electrocardiogram measurement unit 110 in the form of a smart watch or a separate smart device.

[0077]

[0078] Therefore, by configuring the electrocardiogram data noise removal system based on deep learning as described above, it is possible to maintain the electrical signals specific to the heart from the measured electrocardiogram data, and effectively remove noise caused by external electrical signals by reflecting the characteristics of the subject and the characteristics of the measurement method. The noise-removed electrocardiogram data can be used to more accurately predict and diagnose diseases, and the degree of noise can also be quantified by comparing and analyzing the noise-removed electrocardiogram data with the original electrocardiogram data.

[0079]

[0080] The embodiments described in this specification and the configurations shown in the drawings are merely the best embodiment of the present invention and do not fully represent the technical ideas of the present invention, so it should be understood that there may be various equivalents and modifications that can replace them at the time of this application.

[0081] [Explanation of symbols]

[0082] 110 Electrocardiogram measurement unit 120 Style-based ECG generation

[0083] 121 Deep Learning Algorithm for Electrocardiogram Generation 121a encoder

[0084] 121b decoder 121c Generative Network

[0085] 121d Identification Network 130 Health Status Prediction Department

[0086] 140 Warning section

Claims

1. an electrocardiogram measurement unit configured to acquire electrocardiogram data by measuring the electrocardiogram data via a plurality of electrodes in contact with the body of the examinee; a style-based electrocardiogram generator configured to extract an intrinsic style of the electrocardiogram data; The eigenstyle reflects characteristics of a measurement method corresponding to contact positions of the plurality of electrodes based on a pre-trained deep learning algorithm for generating an electrocardiogram; the electrocardiogram generating unit converts the acquired electrocardiogram data into the extracted eigenstyle to generate new noise-removed electrocardiogram data; wherein the eigenstyle includes lead-specific latent characteristics corresponding to features extracted from an electrocardiogram waveform of each lead determined by contact positions of the plurality of electrodes; Here, the electrocardiogram generation deep learning algorithm is constructed by previously learning a set of learning data including a plurality of electrocardiogram data for each lead acquired at a medical institution and information on the eigenstyle of the plurality of electrocardiogram data. A deep learning based electrocardiogram data denoising system.

2. The electrocardiogram data noise removal system includes a plurality of electrocardiogram generation deep learning algorithms corresponding to respective leads; Each of the plurality of electrocardiogram generation deep learning algorithms is trained based on the plurality of electrocardiogram data included in the training data set and information on the eigenstyle of the plurality of electrocardiogram data. The deep learning-based electrocardiogram data denoising system according to claim 1.

3. The deep learning-based electrocardiogram data denoising system according to claim 2 , wherein the deep learning algorithm for generating an electrocardiogram for each lead learns electrocardiogram data having one or more eigenstyles.

4. The deep learning-based electrocardiogram data denoising system of claim 1, wherein the electrocardiogram generation deep learning algorithm is implemented as an autoencoder or a generative adversarial network, either singly or in combination, to extract the eigenstyle of the plurality of electrocardiogram data.

5. The deep learning-based electrocardiogram data denoising system of claim 4, characterized in that the autoencoder is composed of an encoder that represents the eigenstyle of the electrocardiogram data and a decoder that restores the eigenstyle to the original electrocardiogram data, and learns the eigenstyle of the electrocardiogram data.

6. 5. The deep learning-based electrocardiogram data denoising system of claim 4, wherein the adversarial generative network comprises a generative network that generates synthetic electrocardiogram data using randomly generated variables as input, and a discriminative network that classifies whether the synthetic electrocardiogram data is similar to actual electrocardiogram data, and the generative network learns the eigenstyle of the electrocardiogram data.

7. The deep learning-based electrocardiogram data denoising system according to claim 5 or 6, characterized in that the generative adversarial network is converted into the form of the autoencoder and the autoencoder and the generative adversarial network are merged.

Images