Method and system for supporting analysis of electrocardiogram signals

Abstract

According to one aspect of the present invention, provided is a method for supporting analysis of electrocardiogram signals of a subject, the method comprising the steps of: generating symbolic bits and a plurality of unit bits on the basis of electrocardiogram signals acquired from a subject; and inputting the generated symbolic bits and plurality of unit bits into a personalized analysis model so as to infer whether each of the plurality of unit bits corresponds to an abnormal bit.

Description

Method and system for supporting the reading of electrocardiogram signals

[0001] The present invention relates to a method and system for supporting the reading of an electrocardiogram signal.

[0002] Atrial fibrillation (AF) is one of the most common forms of cardiac arrhythmia and requires proper diagnosis and management. Based on its onset pattern and duration, atrial fibrillation can be classified into newly diagnosed, paroxysmal, persistent, long-persistent, and permanent types. Among these, paroxysmal AF is a type that is particularly difficult to detect and evaluate due to its high frequency. Paroxysmal AF is characterized by intermittent occurrence and a return to a normal rhythm after a relatively short duration, making diagnosis difficult. Furthermore, accurate diagnosis is further hampered by the fact that many patients do not recognize their symptoms or do not visit a medical institution when symptoms occur.

[0003] Recently, long-term cardiac monitoring (LTCM) has garnered attention as a more effective method for diagnosing paroxysmal atrial fibrillation than Holter monitors and other portable cardiac monitoring devices. However, when analyzing electrocardiogram (ECG) signal data collected over a long period using deep learning-based models, the sheer volume of data leads to a high frequency of errors caused by various factors such as noise and signal distortion. Consequently, correcting these errors requires significant time, resulting in a degradation of the overall model's performance. Furthermore, the aforementioned conventional deep learning models and algorithms do not account for the unique characteristics of ECG waveforms specific to each individual, which has led to a problem of low prediction accuracy. Consequently, there is a strong industry demand for the development of a technique that can account for individual ECG waveform characteristics while accurately detecting atrial fibrillation, particularly paroxysmal atrial fibrillation, in long-term ECG signals collected using LTCM devices.

[0004] Accordingly, the inventor(s) propose a technology that can maximize the diagnostic accuracy of atrial fibrillation by increasing the efficiency of reading electrocardiogram signals using a deep learning model designed to reference an individual's unique electrocardiogram waveform.

[0005] The present invention aims to solve all the problems of the aforementioned prior art.

[0006] In addition, the present invention has another objective of generating a symbol bit and a plurality of unit bits based on an electrocardiogram signal comprising a plurality of bits obtained from a subject, and inputting the generated symbol bit and a plurality of unit bits into a personalized reading model to infer whether each of the plurality of unit bits corresponds to an abnormal bit.

[0007] In addition, another objective of the present invention is to provide a personalized electrocardiogram signal reading model that improves the classification accuracy of abnormal signals included in an electrocardiogram signal by generating symbol bits representing the subject's unique electrocardiogram signal characteristics based on an electrocardiogram signal acquired from the subject, and utilizing these bits during the learning process of the electrocardiogram signal reading model.

[0008] A representative configuration of the present invention for achieving the above objective is as follows.

[0009] According to one aspect of the present invention, a method is provided comprising the steps of generating a symbol bit and a plurality of unit bits based on an electrocardiogram signal obtained from a subject, and inputting the generated symbol bit and the plurality of unit bits into a personalized reading model to infer whether each of the plurality of unit bits corresponds to an abnormal bit.

[0010] According to another aspect of the present invention, a system is provided comprising a bit generation unit that generates a symbol bit and a plurality of unit bits based on an electrocardiogram signal from a subject, and an inference unit that inputs the generated symbol bit and a plurality of unit bits into a personalized reading model and infers whether each of the plurality of unit bits corresponds to an abnormal bit.

[0011] In addition to this, other methods for implementing the present invention, other systems, and non-transient computer-readable recording media for recording a computer program for executing said methods are further provided.

[0012] According to the present invention, a symbol bit and a plurality of unit bits are generated based on an electrocardiogram signal comprising a plurality of bits obtained from a subject, and the generated symbol bit and a plurality of unit bits are input into a personalized reading model to infer whether each of the plurality of unit bits corresponds to an abnormal bit.

[0013] In addition, according to the present invention, a symbol bit representing the unique electrocardiogram signal characteristics of a subject is generated based on an electrocardiogram signal obtained from a subject, and by utilizing this in the learning process of an electrocardiogram signal reading model, it is possible to provide a personalized electrocardiogram signal reading model that improves the classification accuracy of abnormal signals included in the electrocardiogram signal.

[0014] FIG. 1 is a diagram showing the schematic configuration of an overall system for supporting the reading of an electrocardiogram signal of a subject according to one embodiment of the present invention.

[0015] FIG. 2 is a drawing illustrating in detail the internal configuration of a reading support system (200) according to one embodiment of the present invention.

[0016] FIG. 3 is a diagram illustrating the structure of a personalized reading model (400) according to one embodiment of the present invention.

[0017] FIG. 4 is a diagram exemplarily illustrating the process of generating a symbol bit according to one embodiment of the present invention.

[0018] <Explanation of Symbols>

[0019] 100: Communication network

[0020] 200: Read support system

[0021] 210: Bit generation unit

[0022] 220: Inference section

[0023] 230: Communications Department

[0024] 240: Control unit

[0025] 300: Device

[0026] The following detailed description of the invention refers to the accompanying drawings, which illustrate specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that various embodiments of the invention are different but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be modified from one embodiment to another without departing from the spirit and scope of the invention. It should also be understood that the location or arrangement of individual components within each embodiment may be modified without departing from the spirit and scope of the invention. Accordingly, the following detailed description is not meant to be limiting, and the scope of the invention should be understood to encompass the scope claimed by the claims and all equivalents thereof. Similar reference numerals in the drawings indicate identical or similar components across various aspects.

[0027] Hereinafter, in order to enable a person skilled in the art to easily practice the present invention, various preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

[0028] In this specification, embodiments relating to a service supporting the reading of biosignals are described with a primary focus on reading whether each of at least one unit bit included in an electrocardiogram signal is an abnormal signal; however, the reading support service in the present invention should be understood as the broadest concept, encompassing reading support services for all types of biosignals. That is, the reading support service and the method of generating and utilizing symbol bits in the present invention can be applied to all types of biosignals in which individual differences exist among subjects, such as electrocardiogram (ECG) signals, as well as electroencephalogram (EEG), electromyogram (EMG), phonocardiogram (PCG), respiratory curves, photoplethysmography (PPG), and blood pressure waveforms (ABP).

[0029] Configuration of the entire system

[0030] FIG. 1 is a diagram showing the schematic configuration of an overall system for supporting the reading of an electrocardiogram signal of a subject according to one embodiment of the present invention.

[0031] As illustrated in FIG. 1, the entire system according to one embodiment of the present invention may include a communication network (100), a reading support system (200), and a device (300).

[0032] First, a communication network (100) according to one embodiment of the present invention can be configured regardless of the mode of communication, such as wired communication or wireless communication, and can be configured as various communication networks such as a Local Area Network (LAN), a Metropolitan Area Network (MAN), or a Wide Area Network (WAN). Preferably, the communication network (100) referred to in this specification may be the known Internet or the World Wide Web (WWW). However, the communication network (100) may include at least a known wired / wireless data communication network, a known telephone network, or a known wired / wireless television communication network, without being limited thereto.

[0033] For example, the communication network (100) may be a wireless data communication network and may implement conventional communication methods such as WiFi communication, WiFi-Direct communication, Long Term Evolution (LTE) communication, 5G communication, Bluetooth communication (including Bluetooth Low Energy (BLE) communication), infrared communication, ultrasonic communication, etc., in at least a part thereof. As another example, the communication network (100) may be an optical communication network and may implement conventional communication methods such as Light Fidelity (LiFi), etc., in at least a part thereof.

[0034] Next, a reading support system (200) according to one embodiment of the present invention can generate a symbolic beat and a plurality of unit beats based on an electrocardiogram signal including a plurality of beats obtained from a subject, and input the generated symbolic beat and a plurality of unit beats into a personalized reading model (400) to perform a function of inferring whether each of the plurality of unit beats corresponds to an abnormal beat.

[0035] The configuration and function of the reading support system (200) according to the present invention will be examined in detail through the following detailed description.

[0036] Next, the device (300) according to one embodiment of the present invention is a digital device that includes a function to communicate after connecting to a reading support system (200), and any digital device equipped with memory means and equipped with a microprocessor to have computational capabilities, such as a smartphone, tablet, smart watch, smart band, smart glasses, desktop computer, laptop computer, workstation, PDA, web pad, mobile phone, etc., can be adopted as the device (300) according to the present invention.

[0037] In particular, the device (300) may include an application (not shown) that enables a user to receive services such as electrocardiogram signal reading from the reading support system (200). Such an application may be downloaded from the reading support system (200) or an external application distribution server (not shown). Meanwhile, the nature of such an application may generally be similar to the bit generation unit (210), inference unit (220), communication unit (230), and control unit (240) of the reading support system (200) as described below. Here, at least a part of the application may be replaced with a hardware device or firmware device capable of performing substantially the same or equivalent functions as needed.

[0038] Configuration of the reading support system

[0039] Below, we will examine the internal configuration of the reading support system (200) that performs important functions for the implementation of the present invention and the functions of each component.

[0040] FIG. 2 is a drawing illustrating in detail the internal configuration of a reading support system (200) according to one embodiment of the present invention.

[0041] As illustrated in FIG. 2, a read support system (200) according to one embodiment of the present invention may be configured to include a bit generation unit (210), an inference unit (220), a communication unit (230), and a control unit (240). According to one embodiment of the present invention, the bit generation unit (210), the inference unit (220), the communication unit (230), and the control unit (240) may be program modules, at least some of which communicate with an external system (not shown). Such program modules may be included in the read support system (200) in the form of an operating system, an application program module, or other program modules, and may be physically stored in various known storage devices. Additionally, such program modules may be stored in a remote storage device capable of communicating with the read support system (200). Meanwhile, such program modules encompass, but are not limited to, routines, subroutines, programs, objects, components, data structures, etc., that perform specific tasks or execute specific abstract data types as described below according to the present invention.

[0042] Meanwhile, although the reading support system (200) has been described as above, this description is exemplary, and it is obvious to those skilled in the art that at least some of the components or functions of the reading support system (200) may be realized within a device (300) or server (not shown) or included within an external system (not shown) as needed.

[0043] First, a bit generation unit (210) according to one embodiment of the present invention can perform the function of generating a symbol bit and a plurality of unit bits based on an electrocardiogram signal obtained from a subject.

[0044] A subject according to one embodiment of the present invention may refer to a subject who provides an electrocardiogram signal to be read through a reading support system (200) according to one embodiment of the present invention. Specifically, the subject may refer to a subject who measures an electrocardiogram signal, and may refer to a patient or a health monitoring subject who measures an electrocardiogram signal through an electrocardiogram measuring device (electrocardiogram sensor).

[0045] An electrocardiogram (ECG) signal according to one embodiment of the present invention is a biological signal that records the electrical activity of the heart over time, and may refer to data expressed as a waveform by measuring the electrical signal generated during a heartbeat.

[0046] An electrocardiogram signal according to one embodiment of the present invention may be measured by an electrocardiogram measurement method comprising multiple electrodes or a single electrode. Specifically, an electrocardiogram signal according to one embodiment of the present invention may be measured through at least one method among a standard 12-lead ECG, a 3 / 5 / 6-lead ECG, a Holter monitor, long-term cardiac monitoring (LCTM), and a single-lead ECG.

[0047] More specifically, an electrocardiogram signal according to one embodiment of the present invention can be measured by a single electrode LTCM method. For example, a reading support system (200) according to one embodiment of the present invention can perform or support reading of an electrocardiogram signal measured through an electrocardiogram signal based on an attached electrocardiogram measuring device (or, wearable device).

[0048] Meanwhile, the long-term cardiac monitoring method described above may refer to a system or technology that continuously or event-based monitors a patient's cardiac electrical signals (electrocardiogram, ECG) over a long period (e.g., several days to several months).

[0049] A beat according to one embodiment of the present invention is a unit representing a cycle in which the heart contracts and relaxes once, and may refer to a waveform generated according to a single cardiac activity composed of a P wave, a QRS complex, and a T wave, etc., in an electrocardiogram signal, or a time interval containing said waveform. Specifically, the above-described beat may be defined as a part of the electrocardiogram signal comprising a continuous P wave, a QRS complex, and a T wave, and thus the electrocardiogram signal may contain a plurality of beats.

[0050] Meanwhile, a bit according to one embodiment of the present invention may refer to a bit included in an unprocessed original electrocardiogram signal (i.e., raw data) when compared with a unit bit or symbol bit to be described later.

[0051] Meanwhile, even if the bit according to one embodiment of the present invention or the unit bit described below is of the same type, the characteristics may vary depending on the subject. Specifically, such differences among subjects (i.e., unique individual characteristics) may occur depending on physical characteristics that differ from person to person, such as the location and size of the heart, body structure, and body mass index (BMI). Therefore, it may be important to consider the aforementioned individual differences in order to accurately interpret an electrocardiogram signal or to determine whether it corresponds to an abnormal bit.

[0052] A symbol bit according to one embodiment of the present invention is a bit generated based on an electrocardiogram signal measured from a subject, and may represent a representative bit capable of reflecting the subject's unique electrocardiogram signal characteristics (e.g., morphological characteristics of the electrocardiogram signal). The above-described symbol bit can be obtained through a process of extracting a continuous portion of bits from an individual subject's electrocardiogram signal and processing the extracted bits.

[0053] To continue, a symbol bit according to one embodiment of the present invention can be generated by selecting a consecutive portion of unit bits among the aforementioned plurality of unit bits, averaging the selected consecutive portion of unit bits, and adjusting their size.

[0054] To continue, processing a series of unit bits according to one embodiment of the present invention may mean removing outliers from the selected series of unit bits, averaging a plurality of unit bits from which outliers have been removed to obtain a single unit bit, removing a residual T-wave or a residual P-wave from the obtained single unit bit, and then adjusting the overall size.

[0055] Specifically, the aforementioned symbol bit can be obtained by performing at least one of the following processes: extracting a continuous portion of bits from the total number of bits included in the electrocardiogram signal measured from the subject, selecting bits that satisfy a predetermined criterion for the acquired portion of bits, and performing averaging.

[0056] Referring to FIG. 4, the process of acquiring the aforementioned symbol bit may include the following steps.

[0057] (1) Extraction process (510): In this process, a series of bits can be extracted from an electrocardiogram signal obtained from a subject. Specifically, the size of the series of bits to be extracted can be defined as a predetermined time or a predetermined number of bits. Meanwhile, as described above, the size of the bits to be extracted can be limited to a predetermined level, taking into account that the unique characteristics of the waveform may change when measuring an electrocardiogram signal for a long period of time even within a single subject. For example, the predetermined level limited as described above can be a series of bits with a length of 1 hour or 3,600 series of bits.

[0058] (2) Selection process (520): In this process, multiple bits to be used for generating symbol bits can be selected by removing outliers from the consecutive bits extracted as described above. Specifically, the characteristics of each bit can be analyzed, and bits whose characteristics differ from a reference value (e.g., mean) by a predetermined level or more can be defined as outliers and removed. Specifically, the characteristics of the bits described above may be at least one of the RR interval and amplitude, but are not limited thereto. For example, bits with an RR interval of less than 20 percentile and greater than 80 percentile may be removed as outliers, or bits with a maximum amplitude exceeding 130% of the total bits may be removed as outliers (it is also possible to remove only bits satisfying both conditions as outliers).

[0059] (3) Alignment process (not shown): In this process, a plurality of selected bits can be aligned to a window of the same predetermined length as described above. Specifically, the plurality of bits described above can be aligned based on an R wave (or R peak). For example, the predetermined length described above may be 2 seconds, but is not limited thereto.

[0060] (4) Averaging process (530): In this process, multiple bits aligned as described above can be averaged. The averaging described above can be performed through an "ensemble averaging" method, which averages the multiple bits aligned at corresponding points (e.g., the same point in time for each bit) to create one representative bit.

[0061] (5) Size adjustment process (540): In this process, residual P waves or T waves can be removed. Specifically, when aligning multiple beats, the positions of the P wave, QRS, and T wave may not be perfectly aligned, such as when the heart rate is fast and the P wave of one beat overlaps with the T wave of the previous beat, and thus outliers may occur during the averaging process. Therefore, the occurrence of such outliers can be prevented through the process of removing residual P waves or T waves. Specifically, the above-described process of removing residual P waves or T waves can be performed by smoothly attenuating the signal values ​​at both ends of the beat obtained through averaging. The above-described attenuation method can be performed through various windowing methods, and windowing methods include, for example, a Hamming window and a Tukey window, but are not limited thereto. Next, size adjustment can be performed on the beat from which the residual P waves or T waves have been removed so that the symbol beat has a predetermined time length. Specifically, the size can be adjusted by zero-padding the insufficient section in response to the length of the bit from which the residual P-wave or T-wave has been removed being shorter than a predetermined time length, or by removing the excess section in response to the length of the bit from which the residual P-wave or T-wave has been removed being longer than a predetermined time length. The aforementioned predetermined time length may be, for example, 2 seconds, but is not limited thereto.

[0062] To continue, a plurality of unit bits according to one embodiment of the present invention can be generated by dividing each bit based on the R peak of each bit included in the electrocardiogram signal and adjusting the size of each divided bit.

[0063] A unit beat according to one embodiment of the present invention may refer to a plurality of bits obtained by processing each of the plurality of bits included in an electrocardiogram signal obtained from a subject in a predetermined manner. Specifically, the unit beat may be obtained by normalizing the aforementioned plurality of bits so that each of them has a fixed length. As described below, the aforementioned unit beat may refer to a minimum unit for determining whether a bit is abnormal.

[0064] Specifically, the aforementioned normalization can be performed by extracting the R-wave from the previous R-wave to the immediately following R-wave from the electrocardiogram signal, using the R-wave included in the extracted electrocardiogram signal as a reference point to align the R-wave of the cut section so that it is positioned at the center of the token, and then, in response to the length of the cut electrocardiogram signal being shorter than a predetermined fixed length, adjusting the length of the insufficient portion through zero-padding, and in response to the length of the cut electrocardiogram signal exceeding the predetermined fixed length, removing the excess portion. That is, while the aforementioned symbol bit obtains a single symbol bit from multiple bits, the unit bit can obtain a single unit bit by processing a single bit.

[0065] The fixed length of the unit bit described above can be experimentally determined as a length that can effectively extract the complexity of the waveform while maintaining the uniformity of the entire data set. For example, the predetermined fixed length described above may be 2 seconds, but is not limited thereto.

[0066] Next, the inference unit (220) according to one embodiment of the present invention can input the generated symbol bit and a plurality of unit bits into a personalized reading model (400) and perform the function of inferring whether each of the plurality of unit bits corresponds to an ideal bit.

[0067] An abnormal beat according to one embodiment of the present invention may refer to a beat (or unit beat) that exhibits abnormal characteristics in terms of shape, duration, rhythm interval (RR interval), and amplitude when compared to a beat of a general (or normal) electrocardiogram signal.

[0068] Specifically, the types of abnormal beats described above can be classified into rhythmic abnormalities, premature beats, conduction abnormalities, and waveform abnormalities.

[0069] Abnormal beats caused by the aforementioned rhythm abnormalities include, for example, atrial fibrillation (AF), an arrhythmia in which the RR interval is completely irregular; atrial flutter (AFL), an arrhythmia in which the atria contract rapidly (e.g., 250 to 400 times per minute) regularly; and sinus arrhythmia, an arrhythmia in which the RR interval changes periodically with respiration, but are not limited thereto.

[0070] Abnormal beats caused by the aforementioned premature beats include, for example, premature atrial contraction (PAC) in which the QRS wave is normal while the P wave occurs prematurely, and premature ventricular contraction (PVC) in which a wide abnormal QRS wave occurs, but are not limited thereto.

[0071] Abnormal beats caused by the aforementioned conduction abnormalities include, for example, an atrioventricular block (AV block) exhibiting an abnormal PR interval and a branch block (BBB) ​​exhibiting an abnormal QRS waveform, but are not limited thereto.

[0072] Abnormal beats caused by the waveform abnormalities described above include, for example, T-wave inversions showing abnormal T-waves bent downward, ST abnormalities where the interval between the end of the QRS wave and the beginning of the T wave is shifted upward or downward, and high-amplitude QRS waves where the amplitude of the QRS wave is abnormally large, but are not limited thereto.

[0073] More specifically, the abnormal beat according to one embodiment of the present invention may be atrial fibrillation (AF). The atrial fibrillation described above may be classified into paroxysmal atrial fibrillation (paroxysmal AF), persistent atrial fibrillation (persistent AF), long-standing persistent atrial fibrillation (long-standing persistent AF), and permanent atrial fibrillation (permanent AF).

[0074] More specifically, the abnormal beat according to one embodiment of the present invention may be paroxysmal atrial fibrillation (paroxysmal AF). The aforementioned paroxysmal AF may be difficult to diagnose because it occurs intermittently and naturally returns to a normal rhythm after lasting for a short period of time.

[0075] However, the ideal bit according to one embodiment of the present invention is not limited to those listed above and may be varied within the scope of achieving the purpose of the present invention.

[0076] Meanwhile, a unit bit according to one embodiment of the present invention may be the minimum unit for determining whether a specific bit corresponds to an abnormal bit. That is, the fact that the reading support system (200), the inference unit (220), or the personalized reading model (400) according to one embodiment of the present invention infers whether each of the plurality of unit bits corresponds to an abnormal bit may mean that for each of the plurality of unit bits (i.e., beat-wise), the question of whether it corresponds to an abnormal bit or the probability of it corresponding to an abnormal bit is calculated independently.

[0077] Continuing, the inference unit (220) according to one embodiment of the present invention can independently calculate the probability that each of the plurality of unit bits corresponds to an ideal bit.

[0078] Specifically, the inference unit (220) according to one embodiment of the present invention can quantitatively express (calculate) the probability that a specific unit bit corresponds to an abnormal bit of a specific type (e.g., atrial fibrillation). Meanwhile, the inference unit (220) according to one embodiment of the present invention can determine whether a specific unit bit corresponds to an abnormal bit of a specific type based on the probability derived as described above. For example, if the calculated probability of corresponding to an abnormal bit is equal to or greater than a predetermined threshold, a result indicating that it corresponds to an abnormal bit of a specific type can be calculated, or if the calculated probability of corresponding to an abnormal bit is equal to or less than a predetermined threshold, a result indicating that it does not correspond to an abnormal bit of a specific type can be calculated. The aforementioned predetermined threshold may be changed by considering various factors such as the performance of the model (accuracy, sensitivity, and specificity, etc.), the patient's condition, the monitoring environment, and the medical environment, but the factors that change the threshold are not limited to those listed above.

[0079] Continuing, a personalized reading model (400) according to one embodiment of the present invention may include a feature extraction unit (410) that extracts features from a symbol bit and a plurality of unit bits, an output unit (420) that generates an output value based on information related to the extracted features, and a result unit (430) that generates a result value based on the generated output value.

[0080] A feature extraction unit (410) according to one embodiment of the present invention may be a layer that can be included in a deep learning model (e.g., the personalized reading model (400) described above), and may mean a layer that performs the function of automatically extracting meaningful patterns or attributes (features) from input data and expressing them so that a subsequent layer can perform higher-dimensional tasks such as classification or prediction based on them.

[0081] Meanwhile, the input data (411) that can be input to the feature extraction unit (410) described above may be at least one of the symbol bit and unit bit described above. Specifically, the input data (411) that can be input to the feature extraction unit (410) described above may be one symbol bit and a plurality of unit bits. Meanwhile, one symbol bit included in the input data (411) described above may be generated one by one for each subject.

[0082] Meanwhile, symbol bits are input data without labels, and unlike unit bits, they may not be assigned a classification target (label). That is, symbol bits are not objects for which the presence or absence of an anomalous bit or the probability of corresponding to an anomalous bit is determined, and can be used only as reference information. Furthermore, when the personalized reading model (400) according to one embodiment of the present invention updates weights through a loss function during the learning process, symbol bits may not be involved in backpropagation as they are not subject to weight updates. In other words, symbol bits are not subject to modification during the learning process and can be used only as fixed reference values. By using symbol bits through the method described above, electrocardiogram signal characteristics unique to the subject can be referenced through the symbol bits.

[0083] Continuing, a feature extraction unit (410) according to one embodiment of the present invention may include a residual neural network.

[0084] A residual neural network according to one embodiment of the present invention may refer to a deep learning structure that includes a residual connection, which is a type of convolutional neural network (CNN) that skips the input and adds it to the output instead of simply passing the input to the next layer. The above-described residual neural network can enable effective learning even in deep neural network structures by mitigating the vanishing gradient problem.

[0085] Meanwhile, the feature extraction unit (410) according to one embodiment of the present invention can perform an embedding function by applying a pooling operation (e.g., global mean pooling, etc.) to the extracted feature (high-dimensional feature map) as described above and converting it into a low-dimensional vector.

[0086] Specifically, the residual neural network according to one embodiment of the present invention may be a ResNet-34. The above-described ResNet-34 may refer to a medium-depth CNN model composed of a total of 34 layers (mainly convolutional layers). More specifically, the residual neural network according to one embodiment of the present invention may be a 1-D ResNet-34. The above-described 1-D ResNet-34 may refer to a residual neural network that uses a one-dimensional signal (e.g., an electrocardiogram signal with only a time axis) as input data and performs a convolutional operation to extract local patterns while sliding a kernel (filter) in a one-dimensional direction. Meanwhile, sliding the above-described kernel in a one-dimensional direction may mean moving the kernel only in the direction of the time axis in time-series data (e.g., an electrocardiogram signal).

[0087] Specifically, the above-described residual neural network, ResNet-34, or 1-D ResNet-34 can extract multiple features by using multiple types of filters, each of which outputs a separate channel as a filter result. That is, one filter can output one filter result (output channel), and multiple filter results can be combined to produce one high-dimensional vector (i.e., the information extracted by each filter is summarized into a single number, and the numbers are combined to form one high-dimensional embedding vector), and this high-dimensional vector can be input to the output unit (420) to be described later or to the transformer model included in the output unit (420).

[0088] Meanwhile, according to one embodiment of the present invention, all input data (411) (e.g., one symbol bit and a plurality of unit bits) input to the feature extraction unit (410) may share the weights of a single identical layer (e.g., a residual neural network) included in the feature extraction unit (410). That is, a single identical feature extraction layer may be repeatedly applied to all input data (411) (e.g., each of the plurality of unit bits included in the electrocardiogram signal that is the input data), and accordingly, only one residual neural network is trained during the learning process, and the same residual neural network may be applied to all input data (411). Through this, features can be extracted based on consistent criteria, and the model can be made lighter by sharing operations. Meanwhile, as described above, the learning of the residual neural network may mean learning which features to extract to be most useful for reducing loss, instead of extracting features in a fixed manner.

[0089] To continue, information related to a feature according to one embodiment of the present invention may refer to information generated by tokenizing a symbol bit and a plurality of unit bits and then extracting a feature, or information generated by first extracting a feature and then tokenizing the extracted feature.

[0090] Tokenization according to one embodiment of the present invention may mean dividing the entire time series data (input data) into small and meaningful interval units. Specifically, the small and meaningful interval units described above may refer to interval units having a length that can be processed (with high processing efficiency) by the output unit (420) or the transformer model included in the output unit (420) described later. In the case of the transformer model described above, since the amount of computation (computational complexity) increases squarely with respect to the length of the input data, by adjusting the length of the input data through tokenization, memory shortage can be prevented and computational efficiency can be increased.

[0091] Meanwhile, the aforementioned tokenization may be performed by first performing tokenization on the input data and then extracting features from the obtained tokens, or by first extracting features from the input data and then tokenizing the extracted features.

[0092] Specifically, a method of first performing tokenization on input data and then extracting features from the acquired tokens may mean a method of performing tokenization by dividing time-series data (e.g., electrocardiogram signals) into meaningful units (e.g., bits or time windows) and applying the aforementioned residual neural network to each divided segment to extract features, thereby using the vector extracted for each segment as a single token.

[0093] Furthermore, the method of first extracting features from input data and then tokenizing the extracted features may refer to a method of applying the aforementioned residual neural network to the entire time series data to extract vectors, and performing tokenization by dividing the extracted vectors (feature maps) into certain units and using each unit obtained as a token.

[0094] Continuing, the output unit (420) according to one embodiment of the present invention may include at least one transformer. The transformer described above may be an expert mixing transformer or a general transformer.

[0095] A transformer according to one embodiment of the present invention can mean a deep learning model capable of calculating the relationships between all elements within an input sequence (input data) at once through an attention mechanism, and capable of parallel processing and long-range dependency learning.

[0096] The aforementioned attention mechanism can refer to a method of regulating the flow of information by examining all locations in an input sequence at once and dynamically calculating the interrelationships (importance) between each input. By utilizing the attention mechanism, the Transformer model can easily learn the relationships between distant components within the input data and possess various advantages, such as fast learning speeds, as it can process the entire sequence of the input data in parallel and simultaneously.

[0097] Specifically, the output unit (420) according to one embodiment of the present invention can refer to the patient's unique ECG shape information within the relationships between the components (i.e., the tokens described above) included in the temporally sequential input data by using a transformer model, thereby enabling more accurate analysis. That is, information regarding the unique shape of the electrocardiogram waveform for each subject, information regarding the relationship between each waveform appearing in a continuous electrocardiogram waveform, and information regarding differences in the electrocardiogram waveform caused by minute differences in the measurement environment, such as minute differences in electrode position, may vary. The output unit (420) according to one embodiment of the present invention can refer to the subject's unique information (i.e., the personal unique characteristics and minute environmental differences described above) by learning the symbol bits and unit bits through a transformer model, thereby maximizing the accuracy of the electrocardiogram signal reading.

[0098] Specifically, an output unit (420) according to one embodiment of the present invention may include one mixture of experts (MoE) transformer and one general transformer.

[0099] An expert blending transformer according to one embodiment of the present invention may refer to a high-efficiency transformer model that selectively activates a customized path in response to input data by introducing an expert blending layer into a transformer model. Specifically, the expert blending transformer model may refer to a deep learning model that increases expressiveness while maintaining computational power by replacing the FFN layer included in a general transformer model with multiple experts and selectively activating only some experts for each input. By using such an expert blending transformer, the capacity to process various and complex patterns within the data can be improved.

[0100] Specifically, the experts to be used may be selected using a gating network that applies a top-k mechanism. More specifically, the gating network described above may select experts by calculating a relevance score for all experts corresponding to each input data and activating only the top k experts in order of highest relevance scores. More specifically, the total number of experts used may be 8, and the k described above may be 2 (i.e., selecting the top 2 experts based on relevance scores), but matters such as the method of selecting experts, the total number of experts, and the number of experts selected may be changed without being limited to these.

[0101] Specifically, in accordance with one embodiment of the present invention, the output unit (420) includes one mixture of experts (MoE) transformer and one general transformer, and the general transformer can use the output value of the mixture of experts transformer as an input value.

[0102] That is, referring to FIG. 3, an output unit (420) according to one embodiment of the present invention may include one expert mixed transformer model (421) that receives a feature (412) extracted from a feature extraction unit (410) and one general transformer model (422) that uses the output value of the expert mixed transformer described above as an input value.

[0103] Continuing with reference to FIG. 3, the extracted features (412) can be processed sequentially by an expert mixed transformer model (421) and a general transformer model (422). First, the expert mixed transformer model can increase computational efficiency by dynamically selecting and utilizing experts with high relevance to each extracted feature (412). Meanwhile, since the general transformer model (422) uses one symbol bit and multiple unit bits as training data or input data, it can refer not only to the temporal relationship between the bits included in the multiple unit bits but also to the unique information of the subject included in the symbol bit. Meanwhile, if the unique information about the subject included in the aforementioned symbol bit is not useful (i.e., does not help reduce loss), the personalized reading model (400) or output unit (420) according to one embodiment of the present invention may refer to this unique information less.

[0104] Meanwhile, at least one of the expert mixed transformer model (421) and the general transformer model (422) according to one embodiment of the present invention may use pre-layer normalization. The aforementioned pre-layer normalization may refer to a method of applying layer normalization to the input of a block, and may improve learning stability, learning speed, and convergence performance.

[0105] Continuing, the result unit (430) according to one embodiment of the present invention can generate a result value (431) based on the output value (423) generated by the output unit (420). Specifically, the result unit (430) according to one embodiment of the present invention can generate a result value (431) by linearly transforming the output value (423) generated by the output unit (420) described above, and then converting it into a probability value through an activation function.

[0106] Meanwhile, the output value (423) described above may refer to output data generated by at least one transformer model included in the output unit (420) described above based on input data (411) or extracted features (412).

[0107] Specifically, the above-described result unit (430) may exist only once. That is, the same result unit (430) can be applied collectively to the output value for each unit bit (i.e., by applying the same linear transformation function and activation function) to produce the result value.

[0108] Specifically, the linear transformation described above may mean transforming (mapping) the output value (i.e., the final output vector of the transformer) to a desired output dimension (e.g., one dimension).

[0109] Specifically, the activation function described above may refer to a function that transforms the output value non-linearly. By performing a non-linear transformation through the activation function described above, the model can be enabled to learn complex patterns and non-linear boundaries. For example, the sigmoid function, Tanh function, ReLU function, leaky ReLU function, softmax function, swish function, or GELU function may be used as the activation function, but are not limited thereto.

[0110] Continuing, a result unit (430) according to one embodiment of the present invention can calculate, for a plurality of unit bits included in input data (411), whether each unit bit corresponds to an abnormal bit or the probability that it corresponds to an abnormal bit as a result value. Specifically, whether the unit bit described above corresponds to an abnormal bit can be determined by first calculating the probability that the unit bit corresponds to an abnormal bit and comparing the calculated probability value with a predetermined standard (threshold value). For example, if the predetermined standard is assumed to be 0.5, if the result value for a specific unit bit is 0.3 (i.e., the probability that the specific unit bit is an abnormal bit is determined to be 0.3), it can be determined that it is not an abnormal bit, and if the result value for another unit bit is 0.7, it can be determined that it corresponds to an abnormal bit.

[0111] Specifically, the probability of corresponding to the above-described abnormal bit can be calculated as an independent probability value (e.g., a real number between 0 and 1) for each unit bit by the result unit (430) performing a linear transformation and applying an activation function bit by bit on each unit bit output value (423) generated by the output unit (420) according to one embodiment of the present invention.

[0112] Meanwhile, the result value (431) calculated by the result unit (430) according to one embodiment of the present invention may be further readjusted (such readjustment process is not illustrated in FIG. 3). Specifically, in response to the result value (431) being calculated by the result unit (430) as the probability that each unit bit corresponds to an abnormal bit, a readjustment process may be performed to remove misclassification results or smooth the result value by applying a filter to the result value (431). More specifically, the readjustment process may be performed by applying a filter to the binary classification result that determines whether each unit bit is an abnormal bit or not. The filter used for the readjustment described above may be at least one of a median filter, a moving average filter, a Gaussian filter, a mode filter, an adaptive filter, and a condition-based rule filter, but is not limited thereto. For example, the readjustment process may be performed by applying a median filter having a predetermined tap size (e.g., 5) to the result value (431) described above.

[0113] Meanwhile, regarding the result value that has undergone the readjustment process as described above, an additional process of modifying the result value based on the length of the unit bit can be performed (such modification process is not shown in FIG. 3). That is, even if a unit bit is determined to be an abnormal bit after undergoing the readjustment process, if the length of the unit bit is smaller than a predetermined level, the unit bit can ultimately be determined not to be an abnormal bit.

[0114] For example, in response to the above-mentioned abnormal beat being determined to be atrial fibrillation (AF), considering the fact that significant atrial fibrillation lasts for 30 seconds or longer, a predetermined level may be set to 30 seconds, and among the unit beats determined to be abnormal beats after readjustment, unit beats with a length of less than 30 seconds may be re-determined as not being abnormal beats.

[0115] In conclusion, a personalized reading model (400) according to one embodiment of the present invention can calculate, for a plurality of unit bits included in input data (411), whether each unit bit corresponds to an abnormal bit (i.e., binary classification result) or the probability of corresponding to an abnormal bit as a result value (431). Meanwhile, the personalized reading model (400) described above can read unit bits by additionally using symbol bits as input data (411), taking into account individual characteristics that differ for each subject. Specifically, the probability of corresponding to an abnormal bit for each unit bit can be calculated as a result value even with input data (411) that includes only a plurality of unit bits without including symbol bits, but the result value can be calculated with higher accuracy (i.e., classification accuracy for normal bits and abnormal bits) through input data (411) that additionally includes symbol bits.

[0116] Continuing, a personalized reading model (400) according to one embodiment of the present invention can be learned by adjusting at least one of the weights of the feature extraction unit (410), the output unit (420), and the result unit (430) based on the result value described above.

[0117] Specifically, a personalized reading model (400) according to one embodiment of the present invention can perform learning through a method comprising: (1) a forward propagation process that processes input data through at least one transformer model (general transformer model or expert mixed transformer model) included in the personalized reading model (400) and generates a predicted value through an activation function; (2) an error calculation process that calculates the difference (error) between the predicted value and the actual correct answer (label) using a loss function; (3) a backpropagation process that calculates a gradient for at least one weight (e.g., a weight of a residual neural network) included in a feature extraction unit (410), at least one weight (e.g., a weight included in a transformer model) included in an output unit (420), and at least one weight (e.g., a weight of a linearization function) included in a result unit (430) based on the calculated error (loss value); and (4) a weight update process that adjusts at least one of the weights mentioned above based on the calculated gradient.

[0118] In addition, a personalized reading model (400) according to one embodiment of the present invention may additionally use symbol bits as input data in the learning process described above. As described above, by not producing an output value for the symbol bits and not backpropagating them, the symbol bits are not modified during the learning process and can be used as a fixed reference value.

[0119] Next, the communication unit (230) according to one embodiment of the present invention can perform the function of enabling data transmission and reception from / to the bit generation unit (210) and the inference unit (220).

[0120] Finally, a control unit (240) according to one embodiment of the present invention can perform the function of controlling the flow of data between a bit generation unit (210), an inference unit (220), and a communication unit (230). That is, by controlling the flow of data from / to / from the outside of the read support system (200) or the flow of data between each component of the read support system (200), the control unit (240) according to one embodiment of the present invention can control the bit generation unit (210), the inference unit (220), and the communication unit (230) to perform their respective unique functions.

[0121] The embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., either individually or in combination. The program instructions recorded on the computer-readable recording medium may be those specifically designed and configured for the present invention or those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; 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. Examples of program instructions 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. Hardware devices may be modified into one or more software modules to perform processing according to the present invention, and vice versa.

[0122] Although the present invention has been described above with reference to specific details such as specific components, limited embodiments, and drawings, this is provided only to aid in a more comprehensive understanding of the invention, and the invention is not limited to the above embodiments, and a person skilled in the art to which the invention belongs can make various modifications and changes from this description.

[0123] Accordingly, the scope of the present invention should not be limited to the embodiments described above, and all scopes equivalent to or equivalently modified from the claims set forth below, as well as the claims set forth below, shall be considered to fall within the scope of the concept of the present invention.

Claims

1. A method for supporting the interpretation of an electrocardiogram signal of a subject, A step of generating a symbol beat and a plurality of unit beats based on an electrocardiogram signal obtained from a subject, and The method includes the step of inputting the generated symbol bit and a plurality of unit bits into a personalized reading model to infer whether each of the plurality of unit bits corresponds to an anomalous bit. method.

2. In Paragraph 1, The above symbol bit is generated by selecting a consecutive subset of unit bits among the plurality of unit bits and processing the selected consecutive subset of unit bits. method.

3. In Paragraph 2, Processing the above consecutive partial unit bits involves removing outliers from the selected consecutive partial unit bits, removing residual T-waves or residual P-waves from the unit bits from which outliers have been removed, and then adjusting the size. method.

4. In Paragraph 1, The above plurality of unit bits are generated by dividing each bit based on the R peak of each bit included in the electrocardiogram signal and adjusting the size of each divided bit. method.

5. In Paragraph 1, The above personalized reading model includes a feature extraction unit that extracts features from the symbol bits and a plurality of unit bits, an output unit that generates an output value based on information related to the extracted features, and a result unit that generates a result value based on the generated output value. method.

6. In Paragraph 5, The above feature extraction unit includes a residual neural network. method.

7. In Paragraph 5, The information related to the above features is information generated by tokenizing the symbol bits and a plurality of unit bits and then extracting the features, or information generated by first extracting the features and then tokenizing the extracted features. method.

8. In Paragraph 5, The above output unit includes a mixture of experts (MoE) transformer and a general transformer, wherein the general transformer uses the output value of the mixture of experts transformer as an input value. method.

9. In Paragraph 5, The above result unit linearly transforms the output value generated by the above output unit and then converts it into a probability value through an activation function. method.

10. In Paragraph 5, The above-described personalized reading model is learned by adjusting at least one of the weights of the feature extraction unit, the output unit, and the result unit based on the above-described result value. method.

11. In Paragraph 1, In the above inference step, the probability that each of the plurality of unit bits corresponds to an anomalous bit is independently calculated. method.

12. A non-transient computer-readable recording medium for recording a computer program for executing the method according to paragraph 1.

13. As a system for supporting the interpretation of a subject's electrocardiogram signal, A beat generation unit that generates symbol beats and multiple unit beats based on electrocardiogram signals acquired from a subject, and The inference unit inputs the generated symbol bit and a plurality of unit bits into a personalized reading model and infers whether each of the plurality of unit bits corresponds to an anomalous bit. System.

14. In Paragraph 13, The above symbol bit is generated by selecting a consecutive subset of unit bits among the plurality of unit bits and processing the selected consecutive subset of unit bits. System.

15. In Paragraph 14, Processing the above consecutive partial unit bits involves removing outliers from the selected consecutive partial unit bits, removing residual T-waves or residual P-waves from the unit bits from which outliers have been removed, and then adjusting the size. System.

16. In Paragraph 13, The above plurality of unit bits are generated by dividing each bit based on the R peak of each bit included in the electrocardiogram signal and adjusting the size of each divided bit. System.

17. In Paragraph 13, The above personalized reading model includes a feature extraction unit that extracts features from the symbol bits and a plurality of unit bits, an output unit that generates an output value based on information related to the extracted features, and a result unit that generates a result value based on the generated output value. System.

18. In Paragraph 17, The above feature extraction unit includes a residual neural network. System.

19. In Paragraph 17, The information related to the above features is information generated by tokenizing the symbol bits and a plurality of unit bits and then extracting the features, or information generated by first extracting the features and then tokenizing the extracted features. System.

20. In Paragraph 17, The above output unit includes a mixture of experts (MoE) transformer and a general transformer, wherein the general transformer uses the output value of the mixture of experts transformer as an input value. System.

21. In Paragraph 17, The above result unit linearly transforms the output value generated by the above output unit and then converts it into a probability value through an activation function. System.

22. In Paragraph 17, The above-described personalized reading model is learned by adjusting at least one of the weights of the feature extraction unit, the output unit, and the result unit based on the above-described result value. System.

23. In Paragraph 13, The above inference unit independently calculates the probability that each of the plurality of unit bits corresponds to an abnormal bit. System.

Images