KR102458657B1 - Artificial Intelligence Based ECG Analysis and Application for Automated External Defibrillator - Google Patents

Abstract

The present invention relates to an electrocardiogram analysis and application method of a complex ECG feature extraction method for an AI-based automatic defibrillator. In this technology, only those patients who need an electric shock are given an electric shock, and the reliability of an automatic defibrillator is improved through a highly accurate ECG extraction method. That is, the present invention provides an electrocardiogram analysis and application method for an AI-based automatic defibrillator of a system having a computer device for machine learning consisting of an ECG data storage, a first ECG feature extraction unit, and an ECG machine learning unit, and an automatic defibrillator connected thereto and consisting of an electrocardiogram measurement unit, a second ECG feature extraction unit, and an ECG AI engine.

Description

Electrocardiogram analysis and application method of complex ECG feature extraction method for artificial intelligence-based automatic defibrillator {Artificial Intelligence Based ECG Analysis and Application for Automated External Defibrillator}

The present invention relates to an electrocardiogram analysis and application method of a complex ECG feature extraction method for an AI-based automatic defibrillator. Shock) and non-shock (No-Shock) are accurately judged, so that only the patients who need an electric shock can receive an electric shock, and the high-accuracy ECG extraction method is used to improve the reliability of the AED. is about

In general, an automatic defibrillator (or Automated External Defibrillator, AED) is a device that automatically removes cardiac fibrillation, which is essential for vital activity as blood flow to each tissue is interrupted due to arrhythmia patients with irregular heartbeats or cardiac arrest. It is a device that restores normal heartbeat by giving an electric shock from the outside when sudden cardiac arrest (SCA: Sudden Cardiac Arrest), a phenomenon in which the supply of oxygen is interrupted, occurs.

In the process of using the automatic defibrillator, the patient's electrocardiogram (ECG) is measured through a pad attached to the patient's body, and the device decides whether to select Shock or No-Shock for the patient. Thus, an electric shock is made.

Therefore, accurate judgment is required because what is judged by the AED determines whether or not an electric shock is present. There was a problem with losing.

In addition, since the AED uses a microcontroller unit (MCU) with lower performance than a high-performance central processing unit (CPU) or application processor (AP), it is necessary to calculate the An algorithm with a low load should be used.

According to the determination accuracy of the AED, a sensitivity related to a rhythm requiring a shock may be calculated, and a specificity related to a rhythm that does not require a shock may be calculated. As the algorithm having high sensitivity and specificity has high determination accuracy, specificity related to a rhythm that does not require a shock is more important than sensitivity.

The ECG feature extraction algorithm for cardiac rhythm analysis used in existing AEDs is as follows.

- TCSC: Threshold Crossing Sample Count Algorithm

- VFLEAK: VF Filter Leakage Algorithm

- PSR: Phase Space Reconstruction Algorithm

- HILB: Hilbert Transform Algorithm

- bWT/bCP/bW: QRS complex identification algorithm

References of the ECG feature extraction algorithm and parameters have been known through the following publications.

- TCSC: "A simple time domain algorithm for the detection of ventricular fibrillation in electrocardiogram", Arafat M, Chowdhury A and Hasan M, Signal, Image and Video Processing, 5: 1-10, 2011.

- VFLEAKL: (1) "Computer detection of ventricular fibrillation", S. Kuo, D. Dillman, Computers in Cardiology, 1978, 2747-2750. (2) "Reliability of old and new ventricular fibrillation detection algorithms for automated external defibrillators", A. Amann, R. Tratning, and K. Unterkofler, Biomed Eng Online, 4(60), 2005.

- PSR/HILB: (1) "Detecting ventricular fibrillation by time-delay methods", A. Amann, R. Tratning, and K. Unterkofler, IEEE Trans Biomed Eng, 54(1): 174-77, 2007. (2 ) "A new ventricular fibrillation algorithm for automated external defibrillators", A. Amann, R. Tratning, and K. Unterkofler, Computers in Cardiology, 559-562, 2005.

- bWT/bCP/bW: Irusta U, Ruiz J,Aramendi E, Ruiz de Gauna S, Ayala U, Alonso E A high-temporal resolutionalgorithm to discriminate shockable from nonshockable rhythms in adults and children Resuscitation 2012 Sep;83(9):1090 -7

However, the above algorithm has each problem, and the problems will be described as follows.

- The Threshold Crossing Sample Count (TCSC) algorithm satisfies the sensitivity and specificity required by the AED standard, but in the case of a heart rhythm with an ECG signal amplitude of 0.1mV, shock Although this is not necessary, it has a disadvantage in that the false positive (FP), which is incorrectly judged as a shock heart rhythm, is high.

- The VF filter (VF Filter Leakage) algorithm has the advantage of high specificity, but has a disadvantage in that the false negative (FN) for ventricular tachycardia (VT) is high.

- The Phase Space Reconstruction (PSR) algorithm and the Hilbert Transform algorithm have high sensitivity and high True Negative TN for normal sinus rhythm (NSR). , A heart rhythm with an ECG signal amplitude of 0.1 mV does not require a shock, but has a disadvantage in that false positives (FP), which are incorrectly judged as shock heart rhythms, are high.

- When the QRS complex identification algorithm is used for machine learning, it has the advantage of high specificity, but it has the disadvantage of high sensitivity, especially false positive (FP) for ventricular tachycardia (VT).

Registered Patent No. 10-2143705 (Attachment guide system for automatic defibrillators and a guide method for attaching the same)

The present invention has been devised to solve the above problems, and the electrocardiogram analysis of the automatic defibrillator is actually applied through machine learning using deep learning, shock of the automatic defibrillator. An artificial intelligence that can make the judgment of and no-shock accurate so that only the patient who needs an electric shock can receive the electric shock, and make the feature extraction of the ECG signal more complex so that the machine learning results can be trusted The purpose of this study is to provide an electrocardiogram analysis and application method of a complex ECG feature extraction method for an automatic defibrillator.

The present invention provides a machine learning computer device comprising an ECG data storage, a first ECG feature extracting unit, and an ECG machine learning unit, and an automatic defibrillator connected thereto and comprising an electrocardiogram measuring unit, a second ECG feature extracting unit, and an ECG artificial intelligence engine. In the AI-based electrocardiogram analysis and application method for an automatic defibrillator, the method comprises: collecting ECG data and storing it in an ECG data storage of a computer device for machine learning (S1); Creating a step (S2), performing primary labeling in the form of recording the type of cardiac rhythm for each piece of the data set, and a second labeling step (S3) of recording whether or not a heart shock is present for each of the first labeled pieces; removing noise by applying a digital filter to the labeled data set (S4), analyzing the ECG signal from the data set to which the digital filter is applied and extracting features (S5); ECG machine learning step (S6) of performing deep learning with a supervised learning method and deep neural network (DNN) by using the extracted training data set as an input value and setting the secondary labeled heart shock as a classification result, and the ECG Using an artificial intelligence model such as neural network structure, activation function, and weight created through machine learning, a test data set from which ECG signal characteristics are extracted is an input value, and secondary labeled cardiac shock is set as the classification result. Conducting ECG machine learning test (S7), and evaluating the AI-based ECG analysis performance using the digital filter, extraction after ECG signal analysis, ECG machine learning test results, and the primary labeled heart rhythm data set (S8) and the ECG analysis artificial intelligence model including the neural network structure, activation function, weight, and shock determination threshold finally selected through the evaluated AI-based ECG analysis performance evaluation, digital filter, and ECG feature extraction algorithm automatically The same applies to the defibrillator (S9); and by analyzing the ECG signal, In the step of extracting the features (S5), it is characterized in that the features of the ECG signal are extracted through a plurality of algorithms.

In addition, the ECG data collected in the step (S1) of storing the ECG data storage is characterized in that it consists of the electronically synthesized ECG data made by synthesizing the actual ECG data and electronic signals obtained from the patient.

In addition, in the step (S1) of storing the ECG data storage, the ECG data is changed to the same sampling frequency as the electrocardiogram measuring unit of the automatic defibrillator and stored.

In addition, in the step (S2) of creating a data set by dividing it into pieces, the ECG artificial intelligence engine of the AED divides the data into pieces that can be analyzed.

In addition, in the step (S2) of creating a data set by dividing the pieces into pieces, the data set is a supervised learning method, characterized in that it consists of a learning data set for deep learning and a test data set for a machine learning test.

In addition, the algorithm applied in the step (S5) of extracting features by analyzing the ECG signal is TCSC: Threshold Crossing Sample Count algorithm, VFLEAK: VF Filter Leakage algorithm, PSR: Phase space reconstruction (Phase Space Reconstruction) algorithm, HILB: Hilbert Transform algorithm, bWT/bCP/bW: QRS complex identification algorithm characterized in that it consists of a combination of at least two or more algorithms.

The present invention has the effect of increasing the accuracy of the determination of shock and non-shock of an automatic defibrillator that is actually applied through machine learning using ECG signals, and extracting features of ECG signals By making this more complex, the ECG feature extraction with higher accuracy compared to the conventional simple algorithm has the effect of trusting the machine learning results and the judgment results of the AED.

1 is a diagram showing the connection configuration of a computer device for machine learning and an automatic defibrillator in the method for analyzing and applying an electrocardiogram for an artificial intelligence-based automatic defibrillator according to the present invention;
2 is a view showing an electrocardiogram analysis and application method for an artificial intelligence-based automatic defibrillator of the present invention;
3 is a flowchart showing the process of an electrocardiogram analysis and application method for an artificial intelligence-based automatic defibrillator according to the present invention;
4 is a two-dimensional phase space diagram created using a phase space reconstruction algorithm among the ECG extraction algorithms applied to the present invention.
5 is a two-dimensional phase space diagram created using the Hilbert Transform algorithm among the ECG extraction algorithms applied to the present invention.
6 is a flowchart showing a deep learning method in the present invention

Hereinafter, preferred embodiments of the present invention will be described in detail. In the description of the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

The present invention provides a machine learning computer device 100 comprising an ECG data storage 110, a first ECG feature extraction unit 120, and an ECG machine learning unit 130, and an electrocardiogram measuring unit 210 connected thereto, and a second ECG feature This is a technology related to an AI-based ECG analysis and application method for an AI-based automatic defibrillator in a system consisting of an automatic defibrillator 200 consisting of an extraction unit 220 and an ECG artificial intelligence engine 230 .

The computer device for machine learning 100 includes an ECG data storage 110 in which various ECG data is stored, a first ECG feature extraction unit 120 that extracts ECG features through a specific algorithm, and a deep learning data and test data. It consists of an ECG machine learning unit 130 that performs running.

In addition, the automatic defibrillator 200 includes an electrocardiogram measuring unit 210 that measures an actual patient's electrocardiogram, a second ECG feature extracting unit 220 that extracts ECG characteristics through an already input algorithm, and ECG machine learning generated through It consists of an ECG artificial intelligence engine 230 that analyzes ECG characteristics based on the ECG artificial intelligence model and determines whether or not the heart is in shock.

The present invention is a technology for performing machine learning in a computer device for machine learning using ECG data stored in the ECG data storage, and then applying an algorithm and ECG analysis artificial intelligence model, which is the final result, to an automatic defibrillator. The process takes place in a computer device for machine learning. The computer device for machine learning may include a mechanical device capable of performing an automatic calculation function, such as a notebook computer, a desktop computer, a supercomputer, a server, and a personal terminal.

In addition, the computer device for machine learning consists of a single configuration so that all processes can be performed, but since it is configured in plurality, different computer devices for machine learning can process each step.

As shown in FIG. 2, the specific method of the present invention includes the steps of collecting ECG data and storing it in an ECG data storage of a computer device for machine learning (S1), and creating a data set by dividing the stored ECG data into pieces (S2) and a secondary labeling step (S3) of performing primary labeling in the form of recording the type of cardiac rhythm for each piece of the data set, and recording whether or not cardiac shock is present for each of the first labeled fragments (S3), the labeling removing noise by applying a digital filter to the data set (S4), analyzing the ECG signal from the data set to which the digital filter is applied and extracting features (S5); ECG machine learning step (S6) of performing deep learning with a supervised learning method and deep neural network (DNN) by using a training data set as an input value and setting the secondary labeled heart shock as a classification result, and the ECG machine learning Using an artificial intelligence model such as neural network structure, activation function, and weight created through Conducting a learning test (S7) and evaluating the AI-based ECG analysis performance using the digital filter, extraction after analysis of the ECG signal, the ECG machine learning test result and the primary labeled heart rhythm data set (S8) ) and the ECG analysis AI model including the neural network structure, activation function, weight, and shock determination threshold finally selected through the evaluated AI-based ECG analysis performance evaluation, digital filter, and ECG feature extraction algorithm to the automatic defibrillator A step of applying the same to (S9); consists of,

In the step (S5) of analyzing the ECG signal and extracting the features, it is characterized in that the features of the ECG signal are extracted through a plurality of algorithms.

The ECG data collected in the step (S1) of storing the ECG data storage consists of electronically synthesized ECG data made by synthesizing the actual ECG data and electronic signals obtained from the patient as shown in FIG. 3 .

Here, the actual ECG data can be applied to the ECG data collected from real patients in a hospital, etc. and the ECG data collected through preclinical trials. Excludes ECG data that is too noisy to determine whether an impact is present or data in which the ECG signal is saturated because it cannot be used for machine learning.

Since the ECG data collected in this way has different sampling frequency, data format, and signal unit, it must be converted to the same standard as the ECG measurement unit of the AED to be applied. That is, it is converted to the same sampling frequency as the analog-digital converter of the automatic defibrillator to which the machine learning model is applied, and the ECG signal is converted into nV unit integer.

As described above, the ECG data collected from the patient is the same as the ECG signal that is actually input when the AED analyzes the heart rhythm, so it is suitable as learning data and test data. It is not enough to learn the boundaries of shock. Therefore, in order to learn the boundary between shock and no-shock, the ECG signal is created and synthesized to further create the synthesized electronic ECG data.

The electronically synthesized ECG data synthesizes signals by variously combining a waveform of a heart rhythm, a heart rate (beats per minute, BPM), and an amplitude.

The waveform of the heart rhythm is Coarse VFib, Fine VFib, Poly VTach, Mono VTach, Normal Sinus, Asystole, Multifocal PVCs, PVC1 LV R On T, PVC1 LV Early, PVC1 Left Vent, PVC2 RV R On T, PVC2 RV Early, It is created by combining the waveform characteristics of PVC2 Right Vent, Nodal PNC, AFib, SupraVTach, Nodal Rhythm, Paroxys-mal ATach, ATach, Missed Beat, Sinus Ar-rhythmia, and Atrial Flutter.

And the amplitude range is 0.1~0.5 mV (△0.05 mV), 1~3.5 mV (△0.5 mV), and the heart rate range is 120~300 BPM (△5 BPM) for Mono VTach and 10~360 for Normal Sinus. It consists of BPM (Δ5 BPM). (where △ means an increase value)

The step (S2) of creating a data set by dividing the data set into pieces includes dividing the actual ECG data that has undergone the conversion process and the newly created electronically synthesized ECG data into small units that the ECG artificial intelligence engine of the AED divides the heart rhythm into small units that can be analyzed. will make

Since the machine learning of the present invention is applied as a supervised learning type, in the step (S2) of creating a data set by dividing it into pieces, the data set is a learning data set required for deep learning and a test for machine learning test separated into data sets.

The training data set uses all of the electronically synthesized ECG data and a part of the collected ECG data. At this time, when composing the training data set, ECG data with an ECG signal amplitude of less than 0.1 mV is excluded.

And the test data set is used in the step (S7) of performing the ECG machine learning test.

The labeling step (S3) of the present invention consists of primary labeling for recording the type of heart rhythm and secondary labeling for recording whether or not the heart is in shock. The primary and secondary labeling is to label each ECG fragment of each data set, and primary and secondary labeling is performed to test the performance of machine learning and cardiac rhythm analysis of the training data set.

In the labeling step (S3), the primary labeling corresponding to the heart rhythm is set through annotation recorded in the ECG database and analysis by an expert. That is, the primary labeling is classified into 5 types, such as VF: ventricular fibrillation, VT: ventricular tachycardia, NSR: normal heartbeat, ASYS: asystole, OTHER: all non-shockable heartbeats, and Select and label.

The OTHER items are: Premature ventricular contractions, Supraventricular tachycardia, sinus bradycardia, Atrial fibrillation/atrial flutter, 2nd/3rd degree atrioventricular block ( Second or third-degree heart block), Accelerated Junctional Rhythm, and Pacemaker Rhythm.

In the above data set, ECG fragments with an ECG signal amplitude of less than 0.1 mV are labeled with ASYS. And among the ECG pieces labeled with VF and VT, if the heart rate is slow and non-shock, it is labeled as OTHER.

And in the labeling step (S3), the secondary labeling is to label the shock with SHOCK or NO-SHOCK according to the heart rhythm type and combination of the primary labeling. That is, in the cardiac rhythm of the ECG fragment, if shock is required, it is labeled as shock (SHOCK), and if shock is not required, it is labeled as non-shock (NO-SHOCK). For example, VF, VT: SHOCK is labeled, or NSR, ASYS, OTHER: NO-SHOCK is labeled.

A digital filter is applied to the labeled data set through the step (S4) of removing the noise. The digital filter is a filter that removes noise from the ECG signal and allows the signal to be smoothly analyzed. The application sequence of the digital filter is as follows.

First, it converts nV-class integer-type ECG signals into mV-class floating-point in the data set. Second, a band-pass filter is applied to remove baseline wander and power-line interference and to easily analyze ECG signals. Third, if necessary, a moving average filter may be applied.

The result of machine learning varies according to the digital filter combination according to the frequency of the data set, the filter order, and the type of digital filter to be applied. Therefore, it is necessary to select the optimal digital filter combination through machine learning evaluation, and to obtain the same performance, it is applied to the AED in the same way as the selected digital filter combination.

In the present invention, in the step (S5) of analyzing the ECG signal and extracting the features, the ECG signal is analyzed from the data set to which the digital filter is applied using a plurality of extraction algorithms to extract feature extraction.

The algorithm applied in the step (S5) of extracting features by analyzing the ECG signal is TCSC: Threshold Crossing Sample Count algorithm, VFLEAK: VF Filter Leakage algorithm, PSR: Phase space reconstruction ( Phase Space Reconstruction) algorithm, HILB: Hilbert Transform algorithm, bWT/bCP/bW: QRS complex identification algorithm, characterized in that it consists of a combination of at least two or more algorithms.

As a preferred combination of the algorithms, as shown below, it is preferable that the QRS complex identification algorithm and the shock rhythm determination algorithm of the defibrillator are used in combination.

- bCP, bWT, bW, PSR

- bCP, bWT, bW, HILB

- bCP, bWT, bW, VFLEAK, PSR

- bCP, bWT, bW, VFLEAK, HILB

- bCP, bWT, bW, TCSC, HILB

- bCP, bWT, bW, TCSC, PSR

- bCP, bWT, bW, PSR, HILB

- bCP, bWT, bW, VFLEAK, TCSC, HILB

- bCP, bWT, bW, VFLEAK, TCSC, PSR

- bCP, bWT, bW, VFLEAK, PSR, HILB

- bCP, bWT, bW, TCSC, PSR, HILB

As described above, in the present invention, bWT/bCP/bW: QRS complex identification algorithm is used as the main algorithm, the remaining TCSC: Threshold Crossing Sample Count algorithm, VFLEAK: VF filter (VF Filter Leakage) algorithm , PSR: Phase Space Reconstruction Algorithm, HILB: Hilbert Transform Algorithm may be variously combined as auxiliary algorithms.

The reason for applying the QRS complex identification algorithm as the main algorithm as described above is that the QRS complex identification algorithm has the highest specificity corresponding to the non-impact diagnosis rate. In addition, the fact that the false positive of ventricular tachycardia (VT), which is a disadvantage of the QRS complex identification algorithm, is determined to be high, can be supplemented through other auxiliary algorithms.

As described above, various combinations of algorithms can be applied, and the combination of algorithms is selected as a combination having high sensitivity and specificity after undergoing machine learning to be described later.

Below, each algorithm applied in the step (S5) of extracting features by analyzing the ECG signal will be described in detail.

(1) TCSC: Threshold Crossing Sample Count Algorithm

The Threshold Crossing Sample Count algorithm operates in the time domain and is based on the percentage of time the ECG signal remains outside a certain threshold. In TCSC, an electrocardiogram segment (Segment) goes through a special cosine window that reduces the value of the segment edge.

을 초과하는 정규화된 심전도 샘플 수를 계산하여 시간을 계산한다.Then a predefined threshold that depends on the absolute maximum of the segment.

Calculate the time by counting the number of normalized ECG samples exceeding

The TCSC algorithm sequence is as follows.

First, each ECG segment is multiplied by a cosine window. Here, the window w(t) is defined in the following way.

은 3초 ECG 세그먼트의 길이를 의미한다.here

denotes the length of the 3-second ECG segment.

And the ECG signal with the window set is normalized to the absolute maximum value of the segment.

과

을 비교해서 0과 1의 이진(Binary) 신호인

(여기서 n은 데이터 포인트 수)로 변환한다. 여기서

이면

이 되고, 아니면

이 된다.The normalized ECG signal is then generated for each data sample.

class

is a binary signal of 0 and 1 by comparing

(where n is the number of data points). here

back side

be this, or

becomes this

을 교차하는 데이터 샘플의 백분율인

은 이진수열

에서 1초 수를 세어서 구한다.

은 백분율로 다음과 같이 구한다.and

is the percentage of data samples that intersect

is a binary sequence

is obtained by counting the number of seconds in

is obtained as a percentage as follows.

ECG 에피소드

에 대해 1-초 단계로

개의 연속 3초 데이터 세그먼트에서 얻은

은

연속 값을 평균화해서 최종값을 결정한다.

에피소드의 평균값으로 TCSC를 구하는 공식은 다음과 같다.then all

ECG episode

about in 1-second steps

obtained from three consecutive 3 second data segments.

silver

The final value is determined by averaging the successive values.

The formula for calculating the TCSC as the average value of the episodes is as follows.

는

번째 3초 단계의

값을 의미한다.here

Is

in the second 3 second step

means value.

(2) VFLEAK: VF Filter Leakage Algorithm

The VF filter (VF Filter Leakage) algorithm extracts features using frequency domain analysis. This is based on the information that the shock rhythm is closer to a square wave than the no-shock rhythm.

The VF filter (VF filter leak) algorithm first finds the average period T through the following equation.

If you know the average period (T) of a segment, you can apply a narrowband filter using the equation below and get the VFLEAK parameter.

(3) PSR: Phase Space Reconstruction Algorithm

The phase space reconstruction algorithm is based on a method used to reconstruct the phase space with a time-delay algorithm method. In general, it is used for methods of analyzing dynamic laws or random signals.

는 다음과 같은 방식으로 표시한다. 먼저 x축은

, y축은

로 표시한다. (여기서 τ는 시간상수를 의미한다) 이러한 표시 방식을 2차원 위상 공간도(Two Dimensional Phase Space Diagram)이라 하며, 도 4에 도시한 바와 같다. signal

is displayed in the following way. First, the x-axis is

, the y-axis is

indicated as (Where τ means a time constant) This display method is called a two-dimensional phase space diagram, as shown in FIG. 4 .

Typically, the ventricular fibrillation (VF) signal fills the region in an irregular fashion, creating a curve in the diagram. The curve is distributed almost uniformly over the entire diagram. However, in normal sinus rhythm (Sinus Rhythm, SR), the curve of the phase space diagram is represented by a regular structure, only a small part of the area is filled, and the curve is concentrated into a limited area.

The PSR is found as the area of the figure filled with curves. In a grid of 40 * 40, count the boxes marked with ECG signals. The PSR calculation method is made by the following equation.

(4) HILB: Hilbert Transform Algorithm

에서 복소수 신호(Complex Valued Signal)

는

식으로 구한다.(여기서

는

의 힐베르트 변환을 의미한다) 그렇게 되면

와 같은 식이 성립되는데, 보통 힐베트르 변환은

를 계산하는 데 사용한다. 따라서 2차원 위상 공간 그림은 다음과 같은 방법으로 만든다. x축에 심전도 신호

를 표시하고, y축에는 힐베르트 변환

로 표시하며, 아래의 식을 통해 계산한다. 여기서 P.V.는 코시 주요값(Cauchy Principal Value)으로 구한다는 것을 의미한다.The Hilbert transform algorithm is based on a method of analyzing nonlinear signals. Real Signal

Complex Valued Signal in

Is

(here

Is

means the Hilbert transform of)

In general, the Hilbert transform is

used to calculate Therefore, a two-dimensional phase space figure is created in the following way. ECG signal on the x-axis

, and the Hilbert transform on the y-axis

and is calculated using the formula below. Here, PV means that it is calculated as a Cauchy Principal Value.

와

의 합성곱(Convolution)으로 간주한다. 합성곱의 특성 때문에

의 푸리에 변환

는

와

의 푸리에 변환의 곱이다. 따라서

일 경우에는

이고,

일 경우에는

이다. 이것은 진폭 반응이 통일적이고 위상 반응이 모든 주파수

조건에서 일정하게

정도 지연이 이상적인 필터에 의해 힐베르트 변환이 실현될 수 있다는 것을 의미한다.In the above equation, the Hilbert transform is

Wow

It is regarded as a convolution of Because of the nature of convolution

Fourier transform of

Is

Wow

is the product of the Fourier transform of therefore

in case of

ego,

in case of

to be. This means that the amplitude response is uniform and the phase response is at all frequencies.

constant under conditions

The degree of delay means that the Hilbert transform can be realized with an ideal filter.

을 기반으로 동 리듬(Sinus Rhythm)과 심실세동(Ventricular fibrillation, VF)을 구분할 수 있다. 40 * 40 그리드에서 ECG 신호로 곡선으로 표시된 상자를 계산하는 방법으로 HILB를 구한다.Normal ECG always displays curve-like circles in the Phase-Space Plot, but is filled in an irregular way with random signals. Phase-Space Plot

Based on the sinus rhythm (Sinus Rhythm) and ventricular fibrillation (Ventricular fibrillation, VF) can be distinguished. HILB is obtained by calculating the box plotted with the ECG signal on a 40*40 grid.

(5) bWT/bCP/bW: QRS complex identification algorithm

The shock rhythm is characterized by normal conduction in the ECG or the absence of narrow QRS complexes. Time, slope, and frequency are used to identify the presence of QRS complexes.

)로 정규화한다. Non-Shock heart rhythm is characterized by a longer Isoelectric Interval than Ventricular Arrhythmia. P-wave and T-wave are suppressed with a 6.5-30Hz band-pass filter, and amplitude 1 (

) is normalized to

샘플의 일정 비율을 유지하는 등전선(Isoelectric Line) 주위의 진폭 범위로 정의하며, 아래의 식과 같이 등전선(Isoelectric Line) 위(

)와 아래(

) 차이로 계산한다.The following bWT is

It is defined as the amplitude range around the isoelectric line that maintains a certain proportion of the sample, and above (

) and below (

) is calculated as the difference.

)로 정규화한다. bCP는 아래의 식과 같이 시간

임계값 (

)보다 낮은 비율로 정의한다.In the QRS complex, when the ECG changes rapidly, the descriptor value of the ECG increases. After squaring the first difference of the ECG signal, the amplitude 1 (

) is normalized to bCP is time as shown below

threshold (

) is defined as a ratio lower than

(여기서

는 ECG 세그먼트(Segment) 시간을 의미한다.)

(here

denotes the ECG segment time.)

으로 정규화한다. 다음 총 전력의 특정 비율을 포함하는 위(

)와 아래(

) 주파수를 계산한다. bW는 ECG 대역폭이며 아래의 식으로 정의된다. The well-defined rhythm of the QRS complex is Broadband Signals, whereas Ventricular Arrhythmia is a Narrow Band Signal. The power distribution in the frequency domain is estimated as the square of the 1024-point Fast Fourier Transform (FFT) size of the ECG segment. and the unit area under the curve

normalize to Above (

) and below (

) to calculate the frequency. bW is the ECG bandwidth and is defined by the following equation.

Then, it goes through the ECG machine learning step (S6) of performing deep learning with the supervised learning method and deep neural network (DNN). In general, supervised learning uses learning data unlike unsupervised learning. It is a deep learning method that, for example, informs (or learns) the meaning of each data in advance and finds out the meaning of newly input data.

In deep learning, ECG feature extraction is input in the step (S5) of extracting features by analyzing the ECG signal, and SHOCK and NO-SHOCK, which are shock labels, are set as classification results.

Deep learning constructs a deep neural network (DNN) as a sequential model. The input layer takes the data of ECG feature extraction as input. The hidden layer uses a hyperbolic tangent function to solve nonlinear data as an activation function, and a softmax function that transforms a vector into a probability distribution is used. The output layer is an activation function and uses a sigmoid function that converts a value between 0 and 1.

The input layer of the deep neural network model consists of the same nodes as the ECG feature extraction, and the output layer consists of one node. ECG feature extraction data is [FE(1), FE(2), … , FE(n)], the input node is [L1(1), L1(2), ... , L1(n)]. Here, n means the total number of ECG feature extraction data. The hidden layer consists of three layers, L2, L3, and L4, and each layer consists of n2, n3, and n4 nodes. And we use hyperbolic tangent and soft max as activation functions. Hidden layers can be added to improve ECG analysis performance. The output layer consists of one node and uses a sigmoid function.

If the ECG machine learning model is constructed using ECG feature extraction in this way, the number of nodes and hidden layers of the deep neural network model decreases, and the ECG AI engine of the AED can be made with fewer resources.

In this way, when machine learning is performed, the ECG machine learning test is performed (S7). Test data from which the ECG signal features are extracted using the ECG artificial intelligence model such as the neural network structure, activation function, and weight created through machine learning. ECG machine learning is tested by taking the set as an input and setting the secondary labeled cardiac shock as the classification result.

The ECG machine learning test first inputs the ECG feature values of the test data set into the ECG artificial intelligence model. The input ECG feature value passes through the neural network node according to the learned deep neural network structure, activation function, and weight, and then goes through the sigmoid function of the output layer to output an AI-VALUE value between 0 and 1. The AI-VALUE value approaches 1 as it approaches 1, and the closer it gets to 0, it approaches non-impact.

Finally, impact and non-impact are determined by the AI-VALUE threshold. If the AI-VALUE exceeds the threshold, the AI-SHOCK AI judges it as a shock, and if the AI-VALUE does not exceed the threshold, the AI-NO-SHOCK AI judges it as a non-shock.

In this way, since the determination of whether or not a shock is impacted depends on the threshold, the ECG machine learning test should be performed according to the combination of the thresholds. Therefore, in the step (S7) of performing the ECG machine learning test, C-DF is the number of combinations of digital filters, C-FE is the number of combinations of extracted ECG features, and C-ML is the ECG artificial intelligence model to determine whether the impact is When the total number of machine learning combinations to be performed according to the threshold value of And among the results, the optimal combination that can be applied to the AED is to be selected.

Then, the AI-based ECG analysis performance is evaluated (S8). The above step is a combination of an ECG artificial intelligence model such as a neural network structure, activation function, weight, etc. created through an algorithm such as digital filter and ECG feature extraction and ECG machine learning and a threshold value that is a criterion for determining an impact by artificial intelligence. In order to select an ECG analysis AI model for ECG analysis, AI-based ECG analysis performance evaluation is carried out.

In the step of evaluating the AI-based ECG analysis performance (S8), the evaluation of algorithms such as digital filters and ECG feature extraction algorithms using the ECG machine learning test results and the primary labeled heart rhythm data set, neural network structure, and activation function It evaluates the performance of the ECG analysis AI model, such as , weight, and shock judgment threshold, according to the matching combination of the shock label and the ECG AI model result. TN), false positive (FP), and false negative (FN).

The true positive (TP) is a case in which an impulse heart rhythm is accurately determined, and a true negative (TN) is a case in which a non-impulsive heart rhythm is accurately determined. In addition, false positive (FP) is a case in which a shock heart rhythm is incorrectly determined in a situation where no shock is required, and a false negative (FN) is a case in which a non-shock heart rhythm is incorrectly determined in a situation in which shock is required.

That is, the step (S8) of evaluating the AI-based ECG analysis performance is to determine the matching combination of the shock label and the machine learning model result in four forms. Specifically, the shock label is SHOCK and the ECG AI model result is AI-SHOCK, true positive (TP), the shock label is SHOCK, and the ECG AI model result is AI-NO-SHOCK, then false negative (FN), the shock label is NO-SHOCK, and the ECG AI model result is AI- If NO-SHOCK is true negative (TN), if the shock label is NO-SHOCK, but the ECG AI model result is AI-SHOCK, it is judged as false positive (FP).

For the AED, the specificity of the rhythm that does not require a shock is more important than the sensitivity of the rhythm that requires a shock. Sensitivity and specificity are calculated as follows using the determination result. Preferably, the sensitivity (Sensitivity) is calculated by the TP / (TP + FN) equation, and the specificity (Specificity) is calculated by the TN / (FP + TN) equation.

In international standards, the sensitivity and specificity of the AED is required as follows. Sensitivity of ventricular fibrillation (VF) is more than 90%, Sensitivity of ventricular tachycardia (VT) is more than 75%, specificity of normal heartbeat (NSR) is more than 99%, and asystole (Asys) ) Specificity is required to be 95% or more, and the specificity of all heartbeats that cannot give a shock is 95%.

Then, the ECG analysis AI model combination to be applied to the AED is selected using the heart rhythm determination criteria according to the heart rhythm label. That is, for example, if the heart rhythm label is VF, the shock label is SHOCK, and the ECG analysis AI model result is AI-SHOCK, then it is determined as TP-VF. If the model result is AI-SHOCK, it is determined as TP-VT, and if the shock label is NO-SHOCK among the heart rhythm label NSR and the ECG analysis AI model result is AI-NO-SHOCK, it is determined as TN-NSR, and cardiac rhythm If the shock label is NO-SHOCK in ASYS and the ECG analysis AI model result is AI-NO-SHOCK, it is determined as TN-ASYS, and the heart rhythm label is NO-SHOCK in OTHER, and the shock label is NO-SHOCK, ECG analysis AI If the model result is AI-NO-SHOCK, it is judged as TN-OTHER.

In the above judgment, TP-VF and TP-VT are the same as Sensitivity, and TN-NSR, TN-ASYS, TN-OTHER are the same as Specificity

Then, from among the heart rhythm labels of the electronically synthesized ECG data set, a combination of ECG analysis AI models that satisfy the following criteria is selected.

- TP-VF = 1

- TP-VT = 1

- TN-NSR = 1

- TN-ASYS = 1

- TN-OTHER > 0.95

Among the selected ECG analysis AI model combinations, a combination of ECG analysis AI models satisfying the following criteria is selected from among the cardiac rhythm labels of the test data set.

- TP-VF > 0.95

- TP-VT > 0.95

- TN-NSR > 0.99

- TN-ASYS > 0.99

- TN-OTHER > 0.95

Then, a combination applicable to the AED is selected from among the selected ECG analysis artificial intelligence model combinations. If the desired performance is not obtained, performance improvement is required, or the label is modified, data set creation is performed again.

Finally, the present invention automatically automates the ECG analysis artificial intelligence model, digital filter, and ECG feature extraction algorithm including the neural network structure, activation function, weight, and shock determination threshold finally selected through the evaluated AI-based ECG analysis performance evaluation. A step of applying the same to the defibrillator (S9); is passed.

That is, the neural network structure, activation function, weight, and shock determination threshold of the ECG analysis artificial intelligence model finally selected by the ECG machine learning unit are provided to the ECG artificial intelligence engine of the AED, and the digital filter of the first ECG feature extraction unit The ECG feature extraction algorithm provides the digital filter conditions and ECG feature extraction algorithm as the second ECG feature extraction unit of the AED. The judgment of No-Shock is made accurately.

Although the present invention has been described above with reference to the above embodiments, it goes without saying that various modifications are possible within the scope of the technical spirit of the present invention.

S1 ~ S9 : ECG analysis and application method for artificial intelligence-based automatic defibrillator
100: computer device for machine learning 110: ECG data storage
120: first ECG feature extraction unit 130: ECG machine learning unit
200: automatic defibrillator 210: electrocardiogram measurement unit
220: second ECG feature extraction unit 230: ECG artificial intelligence engine

Claims (6)

Artificial intelligence of a system consisting of a computer device for machine learning consisting of an ECG data storage, a first ECG feature extraction unit, and an ECG machine learning unit, and an automatic defibrillator connected thereto and consisting of an electrocardiogram measuring unit, a second ECG feature extraction unit, and an ECG artificial intelligence engine. In the electrocardiogram analysis and application method for the based automatic defibrillator,
Collecting ECG data and storing it in the ECG data storage of the computer device for machine learning (S1), creating a data set by dividing the stored ECG data into pieces (S2), heart rhythm for each piece of the data set A secondary labeling step (S3) of performing primary labeling in the form of recording the type of (S4), extracting features by analyzing the ECG signal from the data set to which the digital filter is applied (S5), using the training data set from which the features of the ECG signal are extracted as an input value, and secondary labeled heart ECG machine learning step (S6) of performing deep learning with a supervised learning method and deep neural network (DNN) by setting the impact status as the classification result, and artificial intelligence such as the neural network structure, activation function, and weight created through the ECG machine learning The ECG machine learning test is carried out by using the test data set from which the characteristics of the ECG signal are extracted using the model as the input value, and setting the secondary labeled cardiac shock as the classification result, but C-DF is a combination of digital filters C-DF × C- Step (S7) of performing the ECG machine learning test as much as the number of FE × C-ML, the digital filter, the extraction after analysis of the ECG signal, the ECG machine learning test result and the primary labeled heart rhythm data set for artificial intelligence Evaluating the based ECG analysis performance (S8) and the ECG analysis AI model and digital including the neural network structure, activation function, weight, and shock determination threshold finally selected through the evaluated AI-based ECG analysis performance evaluation The filter and the ECG feature extraction algorithm are equally applied to the AED (S9); consists of;
The extraction algorithm applied in the step (S5) of extracting features by analyzing the ECG signal is: TCSC: Threshold Crossing Sample Count Algorithm, VFLEAK: VF Filter Leakage Algorithm, PSR: Phase Space Reconstruction (Phase Space Reconstruction) Algorithm, HILB: Hilbert Transform Algorithm, bWT/bCP/bW: Composite ECG for artificial intelligence-based AED, characterized in that it consists of a combination of at least two or more algorithms among QRS complex identification algorithms ECG analysis and application method of feature extraction method

The method of claim 1,
The ECG data collected in the step (S1) of storing in the ECG data storage consists of electronically synthesized ECG data made by synthesizing the actual ECG data and electronic signals obtained from the patient. ECG analysis and application method of extraction method
The method of claim 1,
In the step (S1) of storing the ECG data storage, the ECG data is changed to the same sampling frequency as the ECG measuring unit of the AED, and stored. and how to apply
The method of claim 1,
ECG analysis and application of the complex ECG feature extraction method for the AI-based automatic defibrillator, characterized in that in the step (S2) of creating a data set by dividing it into pieces, the ECG AI engine of the automatic defibrillator divides it into pieces that can be analyzed Way
The method of claim 1,
In the step (S2) of creating a data set by dividing the pieces into pieces, the data set is an artificial intelligence-based automatic heart shock, characterized in that it consists of a learning data set for deep learning and a test data set for machine learning test in a supervised learning method. ECG analysis and application method of the complex ECG feature extraction method for Ki-Yong
delete

Images