Apparatus and method for predicting paroxysmal atrial fibrillation using normal sinus rhythm electrocardiogram

The method enhances the prediction of paroxysmal atrial fibrillation by focusing on P-wave intervals in electrocardiograms using neural networks, addressing interpretability and overfitting issues to improve diagnostic reliability and accuracy.

KR103044637B1Active Publication Date: 2026-11-06DEEPCARDIO INC
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
KR1020260078921
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2026-04-30
Publication Date
2026-11-06
Estimated Expiration
2046-04-30

AI Technical Summary

Technical Problem

Existing methods for predicting paroxysmal atrial fibrillation from electrocardiogram signals are limited by the difficulty in early diagnosis during asymptomatic periods, inefficiency in processing subtle electrophysiological features, and lack of interpretability in AI-based predictions, leading to potential overfitting and reduced accuracy in clinical environments.

Method used

A method using an artificial neural network that generates mask information to focus on the time interval of physiological P waves, suppresses weights outside this interval, and combines attention weight maps with loss functions to extract fine morphological and heart rhythm-related features, enhancing prediction reliability and interpretability.

Benefits of technology

The method effectively predicts the risk of paroxysmal atrial fibrillation by focusing on clinically significant P-wave features, reducing overfitting, and maintaining robust performance across environments, improving diagnostic reliability and enabling early intervention strategies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112026053123860-PAT00006_ABST
    Figure 112026053123860-PAT00006_ABST
Patent Text Reader

Abstract

An apparatus and method for predicting paroxysmal atrial fibrillation using a normal sinus rhythm electrocardiogram are disclosed. A method for predicting paroxysmal atrial fibrillation according to one embodiment may include the steps of receiving a normal sinus rhythm electrocardiogram signal; detecting a reference waveform from the electrocardiogram signal and generating mask information including a time interval where a physiological P wave is located based on the reference waveform; inputting the electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted such that the reflection of weights for regions outside the time interval specified by the mask information is suppressed; and calculating the risk of paroxysmal atrial fibrillation occurrence in a subject based on the extracted feature information.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] Embodiments of the present invention relate to an electrocardiogram signal analysis technique using normal sinus rhythm electrocardiogram data, and to an apparatus and method for predicting the risk of occurrence of paroxysmal atrial fibrillation from a normal sinus rhythm electrocardiogram signal. Background Technology

[0002] Atrial fibrillation is one of the representative arrhythmias that can cause serious complications such as stroke, heart failure, and thromboembolism. In particular, paroxysmal atrial fibrillation is characterized by the occurrence of atrial fibrillation for a certain period of time followed by a return to normal sinus rhythm, so there is a problem in that early diagnosis is difficult based solely on an electrocardiogram measured during the asymptomatic period.

[0003] Generally, 12-lead electrocardiograms, Holter monitoring, event recorders, or wearable-based long-term biosignal measurement technologies are used to diagnose atrial fibrillation. However, since 12-lead electrocardiograms only reflect electrical cardiac signals over a short period, they have limitations in that it is difficult to identify abnormal findings when paroxysmal atrial fibrillation has not occurred. In addition, Holter monitoring or long-term measurement methods increase testing time and cost, cause greater discomfort to the patient, and impose the burden of continuously analyzing large amounts of signal data.

[0004] Meanwhile, various attempts have recently been made to predict the likelihood of arrhythmia from electrocardiogram signals using artificial intelligence or deep learning technologies. For example, techniques have been proposed to classify the presence of disease by using the entire electrocardiogram waveform as input data, or to train classification models by extracting features from heart rate intervals, waveform amplitudes, and frequency components. However, since these techniques typically process various waveform information contained throughout the entire electrocardiogram signal in a batch, there are cases where subtle electrophysiological features directly related to early signs of specific diseases are not sufficiently reflected due to other relatively large waveform components or noise.

[0005] Furthermore, while conventional AI-based electrocardiogram (ECG) analysis technologies can provide prediction or classification results, they often fail to clearly indicate which specific parts of the ECG signal the model based its judgment on. Consequently, medical professionals face difficulties in verifying the reliability of these prediction results or interpreting them clinically, limiting their use as an actual diagnostic aid.

[0006] In addition, since electrocardiogram signals may contain various noises depending on the measurement environment, electrode attachment status, patient movement, respiration, and baseline fluctuations, there is a concern that machine learning or deep learning analysis performed on the entire signal may lead to excessive learning of unnecessary signal components. Consequently, even if high performance is observed on specific datasets, there is a possibility that prediction accuracy and generalization performance may deteriorate in actual clinical environments or external validation environments.

[0007] Therefore, there is an increasing need for technology that can more effectively predict the likelihood of paroxysmal atrial fibrillation from electrocardiogram signals of normal sinus rhythm, while also reliably reflecting clinically significant information within the signals and improving the interpretability of the prediction results.

[0008] [Prior Literature Number]

[0009] Korean Patent Publication No. 10-2024-0062901 The problem to be solved

[0010] The present invention provides a device and method for predicting the risk of developing paroxysmal atrial fibrillation from a normal sinus rhythm electrocardiogram signal. means of solving the problem

[0011] A method for predicting paroxysmal atrial fibrillation performed by a computer device may be provided, comprising the steps of: receiving a normal sinus rhythm electrocardiogram signal; detecting a reference waveform in the electrocardiogram signal and generating mask information including a time interval in which a physiological P wave is located based on the reference waveform; inputting the electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted such that the reflection of weights for regions outside the time interval specified by the mask information is suppressed; and calculating the risk of occurrence of paroxysmal atrial fibrillation of a subject based on the extracted feature information.

[0012] According to one aspect, the step of detecting the reference waveform may include the step of detecting an R-peak in the normal sinus rhythm electrocardiogram signal, and the step of generating the mask information may include the step of setting a time interval corresponding to a physiological P-wave by inversely calculating based on the detected R-peak.

[0013] According to another aspect, the step of generating the mask information may be characterized by including the step of variably adjusting the length or position of the time interval corresponding to the physiological P wave based on the heart rate or RR interval calculated from the normal sinus rhythm electrocardiogram signal.

[0014] According to another aspect, the step of extracting the feature information may be characterized by including the step of combining the attention weight map produced by the artificial neural network with the mask information, and learning such that a penalty proportional to the attention weight corresponding to the region outside the time interval is reflected in the loss function.

[0015] According to another aspect, the loss function may be characterized by including a base loss representing the classification error between the predicted value and the correct label, and a penalty loss based on attention weights corresponding to the area outside the time interval.

[0016] According to another aspect, the step of extracting the feature information may be characterized by including: a step of extracting fine morphological features from a segment corresponding to a physiological P wave among the normal sinus rhythm electrocardiogram signal; and a step of extracting heart rhythm-related features from the normal sinus rhythm electrocardiogram signal.

[0017] According to another aspect, the step of calculating the risk of occurrence of paroxysmal atrial fibrillation may be characterized by including the step of calculating the risk of occurrence of paroxysmal atrial fibrillation based on the result of combining the micromorphological features and the heart rhythm-related features.

[0018] According to another aspect, the step of extracting the feature information may be characterized by including the step of training the artificial neural network such that at least some of the plurality of electrocardiogram channels are assigned a relatively higher importance than other channels.

[0019] A computer program stored on a non-transient computer-readable recording medium can be provided to execute the above method on a computer device.

[0020] A non-transient computer-readable recording medium may be provided on which a computer program for executing the above method on a computer device is recorded.

[0021] A computer device may be provided comprising: at least one processor; and a memory for storing instructions executed by the at least one processor; wherein the at least one processor receives a normal sinus rhythm electrocardiogram signal, detects a reference waveform in the electrocardiogram signal, generates mask information including a time interval in which a physiological P wave is located based on the reference waveform, inputs the electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted such that the reflection of weights for an area outside the time interval specified by the mask information is suppressed, and calculates the risk of paroxysmal atrial fibrillation of a subject based on the extracted feature information. Effects of the invention

[0022] According to embodiments of the present invention, by generating mask information that reflects the time interval where the physiological P wave is located from a normal sinus rhythm electrocardiogram signal, and controlling learning so that the attention weights of an artificial neural network are focused on said interval, fine morphological features indicating electrical abnormalities of the atrium can be effectively extracted.

[0023] Furthermore, according to embodiments of the present invention, since the attention weights of an artificial neural network can be structurally induced to focus on the physiological P-wave segment, interpretation information consistent with clinical mechanisms can be provided to medical professionals when visualizing prediction results. For example, since it can be confirmed that the prediction was made based on features related to deviation of the P-wave terminal, changes in P-wave duration, or amplitude reduction, this can contribute to the implementation of explainable artificial intelligence (XAI) and the improvement of diagnostic reliability.

[0024] In addition, by suppressing the weighting of regions outside the P-wave interval at the loss function level, noise and overfitting can be prevented, and robust predictive performance can be maintained even in an external validation environment. Accordingly, the risk of developing paroxysmal atrial fibrillation can be predicted more reliably, the interpretability of prediction results can be improved, and it can be usefully applied to early drug treatment, additional precision examinations, or the establishment of patient-specific management strategies, and it can contribute to the preemptive prevention of fatal complications such as cerebral infarction. Brief explanation of the drawing

[0025] FIG. 1 is a drawing illustrating an example of a network environment according to an embodiment of the present invention. FIG. 2 is a block diagram illustrating an example of a computer device according to an embodiment of the present invention. FIG. 3 is a block diagram illustrating an example of the internal configuration of a paroxysmal atrial fibrillation prediction device according to an embodiment of the present invention. FIG. 4 is a flowchart illustrating an example of a method for predicting paroxysmal atrial fibrillation according to an embodiment of the present invention. Specific details for implementing the invention

[0026] Hereinafter, embodiments will be described in detail with reference to the attached drawings.

[0027] Embodiments of the present invention relate to an apparatus and method for predicting the risk of occurrence of paroxysmal atrial fibrillation from a normal sinus rhythm electrocardiogram signal. An apparatus for predicting paroxysmal atrial fibrillation according to embodiments of the present invention may be implemented by at least one computer device. In this case, a computer program according to one embodiment of the present invention may be installed and run on at least one computer device, and at least one computer device may perform an apparatus for predicting paroxysmal atrial fibrillation according to embodiments of the present invention under the control of the run computer program. The above-described computer program may be stored on a computer-readable recording medium to be combined with at least one computer device to execute the apparatus for predicting paroxysmal atrial fibrillation on a computer.

[0028] FIG. 1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention. The network environment of FIG. 1 illustrates an example including a plurality of electronic devices (110, 120, 130, 140), a plurality of servers (150, 160), and a network (170). FIG. 1 is an example for explaining the invention, and the number of electronic devices or servers is not limited to that shown in FIG. 1. Furthermore, the network environment of FIG. 1 is merely an example of one of the environments applicable to the present embodiments, and the environments applicable to the present embodiments are not limited to the network environment of FIG. 1.

[0029] Multiple electronic devices (110, 120, 130, 140) may be fixed terminals or mobile terminals implemented as computer devices. Examples of multiple electronic devices (110, 120, 130, 140) include smartphones, mobile phones, navigation systems, computers, laptops, digital broadcasting terminals, PDAs (Personal Digital Assistants), PMPs (Portable Multimedia Players), tablet PCs, etc. For example, FIG. 1 shows the shape of a smartphone as an example of an electronic device (110), but in embodiments of the present invention, the electronic device (110) may substantially refer to one of various physical computer devices capable of communicating with other electronic devices (120, 130, 140) and / or servers (150, 160) via a network (170) using a wireless or wired communication method.

[0030] The communication method is not limited and may include not only communication methods utilizing communication networks (e.g., mobile communication networks, wired internet, wireless internet, broadcasting networks) that the network (170) may include, but also short-range wireless communication between devices. For example, the network (170) may include any one or more networks such as a PAN (personal area network), LAN (local area network), CAN (campus area network), MAN (metropolitan area network), WAN (wide area network), BBN (broadband network), and the Internet. Additionally, the network (170) may include any one or more network topologies such as a bus network, star network, ring network, mesh network, star-bus network, tree or hierarchical network, but is not limited thereto.

[0031] Each of the servers (150, 160) may be implemented as a computer device or multiple computer devices that communicate with multiple electronic devices (110, 120, 130, 140) through a network (170) to provide commands, code, files, content, services, etc. For example, the server (150) may be a system that provides services to multiple electronic devices (110, 120, 130, 140) connected through the network (170).

[0032] FIG. 2 is a block diagram illustrating an example of a computer device according to an embodiment of the present invention. Each of the plurality of electronic devices (110, 120, 130, 140) or servers (150, 160) described above can be implemented by the computer device (200) illustrated in FIG. 2.

[0033] As illustrated in FIG. 2, such a computer device (200) may include memory (210), a processor (220), a communication interface (230), and an input / output interface (240). The memory (210) is a computer-readable recording medium and may include a non-perishable mass storage device such as RAM (random access memory), ROM (read only memory), and a disk drive. Here, a non-perishable mass storage device such as a ROM and a disk drive may be included in the computer device (200) as a separate permanent storage device distinct from the memory (210). Additionally, an operating system and at least one program code may be stored in the memory (210). These software components may be loaded into the memory (210) from a computer-readable recording medium separate from the memory (210). This separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, a disk, a tape, a DVD / CD-ROM drive, or a memory card. In another embodiment, software components may be loaded into memory (210) via a communication interface (230) rather than a computer-readable recording medium. For example, software components may be loaded into memory (210) of a computer device (200) based on a computer program installed by files received through a network (170).

[0034] The processor (220) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (220) via memory (210) or a communication interface (230). For example, the processor (220) may be configured to execute instructions received according to program code stored in a recording device such as memory (210).

[0035] The communication interface (230) may provide a function for the computer device (200) to communicate with other devices (e.g., storage devices described above) through the network (170). For example, requests, commands, data, files, etc. generated by the processor (220) of the computer device (200) according to program code stored in a recording device such as memory (210) may be transmitted to other devices through the network (170) under the control of the communication interface (230). Conversely, signals, commands, data, files, etc. from other devices may be received by the computer device (200) through the communication interface (230) of the computer device (200) via the network (170). Signals, commands, data, etc. received through the communication interface (230) may be transmitted to the processor (220) or memory (210), and files, etc. may be stored in a storage medium (the permanent storage device described above) that the computer device (200) may further include.

[0036] The input / output interface (240) may be a means for interfacing with an input / output device (250). For example, the input device may include a device such as a microphone, keyboard, or mouse, and the output device may include a device such as a display or speaker. As another example, the input / output interface (240) may be a means for interfacing with a device in which the functions for input and output are integrated into one, such as a touchscreen. At least one of the input / output devices (250) may be configured as a single device with the computer device (200). For example, it may be implemented in a form in which a touchscreen, microphone, speaker, etc., are included in the computer device (200), such as in a smartphone.

[0037] Additionally, in other embodiments, the computer device (200) may include fewer or more components than the components of FIG. 2. However, it is not necessary to clearly illustrate most of the prior art components. For example, the computer device (200) may be implemented to include at least some of the input / output devices (250) described above, or may include other components such as a transceiver, a database, etc.

[0038] FIG. 3 is a block diagram illustrating an example of the internal configuration of a paroxysmal atrial fibrillation prediction device according to an embodiment of the present invention. The paroxysmal atrial fibrillation prediction device (300) illustrated in FIG. 3 may be implemented by at least one of a plurality of electronic devices (110, 120, 130, 140) or at least one of servers (150, 160) in a network environment as previously described in FIG. 1, and may be implemented by a computer device (200) as described in FIG. 2. That is, the paroxysmal atrial fibrillation prediction device (300) may be implemented on a single computer device (200) or distributed across a plurality of computer devices connected via a network.

[0039] Referring to FIG. 3, a paroxysmal atrial fibrillation prediction device (300) according to embodiments of the present invention may include a data receiving unit (310), a preprocessing unit (320), a feature extraction unit (330), and a prediction unit (340).

[0040] The data receiving unit (310) can receive the subject's normal sinus rhythm electrocardiogram signal. In one embodiment, the data receiving unit (310) can receive a 12-lead electrocardiogram signal and can also receive electrocardiogram signals from a biosignal database transmitted via a hospital diagnostic device, a wearable device, an external storage device, or a network. The data receiving unit (310) can receive not only the electrocardiogram signal in the form of a raw waveform, but also electrocardiogram data preprocessed in a preset format or metadata associated with the electrocardiogram measurement.

[0041] The preprocessing unit (320) can detect a reference waveform in the electrocardiogram signal and generate mask information including a time interval where a physiological P wave is located based on the reference waveform. In one embodiment, the preprocessing unit (320) can detect an R-peak in a normal sinus rhythm electrocardiogram signal and, by backcalculating based on the detected R-peak, set a possible interval for the expression of a physiological P wave corresponding to atrial depolarization. At this time, the preprocessing unit (320) can generate the set time interval in the form of a binary mask, bitmap, index information, or weight map. Additionally, the preprocessing unit (320) can dynamically adjust the length or location of the time interval based on the subject's heart rate, average RR interval, heart rate variability, or measurement environment. Accordingly, the preprocessing unit (320) can generate mask information that corresponds to the physiological characteristics of each subject or heart rate more precisely.

[0042] The feature extraction unit (330) inputs the electrocardiogram signal into an artificial neural network to extract feature information, and can extract feature information such that the reflection of weights for regions outside the time interval specified by the mask information is suppressed. In one embodiment, the feature extraction unit (330) may include one or more artificial neural networks that extract time-series features, morphological features, and channel-specific features from the electrocardiogram signal. At this time, the feature extraction unit (330) can perform learning or inference by combining an attention weight map or corresponding importance information calculated by the artificial neural network with mask information generated by the preprocessing unit (320) so that features of the region corresponding to the time interval where the physiological P wave is located are reflected preferentially. For example, the feature extraction unit (330) can suppress the feature extraction process of the artificial neural network from being excessively biased toward waveforms with relatively large amplitudes, such as QRS complexes or T waves, by ensuring that a penalty proportional to the attention weight corresponding to regions outside the time interval specified by the mask information is reflected in the loss function.

[0043] Additionally, the feature extraction unit (330) may perform a fine feature extraction function that extracts fine morphological features from a section corresponding to a physiological P wave in a normal sinus rhythm electrocardiogram signal, and a rhythm feature extraction function that extracts heart rhythm-related features from the entire normal sinus rhythm electrocardiogram signal or multiple heart rate sections. In one embodiment, the feature extraction unit (330) may include a first neural network that extracts fine morphological features and a second neural network that extracts heart rhythm-related features in parallel, and the output feature vectors of each neural network may be combined and used for subsequent prediction. Additionally, the feature extraction unit (330) may perform learning such that at least some channels among multiple electrocardiogram channels, for example, Lead II and Lead V1, which relatively well reflect the electrical activity of the atrium, are assigned a relatively higher importance than other channels.

[0044] The prediction unit (340) can calculate the risk of paroxysmal atrial fibrillation occurring in a subject based on feature information extracted by the feature extraction unit (330). In one embodiment, the prediction unit (340) can calculate a probability value, risk score, grade information, or a combination thereof indicating the likelihood of future paroxysmal atrial fibrillation occurring in the subject, based on the result of combining micromorphological features and heart rhythm-related features transmitted from the feature extraction unit (330). The prediction unit (340) can provide the calculated result to medical staff or users in a visualized form and can also transmit it to an external server, an electronic medical record system, or a diagnostic support system.

[0045] As such, the paroxysmal atrial fibrillation prediction device (300) according to the embodiments of the present invention takes a normal sinus rhythm electrocardiogram signal as input, generates mask information that reflects the time interval where the physiological P wave is located, and controls feature extraction using the mask information so that the artificial neural network does not rely excessively on the area outside the said time interval, thereby enabling more precise prediction of the risk of occurrence of paroxysmal atrial fibrillation even in a normal sinus rhythm state. In particular, instead of uniformly processing the entire electrocardiogram signal, the embodiments of the present invention can improve both prediction performance and interpretability by strengthening the analysis of minute morphological changes in the P wave interval that may be associated with the precursor of paroxysmal atrial fibrillation, and simultaneously reflecting heart rhythm-related features.

[0046] Meanwhile, each component of FIG. 3 may be a logical component classified based on its functional role and is not limited to physically independent hardware blocks. One or more components may be performed integrally by the same processor, and some functions may be performed by different computer devices in a distributed environment. Additionally, each component of FIG. 3 may be implemented through the cooperation of the processor, memory, and communication interface, etc., described in FIG. 2.

[0047] FIG. 4 is a flowchart illustrating an example of a paroxysmal atrial fibrillation prediction method according to an embodiment of the present invention. The paroxysmal atrial fibrillation prediction method illustrated in FIG. 4 can be performed by a paroxysmal atrial fibrillation prediction device (300), and the paroxysmal atrial fibrillation prediction device (300) can be implemented by at least one computer device. At this time, the paroxysmal atrial fibrillation prediction device (300) can perform each step corresponding to the function of each component described in FIG. 3.

[0048] In step (410), the paroxysmal atrial fibrillation prediction device (300) can receive a normal sinus rhythm electrocardiogram signal. In one embodiment, the paroxysmal atrial fibrillation prediction device (300) can receive a 12-lead electrocardiogram signal measured from a subject, and can also acquire an electrocardiogram signal through an external medical device, data storage, or network.

[0049] In step (420), the paroxysmal atrial fibrillation prediction device (300) can detect a reference waveform in an electrocardiogram signal and generate mask information including a time interval where a physiological P wave is located based on the reference waveform. In one embodiment, the paroxysmal atrial fibrillation prediction device (300) can detect an R-peak and set a possible interval for the occurrence of a physiological P wave by back-calculating based on the detected R-peak. Additionally, the paroxysmal atrial fibrillation prediction device (300) can variably adjust the length or position of the time interval based on the heart rate or RR interval.

[0050] In one embodiment, the preprocessing unit (320) can adjust the start and end times of the time interval where the physiological P wave is located based on the average RR interval or heart rate of the target electrocardiogram. For example, if the heart rate increases and the heart rate cycle becomes shorter, the length of the time interval corresponding to the P wave can be relatively reduced, and if the heart rate decreases and the heart rate cycle becomes longer, the length of the time interval can be relatively expanded. Accordingly, the preprocessing unit (320) can more faithfully reflect the heart rate characteristics of each subject compared to the case where a fixed time axis window is uniformly applied, and can improve the temporal precision of the P wave mask bitmap.

[0051] In step (430), the paroxysmal atrial fibrillation prediction device (300) inputs an electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted in such a way that the reflection of weights for regions outside the time interval specified by the mask information is suppressed. In one embodiment, the paroxysmal atrial fibrillation prediction device (300) may combine the attention weight map calculated by the artificial neural network with the mask information and perform learning such that a penalty proportional to the attention weight corresponding to the region outside the time interval is reflected in the loss function. Additionally, the paroxysmal atrial fibrillation prediction device (300) may extract fine morphological features from the interval corresponding to the physiological P wave and together extract heart rhythm-related features from the entire electrocardiogram signal or multiple heart rate intervals. Furthermore, learning may be performed such that at least some of the multiple electrocardiogram channels are assigned higher importance than other channels.

[0052] In step (440), the paroxysmal atrial fibrillation prediction device (300) can calculate the risk of paroxysmal atrial fibrillation occurring in a subject based on extracted feature information. In one embodiment, the paroxysmal atrial fibrillation prediction device (300) can generate a predicted probability or risk score indicating the likelihood of paroxysmal atrial fibrillation occurring by using the result of combining micromorphological features and heart rhythm-related features.

[0053] Hereinafter, specific implementation examples of the paroxysmal atrial fibrillation prediction device and method according to embodiments of the present invention are described.

[0054] In addition, embodiments of the present invention can control learning so that features corresponding to the physiological P-wave interval are preferentially reflected by reflecting clinical domain knowledge regarding electrocardiogram reading in the form of mathematical constraints in the learning process of an artificial neural network.

[0055] Embodiments of the present invention are characterized by controlling learning to predict the risk of occurrence of paroxysmal atrial fibrillation using normal sinus rhythm electrocardiogram signals, thereby preventing the artificial neural network from relying excessively on QRS complexes or T waves with relatively large amplitudes, and prioritizing the reflection of signal features corresponding to the time interval where physiological P waves, which reflect the electrical state of the atrium, are located.

[0056] To this end, a paroxysmal atrial fibrillation prediction device (300) according to embodiments of the present invention performs heart rate unit division and physiological time interval definition for an input normal sinus rhythm electrocardiogram signal, generates mask information reflecting the defined time interval, and can selectively control the reflection of weights for a specific interval during the feature extraction process of an artificial neural network using the mask information.

[0057] In one embodiment, the preprocessing unit (320) detects an R-peak from a 12-lead normal sinus rhythm electrocardiogram signal and can divide each heartbeat based on the detected R-peak. At this time, the preprocessing unit (320) can define a time interval where the P wave is expected to be located physiologically as a certain interval before the R-peak, for example, a interval in the range of about -250ms to -50ms before the R-peak can be set as the P wave corresponding time interval. This time interval may be set as a fixed value, but in other embodiments, it may be dynamically adjusted according to the subject's heart rate or RR interval.

[0058] For example, the preprocessing unit (320) can calculate the average RR interval of the target electrocardiogram signal and perform dynamic scaling by relatively reducing the length of the P-wave corresponding time interval in tachycardia conditions where the heart rate increases, and expanding the length of the time interval in bradycardia conditions where the heart rate decreases. Accordingly, mask information that more accurately reflects the physiological P-wave interval can be generated even under various heart rate conditions.

[0059] The mask information generated as described above can be represented as a map in the form of binary values ​​or continuous weights, and a first value (e.g., 1) is assigned to the region corresponding to the physiological P-wave interval, and a second value (e.g., 0) is assigned to the other region.

[0060] The feature extraction unit (330) can control the reflection of weights for a specific time interval during the learning or inference process of the artificial neural network using the mask information. In one embodiment, the feature extraction unit (330) can combine the mask information with a spatiotemporal attention weight map generated as the electrocardiogram signal passes through the artificial neural network.

[0061] Specifically, during the learning process, if the model relies on weights in the region other than the physiological P-wave region, i.e., the interval where the P-wave mask bitmap is assigned 0, it can be configured so that an additional attention penalty loss is applied to the basic classification error.

[0062] In one embodiment, the hybrid total loss function (L_total) applied to the feature extraction unit (330) can be defined as shown in the following mathematical formula 1.

[0063] [Mathematical Formula 1]

[0064] L_total = L_CE + * Σ[(1 - M_p(i, t)) * A(i, t)]

[0065] Here, L_total represents the total loss error ultimately calculated, and L_CE represents the basic cross-entropy error between the predicted value and the actual correct label, and represents a weight tuning constant that controls the reflection ratio of the attention penalty. Additionally, i and t represent the spatial channel index and temporal timestep index of the electrocardiogram signal, respectively. M_p(i, t) represents the P-wave mask bitmap value at the spatiotemporal coordinate (i, t), which can be 1 during the physiological P-wave interval and 0 during other intervals. A(i, t) represents the attention weight map value calculated by the artificial neural network for the spatiotemporal coordinate (i, t).

[0066] With this loss function configuration, if the artificial neural network assigns high weights to regions outside the P-wave corresponding interval, such as the QRS complex or the T-wave interval, the total loss value (total loss error) increases, so the model is naturally induced to learn features centered on minute signal changes in the P-wave corresponding interval.

[0067] In another embodiment, the feature extraction unit (330) may be configured to control the importance of multiple electrocardiogram channels in a multidimensional manner. For example, spatial weights may be additionally controlled for Lead II and Lead V1, in which P-wave abnormalities are observed relatively clearly due to their proximity to the anatomical location of the atrium among the 12-lead electrocardiograms. In one embodiment, the feature extraction unit (330) may apply an extended loss function such that the sum of the attention weights assigned to the P-wave regions of Lead II and Lead V1 is greater than or equal to a preset threshold ratio relative to the total sum of attention weights assigned to all 12 channels. Here, the preset threshold ratio may be set to, for example, about 60% or more, but is not limited thereto. Accordingly, the artificial neural network may preferentially learn the features of channels that better reflect electrical abnormalities of the atrium.

[0068] Additionally, the feature extraction unit (330) may include a multi-resolution-based feature extraction structure. In one embodiment, the feature extraction unit (330) may include a first stream that extracts fine morphological features by performing a high-resolution analysis of the physiological P-wave segment, and a second stream that extracts heart rhythm-related features by analyzing the overall rhythm characteristics of the electrocardiogram signal at low resolution.

[0069] The first stream can extract features that reflect fine morphological variations such as the duration, amplitude change, and dispersion characteristics of the P wave, and the second stream can extract rhythm-related features such as the variability of the RR interval and heart rate variability. Subsequently, the prediction unit (340) can calculate the risk of paroxysmal atrial fibrillation by concatenating the embedding vectors extracted from the first stream and the second stream in the fusion layer.

[0070] With such a configuration, the device and method for predicting paroxysmal atrial fibrillation according to the embodiments of the present invention can effectively capture subtle electrophysiological changes related to atrial remodeling even in a normal sinus rhythm state, and can improve the reliability and interpretability of the prediction results by reflecting clinical domain knowledge in the learning process of the artificial neural network.

[0071] In addition, according to embodiments of the present invention, overfitting to noise or irrelevant signals can be suppressed through selective weight control for specific segments of an electrocardiogram signal, and stable predictive performance can be maintained for various external environments and datasets. Furthermore, by detecting signs of atrial remodeling in the early stages, it can be usefully utilized for preemptive management and treatment decision-making regarding paroxysmal atrial fibrillation.

[0072] According to embodiments of the present invention, in predicting the risk of occurrence of paroxysmal atrial fibrillation from a normal sinus rhythm electrocardiogram signal, the artificial neural network can be controlled to suppress learning bias toward large amplitude waveforms corresponding to ventricular depolarization, and to preferentially reflect signal features corresponding to the time interval where the physiological P wave is located. Accordingly, subtle electrophysiological changes in atrial remodeling that may be associated with the precursors of paroxysmal atrial fibrillation can be captured more precisely.

[0073] Furthermore, according to embodiments of the present invention, since the attention weights of an artificial neural network can be structurally induced to be concentrated on the P-wave segment, explanatory information consistent with the clinical mechanism can be provided to medical professionals when visualizing or interpreting prediction results. For example, since it can be confirmed that the prediction results are based on features related to the deviation of the P-wave terminal, changes in P-wave duration, or amplitude reduction, the reliability and interpretability of the prediction results can be improved.

[0074] In addition, according to embodiments of the present invention, by suppressing the reflection of excessive weights for regions other than the P-wave segment at the level of the loss function, overfitting to large amplitudes of the ventricular waveform, baseline fluctuations, noise components, or irrelevant signals can be prevented. Accordingly, excessive adaptation to the training data can be reduced, and more robust predictive performance can be maintained in external validation environments or actual clinical environments.

[0075] Furthermore, according to embodiments of the present invention, by reflecting both micromorphological features extracted from the physiological P-wave segment and heart rhythm-related features, the risk of occurrence of paroxysmal atrial fibrillation can be calculated more stably and precisely than when relying on a single type of feature alone. In particular, by combining micromorphological features and macroscopic rhythm features, it becomes possible to make a prediction that considers both structural and electrical abnormalities of the atrium and the overall pattern of rhythm change.

[0076] Furthermore, according to the embodiments of the present invention, the likelihood of future occurrence of paroxysmal atrial fibrillation can be predicted solely from electrocardiogram signals measured in a normal sinus rhythm state; thus, the efficiency of early screening can be improved compared to conventional methods that relied on long-term Holter monitoring or repeated measurements. Accordingly, this can be usefully applied to identify risk groups in the asymptomatic or early stages, and to establish early drug treatment, additional precision examinations, or patient-specific management strategies.

[0077] The system or device described above may be implemented as a hardware component, or a combination of a hardware component and a software component. For example, the device and component described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0078] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0079] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either individually or in combination. The medium may continuously store a program executable by a computer, or temporarily store it for execution or download. Furthermore, the medium may be various recording or storage means in the form of a single or multiple hardware components, and is not limited to a medium directly connected to a computer system, but may also exist distributed over a network. Examples of media may 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 media configured to store program instructions, including ROM, RAM, and flash memory. Additionally, other examples of media may include recording or storage media managed by app stores that distribute applications or sites and servers that supply or distribute various other software. 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.

[0080] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0081] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A method for predicting paroxysmal atrial fibrillation performed by a computer device comprises: receiving a normal sinus rhythm electrocardiogram signal; detecting a reference waveform in the electrocardiogram signal and generating mask information including a time interval where a physiological P wave is located based on the reference waveform; inputting the electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted such that the reflection of weights for regions outside the time interval specified by the mask information is suppressed; and calculating the risk of paroxysmal atrial fibrillation occurrence in a subject based on the extracted feature information, wherein the normal sinus rhythm electrocardiogram signal is a 12-lead electrocardiogram signal, and the step of extracting feature information comprises combining the mask information with a spatiotemporal attention weight map generated as the electrocardiogram signal passes through the artificial neural network. A method for predicting paroxysmal atrial fibrillation, comprising: a step of training an artificial neural network such that the total loss increases as a higher attention weight is assigned to the region outside the time interval, using a base loss representing a classification error between a predicted value and a correct label, and an attention penalty loss based on an attention weight corresponding to a region outside the time interval; a step of extracting fine morphological features using a first stream that analyzes the time interval where the physiological P wave is located at a high resolution; a step of extracting heart rhythm-related features using a second stream that analyzes the overall rhythm characteristics of the electrocardiogram signal at a lower resolution than the first stream; and a step of combining embedding vectors extracted from the first stream and the second stream, respectively, wherein the spatiotemporal attention weight map is generated such that a relatively higher spatial importance is assigned to the region where the physiological P wave of the second lead (Lead II) and the V1 lead (Lead V1) is located compared to other channels. Claim 2 A method for predicting paroxysmal atrial fibrillation according to claim 1, wherein the step of detecting the reference waveform includes the step of detecting an R-peak in the normal sinus rhythm electrocardiogram signal, and the step of generating mask information includes the step of setting a time interval corresponding to a physiological P-wave by inversely calculating based on the detected R-peak. Claim 3 A method for predicting paroxysmal atrial fibrillation according to claim 2, wherein the step of generating the mask information comprises the step of variably adjusting the length or position of a time interval corresponding to the physiological P wave based on the heart rate or RR interval calculated from the normal sinus rhythm electrocardiogram signal. Claim 4 A method for predicting paroxysmal atrial fibrillation according to claim 1, wherein the step of extracting feature information comprises combining the attention weight map calculated by the artificial neural network with the mask information, and learning such that a penalty proportional to the attention weight corresponding to the region outside the time interval is reflected in the loss function. Claim 5 A method for predicting paroxysmal atrial fibrillation according to claim 4, wherein the loss function comprises a base loss representing a classification error between a predicted value and a correct label, and a penalty loss based on an attention weight corresponding to an area outside the time interval. Claim 6 A method for predicting paroxysmal atrial fibrillation according to claim 1, wherein the step of extracting feature information comprises: a step of extracting fine morphological features from a section corresponding to a physiological P wave among the normal sinus rhythm electrocardiogram signal; and a step of extracting heart rhythm-related features from the normal sinus rhythm electrocardiogram signal. Claim 7 A method for predicting paroxysmal atrial fibrillation according to claim 6, wherein the step of calculating the risk of occurrence of paroxysmal atrial fibrillation comprises the step of calculating the risk of occurrence of paroxysmal atrial fibrillation based on the result of combining the micromorphological features and the heart rhythm-related features. Claim 8 A method for predicting paroxysmal atrial fibrillation according to claim 1, wherein the step of extracting feature information comprises the step of training the artificial neural network such that at least some of the plurality of electrocardiogram channels are assigned a relatively higher importance than other channels. Claim 9 A computer program stored on a non-transient computer-readable recording medium that is combined with a computer device and executes the method of any one of claims 1 to 8 on the computer device. Claim 10 In a computer device, at least one processor; The method includes a memory for storing instructions executed by at least one processor, wherein the at least one processor receives a normal sinus rhythm electrocardiogram signal, detects a reference waveform in the electrocardiogram signal, generates mask information including a time interval where a physiological P wave is located based on the reference waveform, inputs the electrocardiogram signal into an artificial neural network to extract feature information, wherein the feature information is extracted such that the reflection of weights for an area outside the time interval specified by the mask information is suppressed, and calculates the risk of paroxysmal atrial fibrillation of a subject based on the extracted feature information, wherein the normal sinus rhythm electrocardiogram signal is a 12-lead electrocardiogram signal, and to extract the feature information, the at least one processor combines the mask information with a spatiotemporal attention weight map generated as the electrocardiogram signal passes through the artificial neural network, and uses a total loss including a base loss representing the classification error between a predicted value and a correct label, and an attention penalty loss based on attention weights corresponding to an area outside the time interval, wherein the area outside the time interval A computer device characterized by training the artificial neural network such that the total loss increases as a high attention weight is assigned to a region, extracting fine morphological features using a first stream that analyzes the time interval where the physiological P wave is located at high resolution, extracting heart rhythm-related features using a second stream that analyzes the overall rhythm characteristics of the electrocardiogram signal at a lower resolution than the first stream, combining embedding vectors extracted from the first stream and the second stream respectively, and generating the spatiotemporal attention weight map such that a relatively higher spatial importance is assigned to the region where the physiological P wave of the second lead (Lead II) and the V1 lead (Lead V1) is located compared to other channels.

Citation Information

Patent Citations

  • Heart Rate Estimation Method Using video-based physiological signal

    KR1020230081689A

  • Apparatus for Identifying Paroxysmal Atrial Fibrillation by using Single-Lead Electrocardiogram Data

    KR1020250145780A

  • Systems and methods for catheter tip placement using ECG

    US9445746B1