WO2020141807A2

WO2020141807A2 - Method for predicting paroxysmal atrial fibrillation in normal sinus rhythm electrocardiogram state by using deep learning

Info

Publication number: WO2020141807A2
Application number: PCT/KR2019/018561
Authority: WO
Inventors: 이상철; 홍슬기; 백용수; 김대혁; 최원익
Original assignee: 인하대학교 산학협력단
Priority date: 2019-01-03
Filing date: 2019-12-27
Publication date: 2020-07-09
Also published as: WO2020141807A3; KR20200084561A; KR102156087B1

Abstract

Disclosed is a method for predicting paroxysmal atrial fibrillation in a normal sinus rhythm electrocardiogram state by using deep learning. A method for predicting paroxysmal atrial fibrillation according to an embodiment may include: a step for performing pre-processing in which electrical biosignals are converted into input data based on a diagnostic model; a step for training a diagnostic model, configured for the prediction of paroxysmal atrial fibrillation, with the pre-processed input data; and a step for evaluating the potential probability of paroxysmal atrial fibrillation in cases that include a group potentially classified as PAF (PAF normal cases) and a real normal group (Real normal cases), by using the training results learned through the diagnostic model.

Description

A method for predicting paroxysmal atrial fibrillation under normal sinus rhythm ECG using deep learning

The following description relates to a method for predicting paroxysmal atrial fibrillation in a normal rhythmic electrocardiogram state using deep learning.

Atrial fibrillation (AF) is the most common arrhythmia disease and is closely related to mortality and morbidity. In particular, it is a major heart disease that has a significant social and economic burden due to a significant increase in the risk of cerebral infarction and heart failure, which significantly decreases athletic performance and quality of life, and thus increases hospitalization costs. Atrial fibrillation can be classified into paroxysmal atrial fibrillation, persistent atrial fibrillation, long-lasting atrial fibrillation, and permanent atrial fibrillation, depending on the manifestation pattern, duration, and whether atrial fibrillation is voluntarily terminated. Atrial fibrillation is achieved by recording a typical pattern of atrial fibrillation on the electrocardiogram (no clearly distinguishable P-waves and irregular RR intervals). Over time, many patients develop persistent atrial fibrillation and, in some cases, persistence. It may change from paroxysmal to.

Atrial fibrillation is characterized by stroke or heart failure, regardless of pattern type. Therefore, it is very important to diagnose AF in the early stages, but there are some limitations. First, patients with paroxysmal atrial fibrillation have a significantly lower incidence of atrial fibrillation. According to the conventional method, it is possible to accurately judge when atrial fibrillation is present, but it is difficult to diagnose because symptoms sometimes appear at the beginning of the onset. In patients with paroxysmal atrial fibrillation, they are usually seen as normal electrocardiograms, making it almost impossible to distinguish between normal ECGs and normal methods of normal people. As the existing method is difficult to observe symptoms with a short-term test, a method such as an event record that records an ECG of 20-24 hours or monitors it for 1-2 weeks is designed to detect it in everyday life. However, this is a method dependent on device attachment for a long period of time. It is also a limitation of the existing approach to this problem that many studies showing good results focus on the discrimination of normal-ECG and abnormal-ECG.

In order to solve the problems presented above, it is necessary to have a different approach.

We present a new perspective to uncover the differences in ECG between a group potentially diagnosed as PAF (PAF-normal) and a real-normal group among signals diagnosed with normal electrocardiogram (normal-ECG).

To detect PAF patients from a steady state electrocardiogram (ECG) showing normal sinus rhythm, and to devise a screen based on RNN (recurrent neural network) to predict paroxysmal atrial fibrillation. It can provide a method and system.

The paroxysmal atrial fibrillation prediction method performed by the paroxysmal atrial fibrillation prediction system comprises: performing a pre-processing process of converting an electrical biosignal into input data based on a diagnostic model; Learning input data on which the pre-processing has been performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation; And determining a potential probability of paroxysmal atrial fibrillation according to a case including a group (PAF normal case) potentially classified as PAF and a real normal case of learning results learned through the diagnostic model. It can contain.

The step of determining the potential probability of paroxysmal atrial fibrillation may include performing a post-processing process of setting an optimal number of pulses detected as paroxysmal atrial fibrillation to diagnose final paroxysmal atrial fibrillation using a statistical analysis method. It can contain.

The step of performing a pre-processing process of converting the electrical biosignal into input data based on a diagnostic model includes an 8-by-5000 signal with baseline drift removed from the electrical biosignal and an R-peak index for processing at an RR interval. Information can be converted to input data.

The pre-processing process of converting the electrical biosignal into input data based on a diagnostic model may include: removing a baseline drift from the electrical biosignal using a Polynominal curve fitting method; Extracting an R-peak index using a Pan Tompkins QRS detector from the electrical signal from which the baseline drift has been removed; Cutting an electrical biosignal from which the baseline drift has been removed with an R-R interval based on the extracted R-peak index, and removing outliers according to the RRI length; And excluding a sudden change from the signal from which the abnormal point is removed as an abnormal signal in the vicinity of the QRS complex from learning.

The step of learning the input data on which the pre-processing is performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation, is a section of a specified pulse while increasing the section size in the direction including the P-wave section starting from the Q point in the electrocardiogram The method may include inputting a section corresponding to a length as an input signal of the diagnostic model.

The step of learning input data on which the pre-processing has been performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation includes constructing a diagnostic model for predicting paroxysmal atrial fibrillation based on RNN, and the RNN-based paroxysmal atrial fibrillation. The diagnostic model for prediction adds a bi-directional connection for considering the front-rear relationship in the forward and reverse direction of the time axis, and long-term memory (LSTM) and GRU to maintain a series of information in the short and long term. (Gated Recurrent Unit) can be used.

In the step of performing the learning, a plurality of codes are acquired by inputting a value obtained from the first Bi-LSTM layer to the second Bi-LSTM, and the obtained plurality of codes are input to the fully-connected layer to generate a PAF- It may include the step of performing the final classification by passing the binary classification of normal ECG and Real-normal ECG.

The paroxysmal atrial fibrillation prediction system includes: a pre-processing unit that performs a pre-processing process of converting an electrical biosignal into input data based on a diagnostic model; A learning unit for learning input data on which the pre-processing is performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation; And a judging unit for judging the potential probability of paroxysmal atrial fibrillation according to a case including a group (PAF normal case) and a real normal case (PAF normal case) potentially classified as a learning result learned through the diagnostic model. It can contain.

The potential probability of paroxysmal atrial fibrillation is estimated by constructing an RNN-based diagnostic model to detect PAF patients from a steady state ECG showing normal sinus rhythm, and to screen for pre-existing screening. Can be evaluated more accurately. In addition, as a result of optimizing a section that may be a key feature in each pulse, classification accuracy between two groups may be increased when RNN learning is performed including the vicinity of a P-wave.

Reducing the amount of measurement data required for diagnosis by performing classification work through pattern analysis of individual pulses, without dealing with data measured over a long period, assuming that there will be an organic causal relationship between signals in each section even within a single pulse I can do it.

1 is a block diagram illustrating a configuration of a paroxysmal atrial fibrillation prediction system according to an embodiment.

2 is a flowchart illustrating a method for predicting paroxysmal atrial fibrillation in a system for predicting paroxysmal atrial fibrillation according to an embodiment.

3 is a diagram illustrating a kernel size optimization method for learning in a paroxysmal atrial fibrillation prediction system according to an embodiment.

4 is a view for explaining the structure of a diagnostic model using Bi-directional LSRM in a paroxysmal atrial fibrillation prediction system according to an embodiment.

5 is a diagram illustrating an example of detection frequency for post-processing in a paroxysmal atrial fibrillation prediction system according to an embodiment.

6 is a view showing an example of a probability distribution function of PAF-normal and real-normal in a paroxysmal atrial fibrillation prediction system according to an embodiment.

7 is an example for explaining detecting an optimal section in a paroxysmal atrial fibrillation prediction system according to an embodiment.

8 and 9 are examples showing classification accuracy using an input signal in a paroxysmal atrial fibrillation prediction system according to an embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

1 is a block diagram illustrating a configuration of a paroxysmal atrial fibrillation prediction system according to an embodiment, and FIG. 2 is a flowchart illustrating a method for predicting paroxysmal atrial fibrillation in a paroxysmal atrial fibrillation prediction system according to an embodiment.

The processor of the paroxysmal atrial fibrillation prediction system 100 may include a pre-processing unit 110, a learning unit 120, and a determination unit 130. The components of the processor may be representations of different functions performed by the processor according to a control instruction provided by at least one program code. The processor and components of the processor may perform steps 210 to 230 included in the paroxysmal atrial fibrillation prediction method of FIG. 2. For example, the processor and the components of the processor may be implemented to execute instructions according to at least one program code and the code of the operating system included in the memory. Here, the at least one program code may correspond to the code of the program implemented to process the paroxysmal atrial fibrillation prediction method. The paroxysmal atrial fibrillation prediction method may not occur in the illustrated order, and some of the steps may be omitted or additional processes may be further included.

The processor may load the program code stored in the program file for the paroxysmal atrial fibrillation prediction method into memory. At this time, each of the pre-processing unit 110, the learning unit 120, and the determining unit 130 included in the processor and the processor executes the corresponding part of the program code loaded in the memory to execute the subsequent steps (210 to 230). It may be different functional representations of the processor for executing.

In step 210, the pre-processing unit 110 may perform a pre-processing process of converting an electrical biosignal into input data based on a diagnostic model. At this time, the electrical biosignal may be an electrocardiogram. The pre-processing unit 110 may convert 8-by-5000 signals from which baseline drift is removed from electrical bio signals and R-peak index information for processing into R-R intervals into input data. The pre-processing unit 110 may remove the baseline drift from the electrical biosignal using a polynominal curve fitting method. The pre-processing unit 110 may extract an R-peak index using a Pan Tompkins QRS detector from an electrical biosignal from which baseline drift has been removed. The pre-processing unit 110 may cut an electrical biosignal from which baseline drift has been removed based on the extracted R-peak index to an R-R interval and remove outliers according to the RRI length. The pre-processing unit 110 may exclude a rapid change, which is regarded as an abnormal signal in the vicinity of the QRS complex, from the signal from which the abnormal point is removed from learning.

In step 220, the learning unit 120 may train input data on which the pre-processing has been performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation. The learning unit 120 may increase the section size in the direction including the P-wave section, starting from the Q point on the electrocardiogram, and input the section corresponding to the section length of the specified pulse as an input signal of the diagnostic model. have. The learning unit 120 may configure a diagnostic model for predicting atrial fibrillation based on RNN. At this time, the diagnostic model for predicting paroxysmal atrial fibrillation based on RNN adds a bi-directional connection for considering the front-rear relationship in the forward and reverse directions of the time axis, and Long Short- to maintain a series of information in the short and long term. Term Memory (LSTM) and Gated Recurrent Unit (GRU) can be used.

In step 230, the determination unit 130 performs paroxysmal atrium according to a case including a group (PAF normal case) and a real normal case (PAF normal case) potentially classified as a learning result learned through a diagnostic model. The potential probability of fibrillation (PAF-normal) can be determined. The determination unit 130 may perform a post-processing process for setting the optimal number of pulses detected by paroxysmal atrial fibrillation to diagnose final paroxysmal atrial fibrillation using a statistical analysis method.

All data sets include electrical biosignals (eg, electrocardiograms) that are classified into normal groups. These data are labeled in two groups. All diagnostic (true label) and electrocardiogram data can be provided after verification by professional medical personnel. At this time, these data are only examples, and are not limited, and may be provided by various hospitals. For example, 547 cases of data obtained from a patient with a history of PAF diagnosis, and 2506 cases of data obtained from a normal person with no abnormalities in the heart.

All data files are in xml format and can be measured on GE MAC5500 machines. The ECG is originally measured on 12 leads, but according to the data storage method in the device, it contains data of 8 leads except for Lead III, avR, avL, and avF in xml. The four leads can be synthesized by simple arithmetic expressions, and approximating data with these operations is a common method. In the embodiment, only 8 measured lead signals, such as Lead I, II, V1, V2, V3, V4, V5, and V6, can be used. Data was obtained by measuring for 10 seconds simultaneously on each lead. For most signals, the sampling rate is 500 (samples/sec), with 5000 samples per unit of measurement. One xml file is one measurement case, and if Base64-encoded values are read from the <WaveFormData> tag, eight one-dimensional arrays can be obtained per xml file. There are multiple pulses within a 10-second signal, and the number of beats varies according to the heart rate of a person, so that about 10 pulses per person can be acquired. Now, in the course of the experiment, the 8-by-5000 processed data is processed as a bundle.

The paroxysmal atrial fibrillation prediction system may perform a pre-processing process that converts electrical biosignals into input data based on a diagnostic model. The paroxysmal atrial fibrillation prediction system can remove baseline drift from electrical biosignals and detect the position of the R-peak using a QRS detector. After that, the signal is cut by R-R Interval (RRI), and then the outlier can be removed according to the RRI length. Afterwards, we work to exclude the rapid change that is considered an abnormal signal in the vicinity of the QRS complex from learning.

The paroxysmal atrial fibrillation prediction system requires input data for a diagnostic model, 8-by-5000 signals with baseline drift removed, and R-peak index information for processing into R-R intervals. ECG baseline wander is a frequently generated noise and is caused by impedance, breathing, and movement. This noise appears slowly and generally within the signal measured for a predetermined period of time (eg, 10 seconds). In order to process data on a pulse-by-pulse basis, it is first necessary to remove drifts that appear throughout the electrical biosignal. The baseline drift can be removed by using the polynominal curve fitting method. At this time, the dimension parameter is 30, and as the baseline drift is effectively removed, detailed data can be set to the extent that it is not damaged.

In the pre-processing process of extracting an electrical biosignal (electrocardiogram) in pulse units, an R-peak index can be extracted using an R peak detection method, a Pan Tompkins QRS detector, and an R-R interval can be extracted. At this time, since 8 leads were measured at the same time, the R-peak index values of each lead are almost identical. In addition, the signal from the start of the measurement to the first R-peak and the signal from the last peak to the end occur frequently, and can be excluded from learning. Statistically, there was no significant difference in the RRI length between the real-normal and PAF-normal groups, but the heart rate varied in each measurement case.

Two outlier criteria can be established that are an obstacle to learning. A 200 sample, which is about half of the average R-R interval, is set as a threshold value, and all cases including RRIs below it can be regarded as outliers. This is when the heart rate is about 150 or higher or noise is eliminated. In this case, ECG is a standard value determined by determining that the environment is not a stable measurement. The upper limit of RRI was not taken into account, because even if the RRI is long, it is not significantly affected by the length of sections such as P and QRS complex of ECG.

In addition, since the QRS complex has a large variation and a very different size for each signal, it is a factor that should be excluded because it is almost irrelevant to the diagnosis, so the Q-R part can exclude 20 samples in the Q direction from the R-peak point. By performing this pre-processing process, 8-lead pulses extracted from cases excluding outliers can be used as the final data set.

Not all data processed in RRI units are used as input to the diagnostic model. AF, which occurs when the atrium is shaken by a microscopically abnormal electrical signal, is expressed as P-wave irregularity in the electrocardiogram, and QRS in the electrocardiogram indicates depolarization of the ventricle, so the most significant difference in the presence or absence of AF is the QRS complex. It is assumed that the signal is near the P-wave before appears. By reflecting this information, it is possible to designate an optimal sequence length while gradually increasing the section size in the direction including the P-wave section starting from the Q point. For example, referring to FIG. 1, it is possible to observe a trend of classification accuracy by performing an experiment by specifying an interval length while increasing 10 samples within a range of 20-180 samples. This is the inspection range that observes all sections except the Q-R interval 20 samples in the 200 samples, which are the RRI outlier thresholds. The section corresponding to the specific section length of the pulse is then a sequence of input signals of the diagnostic model.

The paroxysmal atrial fibrillation prediction system can use a data set of an input sequence to which a preprocessing process and an optimized section are applied for Bi-directional LSTM learning. By inputting the value obtained from the first Bi-LSTM layer to the second Bi-LSTM, a plurality of (eg, 20) codes can be obtained from a plurality of Bi-LSTM layers. By inputting this into the fully-connected layer, binary classification of PAF-normal ECG and real-normal ECG can be performed. At this time, binary classification may be performed in pulse units. Subsequently, the potential probability of PAF-normal can be evaluated on a case-by-case basis by using binary classified information in a post-processing process.

Specifically, the paroxysmal atrial fibrillation prediction system may construct a diagnostic model based on Recurrent Neural Network (RNN) in processing sequence data. At this time, the RNN is a structure that sequentially receives signals that are continuously generated based on time, and transmits compressed information at each time to be connected to operations in the next time-step. The electrical bio-signals are a series of signals measured in chronological order, and each section is a series of inputs and outputs due to the action of the atrium and ventricle, so it can be accessed from the viewpoint of this data. By closely observing the relationship in the time axis, it is directly related to the physical elements of the heart, and RNN not only captures the relationship in adjacent time, but also considers the input signal as a whole.

Additional considerations using RNN applied to the diagnostic model are as follows. The paroxysmal atrial fibrillation prediction system overlaps the structures composed of forward and backward states simultaneously in order to consider the flow of the time axis in both directions, and learns the first acquired code in both directions again to achieve a higher level feature. Models can be constructed to learn. For example, the forward and backward states of the first layer to which the input sequence is input are 50 respectively. The output generated in the first layer will eventually generate 100 hidden state information through the network. This code is learned as the input of the second layer, and also consists of 10 states per direction. As a result, the output of the bi-directional RNN is 20 codes, and information for determining the difference between the two groups is implied. The codes generated in this way go through the classification layer and perform the final classification.

For example, an experiment may be performed on a Long Short-Term Memory (LSTM) and a Gated Recurrent Unit (GRU) that uses different types of memory blocks stored between each time-step based on RNN. The LSTM can properly store or forget necessary information among the information of the sequence data, so that it is possible to judge and accumulate meaningful information in the classification of the two groups. The LSTM block consists of three gates: input, forget, and output, and calculates the compressed information up to the current time-step. The GRU has a simplified structure while taking advantage of the LSTM.

The paroxysmal atrial fibrillation prediction system shows a difference in the number of pulses classified as paroxysmal atrial fibrillation (PAF-normal) within the case, so for final classification, seizures are used to diagnose the final paroxysmal atrial fibrillation using statistical analysis methods. The optimal number of pulses detected by atrial fibrillation can be set. Referring to FIG. 5, for example, only three pulses out of a total of 12 pulses are detected as PAFs. The case consists of 12 pulses, of which 3 pulses can be detected as paroxysmal atrial fibrillation by a trained diagnostic model. In these cases, 25% of the pulses were observed with paroxysmal atrial fibrillation. If the statistically set detection threshold is 40%, these cases can be diagnosed with a real-normal ECG with no abnormalities in the heart.

As an example, an experiment may be conducted with 5-fold cross validation. One set can be excluded for evaluation, and the other 4 sets can be used to train the diagnostic model. Pre-processing and diagnostic models can be applied to learn pulse-by-pulse. When learning is complete, evaluation can be performed for each case. At this time, each diagnosis and evaluation is performed independently under the same conditions, and since the cases used for learning and the cases to be evaluated are not mixed, an environment in which a completely new electrical biosignal is diagnosed later is created.

Through this process, PAF-normal and real-normal probability distribution functions can be derived for all data as shown in FIG. 6. The horizontal axis of the probability distribution function is the rate of occurrence of PAF-normal when using a diagnostic model ([Number of pulses classified as PAF-normal]/[Number of pulses per case]). The vertical axis is the probability distribution for the frequency of occurrence of each ratio. The graph in dark gray is real-normal, and the light gray is the probability distribution function for PAF-normal. In other words, in the case of real normal, it can be seen that the ratio of pulses detected as PAF-normal in the case is the highest (approximately 0.65) at 0 (not even classified as PAF-normal) and gradually decreases to stabilize at 0.4. In contrast, in the case of PAF-normal, it can be seen that the ratio of pulses detected as PAF-normal in the corresponding case is 1.0 (all pulses are classified as PAF-normal) with a probability of about 0.2. Using this probability distribution function, a decision boundary having a minimum error can be set.

For example, in order to balance learning data, PAF-normal and real-normal can be used by randomly extracting 547 cases and 1000 cases, respectively. The optimal section size was set to 120 samples and can be set from 0.28 seconds to 0.04 seconds of R-peak. Rather than using the original signal, it is possible to obtain a more stable result as well as performance by using a first-order differential value of a continuous signal.

Referring to FIG. 7(a), the length of the signal used for learning starts from R-peak and increases from 20 to 180 in the Q direction, and is the result of the experiment. FIG. 7(b) shows the hyper parameters and learning conditions of the diagnostic model. Information about. The signal to which the first-order differential was applied to the pre-processed signal was used as the input of the diagnostic model. The horizontal axis of the graph is the length of the input section, and the vertical axis is the accuracy of the result of binary classification whether each case is PAF-normal ECG or real-normal ECG when all data is evaluated once with 5-fold cross validation. It can be seen that the accuracy starts to increase when approximately 40 samples are used, and when it is 100 or more, diagnostic performance converges. In terms of time, 60-100 samples are about 0.12-0.2 seconds. Since the starting point is the Q point, if you look at this time on the ECG, it is an interval that contains approximately the P-R interval. When approximately 120 samples or more were used as the input of the diagnostic model, the difference in performance was observed to appear within the experimental error range. Therefore, in evaluating by pulse-by-pulse, it is reasonable to check the 0.24 second interval by using approximately 120 samples as input.

After pre-processing and section size optimization, performance may vary depending on whether differential is applied to the signal. When using the 1-based differential of continuous signals, the overall accuracy was high, and the accuracy, overall accuracy, and threshold values for each group, which were confirmed by experiments several times under the same conditions, were also more stable. Table 1 shows the results obtained through repeated experiments in which the learning was conducted in a diagnostic model under the same conditions, but only differentiation was performed. This process can be performed multiple times because the performance of the classifier may vary with each learning. Table 1 shows the parameters used for classifier learning.

FIG. 8 is a classification result when the original input signal is used for learning, and FIG. 9 is a classification accuracy when the 1-system differential value is used. Overall, an improvement in classification accuracy of about 3-5% can be observed.

The device described above may be implemented with hardware components, software components, and/or combinations of hardware components and software components. For example, the devices and components described in the embodiments may include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors (micro signal processors), microcomputers, field programmable gate arrays (FPGAs). , A programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose computers or special purpose computers. The processing device may run an operating system (OS) and one or more software applications running on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, a processing device may be described as one being used, but a person having ordinary skill in the art, the processing device may include a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that may include. For example, the processing device may include a plurality of processors or a processor and a controller. In addition, other processing configurations, such as parallel processors, are possible.

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and/or data may be interpreted by a processing device, or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. Can be embodied in The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and usable by those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floptical disks. -Hardware devices specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language code that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler.

As described above, although the embodiments have been described by a limited embodiment and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims

In the paroxysmal atrial fibrillation prediction method performed by the paroxysmal atrial fibrillation prediction system,

Performing a pre-processing process of converting the electrical bio-signals into input data based on a diagnostic model;

Learning input data on which the pre-processing has been performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation; And

The potential probability of paroxysmal atrial fibrillation (PAF-normal) according to a case including a group (PAF normal case) and a real normal case (PAF normal case) potentially classified as a learning result learned through the diagnostic model Judging steps

Paroxysmal atrial fibrillation prediction method comprising a.
According to claim 1,

Determining the potential probability of paroxysmal atrial fibrillation,

Performing a post-processing process to set the optimal number of pulses detected by paroxysmal atrial fibrillation to diagnose final paroxysmal atrial fibrillation using a statistical analysis method.

Paroxysmal atrial fibrillation prediction method comprising a.
According to claim 1,

The step of performing a pre-processing process of converting the electrical bio-signals into input data based on a diagnostic model may include:

Converting the 8-by-5000 signal from which baseline drift is removed from the electrical bio-signal and R-peak index information for processing into an R-R interval into input data.

Paroxysmal atrial fibrillation prediction method comprising a.
According to claim 3,

The step of performing a pre-processing process of converting the electrical bio-signals into input data based on a diagnostic model may include:

Removing baseline drift from the electrical biosignal using a Polynominal curve fitting method;

Extracting an R-peak index using a Pan Tompkins QRS detector from the electrical signal from which the baseline drift has been removed;

Cutting an electrical biosignal from which the baseline drift has been removed with an R-R interval based on the extracted R-peak index, and removing outliers according to the RRI length; And

The step of excluding the rapid change, which is regarded as an abnormal signal in the vicinity of the QRS complex, from the signal from which the abnormality is removed from the learning.

Paroxysmal atrial fibrillation prediction method comprising a.
According to claim 1,

The step of learning the input data on which the pre-processing is performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation,

Inputting an input sequence as an input signal of the diagnostic model into a section corresponding to a section length of a specified pulse while increasing the section size in a direction including the P-wave section starting from the Q point in the electrocardiogram.

Paroxysmal atrial fibrillation prediction method comprising a.
The method of claim 5,

The step of learning the input data on which the pre-processing is performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation,

And constructing a diagnostic model for predicting atrial fibrillation based on RNN,

The RNN-based diagnostic model for predicting paroxysmal atrial fibrillation,

A bi-directional connection is added to take account of the front-rear relationship in the forward and reverse direction of the time axis, and Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are used to maintain a series of information in the short and long term. doing

Seizure atrial fibrillation prediction method, characterized in that.
The method of claim 6,

The step of performing the learning,

As the values obtained from the first Bi-LSTM layer are input to the second Bi-LSTM, a plurality of codes are acquired, and the obtained plurality of codes are input to the fully-connected layer to generate the PAF-normal ECG and Real-normal ECG. Step to perform final classification through binary classification

Paroxysmal atrial fibrillation prediction method comprising a.
In the paroxysmal atrial fibrillation prediction system,

A pre-processing unit to perform a pre-processing process for converting electrical bio signals to input data based on a diagnostic model;

A learning unit for learning input data on which the pre-processing is performed in a diagnostic model configured for predicting paroxysmal atrial fibrillation; And

A judging unit that determines the potential probability of paroxysmal atrial fibrillation according to a case including a group (PAF normal case) potentially classified as PAF and a real normal case as a result of learning learned through the diagnostic model

Paroxysmal atrial fibrillation prediction system comprising a.