Method, computing device and computer program for deriving cardiac disease indicators through multi-channel biological information analysis

Abstract

A method for deriving heart disease indicators through multi-channel bio-information analysis, a computing device, and a computer program are provided. A method for deriving heart disease indicators through multi-channel bio-information analysis according to various embodiments of the present disclosure comprises, in a method performed through a computing device, a step of acquiring bio-information corresponding to a patient's heart condition and a step of deriving an indicator related to heart disease using the acquired bio-information.

Description

Technology Field

[0001] Various embodiments of the present disclosure relate to a method for deriving heart disease indicators through multi-channel bio-information analysis, a computing device, and a computer program. Background Technology

[0002] Heart disease is known to be one of the leading causes of death worldwide, and early detection and risk assessment play a crucial role in improving patient prognosis. These heart diseases are characterized by the complex interaction of various physiological factors, such as the electrical activity of the heart, mechanical movement, and hemodynamic changes.

[0003] In clinical practice, various biosignals such as electrocardiograms, heart sounds, blood pressure, and pulse waves are utilized to evaluate the condition of the heart. In particular, electrocardiograms have been widely used for a long time because they can non-invasively measure the electrical activity of the heart. Along with this, biosignals that reflect the mechanical movement of the heart or changes in blood flow are also used to indirectly assess the state of cardiac function.

[0004] Recently, as the recognition has spread that a single biological signal alone is insufficient to adequately reflect the complex state of the heart, there has been an increasing number of attempts to utilize multiple biological signals with distinct characteristics together. Since these multichannel biological signals differ in temporal resolution, spatial characteristics, and physical measurement principles, they possess the potential to provide richer physiological information compared to single signals.

[0005] Meanwhile, the form and characteristics of multichannel biometric data can vary significantly depending on the measurement environment, sensor type, signal quality, and physiological differences between individuals. Consequently, effectively utilizing multichannel biometric data requires preprocessing steps such as signal normalization, alignment, and noise reduction, as well as analysis techniques to extract meaningful information.

[0006] With the advancement of computing technology, various information processing methods capable of processing and analyzing large volumes of biological data are being proposed, leading to attempts to derive specific indicators or estimate states from biological signals. These technologies offer the potential to utilize the vast amount of biological data generated in the medical field more systematically.

[0007] However, since heart disease possesses complex characteristics involving closely interrelated anatomical structure, functional status, and physiological responses, there may be limitations in quantitatively representing the state of the heart through the simple analysis of biological information alone. Accordingly, there is a continuously being raised need for technologies capable of efficiently analyzing various heart-related biological data to derive indicators associated with heart disease.

[0008] The aforementioned background technology is one that the inventor possessed or acquired in the process of deriving the content of the present disclosure, and it cannot be considered as prior art disclosed to the general public prior to the filing of this application. The problem to be solved

[0009] The problem that the present disclosure aims to solve is to provide a method for deriving heart disease indicators through multichannel bio-information analysis, a computing device, and a computer program, which, for the purpose of resolving the aforementioned conventional problems, can non-invasively derive various indicators related to heart disease by integrally analyzing multichannel bio-information and can express the heart condition more systematically and quantitatively through an information processing structure that reflects anatomical structure.

[0010] The problems that this disclosure aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0011] The present disclosure may be implemented in various ways, including a method, an apparatus (system), or a computer program stored on a readable storage medium.

[0012] A method for deriving heart disease indicators through multi-channel bio-information analysis according to one embodiment of the present disclosure for solving the above-described problem may include, in a method performed through a computing device, a step of acquiring bio-information corresponding to a patient's heart condition and a step of deriving indicators related to heart disease using the acquired bio-information.

[0013] In various embodiments, the acquired bio-information is collected through multiple different channels and may be multi-channel bio-information including at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to the heart condition.

[0014] In various embodiments, the step of deriving the indicator may include: extracting a first feature vector corresponding to electrical features related to heart condition from the acquired bio-information; generating a first latent space representation by mapping the extracted first feature vector to a predefined electrical-anatomical latent space; generating a coronary artery structure graph by structuring the generated first latent space representation based on the anatomical structure of the coronary artery; and deriving a first indicator related to heart disease using the generated coronary artery structure graph.

[0015] In various embodiments, the acquired bio-information includes a plurality of electrophysiological signals having different temporal and spatial characteristics, and the step of extracting the first feature vector may include the step of extracting the first feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals using a plurality of encoders.

[0016] In various embodiments, the step of extracting a first feature vector from each of the plurality of electrophysiological signals may include preprocessing the plurality of electrophysiological signals based on the characteristics of the plurality of electrophysiological signals and extracting a first feature vector from each of the preprocessed plurality of electrophysiological signals through the plurality of encoders.

[0017] In various embodiments, the step of generating the first latent space representation includes generating the first latent space representation by mapping the extracted first feature vector onto a coordinate system corresponding to the electrical-anatomical latent space, wherein the electrical-anatomical latent space may be a predefined latent coordinate space based on the correspondence between the anatomical structure of the myocardium and coronary artery and the electrical conduction characteristics of the myocardium.

[0018] In various embodiments, the step of generating the coronary artery structure graph comprises generating the coronary artery structure graph by assigning information contained in the generated first latent space representation as a state vector to each of a plurality of nodes included in the tree structure, based on a tree structure predefined in relation to the anatomical structure of the coronary artery, wherein the plurality of nodes may include a plurality of blood vessel nodes corresponding to major blood vessels of the coronary artery and a plurality of segment nodes corresponding to the segmental structure of the coronary artery.

[0019] In various embodiments, the step of generating a coronary artery structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure includes the step of assigning information included in the generated first latent space representation to each of the plurality of nodes using a first weighted mapping matrix, wherein the first weighted mapping matrix is ​​a matrix representing the correspondence between a first feature vector extracted from biological information and the plurality of nodes, and the plurality of elements included in the first weighted mapping matrix may be weights representing the degree to which the first feature vector extracted from biological information contributes to each of the plurality of nodes.

[0020] In various embodiments, the step of generating a coronary artery structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure may include: extracting a global ischemia pattern for the heart based on the extracted first latent space representation and generating a global surveillance node mapped to the extracted global ischemia pattern; defining a reference state based on bio-information obtained from a healthy population without heart disease and generating a normal reference node mapped to the defined reference state; and connecting the generated global surveillance node and the generated normal reference node to at least one of the plurality of nodes.

[0021] In various embodiments, the step of generating a coronary artery structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure may include the step of updating the state vector mapped to each of the plurality of nodes by performing a message passing operation between the plurality of nodes using a graph neural network decoder.

[0022] In various embodiments, the step of deriving the first indicator includes the step of deriving a plurality of first indicators by analyzing each of the generated coronary artery structure graphs using a plurality of prediction heads, and the derived plurality of first indicators may include coronary artery fractional flow reserve, ischemia status, coronary artery stenosis rate, plaque fragility, hemodynamic indicators, and cardiovascular event risk.

[0023] In various embodiments, the step of deriving the first indicator includes the step of deriving the first indicator by analyzing the acquired bio-information through the first indicator derivation module, wherein the first indicator derivation module includes a first encoder that extracts a first feature vector from the bio-information, a first latent space alignment module that generates a first latent space representation, a first graph generation module that generates a coronary artery structure graph, a first graph neural network decoder that updates the coronary artery structure graph, and a first prediction head that derives the first indicator using the coronary artery structure graph, wherein pre-learning is performed based on two or more pairs of indicators of different types, and after the pre-learning is performed, self-supervised learning is performed based on a plurality of multi-channel electrophysiological signals for which no correct answer data exists, and after the self-supervised learning is performed, supervised learning is performed based on a plurality of multi-channel electrophysiological signals for which correct answer data exists.

[0024] In various embodiments, the step of deriving the indicator may include: extracting a second feature vector corresponding to structural and functional features related to the heart condition from the acquired biological information; generating a second latent space representation by mapping the extracted second feature vector to a predefined structural-functional latent space; generating a heart structure graph by structuring the generated second latent space representation based on the anatomical structure of the heart; and analyzing the generated heart structure graph to derive a second indicator related to heart disease.

[0025] In various embodiments, the acquired bio-information includes a plurality of electrophysiological signals having different temporal and spatial characteristics, and the step of extracting the second feature vector may include the step of extracting the second feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals using a plurality of encoders.

[0026] In various embodiments, the step of extracting a second feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals may include preprocessing the plurality of electrophysiological signals based on the input characteristics of each of the plurality of encoders, and extracting a second feature vector from each of the preprocessed plurality of electrophysiological signals through the plurality of encoders.

[0027] In various embodiments, the step of generating the second latent space representation includes generating the second latent space representation by mapping the extracted second feature vector onto a coordinate system corresponding to the structural-functional latent space, wherein the structural-functional latent space may be a predefined latent coordinate space based on the correspondence between the anatomical structure of the heart, cardiac function, and structural disease state.

[0028] In various embodiments, the step of generating the heart structure graph includes generating the heart structure graph by assigning information contained in the generated second latent space representation as a state vector to each of the plurality of nodes included in the tree structure, based on a tree structure predefined for the anatomical structure of the heart, wherein the plurality of nodes may include a plurality of structure nodes corresponding to structural components of the heart and at least one disease node corresponding to structural and functional diseases of the heart.

[0029] In various embodiments, the step of generating a heart structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure includes the step of assigning information included in the generated second latent space representation to each of the plurality of nodes using a second weighted mapping matrix, wherein the second weighted mapping matrix is ​​a matrix representing the correspondence between a second feature vector extracted from biological information and the plurality of nodes, and the plurality of elements included in the second weighted mapping matrix may be weights representing the degree to which the second feature vector extracted from biological information contributes to each of the plurality of nodes.

[0030] In various embodiments, the step of generating a cardiac structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure may include: extracting global structure-disease context features that commonly affect cardiac structure and disease state from clinical / demographic information including at least one of the subject's age, gender, comorbidities, drug history, and examination history, and text information including cardiac image readings, diagnostic records, or examination reports, and generating a global surveillance node mapped with the extracted global structure-disease context features; defining a reference structural state of the heart based on bio-information obtained from a healthy population and generating a normal reference node mapped with the defined reference structural state; and connecting the generated global surveillance node and the generated normal reference node to at least one of the plurality of nodes included in the generated cardiac structure graph.

[0031] In various embodiments, the step of generating a heart structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure may include the step of updating the state vector mapped to each of a plurality of nodes included in the generated heart structure graph by performing a message passing operation between a plurality of nodes included in the generated heart structure graph using a graph neural network decoder.

[0032] In various embodiments, the step of deriving the second indicators includes the step of deriving a plurality of second indicators by analyzing each of the generated cardiac structure graphs using a plurality of prediction heads, and the derived plurality of second indicators may include structural indicators related to the volume, mass, or wall thickness of the left and right ventricles, functional indicators related to left ventricular ejection fraction, ventricular systolic or diastolic function, disease indicators indicating the presence or severity of structural heart disease, and prognostic risk indicators related to the worsening of heart failure, hospitalization, or occurrence of major cardiovascular events within a certain period.

[0033] In various embodiments, the step of deriving the second indicator includes deriving the second indicator by analyzing the acquired bio-information through a second indicator derivation module, wherein the second indicator derivation module includes a second encoder that extracts a second feature vector from the bio-information, a second latent space alignment module that generates a second latent space representation, a second graph generation module that generates a heart structure graph, a second graph neural network decoder that updates the heart structure graph, and a second prediction head that derives an indicator related to heart disease using the heart structure graph, supervised learning is performed using pairs of heart image information and electrophysiological signals, and after the supervised learning is performed, auxiliary learning is performed using clinical / demographic information, and after the auxiliary learning is performed, self-supervised learning and contrastive learning may be performed using large-scale bio-information for which no ground truth data exists.

[0034] A computing device according to another embodiment of the present disclosure for solving the above-described problem comprises a processor, a network interface, a memory, and a computer program loaded into the memory and executed by the processor, wherein the processor can perform a method for deriving heart disease indicators through multi-channel bio-information analysis by executing one or more instructions included in the computer program.

[0035] A computer program according to another embodiment of the present disclosure for solving the above-described problem may be combined with a computing device and stored on a non-transient recording medium readable by the computing device to execute a method for deriving heart disease indicators through multi-channel bio-information analysis.

[0036] Other specific details of the present disclosure are included in the detailed description and drawings. Effects of the invention

[0037] According to various embodiments of the present disclosure, there is an advantage in that various indicators related to heart disease can be derived non-invasively by integrally analyzing multi-channel bio-information, and the heart condition can be expressed more systematically and quantitatively through an information processing structure that reflects anatomical structure.

[0038] The effects of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0039] The following drawings attached to this specification illustrate preferred embodiments of the present disclosure and serve to further enhance understanding of the technical concept of the present disclosure together with the detailed description of the invention; therefore, the present disclosure should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is a diagram illustrating a system for deriving heart disease indicators through multi-channel bio-information analysis according to one embodiment of the present disclosure. FIG. 2 is a diagram illustrating the configuration of an indicator derivation module according to the first embodiment of the present disclosure. FIG. 3 is a flowchart of a method for deriving heart disease indicators through multi-channel bio-information analysis according to the first embodiment of the present disclosure. FIG. 4 is a diagram illustrating the configuration of an indicator derivation module according to a second embodiment of the present disclosure. FIG. 5 is a flowchart of a method for deriving heart disease indicators through multi-channel bio-information analysis according to a second embodiment of the present disclosure. FIG. 6 is a diagram illustrating the functional module configuration of a computing device performing the third and fourth embodiments of the present disclosure. FIG. 7 is a diagram schematically illustrating the hierarchical structure of a prediction model according to various embodiments of the present disclosure. FIG. 8 is a flowchart of a method for predicting myocardial ischemia as an indicator according to a third embodiment of the present disclosure. FIG. 9 is a conceptual diagram illustrating an example in which a myocardial ischemia prediction method according to the third embodiment of the present disclosure is performed in a clinical procedure. FIG. 10 is a flowchart of a method for predicting coronary artery stenosis rates by blood vessel segment as an indicator according to the fourth embodiment of the present disclosure. FIG. 11 is an illustrative diagram showing an example of a coronary artery applicable to various embodiments of the present disclosure. FIG. 12 is a drawing illustrating the hardware configuration of a computing device according to various embodiments of the present disclosure. Specific details for implementing the invention

[0040] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure is complete and to fully inform those skilled in the art of the scope of the present disclosure, and the present disclosure is defined only by the scope of the claims.

[0041] The terms used herein are for describing the embodiments and are not intended to limit the disclosure. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the components mentioned.

[0042] Throughout this specification, the same reference numerals refer to the same components, and "and / or" includes each of the mentioned components and all combinations of one or more thereof. Although terms such as "first," "second," etc., are used to describe various components, they are not limited by these terms. These terms are used merely to distinguish one component from another. Accordingly, the first component mentioned below may be the second component within the technical scope of this disclosure.

[0043] As used herein, the terms “part” or “module” refer to hardware components such as software, FPGAs, or ASICs, and the “part” or “module” performs certain roles. However, the “part” or “module” is not limited to software or hardware. The “part” or “module” may be configured to reside in an addressable storage medium or configured to run on one or more processors. Thus, by example, the “part” or “module” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and “parts” or “modules” may be combined into a smaller number of components and “parts” or “modules,” or further separated into additional components and “parts” or “modules.”

[0044] Spatially relative terms such as "below," "beneath," "lower," "above," and "upper" may be used to facilitate the description of the relationship between one component and other components as illustrated in the drawings. Spatially relative terms should be understood as encompassing different orientations of components during use or operation, in addition to the orientations depicted in the drawings. For example, if a component depicted in a drawing is inverted, a component described as "below" or "beneath" of another component may be placed "above" of that component. Therefore, the exemplary term "below" may encompass both the lower and upper directions. Components may also be oriented in other directions, and accordingly, spatially relative terms may be interpreted according to the orientation.

[0045] Expressions such as "first," "second," or "first," "second" as used in this specification are used to distinguish one object from another when referring to a plurality of objects of the same kind, unless otherwise indicated by the context, and do not limit the order or importance of said objects.

[0046] Expressions used herein such as “A, B, and C,” “A, B, or C,” “A, B, and / or C,” or “at least one of A, B, and C,” “at least one of A, B, or C,” “at least one of A, B, and / or C,” “at least one selected from A, B, and C,” “at least one selected from A, B, or C,” “at least one selected from A, B, and / or C,” etc., may mean each of the listed items or all possible combinations of the listed items. For example, “at least one selected from A and B” may refer to (1) A, (2) at least one of A, (3) B, (4) at least one of B, (5) at least one of A and at least one of B, (6) at least one of A and B, (7) at least one of B and A, and (8) all of A and B.

[0047] As used herein, the expression “based on” is used to describe one or more factors affecting an act or action of a decision or judgment described in the phrase or sentence containing such expression, and such expression does not exclude additional factors affecting said act or action of a decision or judgment.

[0048] As used in this specification, the expression that a certain component (e.g., a first component) is "connected" or "connected" to another component (e.g., a second component) may mean that the said certain component is not only directly connected or connected to the said other component, but is also connected or connected through a new other component (e.g., a third component).

[0049] As used herein, the expression "configured to" may have meanings such as "set to," "capable of," "modified to," "made to," or "capable of." Such expression is not limited to the meaning of "specifically designed in hardware," and, for example, a processor configured to perform a specific operation may mean a generic-purpose processor capable of performing that specific operation by executing software.

[0050] Unless otherwise defined, all terms used herein (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which this disclosure pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0051] In this specification, the term "computer" refers to any type of hardware device comprising at least one processor, and may be understood to include software configurations operating on said hardware device according to the embodiments. For example, the term "computer" may be understood to include smartphones, tablet PCs, desktops, laptops, and user clients and applications running on each of these devices, but is not limited thereto.

[0052] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0053] Each step described in this specification is described as being performed by a computer, but the subject of each step is not limited thereto, and depending on the embodiment, at least some of each step may be performed on different devices.

[0055] FIG. 1 is a diagram illustrating a system for deriving heart disease indicators through multi-channel bio-information analysis according to one embodiment of the present disclosure.

[0056] Referring to FIG. 1, a system for deriving heart disease indicators through multi-channel bio-information analysis according to one embodiment of the present disclosure may include a computing device (100), a user terminal (200), an external server (300), and a network (400).

[0057] Here, the system for deriving heart disease indicators through multi-channel bio-information analysis illustrated in FIG. 1 is according to one embodiment, and its components are not limited to the embodiment illustrated in FIG. 1 and may be added, changed, or deleted as needed.

[0058] In one embodiment, the computing device (100) can provide a service for deriving heart disease indicators through multi-channel bio-information analysis. More specifically, the computing device (100) can acquire bio-information corresponding to the patient's heart condition and can derive indicators related to heart disease using the bio-information.

[0059] Here, the service provided by the computing device (100) may be provided to the user through a dedicated application installed on the user terminal (200), a web page based on a web browser, or other network-accessible interface.

[0060] For example, a computing device (100) may receive user input, commands, and request data on an application or web page from a user terminal (200) and execute a service function corresponding to the request, or perform data processing, analysis, and calculation for providing the service.

[0061] Additionally, the computing device (100) can transmit processing results to a user terminal (200) to update the user interface (UI) or induce subsequent actions, and may include procedures such as user authentication, permission management, access to stored data, and inter-server interaction as needed. In this way, the service provided by the computing device (100) can be implemented in both web-based and app-based environments, and can provide consistent functions in various user terminals and network environments.

[0062] In various embodiments, the computing device (100) may be configured as a server within a hospital, a server located in a data center, or a server system implemented in a cloud environment, and may derive one or more indicators related to heart disease by acquiring and analyzing multi-channel bio-information corresponding to the patient's heart condition.

[0063] In various embodiments, the computing device (100) may be linked with an electrocardiogram device, a blood pressure monitor, a pulse wave measuring device, a heart sound measuring device, a wearable device, or a hospital information system, a picture archiving and transmission system (PACS), etc., to collect the patient's biometric information.

[0064] In one embodiment, the user terminal (200) may receive biometric information from a patient and receive and display indicators derived from the computing device (100). The computing device (100) may transmit the analysis results or derived indicators to the user terminal (200) so that they are displayed through a user interface (UI), and may provide detailed information on the results, trends of change over time, or summary information as needed. In this process, the computing device (100) may include procedures such as user authentication, access control management, data storage, and inter-server linkage.

[0065] Here, the user terminal (200) may refer to any form of entity(s) in a system having a mechanism for communicating with a computing device (100). For example, such a user terminal (200) may include a PC (personal computer), notebook, mobile terminal, smartphone, tablet PC, wearable device, medical workstation, etc., and may include any type of terminal capable of connecting to a wired / wireless network.

[0066] Additionally, the user terminal (200) may include any computing device implemented by at least one of an agent, an API (Application Programming Interface), and a plug-in. Additionally, the user terminal (200) may include an application source and / or a client application.

[0067] In various embodiments, the user terminal (200) may include an input interface and an output interface.

[0068] The input interface is configured to receive patient biometric information (e.g., electrocardiogram (ECG) data) and can collect data through local file uploads, EHR integration, or real-time connection with ECG measurement equipment. For example, medical staff can select and upload locally stored ECG data files (XML, HL7, DICOM, JPG, JPEG, PNG, PDF, etc.) through the input interface. Additionally, ECG data for a specific patient can be automatically retrieved by querying the patient ID through integration with the hospital's electronic medical record system. Alternatively, real-time measured data can be received along with patient identification information (patient ID, name, date of birth, etc.) by directly connecting to ECG measurement equipment.

[0069] Additionally, for example, a web browser-based user interface (UI) or a dedicated application may be executed on the user terminal (200), through which it can communicate with the computing device (100) and perform a request to derive heart disease indicators through multi-channel bio-information analysis. At this time, the user can upload electrocardiogram data through the interface or check the prediction results processed by the computing device (100) in real time.

[0070] The output interface is configured to visualize and provide results derived from the computing device (100). Specifically, through the output interface, the patient's biometric information and indicators derived therefrom can be provided, and the patient's test history or predicted trends can also be provided to support medical staff in making treatment decisions quickly and accurately.

[0071] In one embodiment, the network (400) may refer to a connection structure capable of exchanging information between each node, such as a plurality of terminals and servers. For example, the network (400) may include a Local Area Network (LAN), a Wide Area Network (WAN), the World Wide Web (WWW), a wired / wireless data network, a telephone network, a wired / wireless television network, a Controller Area Network (CAN), and Ethernet.

[0072] Wireless data communication networks may include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth network, NFC (Near-Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc.

[0073] In one embodiment, an external server (300) may be connected to a computing device (100) via a network (400) and may store and manage various information and data necessary for the computing device (100) to perform a method for deriving heart disease indicators through multi-channel bio-information analysis, or may collect, store, and manage various information and data generated as the computing device (100) performs a method for deriving heart disease indicators through multi-channel bio-information analysis. For example, the external server (300) may include, but is not limited to, a database system, a file storage module, a metadata management module, etc., which are separately configured for the storage and efficient management of large-capacity data. Hereinafter, with reference to FIGS. 2 to 11, a method for deriving heart disease indicators through multi-channel bio-information analysis performed by the computing device (100) will be described.

[0075] FIG. 2 is a diagram illustrating the configuration of an indicator derivation module according to the first embodiment of the present disclosure.

[0076] Referring to FIG. 2, in the first embodiment of the present disclosure, a computing device (100) can derive a first indicator related to heart disease using a first indicator derivation module (500). To this end, the first indicator derivation module (500) may include a plurality of first encoders (510), a first latent space alignment module (520), a first graph generation module (530), a first graph neural network decoder (540), and a plurality of first prediction heads (550).

[0077] First, a plurality of encoders (510) can receive multi-channel biometric information as input data and can extract a first feature vector from the input multi-channel biometric information.

[0078] Here, multichannel bio-information is collected through multiple different channels and may include at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to the heart condition.

[0079] For example, multichannel bio-information may include a standard 12-lead electrocardiogram, an 18-lead electrocardiogram, or a multichannel surface electrocardiogram signal including additional chest leads.

[0080] In addition, multichannel bio-information may include high-density electrophysiological signals obtained from a body surface potential mapping system, vector electrocardiograms, and electrophysiological signals of non-standard lead configurations (specific examples such as 1 lead, 6 leads, etc.) obtained from a wearable device or a patch-type device.

[0081] In addition, multichannel bio-information may further include multimodal bio-signals such as pulse waves, heart sounds, and body surface vibration signals as needed.

[0082] Each first encoder (510) may be configured with a different structure depending on the type or characteristics of the input bio-information. For example, a plurality of first encoders (510) may include an encoder that receives a time-domain electrophysiological signal as input and extracts features reflecting the shape, interval, and change of the waveform, an encoder that receives a signal to which time-frequency conversion has been applied as input and extracts features reflecting the energy distribution by frequency band or minute signal changes, and / or independent encoders corresponding to different bio-signals.

[0083] Each of these multiple first encoders (510) can be configured in parallel, and each can generate a first feature vector that reflects various characteristics of multichannel bio-information.

[0084] Next, the first latent space alignment module (520) can receive a first feature vector extracted by a plurality of first encoders (510) as input and generate a first latent space representation.

[0085] The first latent space alignment module (520) can map a first feature vector to an electrical-anatomical latent space, which is a predefined latent coordinate space based on the correspondence between the anatomical structure of the myocardium and coronary artery and the electrical conduction characteristics of the myocardium. Through this, electrical features extracted from different channels and leads can be converted into a latent space representation aligned based on anatomical location or coronary artery segment. For example, the first latent space alignment module (520) can align electrical features in a distributed form to regions corresponding to the proximal, intermediate, or distal parts of the coronary artery.

[0086] Next, the first graph generation module (530) can generate a coronary artery structure graph based on the first latent space representation generated by the first latent space alignment module (520).

[0087] The first graph generation module (530) can generate a graph using a predefined coronary artery structure template. For example, the coronary artery structure graph may include hub nodes corresponding to coronary arteries and segment nodes corresponding to coronary artery segments, and each node may be mapped to a node state vector assigned from a first latent space representation.

[0088] Next, the first graph neural network decoder (540) can receive the coronary artery structure graph as input and update the node state vector.

[0089] The first graph neural network decoder (540) can update the node state vector by performing message passing operations between nodes on the coronary artery structure graph, thereby reflecting the interrelationships between coronary artery segments and global patterns. Through this, a representation that considers not only local electrical abnormalities but also the entire coronary artery structure can be formed.

[0090] Next, a plurality of first prediction heads (550) can analyze a coronary artery structure graph updated by a first graph neural network decoder (540) to derive a plurality of first indicators. For example, each of the plurality of first prediction heads (550) can derive at least some of coronary artery fractional flow reserve, ischemia status, coronary artery stenosis rate, plaque fragility, hemodynamic indicators, and cardiovascular event risk. In this way, different types of indicators can be calculated in parallel through the plurality of prediction heads.

[0091] The components included in the above-mentioned first indicator derivation module (500) are models that perform artificial intelligence-based data processing and can perform pre-learning, self-supervised learning, and supervised learning in stages or in parallel.

[0092] First, in the pre-training phase, synthetic data can be generated and used so that the model can learn the relationship between coronary artery status and electrophysiological signals even when there is insufficient actual patient data. For example, multiple scenarios can be defined in which the location of coronary artery stenosis, the stenosis rate, blood flow conditions, and plaque characteristics are set differently.

[0093] For this scenario, an electrophysiological signal may be generated using an electrical conduction model or a hemodynamic model, and the initial parameters of the first encoder (510), the first latent space alignment module (520), the first graph neural network decoder (540), and the first prediction head (550) may be learned using the generated electrophysiological signal and the corresponding pair of indicators such as FFR, stenosis rate, or plaque fragility.

[0094] Here, the multichannel electrophysiological signals used for pre-training can be generated through a reverse generative model. The reverse generative model is implemented as a generative neural network that takes node states on a coronary artery graph as input and generates surface electrocardiograms or other electrophysiological signals, and through this, synthetic data for various virtual lesion combinations can be generated and used to pre-train or normalize the inference model of the present invention.

[0095] Next, self-supervised learning can be performed using multi-channel electrophysiological signals for which no ground truth data exists after pre-training. For example, contrastive learning can be performed by setting different beat intervals, different lead combinations, or temporally adjacent signal intervals of the same patient as positive samples, and setting signals from different patients as negative samples.

[0096] In addition, a task can be performed to learn to restore some channels or time intervals after masking them, through which general patterns and structures of biosignals can be learned without labels.

[0097] Next, after self-supervised learning is performed, supervised learning using multi-channel electrophysiological signals where ground truth data exists can be performed. For example, using indicators obtained from invasive FFR tests, coronary angiography, or imaging tests as ground truth data, the model can be trained to predict first indicators such as coronary artery fractional flow reserve, ischemia, stenosis rate, and plaque fragility. At this stage, fine-tuning can be performed around the first prediction head (550) while maintaining the representation learned through pre-training and self-supervised learning.

[0098] In one embodiment, considering that a prediction error near a threshold of a specific indicator may have a clinically significant impact, a learning or correction procedure to correct the reliability of the prediction result may be additionally performed for the first indicator derivation module (500). For example, a method of assigning different loss weights to the interval around the threshold or estimating prediction uncertainty together may be applied. Through this, more stable prediction results can be provided for indicators that require threshold-based interpretation, such as coronary artery ischemia judgment.

[0100] FIG. 3 is a flowchart of a method for deriving heart disease indicators through multi-channel bio-information analysis according to the first embodiment of the present disclosure.

[0101] Referring to FIG. 3, in step S110, the computing device (100) can acquire multichannel bio-information related to the heart condition.

[0102] Here, multichannel bio-information is collected through multiple different channels and may include multichannel bio-information including at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to the heart condition.

[0103] For example, the computing device (100) can acquire a multi-channel surface electrocardiogram signal including a standard 12-lead electrocardiogram, an 18-lead electrocardiogram, or additional chest leads.

[0104] Additionally, the computing device (100) can obtain high-density electrophysiological signals obtained from a body surface potential mapping system, a vector electrocardiogram, or electrophysiological signals from a wearable device.

[0105] Here, multichannel bio-information is not limited to the information mentioned above and, if necessary, may further include various types of bio-information such as pulse waves (PPG) time-synchronized with the electrocardiogram, cardiac vibration signals (seismocardiogram) and whole-body vibration signals (ballistocardiogram), heart sounds (phonocardiogram), blood pressure waveforms, and respiration signals, and these bio-information may be used together with the electrocardiogram.

[0106] In step S120, the computing device (100) can extract a first feature vector corresponding to electrical features related to the heart condition using multichannel bio-information obtained through step S110.

[0107] In various embodiments, the computing device (100) can extract a first feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals using a plurality of encoders.

[0108] Here, the first feature vector is a vector that numerically represents electrical features related to the heart state extracted from multichannel bio-information, and may include the shape of the electrocardiogram waveform, temporal changes, frequency components, inter-channel relationships, or a combination thereof.

[0109] For example, a computing device (100) may receive an electrocardiogram signal as a time-domain electrophysiological signal through a time-domain encoder among a plurality of encoders and extract a first feature vector that reflects the morphological features of the electrocardiogram waveform. At this time, the first feature vector may include the position of the R-peak, the RR interval, the rising and falling slopes of the waveform, the duration of the waveform, the amount of change in signal amplitude, time delay information between channels, or a combination thereof.

[0110] As another example, a computing device (100) may receive a time-frequency representation generated through a short-time Fourier transform or a wavelet transform as input through a time-frequency encoder among a plurality of encoders, and extract a first feature vector that reflects the change in frequency components of an electrocardiogram signal and the energy distribution over time. At this time, the first feature vector may include the energy magnitude of a specific frequency band, the energy ratio between frequency bands, a spectrum change pattern over time, or a combination thereof.

[0111] As another example, the computing device (100) may receive additional biosignals, such as pulse wave signals, heart sound signals, or vibration signals, as input through an additional modal encoder among a plurality of encoders, and extract a first feature vector corresponding to the characteristics of each signal. At this time, for the pulse wave signal, features related to the slope of the waveform, the reflection wave index, and the pulse wave propagation time may be included, and for the heart sound signal, features related to the energy distribution by frequency band and the interval between the timing of heart sound occurrence may be included.

[0112] In various embodiments, the computing device (100) can preprocess a plurality of electrophysiological signals and extract a first feature vector from each of the preprocessed plurality of electrophysiological signals.

[0113] At this time, the computing device (100) can preprocess the plurality of electrophysiological signals based on the characteristics of the plurality of electrophysiological signals.

[0114] For example, the computing device (100) can apply a notch filter, a low-pass filter, or a baseline variation removal filter to the input electrophysiological signal to remove noise and perform a channel-specific amplitude normalization or standardization process. This allows for the correction of scale differences between electrophysiological signals acquired under different measurement environments or sensor conditions.

[0115] As another example, the computing device (100) can resample multiple electrophysiological signals having different sampling frequencies to a common sampling frequency and align the time axes between the signals. Additionally, by aligning the signals based on a heart rate cycle or a reference event, a temporal correspondence between the multiple signals can be secured.

[0116] As another example, the computing device (100) may divide the electrophysiological signal into segments of a certain length or divide it based on a specific heart cycle or time window to use as an analysis unit. In this case, the segments may be configured to correspond to a single beat, multiple beats, or a certain time interval.

[0117] In step S130, the computing device (100) can generate a first latent space representation by mapping the first feature vector extracted through step S120 to a predefined electrical-anatomical latent space.

[0118] Here, the electrical-anatomical latent space may be a predefined latent coordinate space based on the correspondence between the anatomical structures of the myocardium and coronary arteries and the electrical conduction characteristics of the myocardium. The electrical-anatomical latent space may include multiple latent coordinates defined based on the anatomical structures of the myocardium and coronary arteries. The latent coordinates may be configured to correspond to anatomical regions including the left ventricle, right ventricle, interventricular septum, atrial region, and major branch regions of the coronary arteries, and each region may be divided into regions with different electrical conduction characteristics.

[0119] In various embodiments, the computing device (100) can align electrical features extracted from different channels based on anatomical location by mapping a first feature vector onto a potential coordinate system corresponding to an electrical-anatomical potential space.

[0120] More specifically, the computing device (100) can non-linearly map a first feature vector to an electrical-anatomical position in a latent coordinate system using a latent warping network. In this case, the latent warping network can align electrical features extracted from different channels based on anatomical positions by learning complex non-linear correspondences between the input feature space and the latent coordinate space.

[0121] For example, a latent spatial matching neural network can be implemented as a coordinate-based neural network structure that includes the inverse problem solving of a cardiac conduction model. In this case, the network can be trained to estimate the spatial distribution of electrical activity within the myocardium from observed electrical features.

[0122] As another example, a latent space matching neural network can be implemented in the form of a U-Net having an encoder-decoder structure, wherein the encoder extracts a latent space representation from a first feature vector, and the decoder can restore the latent space representation into a latent vector field corresponding to an anatomical coordinate system.

[0123] As another example, a latent space matching neural network can be implemented by combining a coordinate-based multilayer perceptron with an attention mechanism or a transformer structure. In this case, the network can align first feature vectors onto a latent coordinate system by considering inter-channel interactions and anatomical adjacency.

[0124] In addition, the latent spatial matching neural network may include a spatiotemporal neural network structure to simultaneously process temporal information, thereby enabling the simultaneous reflection of temporal changes and spatial distributions of electrical activity associated with heartbeats.

[0125] The output of the latent space matching neural network may be provided as a first latent space representation indicating the electrical activity state of the myocardial and coronary artery segment regions. In this case, the first latent space representation may be configured in the form of a latent vector field, and each latent vector may include the degree of electrical activity or conduction characteristics at a specific anatomical location.

[0126] In step S140, the computing device (100) can generate a coronary artery structure graph by structuring the first latent space representation generated through step S130.

[0127] In various embodiments, the computing device (100) can generate a coronary artery structure graph based on a predefined tree structure corresponding to the anatomical structure of the coronary artery.

[0128] Here, multiple nodes included in the tree structure for generating a coronary artery structure graph may include multiple vascular nodes corresponding to major coronary arteries and multiple segment nodes corresponding to the segmental structure of the coronary arteries. For example, the vascular nodes may include nodes corresponding to the left anterior descending artery, the left circumflex artery, and the right coronary artery, respectively. Additionally, the segment nodes may be connected to the vascular nodes and include nodes corresponding to anatomical segments, such as the proximal, intermediate, and distal parts of each coronary artery. For example, the segment node corresponding to the left anterior descending artery may include nodes corresponding to the proximal left anterior descending artery segment, the intermediate left anterior descending artery segment, and the distal left anterior descending artery segment, and similar segmental structures may be defined for the left circumflex artery and the right coronary artery.

[0129] Here, according to various embodiments of the present disclosure, a tree structure for generating a coronary artery structure graph is described as being generated and defined in advance, and a coronary artery structure graph is generated based thereon, but is not limited thereto.

[0130] In various embodiments, the computing device (100) can use a first weighted mapping matrix to assign information included in the first latent space representation to each of a plurality of nodes as a state vector.

[0131] Here, the first weighted mapping matrix is ​​a matrix representing the correspondence between the first feature vector extracted from the bio-information and the plurality of nodes, and the plurality of elements included in the first weighted mapping matrix may represent weights representing the degree to which the first feature vector extracted from the bio-information contributes to each of the plurality of nodes.

[0132] In various embodiments, the computing device (100) can extract a global ischemia pattern for the heart based on a first latent space representation, generate a global monitoring node mapped to the global ischemia pattern, and connect the generated global monitoring node to at least one of a plurality of nodes.

[0133] By connecting global monitoring nodes to the coronary artery structure graph in this manner, it is possible to reflect not only local electrical abnormalities observed in individual coronary artery segments but also global ischemic tendencies occurring throughout the heart. Through this, global ischemic patterns are reflected in the state vector update process of each segment node, allowing for a more coherent representation of interactions between nodes on the coronary artery structure graph.

[0134] In various embodiments, the computing device (100) can define a reference state based on bio-information obtained from a group of healthy people without heart disease, generate a normal reference node mapped to the defined reference state, and connect the generated normal reference node to at least one of a plurality of nodes.

[0135] In this way, by connecting normal reference nodes to the coronary artery structure graph, the relative deviation can be learned and represented by comparing the state of each coronary artery segment node with the normal reference state. Through this, the effects of physiological differences between patients or variations in the measurement environment can be mitigated, and the degree of abnormality for each coronary artery segment can be evaluated more stably and consistently.

[0136] In various embodiments, the computing device (100) can update the state vector mapped to each node by performing a message passing operation between multiple nodes included in the coronary artery structure graph using a graph neural network decoder. In this process, the adjacency relationship between coronary artery segments and information transmitted from a global surveillance node or a normal reference node may be reflected, thereby forming a node state representation that considers both local electrical abnormalities and global ischemia patterns.

[0137] Here, the graph neural network decoder can process each node included in the coronary artery structure graph using a graph convolutional neural network, a graph attention network, or a message passing network structure. The graph neural network decoder can iteratively update the state vector mapped to each node by performing message passing operations based on the connection relationships between nodes.

[0138] In various embodiments, the graph neural network decoder can model the pattern of local electrical abnormalities propagating to adjacent segments by transmitting messages that reflect the relationships between adjacent coronary artery segment nodes. Additionally, by including a global surveillance node or a normal reference node in the graph, information regarding global ischemia patterns can be reflected in the state updates of each segment node.

[0139] In addition, the graph neural network decoder may include in the loss function a smoothing continuity constraint to prevent excessive fluctuation in electrical activity values ​​between adjacent segments within the same coronary artery during the learning process, or a constraint to maintain the monotonicity of electrical activity in response to changes in ischemic state.

[0140] In various embodiments, the computing device (100) may modify a coronary artery structure graph to reflect the anatomical structure and electrical characteristics of the coronary artery. In this case, the coronary artery structure graph may be configured in different forms depending on the purpose of analysis or processing step.

[0141] For example, a coronary artery structure graph can be implemented as a static graph structure. In this case, the graph has a constant node and edge structure regardless of time, where each node represents a segment of the coronary artery and the edges represent anatomical connections between segments.

[0142] As another example, a coronary artery structure graph can be implemented as a dynamic graph structure. In this case, the state of nodes or the weights of edges can be configured to change according to the cardiac cycle or time interval, thereby reflecting changes in electrical activity or ischemia patterns over time.

[0143] As another example, a coronary artery structure graph can be implemented as a bipartite graph structure. In this case, the graph may be configured to include coronary artery segment nodes and electrical feature or channel nodes in different sets, and to map electrical information to coronary artery segments through edges between said sets.

[0144] As another example, a coronary artery structure graph can be implemented as a hypergraph structure. In this case, a single edge is configured to connect multiple nodes simultaneously, allowing for the representation of high-order correlations between coronary artery bifurcations or myocardial regions and multiple coronary artery segments.

[0145] As another example, a coronary artery structure graph can be implemented as a multi-graph structure that combines coronary artery graphs with other vascular graphs in parallel or hierarchically. For instance, by combining coronary artery graphs with carotid artery graphs or peripheral artery graphs, the correlation between the systemic vascular system and the myocardial electrical state can be analyzed together.

[0146] In step S150, the computing device (100) can derive a first indicator related to heart disease using the coronary artery structure graph generated and updated through step S140.

[0147] In various embodiments, the computing device (100) can derive a plurality of first indicators by analyzing each coronary artery structure graph using a plurality of prediction heads.

[0148] For example, the computing device (100) can generate a first indicator that quantitatively indicates the degree of reduced blood flow due to coronary artery stenosis by using an FFR prediction head to estimate the FFR value corresponding to each coronary artery segment node or branch node in a regression manner.

[0149] As another example, the computing device (100) can generate a first indicator that binary classifies the presence of ischemia or classifies the severity of ischemia into multiple intervals based on the predicted FFR value or structural / functional indicator using an ischemia classification head.

[0150] As another example, the computing device (100) can derive a first indicator indicating the severity of a structural lesion by using a stenosis rate or degree of stenosis prediction head to estimate the stenosis rate or degree of reduction of effective cross-sectional area of ​​each coronary artery segment as a continuous value.

[0151] As another example, the computing device (100) can generate a first indicator related to the likelihood of an acute cardiovascular event by calculating the plaque vulnerability score or whether each segment is a high-risk plaque using a plaque vulnerability prediction head.

[0152] As another example, the computing device (100) can generate a first indicator that reflects the cardiovascular functional state in multiple ways by using an additional hemodynamic indicator prediction head to predict CFR, IMR, left ventricular ejection fraction, left ventricular end-stage volume, left ventricular filling pressure estimate, or other hemodynamic variables.

[0153] As another example, the computing device (100) can generate a first indicator that predicts the likelihood of a major cardiovascular event occurring within a certain period, the risk of worsening heart failure, the need for revascularization, or the risk of death using a prognosis and risk prediction head.

[0154] Additionally, the computing device (100) can use a virtual image generating head to generate a first indicator in the form of a virtual CCTA vascular map or an equivalent image that reflects coronary artery stenosis or plaque distribution.

[0155] Additionally, the computing device (100) can calculate a reliability or uncertainty score for each prediction result using an uncertainty estimation and correction head, and generate an auxiliary indicator to induce additional inspection or re-evaluation when the prediction uncertainty is high.

[0157] FIG. 4 is a diagram illustrating the configuration of an indicator derivation module according to a second embodiment of the present disclosure.

[0158] Referring to FIG. 4, in a second embodiment of the present disclosure, a computing device (100) can derive a second indicator related to structural heart disease through a second indicator derivation module (600). To this end, the second indicator derivation module (600) may include a plurality of second encoders (610), a second latent space alignment module (620), a second graph generation module (630), a second graph neural network decoder (640), and a plurality of second prediction heads (650).

[0159] First, a plurality of second encoders (610) can receive multi-channel biometric information as input data and can extract a second feature vector from the input multi-channel biometric information.

[0160] Here, multichannel bio-information is collected through multiple different channels and may include at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to the heart condition.

[0161] For example, multichannel bio-information may include a standard 12-lead electrocardiogram, an 18-lead electrocardiogram, or a multichannel surface electrocardiogram signal including additional chest leads.

[0162] In addition, multichannel bio-information may include high-density electrophysiological signals obtained from a body surface potential mapping system, vector electrocardiograms, and electrophysiological signals of non-standard lead configurations obtained from a wearable device or a patch-type device.

[0163] In addition, multichannel bio-information may further include multimodal bio-signals such as pulse waves, heart sounds, and body surface vibration signals as needed.

[0164] In addition, multichannel bioinformation may include echocardiogram images, cardiac MRI images, or structural or functional features extracted therefrom.

[0165] In addition, multichannel biometric information may further include the patient's clinical / demographic information or text information as needed.

[0166] Here, clinical / demographic information may include the patient's age, sex, height, weight, body mass index (BMI), body surface area (BSA), blood pressure, heart rate, blood glucose, lipids, creatinine, and other laboratory test results, smoking status, comorbidities such as diabetes, hypertension, chronic kidney disease, and coronary artery disease, history of past procedures and surgeries, diagnostic codes (ICD, etc.), test / prescription history, and summaries of past echocardiogram / CT / MRI results.

[0167] In addition, the text information provided here may include natural language text information such as electrocardiogram machine reading text reports, echocardiogram reading summaries, discharge summaries, and free-text memos within electronic medical records. Such text can provide high-dimensional information that complements structural or functional abnormalities not directly observed in electrocardiogram waveforms, clinician judgments, and uncertainties.

[0168] Each second encoder (610) may be configured with a different structure depending on the type or characteristics of the input bio-information. For example, a plurality of second encoders (610) may include an encoder that receives a time-domain electrophysiological signal as input and extracts features reflecting the shape of the electrocardiogram waveform, the interval between beats, and temporal changes; an encoder that receives a time-to-frequency converted bio-signal as input and extracts features reflecting vibration characteristics or energy distribution by frequency band; an encoder that receives a heart image as input and extracts features reflecting the volume, wall thickness, or structural shape of the left and right ventricles; and / or an encoder that receives clinical and text information as input and extracts features reflecting the patient's overall clinical context.

[0169] Each of these multiple second encoders (610) can be configured in parallel, and each can generate a second feature vector that reflects structural and functional characteristics.

[0170] Next, the second latent space alignment module (620) can receive the second feature vector extracted by the plurality of second encoders (610) as input and generate a second latent space representation.

[0171] The second latent space alignment module (620) is a predefined structural-functional latent space based on the correspondence between the anatomical structure of the heart and the mechanical or hemodynamic functional characteristics of the heart, and the second latent space alignment module (620) can map a second feature vector to the latent space. Through this, structural and functional features extracted from different modalities can be converted into a latent space representation aligned on a common latent coordinate system. For example, features related to the degree of left ventricular expansion, contraction, or relaxation function can be integrally represented within the same latent coordinate system.

[0172] Next, the second graph generation module (630) can generate a heart structure graph based on the second latent space representation generated by the second latent space alignment module (620).

[0173] Next, the second graph generation module (630) can generate a heart structure graph based on the second latent space representation generated by the second latent space alignment module (620).

[0174] The second graph generation module (630) can generate a graph using a predefined cardiac anatomical structure template. For example, the cardiac structure graph may include multiple structure nodes corresponding to the left ventricle, right ventricle, left atrium, right atrium, interventricular septum, heart valves, etc., and a node state vector assigned from the second latent space representation may be mapped to each node.

[0175] Next, the second graph neural network decoder (640) can receive the heart structure graph as input and update the node state vector.

[0176] The second graph neural network decoder (640) can update the node state vector by performing message passing operations between structure nodes on the heart structure graph, thereby reflecting the interrelationships between heart structures and global functional patterns. Through this, a representation can be formed that considers not only local abnormalities of specific structures but also the structural and functional state of the entire heart.

[0177] Next, a plurality of second prediction heads (650) can analyze a heart structure graph updated by a second graph neural network decoder (640) to derive a plurality of second indicators, respectively. For example, each of the plurality of second prediction heads (650) can derive at least some of the following: structural indicators related to the volume, mass, or wall thickness of the left and right ventricles; functional indicators related to left ventricular ejection fraction, ventricular contraction or relaxation function; disease indicators indicating the presence or severity of structural heart disease; and prognostic risk indicators related to the worsening of heart failure, hospitalization, or occurrence of major cardiovascular events within a certain period. In this way, different types of indicators can be generated in parallel through the plurality of prediction heads.

[0178] The components included in the above second indicator derivation module (600) can sequentially perform supervised learning, auxiliary learning, self-supervised learning, and contrastive learning.

[0179] The indicator derivation module (600) can perform supervised learning using pairs of cardiac image information and electrophysiological signals. For example, the encoder (610), latent space alignment module (620), graph neural network decoder (640), and prediction head (650) can be trained to predict a second indicator by utilizing structural and functional indicators obtained from an echocardiogram image or a cardiac MRI image as ground truth data.

[0180] Next, supplementary learning using clinical / demographic information or text information can be performed following supervised learning. Through this, the patient's overall clinical context, which is difficult to reflect using only images and biosignals, can be integrated into the model.

[0181] Next, after auxiliary learning is performed, self-supervised or comparative learning can be carried out using large-scale biometric data where ground truth data does not exist. Through this, generalization performance for various patient groups and measurement environments can be improved.

[0183] FIG. 5 is a flowchart of a method for deriving heart disease indicators through multi-channel bio-information analysis according to a second embodiment of the present disclosure.

[0184] Referring to FIG. 5, in step S210, the computing device (100) can acquire multichannel bio-information related to the heart condition.

[0185] Here, multichannel bio-information is collected through multiple different channels and may include at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to the heart condition. For example, the computing device (100) may acquire surface electrocardiogram signals, heart sound signals, pulse wave signals, or heart vibration signals. Additionally, echocardiogram images, cardiac MRI images, or structural or functional features extracted from these may be acquired together.

[0186] Additionally, the computing device (100) may selectively obtain clinical / demographic information such as the patient's age, gender, comorbidities, medication history, and examination history, or text information including cardiac imaging findings, diagnostic records, and examination reports.

[0187] In step S220, the computing device (100) can extract a second feature vector corresponding to structural and functional features related to the heart condition using multichannel bio-information obtained through step S210.

[0188] Here, the second feature vector is a vector that numerically represents structural and functional features related to the heart state extracted from multichannel bio-information, and may include morphological characteristics of cardiac anatomical structures, functional characteristics related to cardiac contraction and relaxation, correlations between electrical and mechanical signals, temporal changes, or combinations thereof.

[0189] For example, a computing device (100) can receive a multi-channel electrocardiogram waveform as input through a time domain encoder among a plurality of encoders and extract a second feature vector that reflects functional features related to the heart contraction and relaxation cycles. At this time, the second feature vector may include a heart beat cycle, the duration of the ventricular contraction and relaxation intervals, inter-beat variability, a temporal pattern related to conduction delay, or a combination thereof.

[0190] As another example, the computing device (100) may receive a time-frequency spectrogram generated from an electrocardiogram signal as input through a time-frequency encoder among a plurality of encoders and extract a second feature vector that reflects the frequency domain characteristics of cardiac electrical activity. In this case, the second feature vector may include energy distributions in low-frequency and high-frequency bands, mutual ratios between frequency bands, changes in frequency components over time, or a combination thereof.

[0191] As another example, the computing device (100) may use a heart sound encoder among a plurality of encoders to receive a heart sound waveform or a heart sound spectrogram as input and extract a second feature vector that reflects mechanical features related to heart valve operation. In this case, the second feature vector may include the timing of occurrence of the first heart sound (S1) and the second heart sound (S2), the time interval between heart sounds, the heart sound energy distribution, or a combination thereof.

[0192] As another example, the computing device (100) may receive a pulse wave signal as input through a pulse wave encoder among a plurality of encoders and extract a second feature vector that reflects the mechanical response characteristics of the cardiovascular system. In this case, the second feature vector may include pulse wave rise and fall speeds, a reflected wave index, a pulse wave propagation time, or a combination thereof.

[0193] As another example, a computing device (100) may use a bit token transformer encoder to receive a bit token sequence composed of beat units or time interval units as input and extract a second feature vector that reflects cardiac structural patterns and long-term time dependencies. In this case, the second feature vector may include repeating beat structures, the frequency of appearance of abnormal patterns, long-term trend information, or a combination thereof.

[0194] Additionally, the computing device (100) may receive the patient's clinical information and medical text as input through a clinical / demographic encoder and a text encoder, and may additionally generate a second feature vector that reflects exogenous factors affecting the interpretation of the heart condition. At this time, the second feature vector may include age, gender, underlying disease, summary information of test results, or a combination thereof.

[0195] In various embodiments, the computing device (100) can preprocess a plurality of electrophysiological signals and extract a second feature vector from each of the preprocessed plurality of electrophysiological signals.

[0196] At this time, the computing device (100) can preprocess multiple electrophysiological signals based on the input characteristics of each of the multiple encoders.

[0197] For example, the computing device (100) can perform spatiotemporal alignment based on the heart cycle and heart sound. For instance, the computing device (100) can detect beat boundaries through R-peak detection in the electrocardiogram signal and detect S1 and S2 in the heart sound signal. Additionally, electrical events (QRS, ST-T, etc.) of the electrocardiogram signal can be matched with heart sounds or mechanical signals (systole, diastole, valve opening / closing intervals) to synchronize them on the time axis. Through this, a spatiotemporal representation in which electrical and mechanical events corresponding to the same heart cycle are aligned can be generated.

[0198] As another example, the computing device (100) can perform shape-based bit tokenization and clustering. For instance, the computing device (100) can generate beat-unit embeddings by extracting features such as QRS width, ST-T shape, presence of P wave, potential change, and time-frequency pattern for each beat. Subsequently, based on these embeddings, the beat pattern can be quantized into multiple bit tokens or clustered to form a token distribution corresponding to structural patterns associated with ventricular hypertrophy, ventricular overload, abnormal repolarization, atrial fibrillation, or atrial flutter. This token distribution can then be used as input for structural-functional latent space alignment.

[0199] As another example, the computing device (100) can perform noise section detection and removal. For instance, the computing device (100) can simultaneously analyze electrocardiogram and heart sound signals to detect noise sections such as electromyogram, movement, and poor electrode contact. Additionally, high-frequency or abnormal energy patterns distinct from the band of normal heart signals can be identified through time-frequency analysis, and masking, interpolation, or weight reduction processing can be performed on these noise sections. Through this, the impact of noise can be mitigated in the subsequent feature extraction step.

[0200] As another example, the computing device (100) can perform clinical, demographic, and text information preprocessing. For instance, the computing device (100) can classify clinical and demographic variables such as age, gender, body mass index, blood pressure, and blood sugar into continuous, discrete, or categorical types, perform outlier processing and missing value imputation, and then normalize them. Additionally, for text-based clinical information, tokenization, subword splitting, or sentence-level segmentation can be performed, and then converted into fixed-length embedding vectors through a pre-trained language model or a dedicated text encoder. The embedding vectors can be used as the initial state of a global context node or integrated as auxiliary inputs injected into a structure node or a disease node.

[0201] In step S230, the computing device (100) can generate a second latent space representation by mapping the second feature vector extracted through step S220 to a predefined structural-functional latent space.

[0202] Here, the structural-functional latent space may be a predefined latent coordinate space based on the correspondence between the anatomical structure of the heart, cardiac function, and structural disease state.

[0203] In a structural-functional latent space, each axis or region of the latent coordinate system can be defined based on structural and functional parameters such as left and right ventricular volume, wall thickness, mass, ejection fraction, systolic and diastolic characteristics, and valve function extracted from cardiac imaging information.

[0204] At this time, the structural-functional latent space can be defined independently of the biological information used as input in the inference stage and may be a coordinate space learned in advance based on reference data including cardiac image information. For example, cardiac image information and electrophysiological signals may be input as pairs in the learning stage, and the features of the electrophysiological signal corresponding to the same cardiac structural and functional state may be learned to be aligned with the cardiac image-based representation on the structural-functional latent space.

[0205] Through this, even if cardiac imaging information is not provided as input during the actual inference step, a second feature vector extracted from non-imaging biological information, such as electrocardiograms, heart sounds, pulse waves, or clinical information, can be mapped to meaningful coordinates on the structural-functional latent space. That is, the structural-functional latent space may be a representation space defined to indirectly represent the structural and functional state of the heart using only non-imaging biological information.

[0206] In various embodiments, the computing device (100) can represent structural and functional features extracted from different modalities on a common basis by aligning a second feature vector on the latent coordinate system. For example, features related to the degree of left ventricular enlargement, or the decline in systole or diastolic function, can be represented integrally within the same latent coordinate system.

[0207] In step S240, the computing device (100) can generate a heart structure graph by structuring the second latent space representation generated through step S230.

[0208] In various embodiments, the computing device (100) can generate a heart structure graph by assigning information included in a second latent space representation as a state vector to each of a plurality of nodes based on a predefined tree structure corresponding to the anatomical structure of the heart.

[0209] Here, a plurality of nodes included in the tree structure for generating a cardiac structure graph may include structural nodes corresponding to structural components such as the left ventricle, right ventricle, left atrium, right atrium, interventricular septum, or heart valves, and disease nodes corresponding to structural or functional heart diseases.

[0210] Here, according to various embodiments of the present disclosure, a tree structure for generating a heart structure graph is described as being generated and defined in advance, and the heart structure graph is generated based thereon, but is not limited thereto.

[0211] In various embodiments, the computing device (100) can weightly assign information included in the second latent space representation to each node by reflecting the correspondence between the second feature vector and each node using the second weighted mapping matrix.

[0212] Here, the second weighted mapping matrix is ​​a matrix representing the correspondence between the second feature vector extracted from the biometric information and the plurality of nodes, and the plurality of elements included in the second weighted mapping matrix may be weights representing the degree to which the second feature vector extracted from the biometric information contributes to each of the plurality of nodes.

[0213] In various embodiments, the second weighted mapping matrix may be configured to reflect the hypergraph structure. In this case, the second weighted mapping matrix may be defined such that a single second feature vector component corresponds to a hyperedge that simultaneously connects multiple structure nodes and disease nodes, thereby expressing a relationship in which a single feature contributes jointly to multiple sets of nodes.

[0214] Specifically, a single hyperedge can be configured to simultaneously connect a specific set of coronary artery segments with a specific disease state or functional feature. In this case, the weights of the second weighted mapping matrix may be applied jointly to multiple nodes connected through the hyperedge, or applied differentially to each node.

[0215] In addition, by utilizing a hypergraph structure, global or regional physiological characteristics affecting multiple coronary artery segments simultaneously can be effectively reflected, and complex interaction relationships between multiple structural nodes and disease nodes can be expressed.

[0216] In various embodiments, the computing device (100) may extract global structure-disease context features that commonly affect cardiac structure and disease state from clinical and demographic information including at least one of the subject's age, gender, comorbidities, drug history, and examination history, and text information including cardiac image readings, diagnostic records, or examination reports, and may generate a global sentinel node mapped with said global structure-disease context features, and may connect the generated sentinel node to at least one of a plurality of nodes included in a cardiac structure graph.

[0217] By connecting global surveillance nodes to the cardiac structure graph in this manner, not only local structural or functional changes observed at individual cardiac structure nodes but also the patient's clinical background, disease history, and overall structure-disease context can be reflected. Through this, global structure-disease context features are incorporated into the state vector update process of each structure node and disease node, allowing for a more coherent representation of interactions between nodes on the cardiac structure graph.

[0218] In various embodiments, the computing device (100) may define a reference structural state of the heart based on bio-information obtained from a group of healthy people without heart disease, generate a virtual healthy node mapped to the reference structural state, and connect the generated virtual healthy node to at least one of a plurality of nodes included in a heart structure graph.

[0219] By connecting normal reference nodes to the cardiac structure graph in this manner, the state of each structural node can be compared with the normal reference structural state to learn and represent relative deviations. This allows for the mitigation of the effects of physiological differences among patients, as well as variations in age, body type, and measurement environments, and enables a more stable and consistent assessment of the severity of cardiac structural abnormalities or structural heart disease.

[0220] In various embodiments, the computing device (100) can update the state vector mapped to each node by performing a message passing operation between a plurality of nodes included in the heart structure graph using a graph neural network decoder.

[0221] More specifically, the graph neural network decoder may be configured to iteratively update the state vector mapped to each node by performing a message passing operation based on the connection relationships between multiple nodes included in the heart structure graph. In this case, the heart structure graph may include structure nodes, disease nodes, and global context nodes, and the graph neural network decoder may update the node state by simultaneously considering the interactions between the nodes.

[0222] A graph neural network decoder can reflect the impact of disease states on cardiac structure through message transmission between structural nodes and disease nodes. In this process, messages transmitted from disease nodes may include information indicating the severity, location of occurrence, or progression of the disease, and the state vectors of structural nodes can be adjusted to reflect this information.

[0223] Furthermore, the graph neural network decoder includes a global context node to reflect global information, such as the patient's age, comorbidities, and medication history, in the state updates of structural and disease nodes. Through this, node state representations can be formed so that even the same structural state can be interpreted differently depending on the patient's specific background information.

[0224] Additionally, graph neural network decoders can include physical normalization terms in the loss function to reflect physiological relationships during the learning process. For example, constraints reflecting the pressure-volume relationship, where ejection fraction generally increases as left ventricular volume decreases, can be applied. As another example, structural constraints can be applied to ensure that changes in cardiac structure due to myocardial hypertrophy or ventricular dilation maintain anatomical consistency.

[0225] In addition, the graph neural network decoder can be trained to maintain physical consistency between cardiac structure and function by including normalization constraints that reflect the correlation between physiological indicators such as myocardial wall thickness, ventricular volume, and blood pressure.

[0226] In addition, the graph neural network decoder can process time-phase condition graphs to sequentially connect cardiac structure-disease states corresponding to different cardiac cycles or time intervals of the same patient. In this case, by sequentially modeling node states that change over time, temporal continuity can be reflected, ensuring that the cardiac structure or disease state does not change abruptly over a short period.

[0227] In step S250, the computing device (100) can derive a second indicator related to heart disease using the heart structure graph generated and updated through step S240.

[0228] For example, a computing device (100) can use a structural indicator prediction head to estimate structural indicators, such as left ventricular ejection fraction, end-terminal volume, end-terminal systolic volume, left and right ventricular mass, myocardial wall thickness, atrial size, or ventricular volume change, from the state vector of a structural node in the form of continuous values ​​or interval values. At this time, the structural indicators can be used as a second indicator that quantitatively represents the cardiac structure and functional state.

[0229] As another example, a computing device (100) can classify the presence or severity of structural heart disease from the state vector of a disease node or a structure node into binary or multi-grades using a disease diagnosis and severity prediction head. In this case, the disease diagnosis and severity prediction head can provide the possibility of the presence of reduced left ventricular ejection fraction, severe aortic stenosis, ischemic heart disease, significant myocardial dysfunction, or significant atrial effusion as separate outputs, respectively.

[0230] As another example, the computing device (100) can generate predictive indicators regarding the likelihood of a cardiovascular event occurring within a certain period, the risk of hospitalization for heart failure, the risk of cardiovascular death, or the need for a heart transplant or assistive device by considering the patient's clinical / demographic information and global context together using a prognosis and test recommendation head.

[0231] As another example, the computing device (100) can calculate the prediction reliability or uncertainty score for each structural indicator or disease prediction result using an uncertainty estimation head. In this case, the uncertainty estimation head may include output values ​​in the form of the variance, confidence interval, or probability distribution of the prediction result, and may generate an auxiliary signal that induces additional testing or re-measurement if the uncertainty exceeds a predetermined standard.

[0233] FIG. 6 is a diagram illustrating the functional module configuration of a computing device performing the third and fourth embodiments of the present disclosure, and FIG. 7 is a diagram schematically illustrating the hierarchical structure of a prediction model according to various embodiments of the present disclosure.

[0234] Referring to FIGS. 6 and 7, the computing device (100) is responsible for the entire artificial intelligence processing pipeline from the input of electrocardiogram data to the output of results, and may include data reception, preprocessing, prediction, judgment, model management, and result provision functions.

[0235] The computing device (100) may include a data receiving module (110), a preprocessing module (120), a prediction module (130), a judgment module (140), a model management module (150), and a result providing module (160). Each module may be implemented in software or hardware and may be executed by a processor (not shown) of the computing device.

[0236] The data receiving module (110) is configured to receive electrocardiogram data from a user terminal or electrocardiogram measuring device.

[0237] In one embodiment, the data receiving module (110) can securely receive data through an encrypted communication protocol such as HTTPS or TLS. In addition, it can perform hash verification or transmission log recording functions to verify the integrity of the received data. Through this, data can be securely received even in a remote medical environment, and the risk of patient data leakage can be minimized.

[0238] In another embodiment, the data receiving module (110) can support medical data standard formats such as HL7 and FHIR to ensure interoperability with the hospital's electronic medical record (EHR) system.

[0239] In addition, when the data receiving module (110) communicates directly with the edge AI device, it may also receive electrocardiogram signals sequentially through a real-time streaming data interface.

[0240] Meanwhile, the received electrocardiogram data can be temporarily stored in internal memory and then transferred to a preprocessing module (120) to undergo standardization and normalization processes.

[0241] The preprocessing module (120) is configured to standardize and normalize electrocardiogram data input through the data receiving module (110) so that the prediction model (e.g., FFR prediction model and / or coronary artery stenosis rate prediction model) described later can operate stably.

[0242] In one embodiment, the preprocessing module (120) can parse electrocardiogram data of various formats (XML, HL7, DICOM, JPG, JPEG, PNG, PDF, etc.) and convert them into an internal data structure. For example, the preprocessing module (120) can extract electrocardiogram signals and metadata (measurement time, sampling rate, patient identification information, etc.) for each lead.

[0243] In addition, the preprocessing module (120) performs standardization and normalization on the electrocardiogram signal of each lead, so that the scale of the data collected from different electrocardiogram equipment or measurement environments can be unified.

[0244] Here, standardization is the process of converting the mean of each lead signal to 0 and the standard deviation to 1, and normalization is the process of scaling the standardized data to a specific range (e.g., 0 to 1).

[0245] In another embodiment, the preprocessing module (120) may further perform additional signal purification processes such as filtering, resampling, and denoising.

[0246] For example, in the filtering process, a high-pass filter to remove low-frequency baseline fluctuations (e.g., 0.5 Hz cutoff), a notch filter to remove power noise (e.g., 50 Hz or 60 Hz), and a low-pass filter to remove electromyography noise (e.g., 40 Hz cutoff) may be applied.

[0247] In addition, resampling is a process of converting data measured at different sampling rates into a unified sampling rate (e.g., 500 Hz), and methods such as linear interpolation or cubic spline interpolation can be used.

[0248] In addition, noise removal can be performed using wavelet transform or signal averaging techniques.

[0249] In another embodiment, the preprocessing module (120) can convert the electrocardiogram data into a two-dimensional spectrogram through a time-frequency transform.

[0250] For example, a one-dimensional time series signal can be transformed into a two-dimensional time-frequency representation using the Short-Time Fourier Transform (STFT) or Wavelet Transform. In this case, the prediction model can analyze the corresponding spectrogram using a 2D CNN or Vision Transformer.

[0251] In the present invention, the type of preprocessing, filter parameters, sampling rate, etc., can be adjusted through a setting file, and some steps may be omitted depending on the quality of the input data or the characteristics of the measurement equipment.

[0252] Meanwhile, the preprocessing module (120) according to the present invention can dynamically change the parameters of each processing step through a configuration file or an administrator interface, so it can flexibly respond to various medical environments and types of equipment.

[0253] The prediction module (130) is configured to calculate the degree of coronary artery stenosis and fractional flow reserve (FFR) values ​​using the preprocessed patient's electrocardiogram data, or to calculate the coronary artery stenosis rate using the preprocessed patient's electrocardiogram data and the electrocardiogram data and patient data together.

[0254] In one embodiment, the prediction module (130) can calculate the degree of coronary artery stenosis and the coronary artery stenosis rate as a percentage (%) before calculating the FFR value.

[0255] Generally, the degree of stenosis is closely correlated with the amount of change in electrocardiogram signals, and the more severe the stenosis, the more pronounced the electrical changes caused by ischemia (ST segment depression, T-wave inversion, etc.) become.

[0256] When actual stenosis rate data measured by coronary angiography and electrocardiogram data are provided together during the learning process, the prediction module (130) can learn the relationship between the electrocardiogram pattern and the stenosis rate through this.

[0257] For example, the calculation of the stenosis rate can be performed using a regression neural network based on morphological features of the electrocardiogram waveform, such as ST segment depression, T-wave inversion, inter-lead phase difference, and potential gradient. In particular, the stenosis rate for each vessel can be quantitatively estimated by learning the correlation between the time-frequency domain pattern per lead and the coronary angiography (CAG) stenosis rate.

[0258] At this time, the prediction module (130) can classify the degree of stenosis into mild (stenosis rate <40%), moderate (stenosis rate 40~70%), and severe (stenosis rate >70%) based on the result of calculating the stenosis rate.

[0259] If the calculated stenosis rate is determined to be a moderate lesion of 40 to 70%, the prediction module (130) can additionally calculate the FFR value to determine the actual necessity of a coronary artery procedure (PCI). This is because, in the case of a moderate lesion, it is difficult to determine whether to perform the procedure based solely on the stenosis rate, and if the FFR is less than 0.75, the procedure is necessary, and if it is 0.8 or higher, drug treatment may be appropriate.

[0260] In addition, the prediction module (130) can analyze ischemia-related electrophysiological patterns from input electrocardiogram data and quantitatively predict the degree of myocardial ischemia through a learned prediction model.

[0261] In addition, the prediction module (130) can analyze electrophysiological characteristics associated with coronary artery stenosis from input electrocardiogram data and quantitatively predict the stenosis rate of each vascular section of the coronary artery through a learned prediction model.

[0262] Hereinafter, the hierarchical structure of a prediction model and the data processing process according to an embodiment of the present invention will be described with reference to FIG. 7.

[0263] Referring to FIG. 7, the prediction model (700) used in the prediction module (130) may include a pre-trained learning model (710), a transfer learning layer (720), an output layer (730), and a correction layer (340).

[0264] The learning model (710) may be a model trained to receive electrocardiogram signals and related patient information as input and generate a feature representation representing the heart condition.

[0265] In various embodiments, the learning model (710) may include at least one model among a machine learning (ML) model, a deep learning model, and a foundation model, depending on the implementation method.

[0266] This learning model (710) may be implemented as a single model or as a combination of two or more models. For example, it may be configured to derive a final prediction result by combining the output of a machine learning model that uses features extracted from an electrocardiogram signal as input and the output of a deep learning model that uses an electrocardiogram raw signal as input.

[0267] First, the machine learning model may be one that utilizes feature vectors extracted from raw electrocardiogram signals as input data. For example, the machine learning model may receive inputs such as time-domain or frequency-domain features, waveform shape features, or inter-lead relationship features extracted from the electrocardiogram waveform; additionally, clinical information such as the patient's age, gender, and underlying disease information (e.g., presence of diabetes, hypertension, etc.) may also be input.

[0268] Next, the deep learning model may be configured to automatically learn features by directly receiving raw electrocardiogram (ECG) signals as input. For example, the deep learning model can use matrix-shaped data (e.g., 12×5000) obtained by sampling 12-lead ECG data, which is typically measured in hospitals, over a certain period of time. Additionally, these ECG signals can be transformed in various ways as needed and used as input to the deep learning model. For instance, the ECG waveform can be represented as a two-dimensional image or converted into a time-series, frequency-based representation such as a spectrogram to be utilized as input data for the deep learning model. Furthermore, in addition to the ECG signals, the patient's clinical information may be combined as additional input to the deep learning model.

[0269] Next, the foundation model (710) can be configured to learn the temporal continuity of electrocardiogram waveforms and the interrelationships between leads from a large amount of electrocardiogram data.

[0270] In one embodiment, the foundation model (710) can be learned using self-supervised learning and can be implemented using a masked autoencoder, contrastive learning, or a combination thereof.

[0271] For example, a Masked Autoencoder can learn the temporal pattern of a signal by masking a portion (e.g., 15–30%) of the input electrocardiogram signal and learning to restore it, and contrast learning can improve the separation of the feature space by maximizing similarity between leads of the same patient and maximizing difference between signals of different patients.

[0272] In various embodiments, the prediction model (710) may be implemented in a hybrid structure that combines a foundation model and a machine learning model, that is, by utilizing a feature vector generated as the output of the foundation model as an input feature of the machine learning model. As such, the learning model (710) is not limited to a single algorithm and may be composed of one or more of a machine learning model, a deep learning model, and a foundation model, and through these various model structures, more stable and accurate heart condition estimation can be performed from electrocardiogram data.

[0273] In various embodiments, the prediction model (710) may derive results predicting FFR values ​​and coronary artery stenosis rates as result data. Here, the predicted FFR values ​​and predicted coronary artery stenosis rates may be in the form of numerical values, but depending on the case, they may be in the form of classification results classified based on a specific threshold.

[0274] First, the prediction model (710) can be implemented as a regression model that predicts FFR values ​​and coronary artery stenosis rates in the form of continuous numerical values. For example, the prediction model (710) can be trained to directly estimate continuous numerical values, such as the patient's FFR value or coronary artery stenosis rate, by analyzing input electrocardiogram data. If the prediction model (710) is a regression model, the electrocardiogram data and the patient's actual FFR test results or stenosis rate values ​​can be used as ground truth data to train the prediction model, and the trained prediction model can be configured to estimate an indicator value close to the actual value from the electrocardiogram data.

[0275] Next, the prediction model (710) may be implemented as a binary classification model that predicts FFR values ​​and coronary artery stenosis rates in the form of classification results distinguished based on specific reference values. For example, the prediction model (710) may be configured to analyze electrocardiogram data to derive a binary classification result that distinguishes whether FFR values ​​and coronary artery stenosis rates are within the normal range or the abnormal range.

[0276] For example, the prediction model (710) can be trained to output a binary classification result that determines whether the FFR value is greater than or less than 0.8, based on an FFR threshold of 0.8 used as a criterion for determining ischemia. In this case, the prediction model can generate a result that determines the patient's condition, for example, as normal / abnormal or Y / N, based on features extracted from the electrocardiogram data.

[0277] Additionally, the prediction model (710) can be implemented as a binary classification model that distinguishes whether there is stenosis based on a specific clinically meaningful threshold value (e.g., 50% or 70%) for coronary artery stenosis rate, and can also be implemented as a multi-class classification method that divides into multiple intervals as needed.

[0278] In various embodiments, the prediction model (710) may be configured as a combination of a regression model and a binary classification model. For example, the prediction model (710) may first predict a continuous numerical value, such as the patient's FFR value or stenosis rate, based on electrocardiogram data using a regression method, and then derive a result classifying the patient's condition by determining whether the predicted value is above a predefined threshold value.

[0279] That is, the prediction model (710) can be implemented in one or more ways, such as a regression model that directly predicts continuous numerical values, a binary classification model that classifies states based on thresholds, or a combination model that predicts numerical values ​​and derives classification results based on them, and these various model structures can be selectively applied depending on the applied data characteristics, learning environment, or clinical application purposes.

[0280] The transfer learning layer (720) is configured to perform fine-tuning on electrocardiogram data labeled as FFR test results using the weights of the learning model as initial values, or to perform fine-tuning on a large amount of unlabeled electrocardiogram data, electrocardiogram data labeled as coronary artery stenosis rate test results, and simulation-based synthetic electrocardiogram data using the weights of the learning model as initial values.

[0281] In this process, gray zone data with FFR values ​​between 0.75 and 0.8 may be excluded from training. This is because the gray zone is a boundary region where clinically, an FFR of less than 0.75 is considered abnormal and 0.8 or higher is considered normal, and including it in training can increase prediction instability near the boundary.

[0282] Therefore, the boundary stability of the model can be improved by training with data with FFR < 0.75 labeled as abnormal (ischemic) and data with FFR ≥ 0.8 labeled as normal (non-ischemic).

[0283] The output layer (730) receives a feature vector extracted from the transfer learning layer (720) and can predict the patient's FFR as a continuous value between 0 and 1 or predict the stenosis rate of the coronary artery section.

[0284] In some embodiments, the coronary artery stenosis rate (%) or myocardial ischemia risk score can be calculated together in the output layer.

[0285] In some embodiments, the output layer (730) can calculate the coronary artery stenosis rate (%), reliability, and decision-making assistance explanation for each blood vessel section.

[0286] In this case, reliability refers to a value that quantitatively represents a clinically reliable range of predicted stenosis rates, and decision aid explanation refers to auxiliary judgment information that medical staff can refer to when interpreting predicted stenosis rate results, such as 'recommendation for additional tests'.

[0287] Meanwhile, a prediction model (700) according to one embodiment of the present invention may further include a calibration layer (740). The calibration layer (740) may include a threshold-aware calibration layer and / or a calibration enhancement criterion-aware calibration layer.

[0288] Here, the threshold-aware correction layer is a configuration designed to mitigate prediction instability that may occur near the FFR threshold (e.g., near 0.75 and 0.80).

[0289] In addition, the correction enhancement criterion-aware correction layer here mitigates predictive instability that may occur near correction enhancement criteria, such as around 50% and 70% which correspond to the boundary region distinguishing normal from abnormal in the interpretation of stenosis rates, and improves the accuracy of the results.

[0290] At this time, the correction enhancement criteria can be set to 50% and 70% stenosis rates, as interpretation is most uncertain when the degree of stenosis is 40–60% and 60–80%, and may refer to representative clinical criteria values ​​used to distinguish between the moderate and severe of coronary artery disease.

[0291] For example, if the stenosis rate is 50% or higher, the likelihood of reduced blood flow begins to increase from this point, and if the stenosis rate is 70% or higher, it can be interpreted as a condition where a percutaneous coronary intervention (PCI) or additional imaging tests should be considered.

[0292] Accordingly, the present invention can be utilized for evaluating the stability of the degree of stenosis predicted based on 50% and 70% stenosis rates, determining whether to perform correction, calculating reliability, and providing decision-making assistance information for each blood vessel segment.

[0293] During the training phase, the calibration strength of the model can be dynamically controlled by increasing the weights of the loss function or adjusting the output scale for samples near a threshold. This enables the model to maintain stable regression results even near FFR thresholds and / or calibration enhancement criteria, thereby improving the reliability of actual clinical decision-making (e.g., determining the necessity of coronary intervention).

[0294] In addition, in another embodiment of the present invention, the prediction module (130) uses a foundation model, and can operate a one-dimensional (1D) foundation model and a two-dimensional (2D) foundation model in parallel to simultaneously learn time series information and frequency characteristics of the electrocardiogram signal.

[0295] For example, a 1D foundation model can calculate a first embedding from a raw electrocardiogram time series, and a 2D foundation model can calculate a second embedding from a spectrogram.

[0296] At this time, the prediction module (130) can form an integrated representation that reflects both temporal and frequency patterns by combining the first and second embeddings through a cross-attention or low-rank pooling-based fusion module, and can improve the accuracy of the prediction of stenosis rate and FFR and / or the accuracy of the prediction of coronary artery stenosis rate by blood vessel segment.

[0297] In addition, in one embodiment of the present invention, 1D and 2D foundation models can perform cross-modal knowledge distillation.

[0298] Specifically, cross-knowledge distillation involves learning to cross-reference feature representations between two models, aligning different representation spaces to form a common embedding space, thereby enabling generalization performance that is robust against distribution differences between data collection equipment or medical institutions.

[0299] Additionally, the prediction module (130) can calculate the uncertainty regarding the predicted FFR value and the predicted coronary artery stenosis rate value for each blood vessel segment in parallel.

[0300] Here, uncertainty can be defined as the variance or entropy of the prediction distribution and can be calculated, for example, using MC Dropout iterative inference, Deep Ensemble mean / variance, or temperature-corrected calibration indicators.

[0301] Accordingly, the result providing module (160) enables medical staff to quantitatively verify the prediction reliability by displaying a 'recommendation for additional testing' or 'FFR retest needed' message through the user terminal (200) when the uncertainty value exceeds a predefined clinical threshold (e.g., 0.75 or 0.80).

[0302] Meanwhile, the prediction model according to the present invention can be implemented as a CNN, RNN, Transformer, or a combination thereof, and the entire 12-lead electrocardiogram or a selected group of leads can be used as input.

[0303] Alternatively, in another embodiment, the prediction model may be implemented in a lightweight form and directly installed not only on a server but also on an edge AI device or an electrocardiogram measuring device.

[0304] With the configuration of such a prediction module (130), the patient's FFR can be quantitatively calculated non-invasively using only electrocardiogram data, and the degree of myocardial ischemia can be evaluated without an invasive FFR test.

[0305] In addition, depending on the configuration of such a prediction module (130), the coronary artery stenosis rate for each blood vessel segment of a patient can be calculated non-invasively by using only electrocardiogram data or by using electrocardiogram data together with patient data such as age, gender and machine automatic readings, and the degree of stenosis can be evaluated without invasive examination.

[0306] The judgment module (140) is configured to determine the risk of myocardial ischemia in a patient by comparing the predicted FFR value with a reference threshold, or to determine a blood vessel section outside the normal range by comparing the predicted coronary artery stenosis rate for each blood vessel section with a reference threshold.

[0307] In one embodiment, the judgment module (140) can classify the FFR value received from the prediction module (130) according to a preset criterion.

[0308] For example, if the FFR value is less than 0.75, it is classified as a high-risk group and determined that a percutaneous coronary intervention (PCI) is required; if it is 0.8 or higher, it is classified as a low-risk group and determined that drug treatment is appropriate; and if it is 0.75 or higher but less than 0.8, it is classified as an intermediate-risk group and determined that additional testing is required. These threshold values ​​are based on clinical guidelines and medical evidence and can be adjusted through a setting file according to the medical institution's policy or the patient's condition.

[0309] In another embodiment, the judgment module (140) can analyze the predicted FFR results to estimate the location of the coronary artery suspected of stenosis and the number of lesions.

[0310] For example, the judgment module (140) can identify areas where strong ischemic signals appear in specific lead groups by analyzing attention weights generated during the learning process of the interaction between leads of the prediction model, and based on this, estimate the location of the blood vessel suspected of stenosis.

[0311] Additionally, the judgment module (140) can estimate the number of coronary artery lesions by analyzing whether ischemic signals are detected independently in different lead groups.

[0312] In one embodiment, the judgment module (140) can analyze the coronary artery stenosis rate results for each blood vessel section received from the prediction module (130) to estimate the location of the coronary artery suspected of stenosis and the number of lesions.

[0313] In another embodiment, the judgment module (140) can analyze the predicted coronary artery stenosis rate results for each blood vessel segment to estimate the location of the coronary artery suspected of stenosis and the number of lesions.

[0314] For example, the judgment module (140) can identify areas where a high stenosis rate is repeatedly predicted in a specific group of leads by analyzing attention weights generated during the learning process of the interaction between leads of the prediction model, and based on this, estimate the location of the blood vessel suspected of stenosis.

[0315] Additionally, the judgment module (140) can estimate the number of coronary artery lesions by analyzing whether high stenosis rates appear repeatedly independently in different lead groups.

[0316] These judgment results can provide practical reference information for medical staff's treatment decisions (e.g., whether to perform PCI, whether to maintain drug treatment), and the calculation may be performed in the aforementioned prediction module (130) or a separate analysis module (not shown).

[0317] The above function may be performed integrally in the judgment module (140) or separated and performed in a separate vascular risk assessment module.

[0318] In another embodiment, the judgment module (140) may further perform a Multi-Vessel Ischemia Risk Evaluation by synthesizing multiple FFR prediction results for each vessel.

[0319] In this process, the judgment module (140) can calculate a comprehensive ischemia risk score for each patient by considering the FFR value for each blood vessel, the number of blood vessels below the threshold, and the correlation between ischemia leads.

[0320] This allows for the identification of patients with multivascular ischemia as well as single-vessel lesions, and enables clinicians to use this information as a reference when determining treatment priorities (vessels to be treated). The above functions may be performed integrally in the judgment module (140) or separated into a separate vascular risk assessment module.

[0321] Finally, the judgment module (140) can generate a prediction result including a predicted FFR value, a predicted coronary artery stenosis rate by blood vessel segment, a risk classification, a treatment judgment result, and information on the suspected location of stenosis, transmit it to a user terminal (200) via a network (400), and simultaneously store it in a database.

[0322] The model management module (150) is a configuration that performs loading, version management, updating, and performance monitoring of the prediction model.

[0323] In one embodiment, the model management module (150) can load the latest version of the prediction model parameters from the model repository of the database and load them into memory when the server starts or when a prediction request is made.

[0324] Additionally, each model version stored in the model management module (150) may include a version number, training date and time, training data information, and performance indicators (accuracy, AUC, sensitivity, specificity, etc.), and may be stored together with metadata.

[0325] The model management module (150) can periodically check the performance of the prediction model stored in the computing device (100) and update it as needed.

[0326] For example, the model management module (150) can automatically or manually perform or trigger model retraining when new FFR label data accumulates to a certain amount or more, and the updated model can be deployed only if the retraining result improves performance compared to the existing model.

[0327] Additionally, the model management module (150) can automatically or manually perform or trigger model retraining when new coronary artery stenosis rate label data by blood vessel segment or simulation-based coronary artery stenosis rate data accumulates to a certain amount or more, and the updated model can be deployed only if the performance is improved compared to the existing model as a result of the retraining.

[0328] Additionally, for example, the model management module (150) can monitor performance by comparing the prediction results with the actual inspection results, and provide a notification to the manager if performance degradation or a change in data distribution (data drift) is detected.

[0329] Through such functions, the computing device (100) can maintain a consistent learning strategy and stable prediction performance of the model.

[0330] The result providing module (160) is configured to transmit the prediction result calculated by the judgment module (140) to a user terminal (200) or an external system.

[0331] In one embodiment, the result providing module (160) can generate result data including a predicted FFR value, a predicted stenosis rate (%) by vascular segment, stenosis reliability, myocardial ischemia risk grade, and treatment recommendations (e.g., drug treatment, whether PCI is needed, etc.) and transmit this to a user terminal (200) in a visualization format.

[0332] For example, the result providing module (160) transmits the result using a REST API, HTTPS, or gRPC-based communication interface, and the user terminal (200) can check it through a web browser-based dashboard or a dedicated application.

[0333] Additionally, the result providing module (160) can provide result data by displaying the FFR value and the reference threshold (0.75, 0.8, etc.) together, or by visualizing it with a color code by risk level (e.g., red-high risk, yellow-medium risk, green-low risk) so that medical staff can intuitively interpret the results.

[0334] In addition, the result providing module (160) can provide result data by displaying the coronary artery stenosis rate by blood vessel section and the correction reinforcement threshold value (e.g., 50%, 70%) together, or by visualizing it with a color code by risk level (e.g., red-high risk, yellow-medium risk, green-low risk) so that medical staff can intuitively interpret the results.

[0335] In another embodiment, the result providing module (160) can link the prediction results with the hospital's electronic medical record (EHR) system or clinical decision support system. Through this, the prediction results can be automatically reflected in the patient's record or used as reference data when medical staff establish a treatment plan.

[0336] Additionally, the result providing module (160) can store the prediction results and judgment history in a database and use them for long-term follow-up analysis or model performance evaluation for the same patient. For example, by storing and displaying the trend of the predicted FFR value and / or the trend of the predicted coronary artery stenosis rate by vascular segment as a time series graph during repeated examinations of the same patient, changes in the cardiovascular condition can be intuitively monitored.

[0337] The result providing module (160) according to the present invention can be extended to a mobile application for patients or a remote monitoring system in addition to a UI for medical staff, thereby enabling monitoring of cardiovascular conditions and customized management even in a non-face-to-face environment. With this configuration, the present invention can rapidly support diagnosis and treatment decision-making by immediately delivering predicted FFR results to medical staff.

[0338] Meanwhile, although each module of the computing device (100) in this specification is described by being divided into functional units for convenience of explanation, in actual implementation, one or more modules may be integrated and executed as a single process, or specific functions may be separated and implemented as separate sub-modules. For example, some functions of the prediction module (130) and the judgment module (140) may be integrated to operate as a single integrated AI model.

[0339] Hereinafter, an embodiment of a method for predicting myocardial ischemia according to a third embodiment of the present disclosure will be described in detail with reference to FIGS. 8 and 9.

[0341] FIG. 8 is a flowchart of a method for predicting myocardial ischemia as an indicator according to a third embodiment of the present disclosure, and FIG. 9 is a conceptual diagram showing an example of a method for predicting myocardial ischemia according to a third embodiment of the present disclosure being performed in a clinical procedure.

[0342] The method of the present embodiment can be performed, for example, by medical staff uploading a patient's electrocardiogram data in a hospital cloud server environment or by automatically retrieving it from an electronic medical record (EHR) system. Below, the procedure for each step of FIG. 8 will be explained in detail with reference to the example in FIG. 9.

[0343] Referring to FIG. 8, a method for predicting myocardial ischemia using electrocardiogram data according to an embodiment of the present invention may include the steps of: collecting electrocardiogram (ECG) data of a patient (S310); standardizing and normalizing the electrocardiogram data (S320); predicting the degree of coronary artery stenosis using the preprocessed electrocardiogram data as input (S330); predicting a fractional flow reserve (FFR) value using the degree of stenosis as input when the predicted degree of stenosis falls within a predetermined range (S340); determining the risk of myocardial ischemia by comparing the predicted FFR value with a reference threshold value (S350); and outputting the prediction result (S360).

[0344] Below, with reference to FIG. 9, each step of FIG. 4 will be explained in detail, assuming a case where patient A visits a medical institution and undergoes an electrocardiogram test.

[0345] Referring to FIG. 9, the present invention can determine the risk of myocardial ischemia in a patient by sequentially performing a stenosis degree prediction function, an FFR prediction function, and a multi-vessel risk assessment function using 12-lead electrocardiogram (ECG) data as input. The functions illustrated in FIG. 9 can be performed stepwise within the prediction module (130) of the aforementioned server, and depending on each result, an appropriate treatment direction among drug treatment, PCI (stent procedure), or CABG (coronary artery bypass graft) can be suggested.

[0346] First, the computing device (100) can collect the patient's electrocardiogram (ECG) data (S310). For example, medical staff can measure a 12-lead electrocardiogram (ECG) and transmit the measured data to the computing device (100) through the hospital's HER system.

[0347] Alternatively, if patient A's 12-lead electrocardiogram data file (XML, HL7, DICOM, JPG, JPEG, PNG and PDF formats, etc.) is stored on a hospital server or user terminal (200), medical staff can select and transmit the file through the web interface of the computing device (100).

[0348] As another example, if the electrocardiogram measuring device is directly connected to a server via a network, the data may be transmitted to the computing device (100) in the form of real-time streaming immediately after the measurement is completed.

[0349] Additionally, 12-lead ECG data can be transmitted to a prediction module (130) through the input interface of a computing device (100). At this time, the data reception process is secured through encrypted communication (HTTPS, TLS, etc.), and metadata such as patient ID, time of examination, and measuring equipment can be stored together.

[0350] At this time, the electrocardiogram data can be temporarily stored in an internal buffer of the computing device (100) and then transferred to a preprocessing step.

[0351] In addition, the computing device (100) according to an embodiment of the present invention can standardize and normalize electrocardiogram data (S320).

[0352] In one embodiment, the computing device (100) can analyze the received electrocardiogram data and perform standardization and normalization to eliminate scale differences between various equipment and measurement environments.

[0353] For example, the signal mean of each read can be converted to 0 and the standard deviation to 1, and then scaled to a range of 0 to 1.

[0354] Alternatively, for example, a low-pass filter can be applied to remove electromyography noise as needed, or a high-pass filter can be used to remove baseline fluctuations. In addition, mutually

[0355] Data collected at different sampling frequencies (e.g., 250 Hz, 500 Hz, etc.) can be resampled to a unified frequency and converted into a time-frequency domain spectrogram using the Short-Time Fourier Transform (STFT).

[0356] The data preprocessed in this way can be normalized into a model input format (e.g., 2D tensor, spectrogram image, etc.).

[0357] Next, the computing device (100) according to an embodiment of the present invention can predict the degree of coronary artery stenosis by using preprocessed electrocardiogram data as input (S330).

[0358] In one embodiment, the computing device (100) can estimate the degree of stenosis as a percentage (%) using preprocessed electrocardiogram data. For example, if ST segment depression and T-wave inversion are observed in the electrocardiogram of patient A, the prediction module (130) can predict that patient A's coronary artery has a stenosis rate of 55%.

[0359] At this time, the computing device (100) can perform an FFR prediction step (S340) if the stenosis rate falls within a predetermined range (e.g., 40 to 70%). On the other hand, if the stenosis rate is less than 40% or greater than 70%, it can be directly branched to a risk determination step (S350) according to the range.

[0360] For example, referring to Figure 9, if the stenosis rate is less than 40%, it is classified as low risk and drug treatment or follow-up may be recommended. Alternatively, for example, if the stenosis rate exceeds 70%, it is classified as high risk and the FFR prediction step can be omitted, and a percutaneous coronary intervention (PCI) or additional imaging examination may be recommended immediately.

[0361] A computing device (100) according to an embodiment of the present invention can reduce unnecessary computations through the branching structure described above and provide a decision path that substantially corresponds to clinical judgment.

[0362] As described above, the computing device (100) according to an embodiment of the present invention can predict a fractional blood flow reserve (FFR) value by using the degree of stenosis as input when the predicted degree of stenosis falls within a predetermined range (S340).

[0363] Referring to FIG. 9, patients with low risk of stenosis (less than 40%) are classified as being under observation, and electrocardiogram data corresponding to the medium or higher risk group can be input into a prediction model loaded into a computing device (100).

[0364] As described above, the prediction model according to the embodiment of the present invention is based on a learning model that is pre-trained using a self-supervised learning method with a large amount of electrocardiogram data, and is a deep learning model generated by transfer learning with FFR labeling data.

[0365] For example, the computing device (100) can call a prediction model to calculate the patient's FFR value.

[0366] If patient A's FFR value is predicted to be 0.83, drug treatment or observation may be recommended according to the condition “eFFR ≥ 0.8” as shown in Figure 5.

[0367] Alternatively, if the FFR is 0.72, it corresponds to the condition “FFR < 0.8”, and you may be advised to undergo coronary angiography (CAG).

[0368] In addition, the computing device (100) of the present invention can determine the risk of myocardial ischemia by comparing the predicted FFR value with a reference threshold value (S350).

[0369] In one embodiment, the computing device (100) can classify the patient's risk group by comparing the FFR value for each blood vessel calculated in the prediction step with a reference threshold.

[0370] For example, if the FFR value is less than 0.75, it may be classified as a high-risk group and determined that a percutaneous coronary intervention (PCI) is necessary; if it is 0.8 or higher, it may be classified as a low-risk group and determined that drug treatment is appropriate; and if it is 0.75 or higher but less than 0.8, it may be classified as an intermediate-risk group and determined that additional testing is necessary.

[0371] In addition, a computing device (100) according to one embodiment of the present invention can further determine the location and number of lesions of coronary arteries suspected of stenosis by analyzing the predicted FFR value and the stenosis rate prediction result together.

[0372] Referring to Fig. 9, by analyzing the correlation between 12-lead electrocardiogram signals and the attention weight distribution, if strong ischemic signals are repeatedly observed in a specific lead group, the possibility of stenosis of the corresponding blood vessel (e.g., left anterior descending artery, right coronary artery, etc.) can be estimated. In addition, if independent ischemic signals are detected in different lead groups, the possibility that the lesion is multi-vessel can be determined.

[0373] In another embodiment, the computing device (100) can evaluate the Multi-Vessel Ischemia Risk by comprehensively analyzing multiple FFR prediction results for each vessel using the stenosis location and number of lesions information as input.

[0374] In this process, the computing device (100) can calculate a comprehensive ischemia risk score for each patient by comprehensively considering the FFR value for each blood vessel, the number of blood vessels below the threshold, the correlation between ischemia leads, and the spatial distribution of ischemia signals.

[0375] For example, if only a single vessel is classified as ischemic, it can be determined as “local ischemia,” and if the FFR value is predicted to be less than 0.75 in two or more major vessels, it can be determined as “high-risk multi-vessel ischemia.” These evaluation results can be used to determine the treatment priority of a patient, as in the example of FIG. 9, or to select vessels for coronary angiography or percutaneous coronary intervention (PCI).

[0376] Finally, the computing device (100) according to an embodiment of the present invention can provide a prediction result including the predicted FFR value and myocardial ischemia risk (S360).

[0377] In one embodiment, the computing device (100) can integrate the predicted FFR value, stenosis rate, risk grade, and recommended treatment information and transmit them to a medical staff terminal.

[0378] For example, prediction results can be visualized using color codes (red-high risk, yellow-medium, green-low risk).

[0379] In addition, the computing device (100) can generate the above judgment result in the form of a patient-specific prediction report and provide it to a user terminal, and by storing it in a database, it can be utilized for future follow-up treatment and model retraining.

[0380] Accordingly, according to the present specification, a patient's fractional flow reserve (FFR) can be predicted non-invasively using electrocardiogram data, allowing for the quantitative assessment of the degree of myocardial ischemia without performing invasive FFR tests or coronary angiography. Consequently, the physical burden on the patient and the risk of complications can be reduced, examination costs and time can be saved, and accessibility to testing is improved, enabling safer and more efficient management of cardiovascular diseases.

[0381] In addition, the present invention can substantially assist clinical decision-making by providing comprehensive information on predicted FFR values ​​and the degree of stenosis, thereby enabling medical staff to select the optimal treatment method between PCI and coronary artery bypass graft (CABG).

[0382] In addition, clinical efficiency can be improved by reducing misdiagnosis / re-examination rates for patients near the threshold value through threshold-cognitive correction and uncertainty-based notifications.

[0383] Hereinafter, with reference to FIGS. 10 and FIGS. 11, an embodiment of a method for predicting coronary artery stenosis rate by blood vessel section according to the fourth embodiment of the present invention will be described in detail. The method of this embodiment can be performed in an environment identical or similar to the system configuration described above.

[0385] FIG. 10 is a flowchart of a method for predicting coronary artery stenosis rates by blood vessel segment as an indicator according to the fourth embodiment of the present disclosure, and FIG. 11 is an illustrative diagram showing examples of coronary arteries applicable to various embodiments of the present disclosure.

[0386] The method of the present embodiment can be performed, for example, by medical staff uploading a patient's electrocardiogram data in a hospital cloud server environment or by automatically retrieving it from an electronic medical record (EHR) system. Below, the procedure for each step of FIG. 10 will be explained in detail with reference to the coronary artery structure illustrated in FIG. 11.

[0387] Referring to FIG. 10, a method for predicting coronary artery stenosis rate by blood vessel segment using electrocardiogram data according to an embodiment of the present invention may include the steps of: collecting electrocardiogram (ECG) data of a patient (S410); standardizing and normalizing the electrocardiogram data (S420); predicting the coronary artery stenosis rate by blood vessel segment using the preprocessed electrocardiogram data as input (S430); determining whether the predicted stenosis rate satisfies a correction enhancement criterion (S440); reinforcing the correction if the predicted stenosis rate satisfies the correction enhancement criterion (S450); applying a basic correction if the predicted stenosis rate satisfies the correction enhancement criterion (S460); and calculating the reliability of the corrected stenosis rate and outputting the prediction result (S470).

[0388] First, a computing device (100) according to an embodiment of the present invention can collect electrocardiogram (ECG) data of a patient (S410). For example, medical staff can measure a 12-lead electrocardiogram (ECG) and transmit the measured data to the computing device (100) through the hospital's HER system.

[0389] Alternatively, if patient A's 12-lead electrocardiogram data file (XML, HL7, DICOM, JPG, JPEG, PNG and PDF formats, etc.) is stored on a hospital server or user terminal (200), medical staff can select and transmit the file through the web interface of the computing device (100).

[0390] As another example, if the electrocardiogram measuring device is directly connected to a server via a network, the data may be transmitted to the computing device (100) in the form of real-time streaming immediately after the measurement is completed.

[0391] Additionally, 12-lead ECG data can be transmitted to a prediction module (130) through the input interface of a computing device (100). At this time, the data reception process is secured through encrypted communication (HTTPS, TLS, etc.), and metadata such as patient ID, time of examination, and measuring equipment can be stored together.

[0392] At this time, the electrocardiogram data can be temporarily stored in an internal buffer of the computing device (100) and then transferred to a preprocessing step.

[0393] In addition, the computing device (100) according to an embodiment of the present invention can standardize and normalize electrocardiogram data (S320).

[0394] In one embodiment, the computing device (100) can analyze the received electrocardiogram data and perform standardization and normalization to eliminate scale differences between various equipment and measurement environments.

[0395] For example, the signal mean of each read can be converted to 0 and the standard deviation to 1, and then scaled to a range of 0 to 1.

[0396] Alternatively, for example, a low-pass filter can be applied to remove electromyography noise as needed, or a high-pass filter can be used to remove baseline fluctuations. In addition, mutually

[0397] Data collected at different sampling frequencies (e.g., 250 Hz, 500 Hz, etc.) can be resampled to a unified frequency and converted into a time-frequency domain spectrogram using the Short-Time Fourier Transform (STFT).

[0398] The data preprocessed in this way can be normalized into a model input format (e.g., 2D tensor, spectrogram image, etc.).

[0399] Next, the computing device (100) according to an embodiment of the present invention can predict the coronary artery stenosis rate for each blood vessel section by using preprocessed electrocardiogram data as input (S430).

[0400] In one embodiment, the computing device (100) can estimate the degree of stenosis as a percentage (%) by using the preprocessed electrocardiogram data alone or by using the preprocessed electrocardiogram data and patient data together. For example, if ST segment depression and T-wave inversion are observed in the electrocardiogram of patient A, the prediction module (130) can predict that patient A's coronary artery has a stenosis rate of 55%.

[0401] Here, the vascular segment refers to a segment subdivided according to medical structure based on the location of the coronary arteries surrounding the heart, and is explained in detail with reference to Fig. 11.

[0402] Referring to FIG. 11, the coronary artery in the present invention may include a right coronary artery (RCA) (810) and a left coronary artery (LCA) (820).

[0403] At this time, the left coronary artery (820) can be subdivided into the left circumflex artery (LCX) (821) and the left anterior descending coronary artery (LAD) (822), and the right coronary artery (810) can be subdivided into multiple vascular sections of the proximal, middle, and distal parts.

[0404] In addition, the computing device (100) can subdivide the blood vessel sections by dividing them into a certain ratio based on the total length of the coronary artery.

[0405] For example, by uniformly dividing the entire length of the RCA into proximal, intermediate, and distal sections based on a certain length, the coronary artery can be subdivided into multiple vascular segments according to the ratio.

[0406] Additionally, the computing device (100) can subdivide the blood vessel sections based on the point where a change in the diameter of the blood vessel appears along the coronary artery.

[0407] In other words, reflecting the anatomical characteristic that there are sections with relatively thicker diameters and sections that become thinner within the same coronary artery, multiple blood vessel sections can be subdivided based on the point of diameter change.

[0408] For example, when a computing device (100) divides the coronary artery into multiple blood vessel sections according to its anatomical structure, the coronary artery stenosis rate for each blood vessel section can be individually predicted based on changes in electrocardiogram signals corresponding to each divided blood vessel section.

[0409] Additionally, the computing device (100) can subdivide the blood vessel sections based on points where the curvature of the blood vessel changes abruptly along the coronary artery.

[0410] For example, by establishing the boundary between the two vascular segments based on a representative sharp bend point between the proximal and intermediate parts of the left anterior descending (coronary) artery, the coronary artery stenosis rate can be predicted individually by dividing the vascular segments based on areas with significant anatomical changes.

[0411] In addition, the computing device (100) can apply a pre-set continuity condition when predicting the stenosis rate of each subdivided blood vessel section.

[0412] Specifically, the computing device (100) can individually predict the coronary artery stenosis rate for each blood vessel section by applying a pre-set continuity condition so that the predicted stenosis rate between adjacent blood vessel sections does not appear abnormally discontinuous, reflecting the medical characteristic that blood flow changes are not abruptly interrupted between continuous blood vessel sections forming a single coronary artery.

[0413] For example, the computing device (100) can be configured so that the stenosis rate is not predicted to be excessively high or low in a specific blood vessel section alone by considering adjacent blood vessel sections, such as the proximal, middle, and distal sections of the right coronary artery, as connected structures and applying conditions that reflect blood flow continuity between adjacent sections.

[0414] Through this, the computing device (100) suppresses physiologically impossible variations in stenosis rates between continuous blood vessel sections and can predict the degree of stenosis of each section of the entire coronary artery more stably and consistently.

[0415] In addition, the computing device (100) can predict the stenosis rate by taking into account anatomical asymmetry in which the left coronary artery and the right coronary artery have different lengths, diameters, and branching shapes, and the clinical impact differs even with the same stenosis rate.

[0416] Specifically, because the left coronary artery has a structural characteristic of rapidly branching into vascular segments after passing through a short main trunk, the stenosis rate may be relatively high when stenosis occurs; therefore, when predicting the stenosis rate, the allowable range of variation for the segment can be applied more finely to predict the stenosis rate.

[0417] For example, the computing device (100) can predict the stenosis rate by determining that the variation in the predicted stenosis rate in the section prior to the branch of the left coronary artery is inconsistent with the surrounding section, and that this is based on a lower tolerance standard than that of the right coronary artery.

[0418] The computing device (100) can individually predict the stenosis rate for each sub-vascular section of the RCA, LCX, and LAD in this way, and quantitatively determine the distribution of stenosis throughout the coronary artery.

[0419] In addition, the computing device (100) can calculate the predicted stenosis rate for each blood vessel section and quantitatively determine the distribution of stenosis throughout the coronary artery.

[0420] Afterwards, the computing device (100) can determine whether the predicted stenosis rate per blood vessel section satisfies the correction reinforcement criteria (S440).

[0421] First, if the predicted stenosis rate is near the correction reinforcement criteria (S440, YES), the computing device (100) can reinforce the correction for the predicted stenosis rate (S450).

[0422] At this time, the vicinity of the correction reinforcement criteria refers to the range included within the allowable error range centered on reference values ​​such as 50% and 70%, where it is clinically difficult to determine moderate stenosis.

[0423] For example, the 50% standard can be set to 40–60%, and the 70% standard can be set to 60–80%, and the predetermined tolerance range can be variably adjusted according to the medical institution's treatment protocol, the distribution of training data, or user settings.

[0424] Through this, it is possible to reduce misclassification and prediction instability that are prone to occur near stenosis rates where clinical judgment is difficult, and to improve the final calculated corrected stenosis rate and reliability.

[0425] Meanwhile, if the predicted stenosis rate is not near the correction reinforcement criteria (S440, NO), the computing device (100) determines that the predicted value is clinically clear or falls within a range where judgment can be made clear, and can perform only the basic correction procedure (S460).

[0426] In this case, the basic correction procedure applies only minimal normalization corrections, such as smoothing the model output, removing outliers, and simple scaling, and does not perform high-intensity corrections required near the reference value.

[0427] Through this, the computing device (100) can improve overall processing efficiency by minimizing computational volume and correction procedures at narrowing rates other than the standard without distorting the predicted value more than necessary.

[0428] A computing device (100) according to an embodiment of the present invention can reduce unnecessary computations through the branching structure described above and provide a decision path that substantially corresponds to clinical judgment.

[0429] As described above, the prediction model according to the embodiment of the present invention is based on a learning model that is pre-trained using a self-supervised learning method with a large amount of electrocardiogram data, synthetic electrocardiogram data generated by simulation, and blood flow simulation data, and is subsequently a deep learning model generated by transfer learning with a large amount of unlabeled electrocardiogram data and labeled data from coronary artery stenosis rate test results.

[0430] For example, the computing device (100) can call a prediction model to calculate the patient's coronary artery stenosis rate.

[0431] Finally, the computing device (100) according to an embodiment of the present invention can provide a prediction result including the predicted coronary artery stenosis rate value and reliability for each blood vessel section (S470).

[0432] In one embodiment, the computing device (100) can integrate the predicted coronary artery stenosis rate, the corrected stenosis rate, the reliability of the stenosis rate, the risk grade, and the recommended treatment information and transmit them to a medical staff terminal.

[0433] For example, prediction results can be visualized using color codes (red-high risk, yellow-medium, green-low risk).

[0434] In addition, the computing device (100) can generate the above judgment result in the form of a patient-specific prediction report and provide it to a user terminal, and by storing it in a database, it can be utilized for future follow-up treatment and model retraining.

[0435] Accordingly, according to the present specification, the coronary artery stenosis rate of a patient can be predicted non-invasively using electrocardiogram data, so the degree of coronary artery stenosis can be quantitatively evaluated without performing an invasive coronary artery stenosis test or coronary angiography.

[0436] As a result, the physical burden on patients and the risk of complications can be reduced, examination costs and time can be saved, and improved accessibility to testing enables safer and more efficient management of cardiovascular diseases.

[0437] In addition, the present invention calculates an independent stenosis rate (%) for each subdivided blood vessel section for each major blood vessel including the right coronary artery (RCA), left circumflex artery (LCX), and left anterior descending artery (LAD), thereby accurately locating the blood vessel where stenosis has occurred and enabling accurate treatment by determining the degree of stenosis differently for each blood vessel section.

[0438] In addition, the present invention can substantially assist in clinical decision-making by providing comprehensive information on the predicted coronary artery stenosis rate and degree of stenosis for each blood vessel segment, as well as the reliability of the results, so that medical staff can select the optimal treatment method between PCI and coronary artery bypass graft (CABG).

[0439] In addition, clinical efficiency can be improved by reducing the misjudgment / re-examination rate of stenosis rates in vascular sections near the correction enhancement criteria through correction-enhanced criteria-cognitive correction, by strongly correcting the stenosis rate for patients near the severe stenosis threshold value and providing decision-making assistance notifications such as recommendations for additional tests including the reliability calculated together.

[0441] The above-described method has been explained with reference to the flowchart illustrated in the drawings. For the sake of simplicity, the method has been illustrated and described using a series of blocks; however, the present disclosure is not limited to the order of said blocks, and some blocks may be performed in a different order or simultaneously than those illustrated and described in this specification. Additionally, new blocks not described in this specification and drawings may be added, or some blocks may be deleted or modified.

[0442] According to various embodiments of the present disclosure, the heart disease-related indicators derived according to the method described above may be provided through a user terminal (200) or a medical information system and may be used as reference information for a user to evaluate cardiovascular conditions or determine the risk associated with coronary artery disease.

[0443] In addition, according to various embodiments of the present disclosure, the heart disease-related indicators derived according to the above-described method can be linked with a Hospital Information System, an image analysis system, or a patient monitoring system and utilized as information to support the analysis or evaluation of a patient's cardiovascular condition.

[0444] In addition, according to various embodiments of the present disclosure, indicators related to heart disease derived according to the method described above can be utilized as information for tracking changes in a patient's cardiovascular condition or for recognizing a risk condition early in a patient monitoring system. For example, a computing device (100) is connected to an external monitoring device and provides indicators related to heart disease to the external monitoring device, thereby monitoring the patient's heart condition through the external monitoring device and outputting a notification as an event is detected. Hereinafter, the hardware configuration of the computing device (100) will be described with reference to FIG. 12.

[0446] FIG. 12 is a drawing illustrating the hardware configuration of a computing device according to various embodiments of the present disclosure.

[0447] Referring to FIG. 12, a computing device (100) according to another embodiment of the present disclosure may include one or more processors (101), a memory (102) for loading a computer program (105A) executed by the processor (101), a bus (103), a communication interface (104), and a storage (105) for storing the computer program (105A). Here, FIG. 12 illustrates only the components related to the embodiments of the present disclosure. Accordingly, a person skilled in the art to which the present disclosure pertains will understand that other general-purpose components may be included in addition to the components illustrated in FIG. 12.

[0448] The processor (101) controls the overall operation of each component of the computing device (100). The processor (101) may be configured to include a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphic Processing Unit), or any form of processor well known in the art of the present disclosure.

[0449] Additionally, the processor (101) may perform operations for at least one application or program for executing the method according to the embodiments of the present disclosure, and the computing device (100) may have one or more processors.

[0450] In various embodiments, the processor (101) may further include Random Access Memory (RAM) (not shown) and Read-Only Memory (ROM) (not shown) for temporarily and / or permanently storing signals (or data) processed within the processor (101). Additionally, the processor (101) may be implemented in the form of a System on Chip (SoC) comprising at least one of a graphics processing unit, RAM, and ROM.

[0451] Memory (102) stores various data, instructions and / or information. Memory (102) may load a computer program (105A) from storage (105) to execute a method / operation according to various embodiments of the present disclosure. When the computer program (105A) is loaded into memory (102), the processor (101) may perform the method / operation by executing one or more instructions constituting the computer program (105A). Memory (102) may be implemented as a volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

[0452] The bus (103) provides communication functions between components of the computing device (100). The bus (103) can be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

[0453] The communication interface (104) supports wired and wireless internet communication of the computing device (100). Additionally, the communication interface (104) may support various communication methods other than internet communication. To this end, the communication interface (104) may be configured to include a communication module well known in the art of the present disclosure. In some embodiments, the communication interface (104) may be omitted.

[0454] Storage (105) can store a computer program (105A) non-temporarily. When performing a process of deriving heart disease indicators through multi-channel bio-information analysis via a computing device (100), storage (105) can store various information necessary to provide the process of deriving heart disease indicators through multi-channel bio-information analysis.

[0455] Storage (105) may be configured to include non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which this disclosure belongs.

[0456] A computer program (105A) may include one or more instructions that cause a processor (101) to perform a method / operation according to various embodiments of the present disclosure when loaded into memory (102). That is, the processor (101) may perform the method / operation according to various embodiments of the present disclosure by executing the one or more instructions.

[0457] In one embodiment, the computer program (105A) may include one or more instructions for performing a method for deriving heart disease indicators through multi-channel bio-information analysis, the step of acquiring bio-information corresponding to a patient's heart condition and the step of deriving indicators related to heart disease using the acquired bio-information.

[0459] The method described above may be provided as a computer program stored on a computer-readable recording medium for execution on a computer. The medium may continuously store a program executable by a computer, or temporarily store it for execution or download. Additionally, the medium may be various recording or storage means in the form of a single or multiple hardware components combined, and may not be limited to a medium directly connected to a computer system but may 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. Furthermore, 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.

[0460] The methods, operations, or techniques of the present disclosure may be implemented by various means. For example, these techniques may be implemented in hardware, firmware, software, or a combination thereof. Those skilled in the art will understand that the various exemplary logical blocks, modules, circuits, and algorithmic steps described in connection with the disclosure herein may be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate such interchangeability between hardware and software, various exemplary components, blocks, modules, circuits, and steps have been generally described above in terms of their functional aspects. Whether such functions are implemented in hardware or in software depends on the design requirements imposed on the specific application and the overall system. Those skilled in the art may implement the functions described in various ways for each specific application, but such implementations should not be construed as departing from the scope of the present disclosure.

[0461] In a hardware implementation, the processing units used to perform the techniques may be implemented in one or more ASICs, DSPs, digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, electronic devices, other electronic units designed to perform the functions described in this disclosure, computers, or a combination thereof.

[0462] Accordingly, the various exemplary logic blocks, modules, and circuits described in connection with the present disclosure may be implemented or performed by any combination of general-purpose processors, DSPs, ASICs, FPGAs or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, or those designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, for example, a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors coupled with a DSP core, or any other combination of configurations.

[0463] In firmware and / or software implementations, techniques may be implemented as instructions stored on a computer-readable medium such as random access memory (RAM), read-only memory (ROM), non-volatile random access memory (NVRAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable PROM (EEPROM), flash memory, compact disc (CD), magnetic or optical data storage devices, etc. The instructions may be executable by one or more processors, and the processor(s) may be enabled to perform specific aspects of the functions described in this disclosure.

[0464] When implemented in software, the techniques described above may be stored on a computer-readable medium as one or more instructions or code, or transmitted through a computer-readable medium. Computer-readable media include both computer storage media and communication media, including any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available media accessible by a computer. As a non-limiting example, such computer-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium accessible by a computer that can be used to transfer or store desired program code in the form of instructions or data structures. Additionally, any connection is appropriately referred to as a computer-readable medium.

[0465] For example, if software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair cable, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, coaxial cable, fiber optic cable, twisted pair cable, digital subscriber line, or wireless technologies such as infrared, radio, and microwave are included within the definition of a medium. As used herein, disk and disc include CD, laser disc, optical disc, DVD (digital versatile disc), floppy disk, and Blu-ray disc, wherein disks usually play data magnetically, whereas discs play data optically using a laser. The above combinations should also be included within the scope of computer-readable media.

[0466] The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other known form of storage medium. An exemplary storage medium may be connected to a processor so that the processor can read information from the storage medium or write information to the storage medium. Alternatively, the storage medium may be integrated into the processor. The processor and the storage medium may exist within an ASIC. The ASIC may exist within a user terminal. Alternatively, the processor and the storage medium may exist as separate components within the user terminal.

[0467] Although the embodiments described above have been described as utilizing aspects of the subject matter disclosed herein in one or more standalone computer systems, the present disclosure is not limited thereto and may be implemented in conjunction with any computing environment, such as a network or a distributed computing environment. Furthermore, aspects of the subject matter in the present disclosure may be implemented in a plurality of processing chips or devices, and storage may be similarly affected across a plurality of devices. Such devices may include PCs, network servers, and portable devices.

[0468] Although the present disclosure has been described in relation to some embodiments, various modifications and changes may be made without departing from the scope of the present disclosure as understood by a person skilled in the art to which the invention of the present disclosure pertains. Furthermore, such modifications and changes should be considered to fall within the scope of the claims appended to this specification. Explanation of the symbols

[0469] 100: Computing device 200 : User terminal 300 : External server 400 : Network

Claims

Claim 1 A method for deriving heart disease indicators through multi-channel bio-information analysis, comprising: a step of acquiring bio-information corresponding to a patient's heart condition, wherein the method is performed through a computing device; and a step of deriving an indicator related to heart disease using the acquired bio-information. Claim 2 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein, in claim 1, the acquired bio-information is collected through multiple different channels and is multi-channel bio-information including at least one of electrical characteristics, mechanical characteristics, acoustic characteristics, and clinical characteristics related to heart condition. Claim 3 A method for deriving a heart disease indicator through multichannel bio-information analysis, wherein the step of deriving the indicator comprises: a step of extracting a first feature vector corresponding to an electrical feature related to a heart condition from the acquired bio-information; a step of generating a first latent space representation by mapping the extracted first feature vector to a predefined electrical-anatomical latent space; a step of generating a coronary artery structure graph by structuring the generated first latent space representation based on the anatomical structure of the coronary artery; and a step of deriving a first indicator related to heart disease using the generated coronary artery structure graph. Claim 4 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the acquired bio-information comprises a plurality of electrophysiological signals having different temporal and spatial characteristics, and the step of extracting the first feature vector comprises the step of extracting the first feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals using a plurality of encoders. Claim 5 A method for deriving heart disease indicators through multi-channel bioinformation analysis, wherein the step of extracting a first feature vector from each of the plurality of electrophysiological signals comprises preprocessing the plurality of electrophysiological signals based on the characteristics of the plurality of electrophysiological signals and extracting a first feature vector from each of the preprocessed plurality of electrophysiological signals through the plurality of encoders. Claim 6 In claim 3, the step of generating the first latent space representation comprises generating the first latent space representation by mapping the extracted first feature vector onto a coordinate system corresponding to the electrical-anatomical latent space, wherein the electrical-anatomical latent space is a latent coordinate space predefined based on the correspondence between the anatomical structure of the myocardium and coronary artery and the electrical conduction characteristics of the myocardium, a method for deriving heart disease indicators through multichannel bioinformation analysis. Claim 7 In claim 3, the step of generating the coronary artery structure graph comprises generating the coronary artery structure graph by assigning information contained in the generated first latent space representation as a state vector to each of the plurality of nodes included in the tree structure based on a tree structure predefined in relation to the anatomical structure of the coronary artery, wherein the plurality of nodes include a plurality of blood vessel nodes corresponding to the major blood vessels of the coronary artery and a plurality of segment nodes corresponding to the segmental structure of the coronary artery, a method for deriving heart disease indicators through multi-channel bioinformation analysis. Claim 8 In claim 7, the step of generating a coronary artery structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure comprises the step of assigning information included in the generated first latent space representation to each of the plurality of nodes using a first weighted mapping matrix, wherein the first weighted mapping matrix is ​​a matrix representing the correspondence relationship between a first feature vector extracted from bio-information and the plurality of nodes, and the plurality of elements included in the first weighted mapping matrix are weights representing the degree to which the first feature vector extracted from bio-information contributes to each of the plurality of nodes, a method for deriving heart disease indicators through multi-channel bio-information analysis. Claim 9 In claim 7, the step of generating a coronary artery structure graph by assigning a state vector to each of the plurality of nodes included in the tree structure comprises: a step of extracting a global ischemia pattern for the heart based on the extracted first latent space representation and generating a global surveillance node mapped to the extracted global ischemia pattern; a step of defining a reference state based on bio-information obtained from a healthy population without heart disease and generating a normal reference node mapped to the defined reference state; and a step of connecting the generated global surveillance node and the generated normal reference node to at least one of the plurality of nodes, a method for deriving heart disease indicators through multi-channel bio-information analysis. Claim 10 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the step of generating a coronary artery structure graph by assigning a state vector to each of the plurality of nodes included in the tree structure comprises the step of updating the state vector mapped to each of the plurality of nodes by performing a message passing operation between the plurality of nodes using a graph neural network decoder. Claim 11 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the step of deriving the first indicator includes the step of deriving a plurality of first indicators by analyzing each of the generated coronary artery structure graphs using a plurality of prediction heads, and the plurality of derived first indicators include at least some of coronary artery fractional flow reserve, ischemia status, coronary artery stenosis rate, plaque fragility, hemodynamic indicators, and cardiovascular event risk. Claim 12 In claim 3, the step of deriving the first indicator includes the step of deriving the first indicator by analyzing the acquired bio-information through a first indicator derivation module, wherein the indicator derivation module includes: a first encoder that extracts a first feature vector from the bio-information; a first latent space alignment module that generates a first latent space representation; a first graph generation module that generates a coronary artery structure graph; a first graph neural network decoder that updates the coronary artery structure graph; and a first prediction head that derives the first indicator using the coronary artery structure graph, wherein pre-learning is performed based on two or more pairs of indicators of different types, and after the pre-learning is performed, self-supervised learning is performed based on a plurality of multi-channel electrophysiological signals for which no correct answer data exists, and after the self-supervised learning is performed, supervised learning is performed based on a plurality of multi-channel electrophysiological signals for which correct answer data exists. Claim 13 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the step of deriving the indicator comprises: a step of extracting a second feature vector corresponding to structural and functional features related to heart condition from the acquired bio-information; a step of generating a second latent space representation by mapping the extracted second feature vector to a predefined structural-functional latent space; a step of generating a heart structure graph by structuring the generated second latent space representation based on the anatomical structure of the heart; and a step of deriving a second indicator related to heart disease by analyzing the generated heart structure graph. Claim 14 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the acquired bio-information comprises a plurality of electrophysiological signals having different temporal and spatial characteristics, and the step of extracting the second feature vector comprises the step of extracting the second feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals using a plurality of encoders. Claim 15 In claim 13, the step of extracting a second feature vector from each of the plurality of electrophysiological signals by analyzing each of the plurality of electrophysiological signals comprises the step of preprocessing the plurality of electrophysiological signals based on the input characteristics of each of the plurality of encoders, and extracting a second feature vector from each of the preprocessed plurality of electrophysiological signals through the plurality of encoders, a method for deriving heart disease indicators through multi-channel bio-information analysis. Claim 16 In claim 13, the step of generating the second latent space representation comprises generating the second latent space representation by mapping the extracted second feature vector onto a coordinate system corresponding to the structural-functional latent space, wherein the structural-functional latent space is a latent coordinate space predefined based on the correspondence between the anatomical structure of the heart, heart function, and structural disease state, a method for deriving heart disease indicators through multichannel bioinformation analysis. Claim 17 In claim 13, the step of generating the heart structure graph comprises generating the heart structure graph by assigning information included in the generated second latent space representation as a state vector to each of the plurality of nodes included in the tree structure based on a tree structure predefined in relation to the anatomical structure of the heart, wherein the plurality of nodes include a plurality of structure nodes corresponding to structural components of the heart and at least one disease node corresponding to structural and functional diseases of the heart, a method for deriving heart disease indicators through multi-channel bio-information analysis. Claim 18 In claim 17, the step of generating a heart structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure comprises the step of assigning information included in the generated second latent space representation to each of the plurality of nodes using a second weighted mapping matrix, wherein the second weighted mapping matrix is ​​a matrix representing the correspondence relationship between a second feature vector extracted from bio-information and the plurality of nodes, and the plurality of elements included in the second weighted mapping matrix are weights representing the degree to which the second feature vector extracted from bio-information contributes to each of the plurality of nodes, a method for deriving heart disease indicators through multi-channel bio-information analysis. Claim 19 In claim 17, the step of generating a cardiac structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure comprises: extracting global structure-disease context features that commonly affect cardiac structure and disease state from clinical / demographic information including at least one of the subject's age, gender, comorbidity, drug history, and examination history, and text information including cardiac image interpretation findings, diagnostic records, or examination reports, and generating a global surveillance node mapped with the extracted global structure-disease context features; defining a reference structural state of the heart based on bio-information obtained from a healthy population and generating a normal reference node mapped with the defined reference structural state; and connecting the generated global surveillance node and the generated normal reference node to at least one of the plurality of nodes included in the generated cardiac structure graph, a method for deriving cardiac disease indicators through multi-channel bio-information analysis. Claim 20 A method for deriving heart disease indicators through multi-channel bio-information analysis, wherein the step of generating a heart structure graph by assigning a state vector to each of a plurality of nodes included in the tree structure comprises the step of updating the state vector mapped to each of a plurality of nodes included in the generated heart structure graph by performing a message passing operation between a plurality of nodes included in the generated heart structure graph using a graph neural network decoder. Claim 21 In claim 13, the step of deriving the second indicators comprises the step of deriving a plurality of second indicators by analyzing each of the generated heart structure graphs using a plurality of prediction heads, wherein the derived plurality of second indicators include at least some of structural indicators related to the volume, mass, or wall thickness of the left and right ventricles, functional indicators related to left ventricular ejection fraction, ventricular contraction or relaxation function, disease indicators indicating the presence or severity of structural heart disease, and prognostic risk indicators related to the worsening of heart failure, hospitalization, or occurrence of major cardiovascular events within a certain period. Claim 22 In claim 13, the step of deriving the second indicator comprises a step of deriving the second indicator by analyzing the acquired bio-information through a second indicator derivation module, wherein the second indicator derivation module comprises: a second encoder that extracts a second feature vector from the bio-information; a second latent space alignment module that generates a second latent space representation; a second graph generation module that generates a heart structure graph; a second graph neural network decoder that updates the heart structure graph; and a second prediction head that derives an indicator related to heart disease using the heart structure graph, wherein supervised learning is performed using pairs of heart image information and electrophysiological signals, and after the supervised learning is performed, auxiliary learning is performed using clinical / demographic information or text-based clinical information, and after the auxiliary learning is performed, self-supervised learning and contrastive learning are performed using large-scale bio-information where no ground truth data exists. Claim 23 A computing device comprising: a processor; a network interface; a memory; and a computer program loaded into the memory and executed by the processor, wherein the processor performs the method of claim 1 by executing one or more instructions included in the computer program. Claim 24 A computer program that is combined with a computing device and stored on a non-transient recording medium readable by the computing device to execute the method of claim 1.

Images