KR102684410B1 - Apparatus for Generating Arrhythmia by using Diffusion Model - Google Patents

Abstract

Disclosed is an arrhythmia generating device using a diffusion model.
In this embodiment, the input signal is divided into a substitution target section and a conditional observation section according to a plurality of symptoms, the substitution target section requiring training is masked, and noise is generated in the masking portion, and then the remaining sections are divided into the conditional observation section and the conditional observation section. In the combined state, the process of removing noise and performing additional noise prediction is learned to create an arrhythmia generation model. By entering the acute ECG disease to be created along with individual-specific input data into the arrhythmia generation model, the output value for the acute ECG disease is generated. We provide an arrhythmia generating device using a diffusion model that allows the acquisition of arrhythmia.

Description

Arrhythmia generating device using diffusion model {Apparatus for Generating Arrhythmia by using Diffusion Model}

One embodiment of the present invention relates to an arrhythmia generating device using a diffusion model.

The content described below simply provides background information related to this embodiment and does not constitute prior art.

An electrocardiogram reading system has been developed to help medical staff analyze electrocardiograms. Conventional electrocardiogram reading systems detect the R, P, and T peaks of the waveform, and detect and classify arrhythmia based on rules.

A conventional electrocardiogram reading system receives the patient's entire electrocardiogram signal data, analyzes it, and outputs the results. Deep learning technology has recently been widely studied as an electrocardiogram reading algorithm because it has higher accuracy than existing methods.

Only qualified medical staff can determine arrhythmia using an electrocardiogram, but the reality is that there is a shortage of manpower compared to demand. When reading an ECG, it takes a lot of time because the ECG signal must be read from various perspectives, such as calculating the time difference between the shapes and sections of the P, QRS, and T waveforms, and analyzing the ECG rhythm. Medical staff must observe the patient's electrocardiogram in real time to keep an eye on the patient's condition, but continuous monitoring is difficult due to a lack of manpower. Since electrocardiogram analysis is directly related to the patient's life, it must be accurate and operate quickly in the event of an emergency.

Conventional electrocardiogram analysis sometimes uses the end and start points of the P, QRS, and T waveform sections, but the conventional technology only searches for peaks, which reduces its utility. When detecting and classifying arrhythmia, rule-based algorithm design is less accurate due to the diversity of waveforms, and when arrhythmia is added, a new rule-based algorithm must be designed.

The conventional arrhythmia generation algorithm was proposed and developed based on GAN (Generative Adversarial Network).

The GAN-based arrhythmia generation algorithm can extract features that are very similar to the input. When the GAN-based arrhythmia generation algorithm outputs only data similar to the input data, it can accurately classify/predict only the input pattern in the secondary classification/prediction model. However, in the GAN-based arrhythmia generation algorithm, it is important to learn various conditions because arrhythmia accompanied by various conditions may appear even within the detailed classification of arrhythmia.

The GAN-based arrhythmia generation algorithm has a problem in that the stability of the model is low, and detailed indicators within the model, such as hyperparameters, must be carefully adjusted to obtain the desired output.
Prior art document: Korea Patent Publication No. 10-2022-0127680 (published on September 20, 2022)

In this embodiment, the input signal is divided into a substitution target section and a conditional observation section according to a plurality of symptoms, the substitution target section requiring training is masked, and noise is generated in the masking portion, and then the remaining sections are divided into the conditional observation section and the conditional observation section. In the combined state, the process of removing noise and performing additional noise prediction is learned to create an arrhythmia generation model. By entering the acute ECG disease to be created along with individual-specific input data into the arrhythmia generation model, the output value for the acute ECG disease is generated. The purpose is to provide an arrhythmia generating device using a diffusion model that can be obtained.

According to one aspect of the present embodiment, a data input unit that receives electrocardiogram data for each specific time unit among preset windows; a data dividing unit that divides the electrocardiogram data into an imputation target section and a conditional observation section according to a plurality of symptoms; a masking unit that generates masking data that masks the substitution target section; Diffusion, which extracts diffusion embedding features based on the results of learning the arrhythmia pattern while performing a diffusion process that adds noise to the substitution target section and a reverse process that removes the noise and restores the original. Embedding part; An arrhythmia data generator that generates arrhythmia data including an arrhythmia pattern corresponding to newly input input data based on the diffusion embedding feature and the masking data.

As described above, according to this embodiment, the input signal is divided into a substitution target section and a conditional observation section according to a plurality of symptoms, the substitution target section requiring training is masked, and noise is generated in the masking portion. Create an arrhythmia generation model by learning the process of removing noise and performing additional noise prediction while the remaining section is combined with the conditional observation section, and input the acute electrocardiogram disease you want to create along with individually specialized input data into the arrhythmia generation model. It is effective in obtaining output values for acute electrocardiogram disease.

According to this embodiment, it is possible to reduce the time and cost required to collect data for learning about acute electrocardiogram disease, and has the effect of reducing the cost invested in clinical research for data collection.

Figure 1 is a diagram showing bio-signal data processing in the bio-signal processing field according to this embodiment.
Figure 2 is a diagram showing characteristic indicators of P, Q, R, S, and T waves (P wave, QRS complex, T wave) and electrocardiogram according to this embodiment.
Figure 3 is a diagram showing an electrocardiogram reading device using an online machine learning algorithm according to this embodiment.
Figure 4 is a diagram showing the diffusion embedding process for context extraction before arrhythmia generation according to this embodiment.
Figure 5 is a diagram showing the arrhythmia generation process using ECG signal data and diffusion embedding according to this embodiment.
Figure 6 is a flowchart for explaining the arrhythmia generation method using the diffusion model according to this embodiment.

Hereinafter, this embodiment will be described in detail with reference to the attached drawings.

Figure 1 is a diagram showing bio-signal data processing in the bio-signal processing field according to this embodiment.

The arrhythmia generator 300 is a test that records the electrical signal generated when the heart beats as a current waveform. Generate arrhythmia data corresponding to acute or rare diseases needed for learning purposes.

The arrhythmia generator 300 understands the characteristics of the electrocardiogram waveform and changes over time, and based on various characteristic information of the electrocardiogram, atrial fibrillation (AFIB), atrial flutter (AFL), and ventricular tachycardia It produces various arrhythmias such as tachycardia (VT) and atrioventricular block (AV block).

Figure 2 is a diagram showing characteristic indicators of P, Q, R, S, and T waves (P wave, QRS complex, T wave) and electrocardiogram according to this embodiment.

The electrocardiogram waveform appears as a series of beats, and beats can be broadly divided into normal beat (N), supraventricular beat (S), and ventricular beat (V). One beat of the electrocardiogram waveform basically includes the P wave, QRS wave, and T wave.

The arrhythmia generator 300 detects the P wave, Q wave, R wave, S wave, and T wave included in the ECG waveform, detects the characteristics of the ECG waveform and changes over time, and generates information based on various characteristic information of the ECG. Generates various arrhythmias suitable for different conditions.

The arrhythmia generator 300 classifies heart beats into Normal Beat (N), Supraventricular Beat (S), and Ventricular Beat (V). When reading an electrocardiogram waveform, the arrhythmia generator 300 can detect an abnormal state based on localization and classification information.

The arrhythmia generator 300 classifies various arrhythmias such as atrial fibrillation (AFIB), atrial flutter (AFL), ventricular tachycardia (VT), and atrioventricular block (AV block). . When generating an electrocardiogram rhythm, the arrhythmia generator 300 can detect an abnormal state based on localization and classification feature information.

Figure 3 is a diagram showing an electrocardiogram reading device according to this embodiment.

The arrhythmia generator 300 according to this embodiment includes a signal generator 310, a preprocessor 320, a data input unit 330, a data separator 340, a masking unit 350, and a diffusion embedding unit 360. , includes an arrhythmia data generation unit 370. Components included in the arrhythmia generating device 300 are not necessarily limited thereto.

Each component included in the arrhythmia generating device 300 is connected to a communication path connecting software modules or hardware modules within the device and can operate organically with each other. These components communicate using one or more communication buses or signal lines.

Each component of the arrhythmia generating device 300 shown in FIG. 3 refers to a unit that processes at least one function or operation, and may be implemented as a software module, a hardware module, or a combination of software and hardware.

The signal generator 310 generates an ECG signal in units of 10 seconds to be used as an input to the framework.

The preprocessor 320 generates ECG data by preprocessing the ECG signal in 10-second increments to correct baseline wandering. The preprocessing unit 320 inputs preprocessed data into the signal input unit 330 every 10 seconds.

The data input unit 330 receives preprocessed ECG data every 10 seconds from the preprocessor 320. The data input unit 330 receives electrocardiogram data for each specific time unit among preset windows.

The data classification unit 340 divides the ECG data into an imputation target section and a conditional observation section according to multiple symptoms. The data classification unit 340 recognizes the imputation target section as arrhythmia data included in training. The data classification unit 340 recognizes the Conditional Observations section as observation data not included in training. The data classification unit 340 divides the ECG data by dividing it into a substitution target section and a conditional observation section.

The masking unit 350 generates masking data that masks the substitution target section.

The diffusion embedding unit 360 performs diffusion embedding based on the results of learning the arrhythmia pattern while performing a diffusion process that adds noise to the substitution target section and a reverse process that removes noise and restores it to the original. Extract features.

The diffusion embedding unit 360 performs a diffusion process to generate noise targets (Noisy Targets) by generating Gaussian noise only in the substitution target section.

The diffusion embedding unit 360 performs a reverse process to restore the original by removing Gaussian noise using a denoising function while combining a noise target and conditional observation sections with different noise levels. do.

The arrhythmia data generator 370 generates arrhythmia data including an arrhythmia form corresponding to an acute disease or rare disease required for learning corresponding to newly input input data based on diffusion embedding features and masking data.

The arrhythmia data generator 370 compares the diffusion embedding feature and the masking data to confirm the estimate added noise resulting from the addition and removal of the Gaussian noise included in the diffusion embedding feature based on the masking data. Based on the prediction, arrhythmia data containing the arrhythmia type corresponding to the input data is generated.

Figure 4 is a diagram showing the diffusion embedding process for context extraction before arrhythmia generation according to this embodiment.

An electrocardiogram interprets the electrical activity of the heart and is recorded by electrodes attached to the skin and equipment outside the body. An electrocardiogram is used to measure the rate and consistency of heartbeat, and is used for the purpose of diagnosis or research into abnormal heart activity.

Arrhythmia means that the heart beats irregularly, either too fast, too slow, or the pulse is irregular. Various arrhythmias may occur paroxysmal or persistent.

The occurrence of arrhythmia occurs less often than the occurrence of a normal electrocardiogram, and this phenomenon is one of the reasons that makes accurate inference about the occurrence of arrhythmia difficult due to imbalance between data when developing arrhythmia classification and prediction algorithms.

General arrhythmia generation algorithms require the entire arrhythmia pattern to be included in one input, but in actual arrhythmia, various arrhythmias may exist in one input.

The arrhythmia generation model based on GAN (Generative Adversarial Network) has the disadvantage that there is instability in the training process and the model lacks diversity when generating images, so only consistent patterns that are very similar to the input data can be extracted.

Because various arrhythmias may appear simultaneously in a specific section, an algorithm that takes into account arrhythmia patterns is needed.

In addition to the stability of the arrhythmia generation model, arrhythmia data with diversity must be generated to ensure generalization ability when used in a real environment after training a classification and prediction model using the generated data.

Therefore, the arrhythmia generating device 300 according to this embodiment receives as input an ECG signal containing a mixture of various arrhythmia and normal ECG signals, and secures the diversity of arrhythmia while simultaneously generating multiple output data.

The arrhythmia generator 300 can secure the diversity of arrhythmia while producing multiple output data from one input, and can resolve data imbalance when training secondary arrhythmia detection/classification and prediction models, and data collection. Time and cost can be reduced.

The arrhythmia generator 300 builds an arrhythmia data generation model by combining a diffusion-based model that can extract features representing arrhythmias from an input signal (an electrocardiogram signal set to 10 seconds) and the features extracted from the diffusion-based model. do.

The arrhythmia generator 300 learns the pattern of a continuous time series signal using a diffusion-based model and an arrhythmia data generation model to generate output data containing multiple arrhythmias.

A diffusion model is one of the types of generation models. It refers to a model that learns change information in the data while generating noise, and then learns the change information in the data again while removing the noise.

The diffusion model learns noise generation patterns or noise removal patterns and predicts change information when new data is input based on the noise generation patterns or noise removal patterns.

The diffusion model creates arrhythmia in a single-guided electrocardiogram based on an algorithm that learns noise generation patterns or noise removal patterns.

The diffusion model predicts multiple outputs with a single input when using a sliding window (data cut by 10 or 30 seconds out of 100 seconds of data). Diffusion models predict multiple outputs based on multiple inputs.

To learn a diffusion model, a lot of learning data is required. However, there are limitations in acquiring learning data. Although it is possible to obtain large amounts of data for common diseases related to stroke, such as atrial fibrillation, there are difficulties in obtaining training data related to AV block data and acute ventricular fibrillation. Accordingly, when the arrhythmia generator 300 acquires learning data, it generates learning data including preceding signals before the occurrence of an acute electrocardiographic disorder.

A single input means, for example, one data input of 10 seconds. Multiple input means, for example, that 10 seconds of data is input in multiple numbers such as 0 to 10 seconds, 10 to 20 seconds, and 20 to 30 seconds.

The arrhythmia generator 300 generates uniform noise from the entire ECG data using a diffusion model, and then learns by removing the uniform noise. Rather than generating uniform noise throughout, the ECG data is generated according to conditions. Pattern changes in the ECG waveform are learned while generating noise only in a specific section and not generating noise in the remaining sections.

The arrhythmia generator 300 generates a generation model (arrhythmia generation model) learned by changing the pattern of the ECG waveform due to noise generation according to conditions. The arrhythmia generator 300 can output normal data and data with atrial fibrillation based on a generation model (arrhythmia generation model). The arrhythmia generator 300 can generate new training data under various conditions based on the output.

The arrhythmia generator 300 divides input data (eg, 10 seconds of data) into an imputation target section and a conditional observation section based on multiple symptoms.

Here, the imputation target refers to the feature to be created that can be the main symptom. The section subject to college entrance refers to the section included in training. The conditional observation section refers to the section that is not included in training.

The arrhythmia generator 300 generates noise targets (Noisy Targets) by generating Gaussian noise only in the substitution target section. The arrhythmia generator 300 performs estimate added noise using a denoising function by combining a noise target and conditional observation sections with different noise levels.

The arrhythmia generator 300 divides input data (e.g., 10-second data) by cutting out specific sections according to an imputation target section and conditional observations according to multiple symptoms.

The arrhythmia generator 300 predicts additional noise by removing noise using a noise removal function by combining conditional observation sections with different noise levels with a noise target that generates Gaussian noise only in the substitution target section, and predicts additional noise. Based on the predicted pattern, the features of the corresponding section are learned.

The arrhythmia generator 300 predicts and generates what additional noise pattern specific data (main symptom) has within the condition and learns the additional noise pattern.

The masking unit 350 performs masking. The masking unit 350 performs masking by filling the substitution target section with zeros. The masking unit 350 allows the user to check which part Gaussian noise is added based on the masking section when new data is generated by adding noise to the substitution target section.

The arrhythmia generator 300 needs learning data to learn about emergency electrocardiogram diseases, but since the learning data is not sufficient, it artificially generates data to be learned.

The arrhythmia generator 300 divides one input data (e.g., a 10-second electrocardiogram waveform) into an imputation target section and a conditional observation section according to a plurality of symptoms.

Ensure that the sum of the imputation target section and the Conditional Observations section according to multiple symptoms is equal to the input data.

The arrhythmia generator 300 generates noise targets by generating Gaussian noise only in the substitution target section, and has a noise removal function ( Perform additional noise prediction (Estimate Added Noise) using the Denoising Function.

The arrhythmia generator 300 performs additional noise prediction and then restores the noise target to the original data. The arrhythmia generator 300 completes the diffusion process by restoring the noise target to the original data.

The arrhythmia generator 300 cuts and divides specific sections into input data (e.g., 10-second data) into an imputation target section and a conditional observation section according to multiple symptoms.

The arrhythmia generator 300 masks the target section required for training. The arrhythmia generator 300 cuts the substitution target section and then generates noise in the masking portion corresponding to the substitution target section. The arrhythmia generator 300 combines the conditional observation section, which is the remaining part of the input data, with the masking part where noise occurs.

The arrhythmia generator 300 combines the masking part where the noise occurs and the conditional observation section, removes the noise using a noise removal function, and performs additional noise prediction (reverse).

The arrhythmia generator 300 generates learning data by going through a diffusion process and a reverse process. The arrhythmia generator 300 generates diffusion-step embedding, which is a characteristic of the ECG disease to be created based on learning data. The arrhythmia generator 300 inputs separate input data (normal rhythm) and diffusion-step embedding, which is a characteristic of the ECG disease to be created, into a generation model (arrhythmia generation model) to generate an output value for an acute ECG disease. get

In order to determine the target section for substitution, the arrhythmia generator 300 determines the target section (arrhythmia or normal rhythm) among the input data based on data that has already been read. The arrhythmia generator 300 initially distinguishes between a normal beat and an arrhythmia section based on heart rate observation data that has already been read. The arrhythmia generator 300 prepares masking data to add noise to the arrhythmia section. The arrhythmia generator 300 adds Gaussian noise to the arrhythmia section and then performs learning processing.

The arrhythmia generator 300 can specifically distinguish arrhythmia in the section to which Gaussian noise is added after learning is completed. Because the probability of ECG disease occurring is low, there is relatively little ECG disease data and a lot of normal rhythm data. The arrhythmia generator 300 inputs normal rhythm data with different waveforms for each person as input data, and generates diffusion-step embedding, which is a characteristic of the ECG disease to be created, for arrhythmia data according to a specific ECG disease. This allows classification of various types of arrhythmia.

In other words, the arrhythmia generator 300 can generate arrhythmia data that reflects the characteristics of the ECG disease to be created in the normal rhythm data considering the characteristics of the subjects since the waveform of normal rhythm data is also different for each person.

Figure 5 is a diagram showing the arrhythmia generation process using ECG signal data and diffusion embedding according to this embodiment.

The data classification unit 340 divides the input data into observation data not included in training (conditional observation section) and arrhythmia data included in training (substitution target section).

The masking unit 350 generates masking data that masks arrhythmia data (substitution target section) included in training.

The diffusion embedding unit 360 is a diffusion process and performs a process of adding Gaussian noise to the arrhythmia data (substitution target section) included in training.

The diffusion embedding unit 360 is a reverse process and performs a process of removing Gaussian noise from the arrhythmia data (substitution target section) included in training and restoring the original signal.

The arrhythmia data generation unit 370 receives diffusion embedding and ECG data generated by learning from the diffusion embedding unit 360 through a diffusion process and a reverse process.

The arrhythmia data generator 370 generates arrhythmia data by referring to the context information of diffusion embedding.

Figure 6 is a flowchart for explaining the arrhythmia generation method using the diffusion model according to this embodiment.

The signal generator 310 generates an electrocardiogram signal in units of 10 seconds to be used as an input to the framework (S610).

The preprocessor 320 generates preprocessed data by performing preprocessing to correct baseline wandering of the ECG signal in units of 10 seconds (S620).

The preprocessing unit 320 inputs preprocessing data in units of 10 seconds to the data input unit 330 (S630).

The data separator 340 receives preprocessed data in units of 10 seconds from the data input unit 330.

The data classification unit 340 distinguishes between the arrhythmia data portion to be included in training and the portion not included among the 10-second preprocessing data (S640).

The masking unit 350 generates masking data in which the arrhythmia data (substitution target section) included in training is masked with 0 and 1 (S650).

The diffusion embedding unit 360 receives the arrhythmia data portion to be included in training and learns the arrhythmia pattern (S660).

After completing learning, the diffusion embedding unit 360 extracts diffusion embedding features and inputs them together with the masking data to the arrhythmia data generating unit 370 (S670).

The arrhythmia data generation unit 370 completes learning from the diffusion embedding unit 360 and outputs 10 seconds of electrocardiogram data including the arrhythmia pattern (S680).

In Figure 6, steps S610 to S680 are described as being sequentially executed, but the process is not necessarily limited thereto. In other words, the steps shown in FIG. 6 may be applied by modifying them or executing one or more steps in parallel, so FIG. 6 is not limited to a time-series order.

As described above, the arrhythmia generation method using the diffusion model according to the present embodiment shown in FIG. 6 may be implemented as a program and recorded on a computer-readable recording medium. A computer-readable recording medium on which a program for implementing the arrhythmia generation method using the diffusion model according to this embodiment is recorded includes all types of recording devices that store data that can be read by a computer system.

The above description is merely an illustrative explanation of the technical idea of the present embodiment, and those skilled in the art will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are not intended to limit the technical idea of the present embodiment, but rather to explain it, and the scope of the technical idea of the present embodiment is not limited by these examples. The scope of protection of this embodiment should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be interpreted as being included in the scope of rights of this embodiment.

300: Arrhythmia generating device
310: signal generator
320: Preprocessing unit
330: data input unit
340: Data separator
350: Masking unit
360: Diffusion embedding unit
370: Arrhythmia data generation unit

Claims (6)

A data input unit that receives electrocardiogram data for specific time units among preset windows;
a data dividing unit that divides the electrocardiogram data into an imputation target section and a conditional observation section according to a plurality of symptoms;
a masking unit that generates masking data that masks the substitution target section;
Diffusion, which extracts diffusion embedding features based on the results of learning the arrhythmia pattern while performing a diffusion process that adds noise to the substitution target section and a reverse process that removes the noise and restores the original. Embedding part; and
An arrhythmia data generator that generates arrhythmia data including an arrhythmia type corresponding to newly input input data based on the diffusion embedding feature and the masking data,
The diffusion embedding unit
Performing the diffusion process to generate noise targets (Noisy Targets) by generating Gaussian noise only in the substitution target section,
Performing the reverse process to restore the original by removing the Gaussian noise using a denoising function while combining the noise target and the conditional observation section with different noise levels,
The arrhythmia data generator
Compare the diffusion embedding feature and the masking data to confirm an estimate of added noise based on the addition and removal of the Gaussian noise included in the diffusion embedding feature based on the masking data, and estimate the additional noise. An arrhythmia generating device, characterized in that it generates the arrhythmia data including an arrhythmia type corresponding to the input data based on the input data. According to paragraph 1,
The data division unit,
An arrhythmia generating device characterized in that it recognizes the imputation target section as arrhythmia data included in training. According to paragraph 1,
The data division unit,
An arrhythmia generator characterized in that it recognizes the Conditional Observations section as observation data not included in training.

delete delete delete