Method of learning artificial intelligence for arrhythmia diagnosis

Abstract

The present invention proposes a method for training an artificial intelligence capable of diagnosing arrhythmia based on photoplethysmography (PPG) signals. The method may include the steps of: removing a noise region from the photoplethysmography (PPG) signal that is determined to be invalid for arrhythmia diagnosis due to noise introduced during the signal measurement process; applying a filter to the photoplethysmography (PPG) signal from which the noise region has been removed so that it has a frequency response within a pre-set pass band; and training a classification model designed for arrhythmia diagnosis based on a Convolutional Neural Network (CNN) based on the photoplethysmography (PPG) signal to which the filter has been applied.

Description

Technology Field

[0001] The present invention relates to medical artificial intelligence (medical AI). More specifically, it relates to a method for training artificial intelligence (AI) capable of diagnosing arrhythmia based on photoplethysmography (PPG) signals. Background Technology

[0002] Artificial intelligence (AI) is a technology that artificially implements some or all of human learning, reasoning, and perceptual abilities using computer programs. In relation to AI, machine learning is a learning process that optimizes parameters using given data by utilizing a model composed of multiple parameters. Depending on the form of the training data, machine learning is classified into supervised learning, unsupervised learning, and reinforcement learning (RL).

[0003] An Artificial Neural Network (ANN) is a specific methodology in machine learning that refers to an algorithm in the form of a network where neurons—basic computational units corresponding to human nerve cells—are interconnected. Representative ANNs include the Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN).

[0004] Specifically, a multilayer perceptron (MLP) is an artificial neural network (ANN) that has multiple hidden layers between the input layer and the output layer. A recurrent neural network (RNN) is an artificial neural network (ANN) that has a recursive connection structure in which the output of a neuron is input back into the input layer. And, a convolutional neural network (CNN) is an artificial neural network (ANN) that mimics the structure of the human visual nervous system, extracts feature maps from multiple convolutional layers, and extracts only important parts from the feature maps by reducing the dimensionality of the matrix through subsampling.

[0005] Meanwhile, arrhythmia refers to an irregular pulse. Generally, arrhythmia is caused by changes or dysfunction of the intracardiac conduction system. Depending on their symptoms and causes, these arrhythmias include sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (A-FIB), atrial flutter (A-FL), premature atrial complex (PAC), premature ventricular complex (PVC), atrial tachycardia (AT), multifocal atrial tachycardia (MAT), paroxysmal supraventricular tachycardia (PSVT), Wolff-Parkinson-White syndrome, ventricular tachycardia (V-TACH), and ventricular fibrillation. It can be classified into V-FIB, sinus arrest (SA), and atrioventricular block (AV-block).

[0006] Such arrhythmias can be diagnosed by measuring an electrocardiogram (ECG). To diagnose the exact type of arrhythmia, going beyond merely determining whether it is an arrhythmia, a 12-lead electrocardiogram must be performed, which involves attaching 12 leads—six precordial leads and six limb leads—to the body. Furthermore, if arrhythmia symptoms occur intermittently, the presence of the arrhythmia cannot be diagnosed unless the ECG is measured immediately upon the onset of symptoms; therefore, the ECG must generally be monitored for 24 hours. In other words, diagnosing the type of arrhythmia by measuring an ECG entails significant inconvenience.

[0007] Therefore, there is a need for a means to diagnose arrhythmia based on biological signals other than the electrocardiogram (ECG). Prior art literature

[0008] Korean Patent Publication No. 10-2022-0154403, 'Method and apparatus for determining arrhythmia and a recording medium having a program for implementing the same', (Published Nov. 22, 2022) Korean Registered Patent Publication No. 10-2294354, 'System for diagnosing signs of arrhythmia', (Registered Aug. 20, 2021) The problem to be solved

[0009] One objective of the present invention is to provide a method for training an artificial intelligence capable of diagnosing arrhythmia based on photoplethysmography (PPG) signals.

[0010] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0011] To achieve the technical objectives described above, the present invention proposes a method for training an artificial intelligence capable of diagnosing arrhythmia based on photoplethysmography (PPG) signals.

[0012] The above method may include: a step of removing a noise region from a photoplethysmography (PPG) signal that is determined not to be effective for arrhythmia diagnosis due to noise introduced during the signal measurement process; a step of applying a filter to the photoplethysmography (PPG) signal from which the noise region has been removed so that it has a frequency response within a pre-set pass band; and a step of training a classification model designed for arrhythmia diagnosis based on a Convolutional Neural Network (CNN) based on the photoplethysmography (PPG) signal to which the filter has been applied.

[0013] More specifically, the method may further include a step of dividing a continuously measured photoplethysmography (PPG) signal into a plurality of photoplethysmography (PPG) signals corresponding to a preset interval length prior to the step of removing the noise region, and then interpolating based on a cubic spline function to convert each divided photoplethysmography (PPG) signal into a preset frequency.

[0014] The step of removing the noise region above can mask a photoplethysmography (PPG) signal containing a flat line with a number of consecutive identical values ​​exceeding a preset number among the resampled multiple photoplethysmography (PPG) signals, or a photoplethysmography (PPG) signal containing an undefined value (not a number).

[0015] The step of removing the noise region may remove a linear component included in each resampled photoplethysmography (PPG) signal, or remove a direct current component included in each photoplethysmography (PPG) signal from which the linear component has been removed.

[0016] The step of removing the noise region above involves removing noise using a Hanning window on each photoplethysmography (PPG) signal from which the DC component has been removed, and decomposing each photoplethysmography (PPG) signal from which the noise has been removed into frequency components by performing a fast Fourier transform.

[0017] And, the step of removing the noise region can calculate the entropy for each photoplethysmography (PPG) signal based on the probability distribution for the photoplethysmography (PPG) signal decomposed into the frequency components, and then identify the noise region based on the calculated entropy.

[0018] Meanwhile, the step of applying the filter may involve applying a high-cut Butterworth filter and a low-cut Butterworth filter, set based on the passband, to the photoplethysmography (PPG) signal, and then performing normalization so that the photoplethysmography (PPG) signal has a value between 0 and 1.

[0019] The step of training the above classification model involves a total of constituting the dataset of the above photoplethysmography (PPG) signals. N n signal data {y1, … , y N }, above N Labels corresponding to each of the signal data {label1, … , label N }, the loss function l , convolution weights W , encoder model f When that is the case, the convolutional weights of the above classification model can be optimized using the following mathematical formula.

[0020] [Mathematical Formula]

[0021]

[0022] The step of training the above classification model is the actual probability distribution in the probability space X. P , the probability distribution obtained through prediction Q When given, the above loss function is the following mathematical expression l It can be used as.

[0023] [Mathematical Formula]

[0024]

[0025] In addition, the step of training the classification model may calculate an evaluation value using one or more of an F1 score or balanced accuracy based on the type of arrhythmia classified by the classification model and the actual type of arrhythmia of the subject in which the photoplethysmography (PPG) signal was measured, and evaluate the classification model based on the calculated evaluation value.

[0026] Specific details of other embodiments are included in the detailed description and drawings. Effects of the invention

[0027] According to embodiments of the present invention, an artificial intelligence (AI) designed for the diagnosis of arrhythmia can be trained using photoplethysmography (PPG) signals that can be easily obtained by a non-invasive method. In addition, the accuracy of arrhythmia type classification can be improved through precise preprocessing of the photoplethysmography (PPG) signals.

[0028] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. Brief explanation of the drawing

[0029] FIG. 1 is an exemplary diagram showing the process of diagnosing arrhythmia according to one embodiment of the present invention. FIG. 2 is a configuration diagram showing an arrhythmia diagnosis system according to one embodiment of the present invention. FIG. 3 is a logical configuration diagram of an arrhythmia diagnosis server according to one embodiment of the present invention. FIG. 4 is an exemplary diagram showing a resampled photoplethysmography (PPG) signal according to one embodiment of the present invention. FIG. 5 is an exemplary diagram showing a photoplethysmography (PPG) signal with linear components removed according to one embodiment of the present invention. FIG. 6 is an exemplary diagram showing a photoplethysmography (PPG) signal with a Hanning window applied according to one embodiment of the present invention. FIG. 7 is an exemplary diagram showing a photoplethysmography (PPG) signal transformed by a fast Fourier transform according to one embodiment of the present invention. FIG. 8 is an exemplary diagram showing a photoplethysmography (PPG) signal with a Butterworth filter applied according to one embodiment of the present invention. FIG. 9 is an illustrative diagram for explaining the logical structure of a classification model according to one embodiment of the present invention. FIG. 10 is an exemplary diagram illustrating the layer structure between encoders of a classification model according to one embodiment of the present invention. FIG. 11 is an illustrative diagram for explaining the process of performing test point augmentation according to one embodiment of the present invention. FIG. 12 is a hardware configuration diagram of an arrhythmia diagnosis server according to one embodiment of the present invention. FIG. 13 is a flowchart illustrating an artificial intelligence learning method according to one embodiment of the present invention. FIG. 14 is a flowchart illustrating a process for removing noise according to an embodiment of the present invention. FIG. 15 is a flowchart illustrating a method for diagnosing arrhythmia according to one embodiment of the present invention. Specific details for implementing the invention

[0030] It should be noted that technical terms used in this specification are used merely to describe specific embodiments and are not intended to limit the invention. Furthermore, unless specifically defined otherwise in this specification, technical terms used herein shall be interpreted in the sense generally understood by those skilled in the art to which the invention pertains, and shall not be interpreted in an overly broad or overly narrow sense. Additionally, if a technical term used in this specification is an incorrect technical term that fails to accurately express the spirit of the invention, it shall be understood as being replaced by a technical term that can be correctly understood by those skilled in the art. Moreover, general terms used in the invention shall be interpreted according to their prior definitions or the context, and shall not be interpreted in an overly narrow sense.

[0031] Additionally, singular expressions used in this specification include plural expressions unless the context clearly indicates otherwise. In this application, terms such as "composed of" or "have" should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as potentially including some of the components or steps, or including additional components or steps.

[0032] Additionally, terms including ordinal numbers, such as first, second, etc., used herein may be used to describe various components, but said components should not be limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0033] When it is stated that one component is "connected" or "connected" to another component, it may be directly connected or connected to that other component, or there may be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0034] Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. Identical or similar components regardless of drawing symbols are given the same reference number, and redundant descriptions thereof will be omitted. Furthermore, in describing the present invention, if it is determined that a detailed description of related known technology may obscure the essence of the present invention, such detailed description will be omitted. Additionally, it should be noted that the attached drawings are intended only to facilitate an easy understanding of the concept of the present invention and should not be interpreted as limiting the concept of the present invention. The concept of the present invention should be interpreted as extending to all modifications, equivalents, and substitutions other than those shown in the attached drawings.

[0036] Meanwhile, as mentioned above, in order to diagnose the exact type of arrhythmia beyond simply determining whether an arrhythmia is present from an electrocardiogram (ECG), a 12-lead ECG with 12 leads attached to the body must be performed for 24 hours. In other words, diagnosing the type of arrhythmia by measuring an electrocardiogram (ECG) entails a lot of inconvenience.

[0037] To address these difficulties, the present invention proposes a means to easily diagnose arrhythmia using non-invasive methods.

[0039] FIG. 1 is an exemplary diagram showing the process of diagnosing arrhythmia according to one embodiment of the present invention.

[0040] As illustrated in FIG. 1, according to one embodiment of the present invention, arrhythmia can be diagnosed by analyzing a photoplethysmography (PPG) signal, which can be easily measured non-invasively, using artificial intelligence (AI).

[0041] Specifically, a photoplethysmography (PPG) signal is a signal that detects the amount of blood flowing in the blood vessels inside the biological tissue using the optical properties of light projected toward the biological tissue, and represents a measured pulse wave based on the detected blood flow. Such a photoplethysmography (PPG) signal can be measured simply by having the measuring device in close contact with the human body.

[0042] Arrhythmia is an irregularly beating pulse. According to the present invention, it is possible to diagnose whether a basically measured photoplethysmography (PPG) signal corresponds to a normal sinus rhythm (SR) or to an arrhythmia. In particular, if it corresponds to an arrhythmia, even the type of arrhythmia can be diagnosed.

[0043] Excluding normal sinus rhythm (SR), the types of arrhythmias diagnosable according to the present invention may include, but are not limited to, sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (A-FIB), atrial flutter (A-FL), premature atrial complex (PAC), sinus arrest (SA), and premature ventricular complex (PVC), as well as atrial tachycardia (AT), multifocal atrial tachycardia (MAT), paroxysmal supraventricular tachycardia (PSVT), and Wolff-Parkinson-White syndrome. Ventricular tachycardia (V-TACH), ventricular fibrillation (V-FIB), and atrioventricular block (AV-block) may also be included.

[0044] To this end, one embodiment of the present invention provides means for training an artificial intelligence (AI) designed for the diagnosis of arrhythmia using photoplethysmography (PPG) signals collected from various paths. Furthermore, another embodiment of the present invention provides means for diagnosing arrhythmia from photoplethysmography (PPG) measured in real time using the previously trained artificial intelligence (AI).

[0046] The configuration of an arrhythmia diagnosis system according to one embodiment of the present invention, which can provide the means described above, will be explained in more detail.

[0047] FIG. 2 is a configuration diagram showing an arrhythmia diagnosis system according to one embodiment of the present invention.

[0048] As illustrated in FIG. 2, an arrhythmia diagnosis system according to one embodiment of the present invention may be configured to include various types of PPG sensors (100a, 100b, …, 100n; 100) and an arrhythmia diagnosis server (200).

[0049] As such, since the components of an arrhythmia diagnosis system merely represent functionally distinct elements, two or more components may be implemented as an integrated unit in an actual physical environment, or a single component may be implemented separately in an actual physical environment.

[0050] To describe each component, the PPG sensor (100) is a device that can measure photoplethysmography (PPG) by being in close contact with a person's body.

[0051] Specifically, the PPG sensor (100) may have the form of a watch-type sensor (100a) worn on the wrist of a measurement target (i.e., a person), a patch-type sensor (100b) attached to the body of a measurement target, or a clip-type sensor (100n) worn on the finger or toe of a measurement target, but is not limited thereto.

[0052] The PPG sensor (100) can continuously measure the photoplethysmography (PPG) of the measurement target. The PPG sensor (100) can transmit the continuously measured photoplethysmography (PPG) signals directly or indirectly to the arrhythmia diagnosis server (200).

[0053] With the following configuration, the arrhythmia diagnosis server (200) collects various photoplethysmography (PPG) signals measured by the PPG sensor (100) and trains an artificial intelligence (AI), and can diagnose arrhythmia from the photoplethysmography (PPG) signals measured in real time by the PPG sensor (100) using the trained artificial intelligence (AI).

[0054] The arrhythmia diagnosis server (200) may be any one of a fixed computing device such as a desktop, workstation, or server, but is not limited thereto.

[0055] Specific configurations and operations for implementing such an arrhythmia diagnosis server (200) will be described later with reference to FIGS. 3 to 16.

[0056] Meanwhile, the PPG sensor (100) and the arrhythmia diagnosis server (200) can transmit and receive data to and from each other using a network that combines one or more of a line directly connecting the devices, a public wired communication network, or a mobile communication network.

[0057] For example, public wired communication networks may include Ethernet, Digital Subscriber Line (xDSL), Hybrid Fiber Coax (HFC), and Fiber To The Home (FTTH), but are not limited thereto. Additionally, mobile communication networks may include Code Division Multiple Access (CDMA), Wideband CDMA (WCDMA), High Speed ​​Packet Access (HSPA), Long Term Evolution (LTE), and 5th generation mobile telecommunication, but are not limited thereto.

[0059] Hereinafter, the logical configuration of an arrhythmia diagnosis server (200) according to one embodiment of the present invention will be described in more detail.

[0060] FIG. 3 is a logical configuration diagram of an arrhythmia diagnosis server according to one embodiment of the present invention.

[0061] As illustrated in FIG. 3, an arrhythmia diagnosis server (200) according to one embodiment of the present invention may be configured to include a communication unit (205), an input / output unit (210), a data collection unit (215), a signal preprocessing unit (220), an artificial intelligence learning unit (225), and an arrhythmia diagnosis unit (230).

[0062] As such, the components of the arrhythmia diagnosis server (200) merely represent functionally distinct elements, so two or more components may be implemented as a single integrated unit in an actual physical environment, or a single component may be implemented as a separate unit in an actual physical environment.

[0063] To explain each component, the communication unit (205) can transmit and receive data to and from the PPG sensor (100), etc.

[0064] Specifically, the communication unit (205) may receive photoplethysmography (PPG) signals directly or indirectly from one or more PPG sensors (100). Additionally, the communication unit (205) may receive photoplethysmography (PPG) signals from a File Transfer Protocol (FTP) server that stores various biosignals. Here, the FTP server may be a data server operated by a medical service association or a professional organization related to medical services, but is not limited thereto.

[0065] With the following configuration, the input / output unit (210) can receive or output various data related to arrhythmia diagnosis.

[0066] Specifically, the input / output unit (210) can directly receive a photoplethysmography (PPG) signal. The input / output unit (210) can receive various parameters related to the preprocessing of the photoplethysmography (PPG). The input / output unit (210) can receive various parameters related to a classification model designed for arrhythmia diagnosis. Additionally, the input / output unit (210) can output the preprocessed photoplethysmography (PPG) signal in the form of a numerical value or a graph.

[0067] And, the input / output unit (210) can output the output result derived from the classification model. The input / output unit (210) can output the type of the finally determined arrhythmia.

[0068] In the following configuration, the data collection unit (215) can collect photoplethysmography (PPG) signals from various paths to train a classification model designed for arrhythmia diagnosis.

[0069] Specifically, the data collection unit (215) can collect photoplethysmography (PPG) signals directly measured by the PPG sensor (100) through the communication unit (205) or the input / output unit (210). The data collection unit (2150) can collect photoplethysmography (PPG) signals stored in an FTP server, etc., through the communication unit (205). In addition, the data collection unit (215) can collect photoplethysmography (PPG) signals by performing web crawling through the communication unit (205).

[0070] Meanwhile, the data collection unit (215) can de-identify information if the data related to the collected photoplethysmography (PPG) signal contains information that can identify who the measurement target (i.e., person) is.

[0071] In the following configuration, the signal preprocessing unit (220) can preprocess the photoplethysmography (PPG) signals collected to train a classification model and the photoplethysmography (PPG) signals measured in real time for arrhythmia diagnosis so that they can be input into the classification model.

[0072] First, the signal preprocessing unit (220) can divide the continuously measured photoplethysmography (PPG) signal into multiple photoplethysmography (PPG) signals corresponding to a preset interval length.

[0073] For example, the signal preprocessing unit (220) can divide the photoplethysmography (PPG) signal, which is measured continuously in a time series, into multiple photoplethysmography (PPG) signals having a length of 10 seconds.

[0074] The signal preprocessing unit (220) can resample each divided photoplethysmography (PPG) signal.

[0075] FIG. 4 is an exemplary diagram showing a resampled photoplethysmography (PPG) signal according to one embodiment of the present invention.

[0076] Referring to FIG. 4, the signal preprocessing unit (220) can interpolate each divided photoplethysmography (PPG) signal based on a cubic spline function to convert each divided photoplethysmography (PPG) signal into a preset frequency.

[0077] For example, the signal preprocessing unit (220) can arrange each divided 200 Hz photoplethysmography (PPG) signal to approximate the target resolution of 125 Hz, and then obtain a spline function to be used for interpolation using cubic spline interpolation. Then, the signal preprocessing unit (220) can convert each photoplethysmography (PPG) signal to have a frequency of 125 Hz using the obtained spline function.

[0078] More specifically, spline functionS (x) is the n+1 points (x0, y0), … , (x n , y n For consecutively connected subintervals, the continuity condition imposed between the function and its derivative can be determined to satisfy the following Equation 1. In this case, for each subinterval [x i-1 , x i Spline function S( x The degree of the polynomial of ) can be 3.

[0079] [Mathematical Formula 1]

[0080]

[0081] Next, the signal preprocessing unit (220) can remove noise regions from the resampled photoplethysmography (PPG) signal that are determined not to be effective for diagnosing arrhythmia due to noise introduced during the signal measurement process. This is because, since the photoplethysmography (PPG) signal is measured while in close contact with the body of the subject being measured, noise can be introduced even by minute movements of the subject being measured during the measurement process.

[0082] Basically, the signal preprocessing unit (220) can mask the photoplethysmography (PPG) signal with a large amount of noise introduced.

[0083] More specifically, the signal preprocessing unit (220) can mask a photoplethysmography (PPG) signal that includes a flat line with the same value being continuous for more than a preset number of times among a plurality of resampled photoplethysmography (PPG) signals.

[0084] For example, the signal preprocessing unit (220) can compare the array of photoplethysmography (PPG) signals by shifting them one step at a time, and if the same value appears five times consecutively before or after the corresponding index, the corresponding photoplethysmography (PPG) signal can be masked.

[0085] Additionally, the signal preprocessing unit (220) can mask photoplethysmography (PPG) signals containing undefined values ​​(Not A Number, NAN) among a plurality of resampled photoplethysmography (PPG) signals.

[0086] For example, the signal preprocessing unit (220) can determine whether the data value of each index for each photoplethysmography (PPG) signal corresponds to an undefined value (NAN), and if it corresponds to an undefined value, it can mask the photoplethysmography (PPG) signal containing it.

[0087] Next, the signal preprocessing unit (220) can remove photoplethysmography (PPG) signals in which the frequency distribution is close to random by using entropy.

[0088] As a preliminary signal processing for calculating entropy, the signal processing unit (220) may apply a Butterworth filter to each resampled photoplethysmography (PPG) signal.

[0089] For example, the signal preprocessing unit (220) may apply a Butterworth filter of order 3 with a low-cut of 1 or less and a high-cut of 5 or more to detect a systolic peak for each resampled photoplethysmography (PPG) signal.

[0090] Next, the signal preprocessing unit (220) can remove linear detrends included in each resampled photoplethysmography (PPG) signal.

[0091] FIG. 5 is an exemplary diagram showing a photoplethysmography (PPG) signal with linear components removed according to one embodiment of the present invention.

[0092] Referring to FIG. 5, the signal preprocessing unit (220) can remove linear components calculated by the method of least squares for baseline correction of each photoplethysmography (PPG) signal.

[0093] More specifically, the signal preprocessing unit (220) y i Let be the data value of the i-th signal and let n be the length of the signal data, then (i, y i A linear component b+ai approximated by ) can be calculated based on the following mathematical formula 2.

[0094] [Mathematical Formula 2]

[0095]

[0096] Additionally, the signal preprocessing unit (220) can remove the direct current (DC) component included in each photoplethysmography (PPG) signal from which the linear component has been removed.

[0097] As described above, the linear component removal process and the DC component removal process may be applied selectively or redundantly depending on the characteristics of the target photoplethysmography (PPG) signal.

[0098] Next, the signal preprocessing unit (220) can remove noise using a Hanning window on each photoplethysmography (PPG) signal from which the DC component has been removed.

[0099] FIG. 6 is an exemplary diagram showing a photoplethysmography (PPG) signal with a Hanning window applied according to one embodiment of the present invention.

[0100] Referring to FIG. 6, the signal processing unit (220) can calculate a Hanning window for each photoplethysmography (PPG) signal and multiply the calculated Hanning window by the original photoplethysmography (PPG) signal to remove noise.

[0101] More specifically, the signal processing unit (220) can calculate the Hanning window w0(x) for a signal of length L expressed as a function of x based on the following mathematical formula 3.

[0102] [Mathematical Formula 3]

[0103]

[0104] Next, the signal preprocessing unit (220) can decompose each noise-removed photoplethysmography (PPG) signal into frequency components by performing a fast Fourier transform using a Hanning window.

[0105] FIG. 7 is an exemplary diagram showing a photoplethysmography (PPG) signal transformed by a fast Fourier transform according to one embodiment of the present invention.

[0106] Referring to FIG. 7, the signal processing unit (220) provides real discrete signals x0, x1, …, x for each photoplethysmography (PPG) signal. N -1 For , the corresponding frequency is 0 / N , 1 / N , … , ( N -1) / N In decomposing into a signal, the frequency is k / N Amplitude of the signal X k It can be calculated through the following mathematical formula 4.

[0107] [Mathematical Formula 4]

[0108]

[0109] Next, the signal preprocessing unit (220) can convert each photoplethysmography (PPG) signal decomposed into frequency components into a probability density function by multiplying it by a constant such that the total sum is 1.

[0110] The signal preprocessing unit (220) can calculate the Reny entropy for each photoplethysmography (PPG) signal based on the probability distribution for the photoplethysmography (PPG) signal converted into frequency components.

[0111] More specifically, when the discrete random variable is X, the probability density function with respect to it is p(x), and the order is α, the Réni entropy H is obtained using the following Equation 5 or Equation 6. α ( X) can be calculated. The present invention can apply the Renny entropy when α→1.

[0112] [Mathematical Formula 5]

[0113]

[0114] [Mathematical Formula 6]

[0115]

[0116] Additionally, the signal preprocessing unit (220) can identify a noise region based on the calculated Lenny entropy. That is, the signal preprocessing unit (220) can identify and remove the corresponding photoplethysmography (PPG) signal as noise if the entropy calculated for each photoplethysmography (PPG) signal is greater than or equal to a preset threshold value.

[0117] Next, the signal preprocessing unit (220) can refine each photoplethysmography (PPG) signal from which noise has been removed. First, the signal preprocessing unit (220) can apply a filter to the photoplethysmography (PPG) signal from which noise regions have been removed so that it has a flat frequency response within a preset pass band.

[0118] FIG. 8 is an exemplary diagram showing a photoplethysmography (PPG) signal with a Butterworth filter applied according to one embodiment of the present invention.

[0119] Referring to FIG. 8, the signal processing unit (220) has a value when the frequency is 0 G 0, the cut standard is ω c When the order is n, the following mathematical formula 7 can be applied as a high-cut Butterworth filter, and the following mathematical formula 8 can be applied as a low-cut Butterworth filter.

[0120] [Mathematical Formula 7]

[0121]

[0122] [Mathematical Formula 8]

[0123]

[0124] For example, the signal processing unit (220) can apply a high-cut Butterworth filter with a cut criterion of 5 or higher and a low-cut Butterworth filter with a cut criterion of 1 or lower to a photoplethysmography (PPG) signal.

[0125] And, the signal processing unit (220) can complete the preprocessing process of the photoplethysmography (PPG) signal by performing normalization so that the photoplethysmography (PPG) signal to which the Butterworth filter is applied has a value between 0 and 1.

[0126] More specifically, the signal processing unit (220) can perform normalization of the photoplethysmography (PPG) signal by applying the following mathematical formula 9.

[0127] [Mathematical Formula 9]

[0128]

[0129] With the following configuration, the artificial intelligence learning unit (220) can train a classification model based on photoplethysmography (PPG) signals collected from various paths by the data collection unit (215) and preprocessed by the signal processing unit (220).

[0130] FIG. 9 is an illustrative diagram illustrating the logical structure of a classification model according to an embodiment of the present invention. FIG. 10 is an illustrative diagram illustrating the layer structure between encoders of a classification model according to an embodiment of the present invention.

[0131] A classification model according to one embodiment of the present invention, as illustrated in FIGS. 9 and 10, is a model designed for the diagnosis of arrhythmia based on a Convolutional Neural Network (CNN). In FIGS. 9 and 10, C in and C OUT is the number of input and output channels between each encoder, K is the kernel size, and S is the stride size.

[0132] A classification model according to one embodiment of the present invention, as described above, can output probability values ​​for each type of arrhythmia in response to an input photoplethysmography (PPG) signal.

[0133] More specifically, the artificial intelligence learning unit (220) comprises a total dataset of photoplethysmography (PPG) signals. N n signal data {y1, … , y N }, N Labels corresponding to each of the signal data {label1, … , label N }, the loss function l , convolution weights W , encoder model f When that is the case, the convolutional weights of the classification model can be optimized using the following mathematical formula 10. Here, the dataset refers to photoplethysmography (PPG) signals obtained from multiple subjects.

[0134] [Mathematical Formula 10]

[0135]

[0136] The artificial intelligence learning unit (220) is an actual probability distribution in probability space X. P , the probability distribution obtained through prediction Q When let, the following mathematical expression is the loss function l It can be used as.

[0137] [Mathematical Formula 11]

[0138]

[0139] The artificial intelligence learning unit (220) can calculate an evaluation value using one or more of an F1 score or balanced accuracy based on the type of arrhythmia classified by the classification model and the actual type of arrhythmia of the subject whose photoplethysmography (PPG) signal is measured. The artificial intelligence learning unit (220) can evaluate the classification model based on the calculated evaluation value.

[0140] To this end, the artificial intelligence learning unit (220) can calculate precision and recall based on the following mathematical formula 12, where cases predicted as positive by the classification model and actually positive are called True Positive, all cases predicted as positive by the classification model are called Predicted Positive, and all cases actually positive are called Actual Positive.

[0141] [Mathematical Formula 12]

[0142]

[0143] The artificial intelligence learning unit (220) can calculate the F1 score based on the following mathematical formula 13 based on the calculated precision and recall rate.

[0144] [Mathematical Formula 13]

[0145]

[0146] The artificial intelligence learning unit (220) sets the weight of the i-th type of arrhythmia to ω i When that is the case, the balance accuracy can be calculated based on the following mathematical formula 14.

[0147] [Mathematical Formula 14]

[0148]

[0149] And, the artificial intelligence learning unit (220) can learn the classification model by considering the evaluation value of the calculated classification model and using photoplethysmography (PPG) signals collected from various paths by the data collection unit (215).

[0150] In the following configuration, the arrhythmia diagnosis unit (230) can determine the type of arrhythmia corresponding to the photoplethysmography (PPG) signal by using a classification model previously learned by the artificial intelligence learning unit (220).

[0151] Specifically, the arrhythmia diagnosis unit (230) can input the photoplethysmography (PPG) signal, which is measured in real time by the PPG sensor (100) and preprocessed by the signal preprocessing unit (220), into a classification model that has been pre-trained by the artificial intelligence learning unit (220). Then, the arrhythmia diagnosis unit (230) can determine the type of arrhythmia of the subject to which the photoplethysmography (PPG) signal was measured, based on the results derived from the classification model.

[0152] More specifically, the arrhythmia diagnosis unit (230) classifies the output values ​​of the classification model into {x1, … , x N When} is used, the probability of corresponding to the i-th arrhythmia type can be calculated using the following mathematical formula 15 regarding softmax.

[0153] [Mathematical Formula 15]

[0154]

[0155] In order to improve the accuracy of classification (i.e., determination of the type of arrhythmia), the arrhythmia diagnosis unit (230) may have test time augmentation applied.

[0156] FIG. 11 is an illustrative diagram for explaining the process of performing test point augmentation according to one embodiment of the present invention.

[0157] With reference to FIG. 11 for a more detailed explanation, the arrhythmia diagnosis unit (230) can identify multiple photoplethysmography (PPG) signals having the same unit length with different starting points set to have overlapping regions from photoplethysmography (PPG) signals that are continuously measured in a time series. Then, the arrhythmia diagnosis unit (230) can finally determine the type of arrhythmia of the subject to which the photoplethysmography (PPG) signal was measured based on the average of the output results of the classification model for the identified multiple photoplethysmography (PPG) signals.

[0158] More specifically, the arrhythmia diagnosis unit (230) determines the probability value output from the classification model that the photoplethysmography (PPG) signal measured after t seconds corresponds to type a among set A of arrhythmia types. p t (a) Let the output result be from the previously derived photoplethysmography (PPG) signal N Total measured every second up to -1 second N Based on the output results of the above classification model for the photoplethysmography (PPG) signals, the type of arrhythmia pred of the subject can be determined using the following mathematical formula 16.

[0159] [Mathematical Formula 16]

[0160]

[0161] For example, the arrhythmia diagnosis unit (230) can finally determine the type of arrhythmia of the subject in which the photoplethysmography (PPG) signal is measured based on the output result of a classification model for eight photoplethysmography (PPG) signals having a length of 10 seconds with different starting points set at 1-second intervals.

[0163] The hardware of the arrhythmia diagnosis server (200) for realizing the logical components described above will be explained in more detail below.

[0164] FIG. 12 is a hardware configuration diagram of an arrhythmia diagnosis server according to one embodiment of the present invention.

[0165] As illustrated in FIG. 12, the arrhythmia diagnosis server (200) may be configured to include a processor (250), memory (255), transceiver (260), input / output device (265), data bus (270) and storage (275).

[0166] Specifically, the processor (250) can implement the operation and function of the arrhythmia diagnosis server (200) based on instructions according to software (280a) which implements an artificial intelligence learning method and / or an arrhythmia diagnosis method residing in memory (255).

[0167] Software (280b) implementing an artificial intelligence learning method and / or arrhythmia diagnosis method stored in storage (275) may be loaded into the memory (255).

[0168] The input / output device (265) can receive signals necessary for the operation of the arrhythmia diagnosis server (200) or output calculation results to the outside according to the command of the processor (250).

[0169] The data bus (270) is connected to the processor (250), memory (255), transceiver (260), input / output device (265), and storage (275), respectively, and can serve as a passage for transmitting signals between each component.

[0170] Storage (275) may store an Application Programming Interface (API), library files, resource files, etc., necessary for the execution of software (280a) in which an artificial intelligence learning method and / or arrhythmia diagnosis method according to embodiments of the present invention is implemented. Storage (275) may store software (280b) in which an artificial intelligence learning method and / or arrhythmia diagnosis method according to embodiments of the present invention is implemented. Additionally, storage (275) may store a classification model designed based on a convolutional neural network (CNN) for arrhythmia diagnosis.

[0171] According to one embodiment of the present invention, software (280a, 280b) for implementing an artificial intelligence learning method that resides in memory (255) or is stored in storage (275) may be a computer program recorded on a recording medium to execute the steps of: removing a noise region that is determined not to be effective for arrhythmia diagnosis due to noise introduced during the measurement process of a signal in a photoplethysmography (PPG) signal by a processor (250); applying a filter to the photoplethysmography (PPG) signal from which the noise region has been removed so that it has a frequency response within a preset passband by a processor (250); and training a classification model designed for arrhythmia diagnosis based on a convolutional neural network (CNN) based on the photoplethysmography (PPG) signal by a processor (250).

[0172] According to another embodiment of the present invention, software (280a, 280b) for implementing an arrhythmia diagnosis method residing in memory (255) or stored in storage (275) may be a computer program recorded on a recording medium to execute the steps of: a processor (250) receiving a continuously measured photoplethysmography (PPG) signal through a transceiver (260); a step of preprocessing the received photoplethysmography (PPG) signal so that it can be input into a classification model pre-trained for arrhythmia diagnosis based on a convolutional neural network (CNN); and a step of determining the type of arrhythmia of the subject to which the photoplethysmography (PPG) signal was measured based on the output result derived from the classification model after inputting the preprocessed photoplethysmography (PPG) signal into the classification model.

[0173] More specifically, the processor (250) may be configured to include one or more of a Central Processing Unit (CPU), an Application-Specific Integrated Circuit (ASIC), a chipset, and a logic circuit, but is not limited thereto.

[0174] The memory (255) may be configured to include one or more of ROM (Read-Only Memory), RAM (Random Access Memory), flash memory, and memory card, but is not limited thereto.

[0175] The input / output device (260) may be configured to include one or more input devices such as a button, switch, keyboard, mouse, and joystick, and output devices such as an LCD (Liquid Crystal Display), LED (Light Emitting Diode), Organic LED (OLED), Active Matrix OLED (AMOLED), printer, and plotter, but is not limited thereto.

[0176] When the embodiments included in this specification are implemented in software, the above-described method may be implemented as modules (processes, functions, etc.) that each perform the above-described function. Each module may reside in memory (255) and be executed by a processor (250). Memory (255) may exist inside or outside the processor (250) and may be connected to the processor (250) by various well-known means.

[0177] Each component illustrated in FIG. 12 may be implemented by various means (e.g., hardware, firmware, software, or a combination thereof). When implemented by hardware, one embodiment of the present invention may be implemented by one or more ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), DSPDs (Digital Signal Processing Devices), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), processors, controllers, microcontrollers, microprocessors, etc.

[0178] In addition, when implemented by firmware or software, an embodiment of the present invention may be implemented in the form of a module, procedure, function, etc., that performs the functions or operations described above, and may be recorded on a recording medium readable through various computer means. Here, the recording medium may include program instructions, data files, data structures, etc., either alone or in combination.

[0179] The program instructions recorded on the recording medium may be those specifically designed and configured for the present invention, or they may be those known and available to those skilled in the art of computer software. For example, the recording medium includes magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs (Compact Disk Read Only Memory) and DVDs (Digital Video Disks); magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory.

[0180] Examples of program instructions may include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. Such hardware devices may be configured to operate as one or more software to perform the operation of the present invention, and vice versa.

[0182] Hereinafter, the operation of the arrhythmia diagnosis server (200) as described above will be explained in more detail.

[0183] FIG. 13 is a flowchart illustrating an artificial intelligence learning method according to an embodiment of the present invention. FIG. 14 is a flowchart illustrating a process of removing noise according to an embodiment of the present invention.

[0184] As illustrated in FIG. 13, an arrhythmia diagnosis server (200) according to one embodiment of the present invention can collect photoplethysmography (PPG) signals from various sources (S110). Specifically, the arrhythmia diagnosis server (200) can collect photoplethysmography (PPG) signals directly measured by a PPG sensor (100), collect photoplethysmography (PPG) signals stored in an FTP server operated by a medical service association or a medical service professional organization, or collect photoplethysmography (PPG) signals by performing web crawling. Meanwhile, if the data related to the collected photoplethysmography (PPG) signals contains information that can identify who the measurement target (i.e., person) is, the arrhythmia diagnosis server (200) can de-identify such information.

[0185] The arrhythmia diagnosis server (200) can divide a continuously measured photoplethysmography (PPG) signal into multiple photoplethysmography (PPG) signals corresponding to a pre-set interval length (S120). For example, the arrhythmia diagnosis server (200) can divide a continuously measured photoplethysmography (PPG) signal into multiple photoplethysmography (PPG) signals having a length of 10 seconds.

[0186] The arrhythmia diagnosis server (200) can resample each divided photoplethysmography (PPG) signal (S130). Specifically, the arrhythmia diagnosis server (200) can interpolate each divided photoplethysmography (PPG) signal based on a cubic spline function to convert each divided photoplethysmography (PPG) signal to a preset frequency. For example, the arrhythmia diagnosis server (200) can arrange each divided 200 Hz photoplethysmography (PPG) signal to approximate a target resolution of 125 Hz, and then obtain a spline function to be used for interpolation using cubic spline interpolation. Then, the arrhythmia diagnosis server (200) can convert each photoplethysmography (PPG) signal to have a frequency of 125 Hz using the obtained spline function.

[0187] The arrhythmia diagnosis server (200) can mask photoplethysmography (PPG) signals that have a large amount of noise introduced (S140). Specifically, the arrhythmia diagnosis server (200) can mask photoplethysmography (PPG) signals that contain more than a preset number of consecutive values ​​among a plurality of resampled photoplethysmography (PPG) signals. Additionally, the arrhythmia diagnosis server (200) can mask photoplethysmography (PPG) signals that contain undefined values ​​(NAN) among a plurality of resampled photoplethysmography (PPG) signals.

[0188] The arrhythmia diagnosis server (200) can remove photoplethysmography (PPG) signals (i.e., noise) with a frequency distribution that is close to random by using entropy (S150).

[0189] To explain in detail with reference to FIG. 14 for a moment, the arrhythmia diagnosis server (200) may apply a Butterworth filter to each resampled photoplethysmography (PPG) signal as a preliminary signal processing for calculating entropy (S151). For example, the arrhythmia diagnosis server (200) may apply a Butterworth filter of order 3 with a low-cut of 1 or less and a high-cut of 5 or more to each resampled photoplethysmography (PPG) signal to detect a systolic peak.

[0190] The arrhythmia diagnosis server (200) can remove linear components included in each resampled photoplethysmography (PPG) signal (S152). Specifically, the arrhythmia diagnosis server (200) can remove linear components calculated by the least squares method for baseline correction of each photoplethysmography (PPG) signal. The arrhythmia diagnosis server (200) can remove direct current components included in each photoplethysmography (PPG) signal from which linear components have been removed (S153). Such steps for removing linear components (S152) and removing direct current components (S153) may be applied selectively or redundantly depending on the characteristics of the photoplethysmography (PPG) signal to be targeted.

[0191] The arrhythmia diagnosis server (200) can remove noise from each photoplethysmography (PPG) signal from which the DC component has been removed using a Hanning window (S154). Then, the arrhythmia diagnosis server (200) can decompose each photoplethysmography (PPG) signal from which the noise has been removed using a Hanning window into frequency components by performing a Fast Fourier Transform (S155).

[0192] The arrhythmia diagnosis server (200) can convert each photoplethysmography (PPG) signal, which is decomposed into frequency components, into a probability density function by multiplying it by a constant so that the total sum is 1, and calculate the Renny entropy for each photoplethysmography (PPG) signal based on the probability distribution (S156).

[0193] And, the arrhythmia diagnosis server (200) can identify and remove the photoplethysmography (PPG) signal as noise if the entropy calculated for each photoplethysmography (PPG) signal is greater than or equal to a preset threshold value (S157).

[0194] Referring again to FIG. 13, each photoplethysmography (PPG) signal from which noise has been removed can be refined (S160). Specifically, the arrhythmia diagnosis server (200) can apply a filter to the photoplethysmography (PPG) signal from which noise regions have been removed so that it has a flat frequency response within a preset passband. For example, the arrhythmia diagnosis server (200) can apply a high-cut Butterworth filter with an order of 3 and a cut-off criterion of 15.0 and a low-cut Butterworth filter with a cut-off criterion of 0.5 to the photoplethysmography (PPG) signal. Then, the arrhythmia diagnosis server (200) can perform normalization so that the photoplethysmography (PPG) signal to which the Butterworth filter has been applied has a value between a minimum of 0 and a maximum of 1.

[0195] Finally, the arrhythmia diagnosis server (200) can train a classification model based on the preprocessed photoplethysmography (PPG) signals (S170). Specifically, the arrhythmia diagnosis server (200) can train a total of the photoplethysmography (PPG) signal datasets. N n signal data {y1, … , y N }, N Labels corresponding to each of the signal data {label1, … , label N }, the loss function l , convolution weights W , encoder model f When that is the case, the convolutional weights of the classification model can be optimized using the following mathematical formula 10.

[0196] [Mathematical Formula 10]

[0197]

[0198] And, the arrhythmia diagnosis server (200) is the actual probability distribution in the probability space X. P , the probability distribution obtained through prediction Q When let, the following mathematical expression is the loss function l It can be used as.

[0199] [Mathematical Formula 11]

[0200]

[0201] Meanwhile, the arrhythmia diagnosis server (200) can calculate an evaluation value using one or more of the F1 score or balance accuracy based on the type of arrhythmia classified by the classification model and the actual type of arrhythmia of the subject whose photoplethysmography (PPG) signal is measured. In addition, the arrhythmia diagnosis server (200) can learn the classification model by considering the calculated evaluation value of the classification model and using photoplethysmography (PPG) signals collected from various paths by the data collection unit (215).

[0203] FIG. 15 is a flowchart illustrating a method for diagnosing arrhythmia according to one embodiment of the present invention.

[0204] As illustrated in FIG. 15, the arrhythmia diagnosis server (200) can receive photoplethysmography (PPG) signals continuously measured by the PPG sensor (100) (S210).

[0205] The arrhythmia diagnosis server (200) can preprocess the photoplethysmography (PPG) signal received from the PPG sensor (100) so that it can be input into a classification model pre-trained for arrhythmia diagnosis based on a convolutional neural network (CNN) (S220). A detailed description of the process of the arrhythmia diagnosis server (200) preprocessing the photoplethysmography (PPG) signal is identical to the step of dividing the photoplethysmography (PPG) signal (S120) described with reference to FIGS. 13 and 14 up to the step of refining the photoplethysmography (PPG) signal (S170), so it is not described redundantly.

[0206] The arrhythmia diagnosis server (200) can input a preprocessed photoplethysmography (PPG) signal into a pre-trained classification model (S230). Specifically, the arrhythmia diagnosis server (200) can input the output values ​​of the classification model into {x1, … , x N When} is used, the probability of corresponding to the i-th arrhythmia type can be calculated using the following mathematical formula 15 regarding softmax.

[0207] [Mathematical Formula 15]

[0208]

[0209] The arrhythmia diagnosis server (200) may determine the final arrhythmia type by applying test time point augmentation to improve the accuracy of classification (i.e., determination of the arrhythmia type) (S240). Specifically, the arrhythmia diagnosis server (200) can identify multiple photoplethysmography (PPG) signals having the same unit length, with different starting times set to have overlapping regions from photoplethysmography (PPG) signals that are continuously measured in a time series. Then, the arrhythmia diagnosis server (200) can finally determine the type of arrhythmia of the subject in which the photoplethysmography (PPG) signal was measured, based on the average of the output results of the classification model for the multiple identified photoplethysmography (PPG) signals.

[0210] More specifically, the arrhythmia diagnosis server (200) determines the probability value output from the classification model that the photoplethysmography (PPG) signal measured after t seconds corresponds to type a among set A of arrhythmia types. p t (a) Let the output result be from the previously derived photoplethysmography (PPG) signal N Total measured every second up to -1 second N Based on the output results of the above classification model for the photoplethysmography (PPG) signals, the type of arrhythmia pred of the subject can be determined using the following mathematical formula 16.

[0211] [Mathematical Formula 16]

[0212]

[0214] As described above, preferred embodiments of the present invention have been disclosed in this specification and drawings; however, it is obvious to those skilled in the art that other variations based on the technical spirit of the present invention are possible in addition to the embodiments disclosed herein. Furthermore, although specific terms have been used in this specification and drawings, they are used merely in a general sense to facilitate the explanation of the technical content of the present invention and to aid in understanding the invention, and are not intended to limit the scope of the present invention. Accordingly, the detailed description above should not be interpreted restrictively in any respect and should be considered illustrative. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the present invention are included within the scope of the present invention. Explanation of the symbols

[0215] 100: PPG Sensor 200: Arrhythmia Diagnosis Server 205 : Communication unit 210 : Input / Output unit 215: Data acquisition unit 220: Signal preprocessing unit 225 : Artificial Intelligence Learning Unit 230 : Arrhythmia Diagnosis Unit

Claims

Claim 1 An artificial intelligence learning method using an arrhythmia diagnosis server comprising a memory, a transceiver, and a processor for processing instructions residing in said memory, wherein the arrhythmia diagnosis server comprises: a step of dividing a continuously measured photoplethysmography (PPG) signal into a plurality of photoplethysmography (PPG) signals corresponding to a preset interval length; a step of removing a noise region among the divided plurality of photoplethysmography (PPG) signals that is determined to be invalid for arrhythmia diagnosis due to noise introduced during the signal measurement process; a step of the arrhythmia diagnosis server applying a filter to the plurality of photoplethysmography (PPG) signals from which the noise region has been removed so that they have a frequency response within a preset pass band; and a step of the arrhythmia diagnosis server, based on the plurality of photoplethysmography (PPG) signals to which the filter has been applied, a classification model designed for arrhythmia diagnosis based on a Convolutional Neural Network (CNN). A training step; and a step in which the arrhythmia diagnosis server diagnoses an arrhythmia through the trained classification model;The step of removing the noise region includes comparing the array of the plurality of photoplethysmography (PPG) signals by shifting it one step at a time, and masking the photoplethysmography (PPG) signals among the plurality of photoplethysmography (PPG) signals that include a flat line with identical values ​​for more than a preset number of consecutive sections, or photoplethysmography (PPG) signals that include undefined values ​​(not a number); the step of diagnosing includes receiving continuously measured photoplethysmography (PPG) signals, inputting the photoplethysmography (PPG) signals into the classification model, and determining the type of arrhythmia possessed by the subject to which the photoplethysmography (PPG) signals were measured based on the output result derived from the classification model, wherein a plurality of identical unit lengths with different starting points set to have overlapping regions from the photoplethysmography (PPG) signals continuously measured in a time series An artificial intelligence learning method characterized by identifying photoplethysmography (PPG) signals and finally determining the type of arrhythmia possessed by the subject in which the photoplethysmography (PPG) signal is measured, based on the average of the output results of the classification model for the identified plurality of photoplethysmography (PPG) signals. Claim 2 An artificial intelligence learning method according to claim 1, further comprising the step of converting each of the divided plurality of photoplethysmography (PPG) signals into a preset frequency by interpolating based on a cubic spline function prior to the step of removing the noise region. Claim 3 delete Claim 4 An artificial intelligence learning method according to claim 1, wherein the step of removing the noise region is characterized by removing a linear component included in each of the divided plurality of photoplethysmography (PPG) signals, or removing a direct current component included in each photoplethysmography (PPG) signal from which the linear component has been removed. Claim 5 delete Claim 6 delete Claim 7 delete Claim 8 delete Claim 9 delete Claim 10 delete

Images