Electrocardiogram-based tensor generation method and apparatus

The electrocardiogram-based tensor generation method enhances ECG interpretation accuracy and explainability by grouping signals into beat units and supporting rule-based algorithms with visual explanations, addressing the complexity of arrhythmic ECGs and lack of explainable AI systems.

WO2026240267A1PCT designated stage Publication Date: 2026-11-19MIRABELL SOFT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2025/017182
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-08-07
Filing Date
2025-10-27
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Clinical decisions based on electrocardiograms require additional verification due to the complexity of processing arrhythmic ECGs, and existing AI systems lack explainability, necessitating a method for efficient data processing and storage to support rapid and accurate clinical decision-making.

Method used

A method for generating an electrocardiogram-based tensor that groups signals into beat units, aligns with academic rules, and supports explainable AI by visually displaying analysis steps, ensuring scalability for new indicators.

Benefits of technology

Improves the accuracy, explainability, and reliability of ECG interpretation by maintaining structural consistency and enabling clear visualization of judgment processes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2025017182_19112026_PF_FP_ABST
    Figure KR2025017182_19112026_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to an electrocardiogram-based data generation method. The electrocardiogram-based tensor generation method performed by a computer device, according to the present invention, may comprise the steps of: constructing a tensor on the basis of section information obtained by classifying an electrocardiogram into sections according to a plurality of items; modifying or expanding vectors in the constructed tensor on the basis of the section information, such that the vectors have the items arranged in a predefined expected progression order; and adding, on the basis of the expected progression order, a beat value to each vector in the modified or expanded tensor. According to the present invention, different types of variables can be efficiently stored in a single tensor structure, thereby enabling integrated management of various indicators of electrocardiogram data.
Need to check novelty before this filing date? Find Prior Art

Description

ECG-based tensor generation method and device

[0001] The present invention relates to a method for generating electrocardiogram-based data.

[0002] Although machine learning algorithms are widely utilized in the field of clinical diagnosis using electrocardiograms, the reality in actual clinical practice is that clinical decisions regarding disease diagnosis and the overall management of the patient's condition are made only after undergoing an additional verification process based on academically established medical rules.

[0003] A rule-based algorithm is a technology for automating electrocardiogram interpretation based on the aforementioned academically established knowledge. It recognizes each wave (e.g., P wave, QRS complex, T wave, etc.) and their subwaves appearing in the electrocardiogram data, and determines whether standardized conditions for each disease are satisfied through numerical values ​​generated by combining indicators derived from each segment or different segments.

[0004] In this process, not only the indicators measured from each wave (continuous variables) but also the results of classifying the morphology of each wave (categorical variables) are utilized. Additionally, waves existing independently on the electrocardiogram are grouped into specific intervals (beats) to align with the electrophysiological characteristics of the heart inherently reflected by the electrocardiogram, and the number and types of beats are also used in the analysis. Some continuous variables undergo complex conditional calculation processes, such as being extracted only from waves within beats that satisfy specific conditions or calculated only among beats that satisfy the same conditions.

[0005] In electrocardiograms during normal or sinoatrial conductance, each wave is recognized, making it relatively easy to process them as a single beat; however, in the electrocardiograms of patients with dynamic arrhythmias, the order of appearance of waves changes or two or more waves frequently overlap. Even in such cases, processing is required to group each wave into a single beat. Additionally, since continuous variables must be calculated based on the premise that waves are grouped into beat units even when some waves are missing (e.g., a P wave is present but the QRS complex does not appear), a preprocessing step is required to treat the recognized P wave and a certain range where the QRS complex and T wave would normally be located as a single beat.

[0006] To enable clinicians to make rapid and accurate decisions, the time required to review the results of electrocardiogram (ECG) analysis using rule-based algorithms needs to be reduced. To achieve this, an additional process is required to explain the rationale behind the algorithm's judgments to clinicians by visually displaying the results of each analysis step directly on the raw ECG data. In other words, rule-based algorithms need to be operated as part of explainable artificial intelligence (Explainable AI). However, there are currently no commercially available explainable AI systems in the fields of ECG interpretation and arrhythmia differentiation.

[0007] The present invention was devised in response to the aforementioned background technology and aims to provide a method for processing data from an electrocardiogram signal composed of one or more leads and storing them for the efficient operation of explainable artificial intelligence and the successful operation of rule-based algorithms.

[0008] In addition, the present invention enables the storage of information by efficiently grouping circular electrocardiogram signals into beat units.

[0009] In addition, the present invention has a structure that allows the processed and stored information to correspond one-to-one with the sections and points of the original electrocardiogram data that served as the basis for each calculation.

[0010] Through this, the present invention enables an artificial intelligence system equipped with a rule-based algorithm to present a judgment conclusion, and then, in accordance with the requirements of a user (clinician), to recognize which rule was applied to which part of the electrocardiogram data as the basis for the conclusion.

[0011] Furthermore, the present invention guarantees structural scalability so that even if result values ​​related to new indicators or waves are added to existing data, a rule-based algorithm can make a judgment considering the actual order of appearance.

[0012] A method for generating an electrocardiogram-based tensor performed on a computer device according to various embodiments for solving the problem of the present invention may include: a step of constructing a tensor based on segment information in which an electrocardiogram is classified into segments according to a plurality of items; a step of modifying or expanding a vector in the constructed tensor to have a predefined expected progression order for each item based on the segment information; and a step of adding a beat value to each vector in the modified or expanded tensor based on the expected progression order.

[0013] The above interval information includes start and end time values ​​of the above item, and the step of constructing the above tensor may generate item values ​​indexed in ascending order by item based on the start time value of the above item, and construct the above tensor with the generated item values ​​and the above interval information.

[0014] The above modification or expansion step preferably involves adjusting the array of the vector.

[0015] The above modification or expansion step preferably involves adjusting the leading or trailing vector.

[0016] In the above modification or expansion step, it is preferable to add a vector to satisfy the above expected sequence of progress.

[0017] The step of adding an indicator containing induction-related information to a vector within the above tensor may be further included.

[0018] A computer device according to various embodiments for solving the problems of the present invention includes a processor and a memory that communicates with the processor, wherein the memory stores instructions for the processor to perform operations, and the operations may include an operation to construct a tensor based on segment information in which an electrocardiogram is classified into segments according to a plurality of items, an operation to modify or expand a vector in the constructed tensor to have a predefined expected progression order for each item based on the segment information, and an operation to add a beat value to each vector in the modified or expanded tensor based on the expected progression order.

[0019] According to the present invention, since different types of variables can be efficiently stored within a single tensor structure, it becomes possible to manage various indicators of electrocardiogram data in an integrated manner.

[0020] The present invention groups and stores circular electrocardiogram signals by beat unit, and since grouping within the same beat is possible in a manner consistent with academic rules even when retrograde P waves are present or specific waves are absent, the accuracy of beat-unit analysis is improved.

[0021] The present invention is configured so that processed and stored information corresponds one-to-one with sections and points of the original electrocardiogram data, so that when an artificial intelligence system equipped with a rule-based algorithm presents a judgment conclusion, it can clearly indicate to a clinician the basis for each conclusion which rule was applied to which part of the original electrocardiogram data.

[0022] The present invention guarantees structural scalability so that even if result values ​​for new indicators or waves are added between existing data regions, a rule-based algorithm can make a judgment reflecting the actual order of appearance.

[0023] Therefore, the present invention can improve the accuracy, explainability, and reliability of electrocardiogram interpretation for clinicians.

[0024] FIG. 1 is a diagram showing a system for performing an electrocardiogram-based tensor generation method according to one embodiment of the present invention.

[0025] FIG. 2 is a flowchart illustrating a tensor generation method according to one embodiment of the present invention.

[0026] FIG. 3 is a figure illustrating the extraction of waveform-specific segment information through a segment separation model according to an embodiment of the present invention.

[0027] FIG. 4 is a figure illustrating an indexing result according to one embodiment of the present invention.

[0028] FIG. 5 is a figure illustrating a tensor generated according to one embodiment of the present invention.

[0029] FIG. 6 is a flowchart illustrating a specific rearrangement process according to one embodiment of the present invention.

[0030] FIG. 7 is an exemplary diagram illustrating the rearrangement process of vectors within a tensor according to one embodiment of the present invention.

[0031] FIG. 8 is a flowchart illustrating the process of deleting a vector according to one embodiment of the present invention.

[0032] FIG. 9 is an example diagram illustrating the process of searching for and deleting column vectors based on the leading vector according to one embodiment of the present invention.

[0033] FIG. 10 is a diagram showing the processing process of an edited tensor according to one embodiment of the present invention.

[0034] FIG. 11 is an exemplary diagram illustrating the process of adding a vector according to one embodiment of the present invention.

[0035] FIG. 12 illustrates the step of expanding a vector within a tensor generated according to one embodiment of the present invention.

[0036] FIG. 13 illustrates a tensor configured according to one embodiment of the present invention.

[0037] FIG. 14 is an exemplary diagram showing the correspondence between a tensor generated according to one embodiment of the present invention and circular electrocardiogram data.

[0038] FIG. 15 is an exemplary diagram showing the implementation of a system in the form of a computing device according to one embodiment of the present invention.

[0039] Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The following detailed description is provided to facilitate a comprehensive understanding of the methods, apparatuses, and / or systems described herein. However, this is merely illustrative and the present invention is not limited thereto.

[0040] In describing the embodiments of the present invention, detailed descriptions of known technologies related to the present invention are omitted if it is determined that such descriptions may unnecessarily obscure the essence of the invention. Furthermore, the terms described below are defined considering their functions in the present invention, and these definitions may vary depending on the intentions or practices of the user or operator. Therefore, their definitions should be based on the content throughout this specification.

[0041] The terms used in the detailed description are merely for describing embodiments of the invention and should not be limiting in any way. Unless explicitly stated otherwise, expressions in the singular form include the meaning of the plural form.

[0042] In this description, expressions such as “include” or “equipped” are intended to refer to certain characteristics, numbers, steps, actions, elements, parts or combinations thereof, and should not be interpreted to exclude the existence or possibility of one or more other characteristics, numbers, steps, actions, elements, parts or combinations thereof other than those described.

[0043] In addition, terms such as first, second, etc. may be used to describe the compartment elements of the embodiments of the present invention. These terms are used merely to distinguish the compartment element from other compartment elements, and the essence, order, or sequence of the compartment element is not limited by the terms.

[0044] Various embodiments of the present invention will be described in detail below with reference to the attached drawings.

[0045] FIG. 1 is a diagram showing a system for performing an electrocardiogram-based tensor generation method according to one embodiment of the present invention.

[0046] Referring to FIG. 1, the system according to the present embodiment includes an electrocardiogram measuring device, a tensor generating device (300), and an interpretation device to provide a series of functions for measuring electrocardiogram data and processing and interpreting it in a form optimized for the operation of a rule-based algorithm.

[0047] First, the electrocardiogram (ECG) measurement device detects the electrical activity of the heart through multiple electrodes attached to the subject's body surface and acquires ECG signals corresponding to multiple leads. The ECG measurement device may include an A / D converter that converts analog ECG signals into digital signals, and performs noise removal and filtering on the measured ECG signals to generate ECG data of a quality suitable for subsequent data processing and analysis.

[0048] As a representative example, the 12-lead electrocardiogram (12-lead EKG), which is most commonly used in clinical practice, uses a total of 12 standard leads (I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6), which is essentially the same as arranging the fluctuations in potential difference values ​​in a time series by projecting the current trace in three-dimensional space onto twelve-directional one-dimensional spaces. Therefore, the data output after electrocardiogram measurement is conveyed to the clinician by presenting the potential difference values ​​(y) on 12 xy planes as a function of time (x).

[0049] Academically, the entire or partial variation of a potential difference over time is referred to as a waveform, and a function existing within a specific time interval is called a wave. Furthermore, the shape of the graph of a function defined within a specific time interval is classified into various types according to certain criteria, and this is called morphology. If the morphology appears to take on a specific shape to the naked eye, it is sometimes classified as a separate pattern.

[0050] The measured electrocardiogram data is transmitted to a tensor generation device (300). The tensor generation device (300) receives the data received from the electrocardiogram measuring device as input and performs a data processing process to stably maintain the operation of a rule-based algorithm. Specifically, the tensor generation device (300) recognizes each waveform (P wave, QRS complex, T wave, etc.) and its subwaves on the electrocardiogram and generates a multidimensional tensor capable of efficiently storing continuous variable indicators calculated from each wave and shape-based categorical variable classification results.

[0051] Additionally, the tensor generation device (300) includes a function to group the original electrocardiogram signal into beat units, and groups them in a manner that does not violate academically established beat recognition rules even if abnormal findings, such as retrograde P waves or the absence of specific waves, are present. The tensor generated in this process includes an indexing structure so that each data element corresponds one-to-one with the intervals and points of the original electrocardiogram data, and ensures scalability so that structural consistency is maintained even when new indicators or waveform-related result values ​​are added to appropriate locations within the existing tensor.

[0052] Subsequently, the generated tensor is transmitted to an interpretation device. Based on the tensor generated by the tensor generation device (300), the interpretation device interprets the patient's electrocardiogram data using a rule-based algorithm and determines the patient's condition, diagnosis, and whether there is an abnormality. The interpretation device provides the determined diagnosis result to the clinician and includes a function to visually display the rule on which such judgment was made and the section of the electrocardiogram data to which the rule was applied. Through this, the clinician can easily understand the judgment process and basis of the rule-based algorithm, and the interpretation device, as an explainable AI system, can effectively support fast and accurate clinical decision-making.

[0053] FIG. 2 is a flowchart illustrating a tensor generation method according to one embodiment of the present invention.

[0054] Referring to FIG. 2, the tensor generating device (300) according to the present invention first constructs a tensor based on section information in which an electrocardiogram is classified into sections according to a plurality of items (S100).

[0055] The tensor generating device (300) generates a tensor containing multiple items (P, QRS, T, etc.) in segment units based on segment information (onset, offset, type) for each waveform recognized through a segment separation model.

[0056] FIG. 3 is a figure illustrating the extraction of waveform-specific segment information through a segment separation model according to an embodiment of the present invention.

[0057] Referring to FIG. 3, the result of recognizing the type of each wave (P wave, QRS complex, T wave) extracted from the original electrocardiogram data (10) and the start (onset) and end (offset) of each wave through the segment separation model (55) is used as a precondition.

[0058] Here, onset and offset are defined as specific time values, and for example, when the waveform measured for 10 seconds (10,000 ms) of Lead II in a standard 12-lead electrocardiogram is analyzed using a segment separation model, the start and end times for each waveform are derived as [wave type, onset, offset] ordered pairs.

[0059] Specifically, the tensor generating device (300) checks whether there are any ordered pairs containing one or more missing values, excluding the first and last ordered pairs among all ordered pairs.

[0060] If the first and last ordered pairs contain missing values, the tensor generating device (300) excludes the corresponding ordered pair and converts the information for each section (20) in the form of [wave type, onset, offset] for the remaining ordered pairs into a list for reference. For example, based on the analysis results derived from Lead II, the following list is generated.

[0061] list1 = [94.837, 273.973], type: QRS

[0062] list2 = [368.809, 526.87], type: T

[0063] …

[0064] list10 = [3308.746, 3382.508], type: P

[0065] list11 = [3424.658, 3572.181], type: QRS

[0066] list12 = [3656.481, 3835.616], type: T

[0067] …

[0068] list29 = [6722.866, 6952.433], type: QRS

[0069] list30 = [6817.703, 6933.614], type: P

[0070] list31 = [6975.764, 7070.601], type: T

[0071] …

[0072] list50 = [9620.653, 9863.014], type: QRS

[0073] list51 = [9715.49, 9820.864], type: P

[0074] Next, the tensor generating device (300) generates item values ​​indexed in ascending order for each item based on the start time values ​​of the items (P, QRS, T) in the list, and constructs a tensor with the generated item values ​​and interval information.

[0075] FIG. 4 is a figure illustrating an indexing result according to one embodiment of the present invention.

[0076] Referring to FIG. 4, the tensor generating device (300) refers to a list as input to classify output values ​​(Ps, QRSs, Ts) into categories, and then sorts them in ascending order based on the start time value within each category group and outputs them.

[0077] The tensor generating device (300) performs indexing according to the sorted order, and, for example, in the case of P waves, defines modified item values ​​such as P_1, P_2, P_3, ��� in sequential order starting from the waveform with the smallest start time value. The same operation is performed for QRS complexes and T waves, so that indexed item values ​​are generated in the form of QRS_1, QRS_2, ��� and T_1, T_2, T_3, ���, respectively. For example, in the case of P = [1.035, 118.952], P_1 is defined as [1.035, 118.952] after ascending indexing.

[0078] The tensor generating device (300) generates a tensor with a scale of 4 (rows) x N (columns) x XL (height) and inputs the generated item values ​​and interval information. Here, L is the height of the tensor, which corresponds to the number of leads from which electrocardiogram data was acquired, and the output values ​​of the designated step obtained from the L-th lead are input into the matrix of each L-th height. The lead sequence number (L) follows a pre-specified setting, for example, in the case of a standard 12-lead electrocardiogram, it can be set as I=1, II=2, III=3, aVR=4, aVL=5, aVF=6, V1=7, V2=8, V3=9, V4=10, V5=11, V6=12. Also, N is the number of columns, which is equal to the total number of output values ​​generated in the designated step, i.e., N=α+β+γ (α is the total number of P waves, β is the total number of QRS complexes, and γ is the total number of T waves).

[0079] FIG. 5 is a figure illustrating a tensor generated according to one embodiment of the present invention.

[0080] Referring to FIG. 5, the values ​​in rows 1, 2, and 3 of the same column within the generated tensor (30) correspond to the same item. Specifically, in row 1, the item value (Item value) (P_1, QRS_1, T_1, etc.) generated in the designated step is input. In row 2, the onset value of the corresponding item is input, and the input onset value is configured to increase as the sequence number of each column increases. In row 3, an offset value is input, and at this time, the input condition of row 2 (increase in onset value according to increase in sequence number of each column) is also satisfied.

[0081] Each column vector (32) of the tensor (30) generated through this process corresponds to a single wave on the original electrocardiogram data, and the generated tensor can be utilized as an optimized data structure for the operation of rule-based algorithms and explainable artificial intelligence systems.

[0082] Next, in the constructed tensor, the vector is modified or expanded to have a predefined expected order of progress for each item based on interval information (S200).

[0083] The tensor generating device (300) rearranges each vector (item-specific interval data) within the generated tensor to match the expected progression order (P → QRS → T, etc.) defined in the rule-based algorithm. If there is a sequence different from the expected progression order, each vector is edited within the tensor to correct it, thereby ensuring the operational stability of the algorithm.

[0084] FIG. 6 is a flowchart illustrating a specific rearrangement process according to one embodiment of the present invention.

[0085] The tensor generating device (300) according to the present embodiment adjusts an array of vectors having an overlapping time range (time stamp interval).

[0086] Referring to FIG. 6, for the adjustment of a specific array, the tensor generating device (300) compares the start time value according to the item value of an adjacent vector within the tensor (S210).

[0087] First, the tensor generating device (300) compares start time values ​​for pairs of adjacent column vectors within the tensor.

[0088] For example, regarding two adjacent column vectors as the first condition, if the item value in the front (left) column vector contains the character 'QRS' and the item value in the back (right) column vector contains the character 'P', and at the same time the onset value of the front column vector is greater than or equal to the onset value of the back column vector, it can be determined that the condition is satisfied.

[0089] Alternatively, for two adjacent column vectors as a second condition, if the item value in the first (left) column vector contains the character 'P' and the item value in the second (right) column vector contains the character 'T', and at the same time the onset value of the first column vector is greater than or equal to the onset value of the second column vector, it can be determined that the condition is satisfied.

[0090] The tensor generating device (300) exchanges the two column vectors by changing their order if there exists a pair of column vectors that satisfy the comparison result condition (S220). At this time, changing the order means modifying the values ​​of all terms of the n-th column vector and the (n+1)-th column vector to each other.

[0091] FIG. 7 is an exemplary diagram illustrating the rearrangement process of vectors within a tensor according to one embodiment of the present invention.

[0092] Referring to Fig. 7, when rearrangement is performed according to the first condition (30a), the blank value recorded in the 4th row of the first column vector among the rearranged column vectors is modified to 'P4'.

[0093] In addition, if rearrangement is performed in accordance with the second condition, the blank value recorded in the 4th row of the latter column vector among the rearranged column vectors is modified to 'P7'. Here, P4 represents a retrograde P wave and P7 represents a P on T pattern, which can be used to ensure that the data is properly classified and recognized by rule-based algorithms and explainable artificial intelligence (XAI) in the future.

[0094] In addition, a rearrangement such as the first condition allows the sequence of future beats to be determined without error by artificially bringing the P wave to the forefront of the QRS complex to distinguish it when a retrograde P wave is present.

[0095] The rearrangement, such as the second condition, is intended to enable clinicians to explain the presence of an early-onset P wave on the electrocardiogram of a patient with atrial premature beats in future XAIs.

[0096] In the present invention, P4 and P7 are each result values ​​of categorical variables that classify retrograde P wave and P on T pattern, respectively, and by inputting these values ​​into the 4th row of the corresponding vector during the rearrangement process, the data is recognized as reflecting these patterns in the future.

[0097] In addition, this process enables the presence of a retrograde P wave to be recognized by extracting it ahead of the QRS sequence or by reflecting the presence of a P on T pattern in the reading result in an explanatory manner, thereby allowing it to function appropriately in the final beat sequence determination and XAI-based clinical explanation stages.

[0098] Next, the tensor generation device (300) according to the present embodiment can delete at least one vector within the rearranged tensor so that the indices of all column vectors maintain continuity and accurate reference is possible during the operation of a rule-based algorithm and the interpretation process of artificial intelligence in the future.

[0099] Specifically, the tensor generating device (300) can determine whether the progression order is disturbed by the leading or trailing vector and adjust the vector.

[0100] FIG. 8 is a flowchart illustrating the process of deleting a vector according to one embodiment of the present invention.

[0101] Referring to FIG. 8, the tensor generating device (300) deletes at least one vector if the indexing item value of the leading or trailing vector in the rearranged tensor satisfies the deletion condition (S230).

[0102] FIG. 9 is an example diagram illustrating the process of searching for and deleting column vectors based on the leading vector according to one embodiment of the present invention.

[0103] Referring together with FIG. 9, the tensor generating device (300) searches for a column vector containing P_1 as a value when the value recorded in the first row of the leading vector of the tensor (the leftmost column, i.e., the column vector with column number 1) is not P_1.

[0104] All column vectors (30b) to the left of the searched P_1 column vector are deleted. The purpose of this process is to prevent errors in future beat unit indexing and rule-based algorithm operation by correcting the leading vector of the tensor to indicate the starting point of the normal beat.

[0105] In addition, if the item value recorded in the first row of the trailing vector (rightmost column) of the tensor corresponds to a P-wave, that trailing vector is deleted. The purpose of this process is to ensure that if the trailing vector exists as a P-wave alone, it is considered an incomplete beat and is excluded from the computation and interpretation results of the rule-based algorithm.

[0106] After deleting a vector, the tensor generating device (300) sequentially reassigns the indices of the indexed item values ​​within the tensor (S240).

[0107] Specifically, the values ​​recorded in the first row of each column vector are modified so that the sequence numbers of the QRS complex and T wave both start from 1 and continue as consecutive numbers.

[0108] In the deletion process described above, there are cases where the wave referred to by the leading vector as a deletion condition corresponds to a premature ventricular complex (PVC) on the circular electrocardiogram data.

[0109] At this time, since the tensor generating device (300) should not delete the QRS complex corresponding to a PVC when the wave referred to by the leading vector is a QRS complex, the result value determined by a separate machine learning algorithm or the like during the reassignment step can be used to determine whether each QRS complex corresponds to a PVC.

[0110] When the tensor generating device (300) identifies that the corresponding QRS complex corresponds to a PVC as a result of the determination, it modifies the blank value recorded in the 4th row of the column vector corresponding to the corresponding QRS complex to 'QRS2'. Here, QRS2 is one of the result values ​​of a pattern (categorical variable) that classifies QRS complex suggestive of PVC, and indicates that the corresponding QRS complex is a PVC so that it can be properly interpreted in the future rule-based algorithm and explainable artificial intelligence processing.

[0111] In the present invention, QRS2 is one of the result values ​​of a pattern (categorical variable) that classifies QRS complex suggestive of PVC, and by being input into the 4th row of the corresponding vector, items classified as special patterns on the electrocardiogram, such as retrograde P wave, P on T, and PVC, are appropriately filtered, corrected, and interpreted in subsequent algorithm steps.

[0112] FIG. 10 is a diagram showing the processing process of a reassigned tensor according to one embodiment of the present invention.

[0113] Referring to FIG. 10, the tensor generating device (300) adds a vector such that the progression order of item values ​​within the rearranged vector satisfies the expected progression order (S250).

[0114] First, the tensor generating device (300) checks whether the array of item values ​​recorded in the first row of each matrix of the tensor satisfies the expected progression order from left to right, such as P→QRS→T→P→QRS→T→…. Here, the expected progression order is defined as the order of appearance of waveforms constituting one beat (cardiac cycle) on a normal electrocardiogram. If, as a result of the check, the order is not satisfied, the location where the missing item exists is identified, and a column vector is added to that location.

[0115] FIG. 11 is an exemplary diagram illustrating the process of adding a vector according to one embodiment of the present invention.

[0116] Referring to FIG. 11, the tensor generating device (300) inputs the value of the missing item among P, QRS, and T into the first row of each added column vector to satisfy the above order. For example, if QRS and T are not present in a location where QRS and T should be present, QRS and T are input into the first row of the corresponding column vectors (30c).

[0117] In the 2nd and 3rd rows of the added column vector, a '.' (dot) representing a missing value can be entered. This indicates that the waveform was not actually measured, and at the same time, is intended to maintain a structure that allows the rule-based algorithm to operate without error when performing calculations based on the order of progress.

[0118] In the 4th row of the added column vector, if the value entered in the 1st row of the same column vector is P, P5 is entered; if it is QRS, QRS4 is entered; and if it is T, a missing value ('.') is entered. Here, P5 is one of the result values ​​of a pattern (categorical variable) that classifies 'no P wave' and QRS4 is 'no QRS complex' in the development of the present invention, and the tensor generating device (300) can explicitly indicate that there is no actual waveform.

[0119] These P5 and QRS4 inputs are intended to clearly indicate the absence of a P wave or QRS complex when the electrocardiogram data is subsequently passed to the rule-based algorithm for waveform presence determination or pattern classification. In this invention, the system is configured to maintain the expected progression order while systematically managing the presence of missing waveforms through this method.

[0120] Next, the tensor generating device (300) adds beat values ​​to each vector in the rearranged tensor in units of expected progression order (S300).

[0121] The tensor generating device (300) adds a value indicating the order of the beat to which each vector in the rearranged tensor belongs, thereby enabling clear identification of which beat each vector (wave segment data) corresponds to. This addition of beat values ​​is essential information for a rule-based algorithm to perform conditional judgments on a beat-by-beat basis.

[0122] The tensor generating device (300) can set each matrix of the tensor as a new row by adding a new term on top of the existing row 1. Accordingly, the existing rows 1, 2, 3, and 4 of the tensor are sequentially moved to rows 2, 3, 4, and 5, respectively.

[0123] Afterward, the tensor generating device (300) inputs sequential beat numbers from left to right, such as 1→1→1→2→2→2→3→3→3 … for each term of the added row. That is, the same beat number is assigned to the three items of P, QRS, and T that constitute one beat, and the numbers for the next beat items are input in a manner that increases by 1.

[0124] In this embodiment, the tensor generation device (300) can perform missing processing on the newly added column vectors.

[0125] The tensor generating device (300) modifies all values ​​corresponding to the second row of each added column vector to missing values ​​('.'). Here, the column vectors added during the tensor construction process can be identified by this, as the second row values ​​of each column vector appear as unnumbered values ​​such as P, QRS, and T, unlike the existing column vectors which are numbered such as P_1, QRS_3, and T_10.

[0126] In addition, among the column vectors of the absence processing results, if the values ​​of rows 2, 3, and 4 are all missing values ​​('.') and the value of row 5 is P5, QRS4, or a missing value ('.'), these column vectors can be defined as column vectors corresponding to non-existing waves. Here, P5 is one of the pattern (categorical variable) result values ​​in the development of the present invention indicating 'no P-wave' and QRS4 is 'no QRS complex', and is used to specify that an actual waveform does not exist.

[0127] Consequently, after the pulse information input step and the absence processing step are completed, the following data can be stored in each matrix of the generated 5 x K x L tensor.

[0128] Here, K is a natural number representing the number of column vectors in the matrix for each lead, and L is the total number of leads of the original ECG data to be analyzed.

[0129] In the first row, the sequence number of the beat to which each item belongs can be stored; in the second row, the item value or missing value ('.'); in the third row, the onset value or missing value of each item; in the fourth row, the offset value or missing value of each item; and in the fifth row, the result value (categorical variable) or missing value specifying a retrograde P wave or 'corresponds to a pattern of premature ventricular beats', 'no P wave (P5)', or 'no QRS complex (QRS4)' can be stored, respectively.

[0130] In addition, among the values ​​in the 5th row, items for which the result value of the pattern (categorical variable) has not yet been stored remain blank and are configured to be additionally determined and input during the subsequent rule-based algorithm or explainable artificial intelligence interpretation stage.

[0131] The tensor generating device (300) can add indices as variables to a tensor of size 5 x K x L completed up to the previous step.

[0132] FIG. 12 illustrates the step of expanding a vector within a tensor generated according to one embodiment of the present invention.

[0133] The tensor generating device (300) according to the present embodiment provides a structural basis in the tensor for storing various indicators as variables by expanding the size of the vector (S400).

[0134] In this embodiment, the indicators added as variables may be classified by degree according to hierarchical reference relationships; for example, the type or name of the lead corresponding to the sequence number of each lead may be entered as a first-order indicator. This is determined by referencing the [lead]-[sequence number] order pair used when the tensor was first created. Next, the duration of each item may be entered. The duration is calculated by subtracting the onset value from the offset value, but if any of item, onset, or offset is missing, a missing value may also be entered into the corresponding duration term. Additionally, the amplitude value of each item may be entered, which is calculated using various decision algorithms and formulas, and includes, for example, the amplitude of the P-wave, the amplitude of the R-wave, the net amplitude of the QRS complex, and the amplitude of the T-wave. Subsequently, the morphology value of each item may be entered; this is a categorical variable and can be derived through various decision algorithms and classification systems.

[0135] Next, the pattern result value can be input as a primary indicator. For the P wave, one of normal P wave (P1), Flutter wave (P2), fibrillatory wave (P3), retrograde P wave (P4), no P wave (P5), other P wave abnormality (P6), and P on T (P7) can be input, and for the QRS complex, one of normal QRS complex (QRS1), QRS complex suggestive of PVC (QRS2), other QRS abnormality (QRS3), and no QRS complex (QRS4) ​​can be input.

[0136] When primary indicators are added, a tensor of size M x K x L is finally completed. Here, M is the total number of indicators, preferably 9 or more, and the indicators included may be information related to each lead, such as the type or name of the lead corresponding to the sequence number (nominal variable), the sequence number of the beat (discrete variable), the item value or missing value (nominal variable), the onset time value or missing value of each item (continuous variable), the offset time value or missing value of each item (continuous variable), the pattern of each item (categorical variable), the duration or missing value of each item (continuous variable), the amplitude or missing value of each item (continuous variable), or the shape or missing value of each item (categorical variable).

[0137] In addition, secondary indicators are defined in each item but are not directly used for pattern determination and can be calculated based on primary indicators. Examples include q wave duration, P wave notch duration, and QRS R' / R amplitude ratio.

[0138] The third index represents the relationship between multiple items, and includes PR interval, RR interval, QRS amplitude ratio, heart rate variability, and classification of beat types.

[0139] The fourth index represents the relationship between different items or expresses the characteristics of the entire electrocardiogram segment, and may include the overall P wave axis average, the amplitude deviation of a specific beat-in-question, and the RR interval average of the sinal beats.

[0140] In an embodiment of the present invention, the tensor finally generated is composed of a multidimensional array having a three-dimensional structure of rows (A), columns (B), and height (C), as shown in FIG. 13.

[0141] FIG. 13 illustrates a tensor configured according to one embodiment of the present invention.

[0142] Referring to FIG. 13, the tensor (30) is constructed by stacking two-dimensional planes (matrices) corresponding to each height (C). Here, each height corresponds to a lead of the electrocardiogram data, and for example, in the case of a standard 12-lead electrocardiogram, C=12.

[0143] A plane of each height is again composed of rows (A) and columns (B). As shown in the enlarged view on the right side of the figure, a single two-dimensional plane (i.e., a matrix for waveforms within a single lead) has a size of A x B, where A is the number of indices included in each column vector and B is the number of column vectors (items). For example, in the present invention, A is 9 or greater, and each row contains values ​​such as beat number, item, onset, offset, pattern, duration, amplitude, morphology, and lead name.

[0144] Each column vector (30d) within the matrix is ​​defined in the form vj = [a1j, a2j, a3j, ..., aAj], which sequentially contains values ​​from row 1 to row A of the corresponding column. These column vectors correspond to a single wave in the original electrocardiogram data and contain information such as, for example, the P wave, QRS complex, T wave, etc. The column vectors are represented by the vertical values ​​of each column, and it can be seen that they are the smallest unit data structure of a tensor.

[0145] In addition, each matrix is ​​composed of B column vectors of the same structure arranged in a row, and these planes are stacked in the height (C) direction to form the entire tensor.

[0146] In addition, the final tensor of the present invention comprises C derivation-specific matrices as described above, and each matrix is ​​composed of A rows and B column vectors, and each column vector may be multidimensional data corresponding to a single wave, but is not limited thereto. The order and arrangement method of rows, columns, and heights can be changed according to the design, storage structure, or processing efficiency of the data processing system, and it is obvious that various modifications are possible within the spirit and scope of the present invention.

[0147] This tensor structure is an optimized data structure that enables the integrated storage and analysis of various variables of electrocardiogram data (beat sequence number, item type, start time, end time, pattern, duration, amplitude, shape, lead name, etc.) by lead, item, and indicator, and provides a core foundation for the rule-based algorithm and artificial intelligence system of the present invention to operate efficiently.

[0148] FIG. 14 is an exemplary diagram showing the correspondence between a tensor generated according to one embodiment of the present invention and circular electrocardiogram data.

[0149] Referring to FIG. 14, the present invention enables efficient and accurate interpretation in rule-based algorithms and explainable artificial intelligence systems by converting electrocardiogram data into a tensor form and storing and processing it.

[0150] More specifically, referring to FIG. 14, the original electrocardiogram data is recorded as waveforms continuously along the time axis, and each waveform can be classified into P waves, QRS complexes, T waves, etc. In the present invention, the tensor generating device (300) defines each waveform on the electrocardiogram as an individual item and manages various variables such as the start time value, end time value, and pattern of the corresponding waveform by inputting them into a single column vector.

[0151] Each column vector(s) (30e, 30f) of the finally constructed tensor corresponds to the waveform(s) (10e, 10f) on the electrocardiogram data. For example, column vectors such as P_1, QRS_1, T_1, etc. are stored sequentially, and each column vector includes the type of waveform, start time, end time, and pattern result values ​​(P4, QRS2, etc.) used in rule-based algorithms and XAI systems.

[0152] Each item in the column vector within the tensor corresponds one-to-one with the interval and waveform(s) on the original electrocardiogram data. Through this, when the AI ​​system presents a final judgment conclusion after a rule-based algorithm performs analysis, it is possible to clearly visualize and explain to the clinician which waveform and interval of the electrocardiogram data the basis for that conclusion was derived.

[0153] In other words, the tensor-based data processing structure according to the present invention aligns and stores complex time and waveform information of an electrocardiogram signal in a multidimensional structure, thereby supporting rule-based algorithms to stably and quickly perform condition judgments at the beat and waveform levels. At the same time, it enables an explainable artificial intelligence system to trace the basis of the results at the original data interval and waveform levels, thereby significantly improving the reliability of the clinician and the clinical utility value.

[0154] The present invention efficiently stores different types of variables within a single tensor structure, enables accurate grouping of electrocardiogram signals in beat units, and is configured so that the stored information corresponds one-to-one with the original data interval. In addition, since structural scalability is guaranteed even when new indicators or waveforms are added, the accuracy and explainability of electrocardiogram interpretation can be increased, thereby enhancing the confidence of clinicians.

[0155] FIG. 15 is an exemplary diagram showing the implementation of a tensor generation device in the form of a computing device according to one embodiment of the present invention.

[0156] Referring to FIG. 15, in some embodiments of the present invention, the system (300) may be implemented in the form of a computing device.

[0157] One or more of the modules constituting the system (300) are implemented on a general-purpose computing processor and thus may include a processor (308), an input / output I / O (302), a memory (304), an interface (306), and a bus (314). The processor (3008), the input / output device (302), the memory (304), and / or the interface (306) may be connected to each other via the bus (314). The bus (314) corresponds to a path through which data travels.

[0158] Specifically, the processor (308) may include at least one of a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), microprocessor, digital signal processor, microcontroller, application processor (AP), and logic elements capable of performing similar functions.

[0159] The input / output device (302) may include at least one of a keypad, a keyboard, a touchscreen, and a display device. The memory (304) may store data and / or programs, etc.

[0160] The interface (306) can perform the function of transmitting data to a communication network or receiving data from a communication network. The interface (306) may be wired or wireless. For example, the interface (306) may include an antenna or a wired / wireless transceiver, etc. The memory (304) is a volatile operational memory for protecting personal information, which enhances the operation of the processor (308), and may further include high-speed DRAM and / or SRAM, etc.

[0161] Additionally, memory (304) stores programming and data configurations that provide the functions of some or all of the modules described herein.

[0162] A program or application is loaded as a set of instructions including each operation using the above-described language model stored in memory (304), and the processor is enabled to perform each operation.

[0163] The various embodiments described herein may be implemented, for example, in a recording medium readable by a computer or similar device using software, hardware, or a combination thereof.

[0164] According to hardware implementation, the embodiments described herein may be implemented using at least one of ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), processors, controllers, microcontrollers, microprocessors, and other electrical units for performing functions. In some cases, the embodiments described herein may be implemented as the control module itself.

[0165] According to software implementation, embodiments such as the procedures and functions described herein may be implemented in separate software modules. Each of the software modules may perform one or more functions and operations described herein. Software code may be implemented as a software application written in a suitable programming language. The software code may be stored in a memory module and executed by a control module.

[0166] Although various embodiments of the present invention have been described in detail above, those skilled in the art will understand that various modifications can be made to the above-described embodiments without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the claims set forth below as well as equivalents thereof.

Claims

1. In an electrocardiogram-based tensor generation method performed on a computer device, A step of constructing a tensor based on segment information in which the electrocardiogram is classified into segments according to multiple items; In the above-configured tensor, a step of modifying or expanding a vector to have a predefined expected progression order for each item based on the interval information; and A step comprising adding beat values ​​to each vector within the modified or expanded tensor based on the expected progression order, ECG-based tensor generation method.

2. In Paragraph 1, The above interval information includes start and end time values ​​of the above item, and The step of constructing the above tensor is, Generating item values ​​indexed in ascending order by item based on the start time value of the above item, and constructing the above tensor with the generated item values ​​and the above interval information, ECG-based tensor generation method.

3. In Paragraph 2, The above modification or expansion step is, Adjusting the array of the above vectors, ECG-based tensor generation method.

4. In Paragraph 3, The above modification or expansion step is, Adjusting the leading or trailing vector ECG-based tensor generation method.

5. In Paragraph 1, The above modification or expansion step is, Adding a vector to satisfy the above expected sequence of progress, ECG-based tensor generation method.

6. In Paragraph 1, A method further comprising the step of adding an indicator containing induction-related information to a vector within the above tensor. ECG-based tensor generation method.

7. Processor, It includes a memory that communicates with the above processor, and The above memory stores instructions that allow the processor to perform operations, and The above operations are, Operation of constructing a tensor based on segment information classifying an electrocardiogram into segments according to multiple items, In the above-configured tensor, an operation of modifying or expanding a vector to have a predefined expected progression order for each item based on the interval information, and A method comprising adding beat values ​​to each vector within the modified or expanded tensor based on the expected progression order. Computing device.

8. In Paragraph 7, The above interval information includes start and end time values ​​of the above item, and The operation constituting the above tensor is, Generating item values ​​indexed in ascending order by item based on the start time value of the above item, and constructing the above tensor with the generated item values ​​and the above interval information, Computing device.

9. In Paragraph 8, The above-mentioned modification or expansion operation is, Adjusting the array of the above vectors, Computing device.

10. In Paragraph 9, The above-mentioned modification or expansion operation is, Adjusting the leading or trailing vector Computing device.

11. In Paragraph 7, The above-mentioned modification or expansion operation is, Adding a vector to satisfy the above expected sequence of progress, Computing device.

12. In Paragraph 7, Further including the operation of adding an indicator containing induction-related information to a vector within the above tensor, Computing device.