Heart disease prediction model generation method, server, and computer program

Abstract

The embodiment of the invention discloses a heart disease prediction model generation method. The method may include a step of acquiring a plurality of electrocardiogram data and a step of generating a model for predicting a heart disease based on the plurality of electrocardiogram data.

Description

[0001] This application is a divisional application of the patent application with application number 202411186030.6, application date August 27, 2024, and invention name “Method, server and computer program for generating heart disease prediction model”. Technical Field

[0002] Multiple embodiments of the present invention relate to methods for providing a neural network model for predicting a patient's likely future cardiac disease, and more specifically, to techniques for using a deep learning model to evaluate and predict the risk of arrhythmias, including atrial fibrillation, and other cardiac diseases based on electrocardiogram data. Background Art

[0003] Electrocardiogram (ECG) data is a widely used and important tool for diagnosing heart conditions by recording the heart's electrical activity. Existing multi-lead electrocardiographs play a core role in diagnosing arrhythmias, i.e., heart diseases such as atrial fibrillation, by detecting abnormalities in heart rhythm and electrical conduction. Arrhythmias, including atrial fibrillation, are common diseases characterized by irregular or abnormal heart rates. Since these arrhythmias are usually asymptomatic or have mild symptoms, they are difficult to detect without proper examination. If such arrhythmias are not treated in time, they may cause serious health complications such as stroke, heart failure, and sudden cardiac death. In particular, atrial fibrillation is associated with an increased risk of stroke and heart failure, making it one of the important types of arrhythmias. Therefore, early detection and intervention of atrial fibrillation and arrhythmias are very important for effective disease management.

[0004] On the other hand, existing electrocardiographs, while focused solely on diagnosis, have limitations in predicting or stratifying future heart disease risk. This is because existing electrocardiogram analysis methods primarily rely on statistical analysis. Statistical methodologies cannot fully reflect nonlinear and complex patterns, making it difficult to accurately assess each patient's cardiac status and the risk of various arrhythmias.

[0005] Furthermore, statistical models have lower performance when processing large amounts of data compared to probabilistic models based on deep learning. ECG data can be affected by overlapping signals such as muscle movement, breathing, and spot noise from other devices. These external factors can reduce the accuracy of statistical methodologies, ultimately becoming a major obstacle to accurately stratifying and quantifying heart disease risk.

[0006] Prior art literature

[0007] Patent Literature

[0008] Korean Patent Publication No. 10-2022-0104583 Summary of the Invention

[0009] The technical problem to be solved by the present invention is proposed in response to the above background technology. The purpose of the present invention is to provide the following technology, which can predict and stratify the risk level of future heart disease based on electrocardiogram data.

[0010] The technical problems to be solved by the present invention are not limited to the problems mentioned above. Ordinary technicians in the technical field to which the present invention belongs can clearly understand other technical problems not mentioned from the following description.

[0011] One embodiment of the present invention for solving the above problem discloses a method for generating a heart disease prediction model. The method may include the following steps: acquiring a plurality of electrocardiogram data; and generating a model for predicting heart disease based on the plurality of electrocardiogram data.

[0012] In an alternative embodiment, the step of generating a model for predicting the above-mentioned heart disease may include the following steps: constructing a learning data set based on the above-mentioned multiple electrocardiogram data; and learning at least one network function based on the above-mentioned learning data set to generate a heart disease risk level prediction model, the above-mentioned learning data set includes a first learning data set and a second learning data set classified according to different learning purposes, and the above-mentioned heart disease risk level prediction model is stratified based on the patient's electrocardiogram data and provides predictive information related to the risk level of future heart disease.

[0013] In an alternative embodiment, the first learning dataset may include data for learning corresponding to a process of transforming the characteristics of the electrocardiogram data into a characteristic space, and the second learning dataset may include calibration data related to cardiac risk prediction.

[0014] In an alternative embodiment, the step of constructing the above-mentioned learning data set may include the step of performing preprocessing on the above-mentioned multiple electrocardiogram data, and the step of performing the above-mentioned preprocessing includes the following steps: performing noise preprocessing on the above-mentioned multiple electrocardiogram data; dividing the multiple electrocardiogram data after the above-mentioned noise preprocessing into predefined window sizes to generate multiple region of interest (ROI) signals; and extracting heart rate variability (HRV) characteristic information based on the leads of the above-mentioned multiple electrocardiogram data, and the above-mentioned heart rate variability characteristic information includes indicator information related to the variability between relevant heart rate intervals.

[0015] In an alternative embodiment, the step of generating the above-mentioned heart disease risk prediction model may include: an embedding model generation step, generating an embedding model by using the masked autonomous learning of the above-mentioned first learning data set; an embedding execution step, using the above-mentioned embedding model to extract multiple latent vectors (latent vectors) corresponding to multiple region of interest signals respectively, and the above-mentioned multiple region of interest signals correspond to the above-mentioned second learning data set; a clustering execution step, using the clustering model to perform clustering on the above-mentioned multiple latent vectors; and a vector database construction step, constructing a vector database (DB, DataBase) based on the above-mentioned multiple latent vectors and information related to each cluster corresponding to each latent vector.

[0016] In an alternative embodiment, the above-mentioned embedding model generation step may include the following steps: inputting data contained in the above-mentioned first learning data set into the self-reconstruction model to guide self-supervised learning, so that the above-mentioned self-reconstruction model generates an output similar to the above-mentioned input data; and extracting the encoder from the self-reconstruction model that completes the above-mentioned learning and generating the above-mentioned embedding model, the above-mentioned self-reconstruction model serves as a neural network model that masks a part of the input data and restores the masked part, including a dimensionality reduction network function and a dimensionality restoration network function.

[0017] In an alternative embodiment, the clustering model may perform clustering based on similarity distances between multiple latent vectors corresponding to the multiple ROI signals, and the clustering step includes extracting a representative vector for each cluster corresponding to the clustering result.

[0018] In an alternative embodiment, the method may include the following steps: acquiring electrocardiogram data of a predicted object; performing preprocessing on the electrocardiogram data of the predicted object; using the embedding model to generate a plurality of latent vectors corresponding to the preprocessed electrocardiogram data; respectively comparing the plurality of latent vectors with representative vectors corresponding to the plurality of clusters to classify the plurality of latent vectors into one of the plurality of clusters; and generating hierarchical information related to the risk level of heart disease based on the classification results of the plurality of latent vectors into the plurality of clusters.

[0019] In an alternative embodiment, the step of generating the above-mentioned heart disease risk prediction model may include the following steps: obtaining user meta-information corresponding to the above-mentioned multiple electrocardiogram data respectively; obtaining the above-mentioned multiple latent vectors and cluster information corresponding to the above-mentioned multiple latent vectors from the above-mentioned vector database; and training multiple tree models based on the above-mentioned heart rate variability characteristic information, the above-mentioned user meta-information, the above-mentioned multiple latent vectors and the cluster information corresponding to the above-mentioned latent vectors, and performing integrated learning to merge the outputs of each tree model to generate the above-mentioned heart disease risk prediction model.

[0020] In an alternative embodiment, the step of generating a model for predicting the above-mentioned heart disease may also include the following steps: classifying the above-mentioned multiple electrocardiogram data obtained into a training data set and a validation data set; respectively obtaining multiple region-of-interest electrocardiogram data divided into predefined window sizes from the multiple electrocardiogram data classified into the above-mentioned training data set and the validation data set; training the above-mentioned deep learning model accordingly by inputting the multiple first region-of-interest electrocardiogram data obtained from the multiple electrocardiogram data contained in the above-mentioned training data set into the deep learning model for predicting the patient's heart disease, so as to predict the heart disease category of each of the above-mentioned multiple first region-of-interest electrocardiogram data; and using the multiple second region-of-interest electrocardiogram data obtained from the multiple electrocardiogram data contained in the above-mentioned validation data set for the trained deep learning model to determine the threshold for distinguishing the above-mentioned heart disease categories.

[0021] In an alternative embodiment, the step of determining the threshold value for distinguishing the above-mentioned heart disease categories may include the following steps: collecting probability scores associated with multiple region-of-interest electrocardiogram data separated from the same electrocardiogram data for the above-mentioned multiple second region-of-interest electrocardiogram data; averaging the probability scores associated with the above-mentioned multiple second region-of-interest electrocardiogram data collected to derive a final probability score; and fine-tuning the threshold value for distinguishing the above-mentioned heart disease categories based on the derived final probability score.

[0022] In an alternative embodiment, the determined threshold value may be stored together with the weighted value of the deep learning model.

[0023] According to yet another embodiment of the present invention, a server for executing a method for generating a heart disease prediction model is disclosed. The server may include: a memory storing at least one instruction; and a processor executing the at least one instruction stored in the memory. The processor executing the at least one instruction executes the method for generating a heart disease prediction model.

[0024] According to another embodiment of the present invention, the present invention discloses a computer program stored in a computer-readable recording medium. The computer program can be combined with a computer as hardware to execute the method for generating a heart disease prediction model.

[0025] Other details of the present invention are included in the detailed description and the accompanying drawings.

[0026] According to multiple embodiments of the present invention, region-of-interest electrocardiogram data can be extracted from a patient's electrocardiogram data, and the extracted region-of-interest electrocardiogram data can be applied to a deep learning model to improve the accuracy of predicting the patient's heart disease.

[0027] Furthermore, the present invention can provide the following benefits: using multiple machine learning models to predict the risk of future heart disease, allowing patients to understand their health status and proactively prevent heart disease. This allows for early detection of heart disease and prompt treatment, saving patients' lives and improving treatment outcomes.

[0028] Furthermore, the present invention can provide the following effects: through an AI-based automated processing process, the physical time of medical staff required to analyze electrocardiogram data one by one is greatly reduced, thereby improving the effectiveness of screening and reducing the opportunity cost at the medical site.

[0029] Additionally, the present invention can use multiple tree series machine learning models to perform integration to integrate the latent vectors extracted from the embedding model of the heart rate variability characteristic information and metadata such as the patient's age, thereby further improving the prediction accuracy of the risk level of heart disease.

[0030] The effects of the present invention are not limited to the effects mentioned above, and those skilled in the art can clearly understand other effects not mentioned from the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 FIG. 1 is a structural diagram illustrating a computing device for executing a method for generating a heart disease prediction model according to an embodiment of the present invention.

[0032] Figure 2 2 is a flowchart illustrating a learning method of a deep learning model for predicting heart disease according to an embodiment of the present invention.

[0033] Figure 3 FIG. 1 is a diagram illustrating a method for classifying a plurality of electrocardiogram data according to an embodiment of the present invention.

[0034] Figure 4 FIG. 1 is a diagram illustrating a method for subdividing electrocardiogram data into individual heart rates according to an embodiment of the present invention.

[0035] Figures 5a to 5cA diagram illustrating types of deep learning models according to an embodiment of the present invention.

[0036] Figure 6 A diagram illustrating the learning steps of a deep learning model according to an embodiment of the present invention.

[0037] Figure 7 2 is a flowchart illustrating a method for predicting heart disease using a trained deep learning model according to an embodiment of the present invention.

[0038] Figure 8 A diagram illustrating the steps of predicting heart disease using a deep learning model according to an embodiment of the present invention.

[0039] Figure 9 An exemplary flow chart related to a method for generating a heart disease prediction model according to an embodiment of the present invention is shown.

[0040] Figure 10 The flowchart exemplarily shows a process of pre-processing a plurality of electrocardiogram data according to an embodiment of the present invention.

[0041] Figure 11 The flowchart exemplarily shows a process of performing embedding on multiple latent vectors according to an embodiment of the present invention and constructing a vector database based on the execution results.

[0042] Figure 12 The following is a flowchart illustrating a process of generating an embedding model according to an embodiment of the present invention.

[0043] Figure 13 1 is an illustrative diagram for illustrating a self-reconstruction model and a process of generating an embedding model from a trained self-reconstruction model according to an embodiment of the present invention.

[0044] Figure 14 The flowchart is an exemplary flowchart for illustrating a process of generating a heart disease risk prediction model through ensemble learning of multiple tree series models according to an embodiment of the present invention.

[0045] Description of Reference Signs

[0046] 100: Computing device

[0047] 110: Processor

[0048] 120: Memory

[0049] 130: Procedure

[0050] 140: Display DETAILED DESCRIPTION

[0051] The specific structural or functional descriptions of the embodiments are for illustrative purposes only and may be implemented in a variety of forms. Therefore, actual implementations are not limited to the specific disclosed embodiments. The scope of this specification includes modifications, equivalent technical solutions, and alternative technical solutions included in the technical concepts described in the embodiments.

[0052] The terms "first" or "second" can be used to describe multiple structural elements, but these terms are only used to describe two structural elements. For example, the first structural element can be named the second structural element, and similarly, the second structural element can also be named the first structural element.

[0053] When it is mentioned that a structural element is “connected” to another structural element, it can be directly connected or coupled to the other structural element, or other structural elements may be present in between.

[0054] Unless the context clearly indicates otherwise, expressions in the singular include expressions in the plural. In this document, sentences such as "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C", and "at least one of A, B or C" may respectively include one or all possible combinations of the multiple items listed together in the corresponding sentences in the above multiple sentences. In this specification, terms such as "including" or "having" are used to specify the presence of the described features, numbers, steps, actions, structural elements, components, or combinations thereof, and do not preclude the presence or additional possibilities of one or more other features, numbers, steps, actions, structural elements, components, or combinations thereof.

[0055] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meanings as those generally understood by those skilled in the art to which the present invention belongs. Terms defined in commonly used dictionaries have the same meanings as those in the context of the relevant technology and should not be interpreted as ideal or excessive meanings unless explicitly defined in this specification.

[0056] In the following, the embodiments are described in detail with reference to the accompanying drawings. In the description with reference to the accompanying drawings, the same components are given the same reference numerals regardless of the reference numerals, and repeated description thereof is omitted.

[0057] The present invention can acquire multiple electrocardiogram (ECG) data corresponding to multiple patients and generate a model for predicting heart disease based on the acquired ECG data. In an embodiment, the model for predicting heart disease can be a neural network model that extracts important patterns and features from the ECG data and predicts the likelihood of heart disease based on these patterns. The computing device of the present invention processes the multiple ECG data into a form suitable for learning and uses the processed data to train the model for predicting heart disease.

[0058] According to one embodiment, the present invention can generate and provide a deep learning model and a heart disease risk prediction model as models for predicting heart disease. In this embodiment, the present invention can process multiple electrocardiogram (ECG) data into various formats and train neural networks using different methods to generate and provide multiple prediction models. The following describes in detail the method for generating multiple models for predicting heart disease, with reference to various figures.

[0059] Figure 1 FIG. 1 is a structural diagram illustrating a computing device for executing a method for generating a heart disease prediction model according to an embodiment of the present invention.

[0060] like Figure 1 As shown, the computing device 100 may include at least one processor 110 and a memory 120 that may load or store a program 130 executed by the processor 110 . Figure 1 The structural elements included in the computing device 100 are only an example, and those skilled in the art can understand that Figure 1 In addition to the structural elements shown, other commonly used structural elements may also be included.

[0061] The processor 110 controls the overall operation of the various structural elements of the computing device 100. The processor 110 may include a central processing unit (CPU), a microprocessor unit (MPU), a microcontroller unit (MCU), a graphics processing unit (GPU), a neural processing unit (NPU), a digital signal processor (DSP), or at least one of any form of processor known in the technical field of the present invention. In addition, the processor 110 may execute operations on at least one application or program for running the methods / work of various embodiments of the present invention. The computing device 100 may include at least one processor.

[0062] The memory 120 stores at least one or a combination of two or more of various data, instructions, and information used by structural elements (eg, the processor 110) included in the computing device 100. The memory 120 may include a volatile memory and / or a nonvolatile memory.

[0063] The program 130 may include at least one action for implementing the methods / actions of various embodiments of the present invention, and is stored in the memory 120 in the form of software. The action corresponds to the instructions executed in the program 130. For example, the program 130 may include multiple instructions for performing the following actions: an action for classifying a plurality of electrocardiogram data acquired from a plurality of patients into a training dataset, a validation dataset, and an experimental dataset; an action for acquiring a plurality of region-of-interest electrocardiogram data segmented into predefined window sizes from the plurality of electrocardiogram data classified into the training dataset, the validation dataset, and the experimental dataset; an action for training a deep learning model for predicting a patient's heart disease by inputting a plurality of first region-of-interest electrocardiogram data acquired from the plurality of electrocardiogram data included in the training dataset into the deep learning model for predicting a patient's heart disease, so as to predict the heart disease category of each first region-of-interest electrocardiogram data; and an action for applying a plurality of second region-of-interest electrocardiogram data acquired from the plurality of electrocardiogram data included in the validation dataset to the trained deep learning model to determine a threshold for distinguishing heart disease categories.

[0064] According to an embodiment, ROI ECG data segmented into predefined window sizes can represent individual heart rates, but is not limited thereto. For ease of explanation, according to an embodiment, ROI ECG data segmented into predefined window sizes will be described using individual heart rates as an example. However, the types of ROI ECG data applicable to various embodiments are not limited thereto.

[0065] When the program 130 is loaded into the memory 120 , the processor 110 may execute various tasks for implementing the program 130 , thereby performing the methods / tasks of various embodiments of the present invention.

[0066] The running screen of the program 130 can be displayed through the display 140. Figure 1 In the case of a computer, display 140 is a separate device connected to computing device 100. In the case of a computing device 100 such as a portable terminal such as a smartphone or tablet, display 140 may be a component of computing device 100. The screen displayed on display 140 may be before information is input to the program or the result of the program execution.

[0067] Figure 2 2 is a flowchart illustrating a learning method of a deep learning model for predicting heart disease according to an embodiment of the present invention.

[0068] Figure 2 The learning method of the deep learning model shown is achieved by Figure 1 The processor of the computing device shown is executed. In step 210, the processor can classify the multiple electrocardiogram data obtained from the multiple patients into a training data set, a validation data set and an experimental data set. First, the multiple electrocardiogram data obtained from the multiple patients can be grouped into normal sinus rhythm (True-Normal Sinus Rhythm, T-NSR), atrial fibrillation-normal sinus rhythm (Afib-Normal Sinus Rhythm, AF-NSR) and clinically important arrhythmia-normal sinus rhythm (Clinically Important Arrhythmia-Normal Sinus Rhythm, CIA-NSR). In this case, the multiple electrocardiogram data obtained from the multiple patients can be 10 seconds 12-lead electrocardiogram data, but the type of this electrocardiogram data is only an example and is not limited to the above example.

[0069] On the other hand, clinically important arrhythmias (CIA) may include atrial arrhythmias, ventricular arrhythmias, atrial fibrillation, and incomplete bundle branch block (BBB). Atrial arrhythmias, as arrhythmias originating from the atria, refer to non-sustained arrhythmias that last for more than 30 seconds or less than 30 seconds, and may include atrial premature beats or sustained / non-sustained atrial rhythms, etc. Ventricular arrhythmias are arrhythmias originating from the ventricles, and refer to non-sustained arrhythmias that last for more than 30 seconds or less than 30 seconds, and may include ventricular premature beats or sustained / non-sustained ventricular arrhythmias, etc. Atrial fibrillation may refer to an arrhythmia in which the atria do not have regular electrical signals and contractions, thereby causing irregular contractions of the ventricles. Finally, incomplete bundle branch block can refer to an arrhythmia that presents a characteristic electrocardiographic pattern due to blockage of the right or left bundle of cardiac signals through the ventricles.

[0070] The normal sinus rhythm group can be composed of patients without atrial fibrillation or arrhythmia pathology, with normal sinus rhythm ECG within 1 year. Figure 3The atrial fibrillation-normal sinus rhythm group can be composed of paired ECGs of patients with normal sinus rhythm and an ECG of atrial fibrillation or atrial flutter that occurred within 14 days of the corresponding normal sinus rhythm ECG. Similarly, the clinically important arrhythmia-normal sinus rhythm group can be composed of paired ECGs of patients with normal sinus rhythm and an ECG of arrhythmia that occurred within 14 days of the corresponding normal sinus rhythm ECG.

[0071] The processor can classify multiple electrocardiogram data belonging to a normal sinus rhythm group, an atrial fibrillation-normal sinus rhythm group, and a clinically important arrhythmia-normal sinus rhythm group into a training dataset, a validation dataset, and an experimental dataset that can be used in a deep learning model for predicting heart disease.

[0072] As an example, Figure 3 This figure illustrates a method for classifying multiple electrocardiogram (ECG) data sets according to an embodiment of the present invention. A processor can classify multiple ECG data sets 310 into groups of normal sinus rhythm, atrial fibrillation and normal sinus rhythm, and clinically significant arrhythmias and normal sinus rhythm, based on the heart disease they are predicted to belong to. For example, to predict atrial fibrillation, the processor can classify multiple ECG data sets 310 into groups of normal sinus rhythm and atrial fibrillation and normal sinus rhythm. To predict arrhythmias, the processor can classify multiple ECG data sets 310 into groups of normal sinus rhythm and clinically significant arrhythmias and normal sinus rhythm.

[0073] Then, the processor may classify the plurality of electrocardiogram data 320 , 330 classified for each heart disease into a training data set, a validation data set, and an experimental data set based on an arbitrary date when the corresponding plurality of electrocardiogram data 320 , 330 is generated.

[0074] As an example, Figure 3 As shown, when multiple electrocardiogram data 320, 330 are acquired during the period from 2017-05-23 to 2022-05-23, for the multiple electrocardiogram data 320, 330, the processor can classify the multiple electrocardiogram data before the corresponding date into a training data set and a validation data set according to a specified ratio based on an arbitrary date (for example, 2021-06-11), and classify the multiple electrocardiogram data after the corresponding date (including the corresponding date) into an experimental data set. Figure 3 In the example, multiple electrocardiogram data 320, 330 can be classified according to the ratio of training data set (60%), verification data set (20%) and experimental data set (20%). This classification ratio is only an example, and the present invention is not limited to the above example.

[0075] In step 220, the processor may obtain individual heart rates from the plurality of ECG data classified into the training dataset, the validation dataset, and the experimental dataset. To obtain accurate and reliable data, the processor may pre-process the 10-second 12-lead ECG data classified into the training dataset, the validation dataset, and the experimental dataset.

[0076] More specifically, the processor may receive multiple electrocardiogram data sets, classified into a training dataset, a validation dataset, and an experimental dataset, in the form of an Extensible Markup Language (XML) file. The processor may parse the input data into a fixed-data portion consisting of the patient's name, age, and gender, and an unfixed-data portion consisting of a continuous signal. The processor may then perform noise removal preprocessing on the unfixed-data portion consisting of the continuous signal and distinguish key heart rate markers. The processor may then obtain discrete individual heart rates from the unfixed-data portion consisting of the continuous signal using the key heart rate markers distinguished as described above.

[0077] As an example, refer to Figure 4 After decoding 10 seconds of 12-lead ECG data using Base64 encryption, the processor passes the decoded data through an IIR Butterworth SOS filter with a moving average kernel and a power line noise filter for noise removal and purification. The processor then uses a QRS peak detection algorithm to segment the 10 seconds of de-noised 12-lead ECG data into individual heart rates.

[0078] Results, such as Figure 3 As shown, the processor can obtain multiple discontinuous individual heart rates from 10 seconds of 12-lead electrocardiogram data. The individual heart rates obtained as above can be used to train a deep learning model to more accurately predict heart disease.

[0079] In step 230, corresponding to inputting the first individual heart rate obtained from the plurality of electrocardiogram data included in the training data set into the deep learning model for predicting the patient's heart disease, the processor can train the deep learning model to predict the heart disease category of each first individual heart rate.

[0080] In this case, at least one deep learning model of ResNet-18, Conv1D including LSTM (Long Short-Term Memory), and Conv1D including Transformer can be used for training to predict the heart disease category of each first individual heart rate.

[0081] As an example, Figure 5aThe ResNet-18 shown in the figure can be a deep learning model that uses convolutional neural networks (CNNs) and other convolutional operations to extract necessary features from the input. In order to solve the vanishing gradient problem of convolutional neural network architecture, such as Figure 5a As shown in Figure 1, ResNet-18 can perform residual learning through skip connections, which is a way of directly connecting input data to the output layer through multiple layers of the skip network.

[0082] The ResNet-18 deep learning model requires fixed-length input. Therefore, the processor may fix the length of individual heart rates to the average length of all individual heart rates. For example, when the average length of all individual heart rates is 700, the processor may slice individual heart rates with lengths longer than 700 and zero-pad individual heart rates with lengths shorter than 700 to fix their lengths to 700.

[0083] As another example, Figure 5b The Conv1D layer including the Long Short-Term Memory (LSTM) shown can capture both local and long-range temporal patterns from sequential data. In this case, the Conv1D layer excels at detecting local temporal patterns, while the LSTM layer excels at modeling long-range dependencies.

[0084] As another example, Figure 5c The Conv1D layers shown here, each with a transformer, can capture both local patterns and global dependencies in the input data. In this case, the transformer layer excels at modeling global dependencies, while the Conv1D layer effectively detects local patterns. Unlike ResNet-18, which has a fixed input length, the advantage of a Conv1D layer with a transformer is that it can accommodate a wide range of input sizes.

[0085] on the other hand, Figure 6 FIG is a diagram showing the learning steps of a deep learning model according to an embodiment of the present invention. Figure 6 The processor can optimize the parameters of the deep learning model using binary cross entropy with logarithmic loss and an AdamW optimizer with an initial learning rate of 0.0001. In this case, binary cross entropy is a loss function that reduces the difference between the predicted result and the actual answer during the learning process of the deep learning model, and the AdamW optimizer is an algorithm that participates in the update of the actual deep learning model based on this loss function. The processor can apply a sigmoid function to the output of the deep learning model optimized by the above-mentioned binary cross entropy and AdamW optimizer, thereby obtaining a probability value in the range of 0 to 1 for the heart disease category of each first individual heart rate.

[0086] In step 240 , the processor may apply the second individual heart rate obtained from the plurality of electrocardiogram data included in the validation dataset to the trained deep learning model to determine a threshold value for distinguishing the heart disease category.

[0087] More specifically, if Figure 6 As shown, the processor can collect the probability value of each individual heart rate separated from the same electrocardiogram data for each second individual probability value, average the collected probability values ​​of all second individual heart rates and derive a final probability score, and then fine-tune the optimal threshold for distinguishing heart disease categories based on the derived final probability score.

[0088] In this case, the optimal threshold can be obtained by applying a threshold between 0 and 1 in units of 0.01 and achieving the highest F1 score in the inflammation dataset related to cardiac disease categories, i.e., normal sinus rhythm and atrial fibrillation-normal sinus rhythm category or normal sinus rhythm and clinically important arrhythmias-normal sinus rhythm category.

[0089] In this case, the F1 score can be defined by the following formula 1.

[0090] Formula 1

[0091]

[0092] Among them, Precision is Recall

[0093]

[0094] Among them, Precision is the proportion of predicted positives that are accurately classified, which can indicate the accuracy of the detected results, that is, how many actual objects are included in the detection results. Recall represents the proportion of actual positives that are accurately classified, which is an indicator of the degree to which the actual object is captured without being lost. In addition, TruePositive and TruePositive respectively represent samples classified as actual positive and actual negative, while FalsePositive and FalseNegative respectively refer to samples classified as false positive and false negative.

[0095] The processor can store the optimal threshold for the heart disease category together with the weighted value of the deep learning model.

[0096] Figure 7 2 is a flowchart illustrating a method for predicting heart disease using a trained deep learning model according to an embodiment of the present invention.

[0097] use Figure 7 The deep learning model for predicting heart disease is shown in Figure 1 As shown in step 710, the processor receives electrocardiogram data of a patient whose heart disease category needs to be predicted. As shown in step 720, individual heart rates can be obtained from the received electrocardiogram data. In this case, the received electrocardiogram data and the obtained individual heart rates can be compared with Figure 3 The experimental dataset shown corresponds to .

[0098] In step 730 , the processor inputs the acquired individual heart rates into a deep learning model for predicting heart disease of the patient to derive a probability value of predicting a specific heart disease category for each individual heart rate.

[0099] In step 740, the processor may use the probability value derived for each individual heart rate to predict the patient's heart disease. More specifically, the processor compares the average of the probability values ​​derived for all individual heart rates with a threshold value to distinguish the patient's heart disease category.

[0100] As an example, Figure 8 FIG is a diagram illustrating the steps of predicting heart disease using a deep learning model according to an embodiment of the present invention. Figure 8 After loading the weighted values ​​and thresholds of the trained deep learning model, the processor derives the probability value of each individual heart rate. The processor can calculate the average of all the probability values ​​of the individual heart rates derived above.

[0101] As an example, when the average normal sinus rhythm Logit value, which is the probability value for judging as normal sinus rhythm, is greater than the first threshold θ1 and the average clinically important arrhythmia-normal sinus rhythm Logit value, which is the probability value for judging as clinically important arrhythmia-normal sinus rhythm, is less than the second threshold θ2, the processor can judge the corresponding patient's heart disease category as a normal state, that is, normal sinus rhythm.

[0102] Alternatively, when the normal sinus rhythm logit value is greater than a first threshold value θ1 and the clinically important arrhythmia-normal sinus rhythm logit value is greater than a second threshold value θ2, the processor may compare the normal sinus rhythm logit value and the clinically important arrhythmia-normal sinus rhythm logit value and select the larger value as the final prediction. For example, when the clinically important arrhythmia-normal sinus rhythm logit value is greater than the normal sinus rhythm logit value, the patient's heart disease category may be determined to be clinically important arrhythmia-normal sinus rhythm, which indicates a possible arrhythmia.

[0103] Furthermore, when the normal sinus rhythm Logit value is less than a first threshold θ1 and the clinically important arrhythmia-normal sinus rhythm Logit value is less than a second threshold θ2, the processor may compare the normal sinus rhythm Logit value and the clinically important arrhythmia-normal sinus rhythm Logit value to select the larger value as the final prediction.

[0104] Finally, when the normal sinus rhythm Logit value is less than the first threshold θ1 and the clinically important arrhythmia-normal sinus rhythm Logit value is greater than the second threshold θ2, the processor can judge the corresponding patient's heart disease category as clinically important arrhythmia-normal sinus rhythm with the possibility of arrhythmia.

[0105] The four methods for distinguishing between these heart disease categories also apply equally to differentiating between normal sinus rhythm and atrial fibrillation-normal sinus rhythm.

[0106] Below, refer to Figures 9 to 14 , specifically describing a method for using heart rate variability characteristic information, latent vectors, cluster information corresponding to the latent vectors, and metadata, and a method for predicting heart disease by integrating multiple tree model outputs, in accordance with an embodiment of the present invention.

[0107] Figure 9 An exemplary flow chart related to a method for generating a heart disease prediction model according to an embodiment of the present invention is shown.

[0108] According to an embodiment of the present invention, the method for generating a heart disease prediction model may include step S1000 of acquiring a plurality of electrocardiogram data.

[0109] According to an embodiment, multiple ECG data can be obtained from multiple patients and collected in various ways. In one embodiment, the ECG data of the present invention may refer to multi-lead ECG (Electrocardiogram) data related to recording the electrical activity of the heart to evaluate the heart state. Multi-lead ECG data includes cardiac electrical signals obtained through multiple leads (i.e., leads), thereby providing comprehensive information related to heart rhythm, electrical conductivity, and other cardiac functions.

[0110] In an embodiment, the multiple electrocardiogram data obtained from multiple patients may include normal electrocardiogram data, electrocardiogram data related to arrhythmias, and electrocardiogram data related to other cardiac conditions. In a specific embodiment, the electrocardiogram data obtained from the multiple patients can be divided into groups such as normal sinus rhythm (True-Normal Sinus Rhythm, T-NSR), atrial fibrillation-normal sinus rhythm (Afib-Normal Sinus Rhythm, AF-NSR), and clinically important arrhythmia-normal sinus rhythm (Clinically Important Arrhythmia-Normal Sinus Rhythm, CIA-NSR). In this case, the multiple electrocardiogram data obtained from the multiple patients can be 10-second 12-lead electrocardiogram data, but this type of electrocardiogram data is only an example and the present invention is not limited to this.

[0111] In one embodiment, the acquisition of multiple electrocardiogram data may be receiving or loading data stored in the memory 120. The acquisition of multiple electrocardiogram data may be receiving or loading multiple electrocardiogram data from other storage media, other computing devices, or separate processing modules within the same computing device based on a wired or wireless communication unit. As an example, a user (e.g., a patient or medical staff) may access the computing device 100 through a user terminal and receive a user interface for heart disease prediction from the computing device 100, and may transfer the electrocardiogram data to the computing device 100 by dragging and dropping the electrocardiogram data on the provided user interface.

[0112] According to various embodiments, the computing device 100 of the present invention may perform data enhancement on the acquired multiple electrocardiogram data. Specifically, the computing device 100 may perform data enhancement using lead pairs to enhance learning data used to learn a neural network.

[0113] Specifically, ECG data can be acquired by lead. For example, ECG data can be collected in various lead configurations, such as 1-lead, 3-lead, or 12-lead. This allows for a more comprehensive assessment of cardiac status by recording electrical activity in multiple parts of the heart.

[0114] For example, 12-lead signals can acquire many signals in a relatively short period of time, but the total amount of data is limited. Conversely, 1-lead (single-lead) signals, while recorded for a longer period, have limited data diversity. While using these two types of data together can provide a more accurate assessment of cardiac status, in reality, the amount of such paired data is insufficient.

[0115] To address this issue, the computing device 100 of the present invention can generate a reconstruction model that can generate other lead data based on a single lead. The reconstruction model receives single-lead data and generates multi-lead data based on it. In this process, the single-lead data can be expanded to 3 or more leads (12 leads), which has the effect of enhancing the learning data to include more lead data.

[0116] Learning the reconstruction model involves learning the correlation between single-lead and multi-lead data. To this end, the reconstruction model is trained using paired data containing both single-lead and multi-lead data. During the learning process, the model learns how to reconstruct multi-lead data from single-lead data. In this case, it learns the characteristics of the single-lead data and the patterns of the multi-lead signal that can be inferred from it.

[0117] After the model has learned, the reconstructed model can generate multi-lead signals from single-lead data alone, thereby addressing the deficiencies of actual multi-lead data. This increases data diversity and improves the generalization performance of the model. This generated multi-lead data (i.e., enhanced multi-lead data) is used together with existing multi-lead data for neural network learning, allowing the model to learn a more diverse lead structure and improve prediction performance for various cardiac states.

[0118] As a result, data augmentation for model reconstruction substantially increases the amount of learning data, improving the model's predictive accuracy and strengthening the reliability of the heart disease prediction model. This allows medical professionals to more accurately diagnose patients' health conditions and help develop more effective treatment and prevention plans.

[0119] According to an embodiment of the present invention, the method for generating a heart disease prediction model may include step S2000 of constructing a learning dataset based on a plurality of electrocardiogram data.

[0120] In the embodiment, constructing a learning dataset from multiple electrocardiogram data is a process of collecting and structuring various data in order to learn important features and patterns of electrocardiogram data, thereby generating a dataset containing various cases and patterns required for heart disease prediction.

[0121] According to an embodiment, the learning dataset of the present invention includes a variety of data related to the prediction of heart disease, which may include a training dataset for training a neural network model, a validation dataset for evaluating and optimizing model performance, and experimental data for evaluating the generalization performance of the model.

[0122] In one embodiment, the learning dataset may include a first learning dataset and a second learning dataset categorized according to different learning objectives. The first learning dataset and the second learning dataset may be categorized for training different neural network models. The first learning dataset and the second learning dataset may be categorized by the computing device 100.

[0123] For example, the first learning dataset is used to train an embedding model, focusing on extracting and representing key features of electrocardiogram (ECG) data. The second learning dataset is used to train a model for predicting the risk of heart disease, and can be used to calibrate and evaluate the accuracy of the prediction model.

[0124] Specifically, the first learning dataset may include data for learning corresponding to the following process, namely, transforming the characteristics of the electrocardiogram data into a feature space. The first learning dataset is data for electrocardiogram representation learning (ECG Representation Learning), which can be used to extract and embed various characteristics and patterns of the electrocardiogram signal. In addition, the second learning dataset may include data for calibration related to the prediction of cardiac risk level. That is, the second learning dataset, as data for correcting the cardiac risk level, can be used to more accurately match the risk level predicted by the model.

[0125] According to an embodiment, the step of constructing the learning dataset may include the step of performing preprocessing on the plurality of electrocardiogram data.

[0126] Figure 10 The flowchart exemplarily shows a process of pre-processing a plurality of electrocardiogram data according to an embodiment of the present invention.

[0127] Reference Figure 10 The method for preprocessing multiple electrocardiogram (ECG) data may include: step S2100, performing noise preprocessing on the multiple ECG data; step S2200, segmenting the noise-preprocessed multiple ECG data into predefined window sizes to obtain multiple region-of-interest signals; and step S2300, extracting heart rate variability (HRV) characteristic information based on leads of the multiple ECG data. In one embodiment, the heart rate variability characteristic information may include indicator information related to the variability between heart rate intervals.

[0128] In more detail, first, preprocessing is performed to remove noise from the multiple electrocardiogram data obtained in step S2100. Among them, the preprocessing for removing noise includes low-pass filtering and noise removal, which are used to remove the baseline noise of the electrocardiogram signal. Specifically, low-pass filtering improves signal purity by removing high-frequency components from the electrocardiogram data through frequency band limitation, and noise removal uses a timing analysis method to reduce unnecessary noise from the electrocardiogram data to make important signal components clearer. This preprocessing process minimizes the distortion of the electrocardiogram signal and contributes to improving the signal quality for more accurate evaluation of cardiac activity.

[0129] Furthermore, according to an embodiment, the computing device 100 may perform data augmentation to improve the diversity and robustness of model learning. The computing device 100 may perform data augmentation by simulating various noise patterns, thereby enabling the model to learn and adapt to various noise conditions that may occur in real-world environments.

[0130] In a specific embodiment, during the preprocessing of the ECG data, the dataset is enhanced by artificially adding various noises that may occur in actual ECG measurement environments, such as Gaussian noise, jitter noise, random time warping, scaling, amplitude modulation, and phase shifting. This allows the model to learn and adapt to noise that may occur in various environments. This noise addition and removal process has the advantages of improving data quality and enhancing the generalization performance of the model.

[0131] Afterwards, in step S2200, the electrocardiogram data from which noise has been removed is segmented into predefined window sizes and organized into a plurality of region of interest signals. According to an embodiment, the region of interest signals are defined as a range including at least 3 to 5 QRS complexes (QRSComple) (heart rate) as a changeable window size unit including a meaningful heart rate from the electrocardiogram data. In one embodiment, the QRS complex represents the depolarization of the ventricles in the heart's electrical activity, and as a main feature of the electrocardiogram, it helps to identify the start and end of the heart rate. For example, each region of interest signal can be segmented into units from 30 seconds to 1 minute, which is useful for analyzing important patterns in the electrocardiogram signal and identifying the regularity and abnormality of the heart rate.

[0132] The computing device 100 extracts the region of interest signal and organizes it into data that can more accurately analyze the main characteristics of the heart rate. This plays an important role in identifying specific patterns and abnormal symptoms of the electrocardiogram signal and assessing the risk of heart disease.

[0133] Furthermore, in step S2300, the heart rate variability characteristic information extraction process is executed. Heart rate variability characteristic information is an indicator that measures the periodic changes in heart rate over time and is an important signal characteristic reflecting cardiac activity. For example, the computing device 100 divides the electrocardiogram data into 30-second to 1-minute intervals and obtains multiple region-of-interest signals. The computing device 100 then extracts heart rate variability characteristics corresponding to each region-of-interest signal. This allows for more accurate assessment of overall cardiac activity by analyzing data over a longer period than the bit-based analysis time.

[0134] The computing device 100 derives heart rate variability characteristics from each region of interest signal through frequency analysis or time domain analysis. In specific embodiments, the computing device 100 may apply various analysis methods to extract heart rate variability characteristic information corresponding to each region of interest signal. As a specific example, the computing device may extract heart rate variability characteristic information through evaluation of high- and low-frequency components based on spectral analysis and statistical analysis of normal-to-normal intervals.

[0135] Spectral analysis evaluates heart rate variability in the frequency domain. High-frequency components primarily reflect parasympathetic nervous system activity, while low-frequency components indicate a mix of sympathetic and parasympathetic nervous system activity. This helps understand the balance of the autonomic nervous system. Statistical analysis of normal-to-normal intervals quantifies the activity of the cardiac autonomic nervous system by evaluating the variability of heart rate intervals over time. This can be an important indicator for assessing heart rate irregularity, stress response, and heart health.

[0136] In one embodiment, the heart rate variability characteristic information extracted on a lead basis may include successive root mean square differences (RMSSD), standard deviation of successive normal intervals (SDNN), and standard deviation of average normal-to-normal intervals (SDANN). Specifically, successive root mean square differences (RMSSD) represents the square root of the mean square deviation of consecutive heart rate intervals, primarily reflecting the influence of the parasympathetic nervous system. Standard deviation of normal-to-normal intervals represents the standard deviation of all normal heart rate intervals, representing the overall variability of the autonomic nervous system. Standard deviation of average normal-to-normal intervals measures the standard deviation of scored heart rate intervals within a specified period to assess long-term heart rate variability. This indicator plays an important role in evaluating the activity of the heart's autonomic nervous system and comprehensively analyzing heart health. In other words, heart rate variability characteristic information can reflect various aspects of heart rate variability, be used to assess heart rate irregularity, autonomic nervous system activity, and other factors, and is a foundation for predicting the risk of heart disease.

[0137] According to an embodiment of the present invention, the method for generating a heart disease prediction model may include step S3000 of learning at least one network function based on a learning data set to generate a heart disease risk prediction model.

[0138] Figure 11 The flowchart exemplarily shows a process of performing embedding on multiple latent vectors according to an embodiment of the present invention and constructing a vector database based on the execution results.

[0139] Reference Figure 11 The steps of generating a heart disease risk prediction model may include: step S3100, generating an embedding model by using a shielded autonomous learning method of a first learning data set; embedding execution step S3200, using the embedding model to extract multiple latent vectors corresponding to multiple region of interest signals, respectively, and the above-mentioned latent vectors correspond to the above-mentioned second learning data set; step S3300, using a clustering model to perform clustering on the multiple latent vectors; and step S3400, constructing a vector database (DB, DataBase) based on the above-mentioned multiple latent vectors and information related to each cluster corresponding to each latent vector.

[0140] In more detail, in step S3100, the computing device 100 may generate an embedding model based on the first learning dataset.

[0141] The embedding model can be a model trained to represent complex patterns in electrocardiogram data in a low-dimensional space. When multiple region-of-interest signals corresponding to the electrocardiogram data are input, the embedding model can generate a latent vector corresponding to each region-of-interest signal.

[0142] The latent vector can be a low-dimensional vector that compresses and represents the important features of the ECG data. To make the high-dimensional ECG signal easier to understand and analyze, it is transformed into a data representation that represents the signal's primary information. The latent vector can be used to assess cardiac health or provide useful information for detecting disease symptoms based on patterns in the ECG data.

[0143] The detailed instructions for generating the embedding model will be referred to Figure 12 and Figure 13 To explain.

[0144] Figure 12 The following is a flowchart illustrating a process of generating an embedding model according to an embodiment of the present invention. Figure 13 1 is an illustrative diagram for illustrating a self-reconstruction model and a process of generating an embedding model from a trained self-reconstruction model according to an embodiment of the present invention.

[0145] Reference Figure 12 , the step of generating an embedding model may include: step S3110, inputting data contained in a first learning data set into the self-reconstruction model to guide self-supervised learning, so that the self-reconstruction model generates an output similar to the above-mentioned input data; and step S3120, extracting an encoder from the self-reconstruction model that completes the above-mentioned learning and generating the above-mentioned embedding model.

[0146] The self-reconstructing model may be a model that is designed to internally learn important features of data during a learning process through input and output.

[0147] In one embodiment, the self-reconstruction model is a neural network model that masks a portion of the input data and restores the masked portion, and may include an encoder (or a dimensionality reduction network function) and a decoder (or a dimensionality restoration network function). In one embodiment, the self-reconstruction model of the present invention may be a masked autoencoder (MAE) that learns important patterns in electrocardiogram data and accurately restores the masked signal, but is not limited to this.

[0148] The self-reconstruction model can effectively learn important features of the data by actively blocking a portion of the electrocardiogram signal (specifically, the signal of each region of interest) and reconstructing the entire signal based on the remaining portion.

[0149] Self-reconstruction models learn by minimizing the restoration error between input and output, effectively extracting and representing the important characteristics of the signal. In particular, when the data is complex or incomplete, self-reconstruction models have an advantage in understanding and utilizing the inherent structure of the data.

[0150] In more detail, Figure 13 As shown, after reducing the dimensionality of the input data, the self-reconstruction model 800 learns important features by restoring it. The self-reconstruction model may include an encoder 810 (or a dimensionality reduction network function) and a decoder 820 (or a dimensionality restoration network function). The self-reconstruction model 800 transforms high-dimensional data into a low-dimensional latent space through the encoder 810, compressing the core information of the data and reducing noise. Subsequently, the decoder 820 attempts to restore the compressed low-dimensional representation to the original high-dimensional space, thereby learning and reproducing the important patterns of the data.

[0151] According to an embodiment, the computing device 100 can guide the self-reconstruction model 800 to learn using a masked learning approach. In this approach, a portion of the input data is actively masked, allowing the model to recover the masked portion. For example, when unmasked, the self-reconstruction model can focus on learning noise rather than QRS complexes, thereby focusing on learning important features through masking. This allows the model to develop the ability to extract important pattern features from the remaining portion of the input data and accurately recover the masked information. This has the advantage of being able to operate robustly even in noisy or incomplete portions of the data.

[0152] Furthermore, by restoring masked areas, the data context can be understood and key features learned, making it extremely useful for analyzing and interpreting complex medical data such as electrocardiograms. By applying masking, the model can focus on key signals such as the QRS complex, enabling highly reliable data analysis for accurate diagnosis and evaluation of heart disease.

[0153] Specifically, in this invention, to improve reliability during the analysis and interpretation of ECG data, a self-reconstruction model is trained using a masked learning method to focus on important cardiac signals. This allows the model to understand the context of the data and learn key features, thereby accurately grasping and restoring core signals such as the QRS complex.

[0154] The self-reconstruction model aims to reproduce the entire signal. After masking a specific, significant range of the signal, it focuses on restoring that masked portion. For example, after masking a significant portion of the entire ECG signal, the masked region contains approximately three QRS complexes. The self-reconstruction model focuses on learning this region, calculating the loss function only for that region. This approach allows the model to learn the characteristics of the data, focusing on the important signals, thereby more accurately understanding and reproducing the entire context of the data.

[0155] As a result, it can operate robustly even with potential noise and incompleteness in ECG data, enabling highly reliable data analysis necessary for early diagnosis and evaluation of heart disease. This access method minimizes the loss of important information during ECG signal analysis, supporting accurate assessment of cardiac status and significantly improving medical data analysis.

[0156] According to one embodiment, the self-reconstruction model 800 may include at least one hidden layer, and an odd number of hidden layers may be configured between the input and output layers. The number of nodes in each layer is reduced from the number of nodes in the input layer to the middle layer of the bottleneck layer (encoding), and then reduced from the bottleneck layer and symmetrically extended to the output layer (symmetrical with the input layer). In this case, Figure 13 The layers of the encoder 810 and the layers of the decoder 820 are shown to be symmetrical, but the present invention is not limited to this. The nodes of the layers of the encoder 810 and the layers of the decoder 820 can be symmetrical and can also be asymmetrical. The self-reconstruction model 800 can perform nonlinear dimensionality reduction. The number of input layers and output layers can correspond to the number of items of input data remaining after preprocessing the input data. In the structure of the self-reconstruction model 800, the number of nodes of the hidden layer included in the encoder 810 can gradually decrease in the direction away from the input layer. When the number of nodes in the bottleneck layer (the layer with the least nodes located between the encoder and the decoder) is too small, a sufficient amount of information cannot be transmitted, and therefore, a specific number or more (for example, more than half of the input layer, etc.) can be maintained.

[0157] In an embodiment, computing device 100 trains a self-reconstruction model using multiple region-of-interest signals corresponding to a first learning dataset, thereby learning important features and patterns in the electrocardiogram (ECG) data. The trained self-reconstruction model is applied to each of the multiple ECG data sets, thereby extracting and analyzing the key features of the ECG signal by masking and restoring a portion of the input data.

[0158] In summary, the computing device 100 guides the self-reconstruction model through a masked learning approach to ensure high reliability during the analysis and interpretation of ECG data. During this process, the self-reconstruction model focuses on key signals, such as the QRS complex, to learn the inherent distribution of the heart rate. While the self-reconstruction model reproduces the original information of the ECG data, it effectively learns important features (or characteristics) inherent in the ECG data, such as the peak value of the QRS complex.

[0159] Then, in step S3120, the encoder is extracted from the learned self-reconstruction model and an embedding model is generated. In the trained self-reconstruction model, the encoder portion is trained (i.e., pre-learned using the first learning dataset) to compress important features of high-dimensional data and remove noise, thereby outputting a low-dimensional latent vector corresponding to the input data (e.g., multiple region of interest signals included in the second learning dataset).

[0160] According to embodiments, the generated embedding model, because it represents the characteristics of compressed data, can effectively reflect the similarity between the input data in the latent space. For example, corresponding to similar input data, the embedding model outputs latent vectors on similar latent spaces. For example, a second latent vector for a second ROI signal similar to a first ROI signal is located on a latent space similar to the first latent vector corresponding to the first ROI signal. In other words, they are located at similar positions in the latent space.

[0161] In contrast, when the input data is dissimilar (i.e., significantly different), the embedding model places the data at locations farther apart in the latent space. This allows for clear distinction between similarities and differences between data based on the characteristic patterns of ECG data, which can be used in the diagnosis and evaluation of heart disease.

[0162] That is, the present invention uses the first learning data set to guide the autonomous learning of the self-reconstruction model, thereby completing the learned self-reconstruction model to extract the encoder to form the embedding model.

[0163] After using the first learning dataset to generate an embedding model, the computing device 100 can use the embedding model to perform embedding to extract multiple latent vectors corresponding to each region of interest signal, where the multiple region of interest signals correspond to the second learning dataset.

[0164] More specifically, each ROI signal included in the second learning dataset is input into the embedding model via computing device 100. The embedding model compresses the key features of the corresponding signal to output a low-dimensional latent vector. The latent vector simply represents the complex pattern of the ROI signal, reflecting the similarities and differences between the ROI signals in the latent space. For example, signals with similar patterns in the ROI signals of the electrocardiogram data included in the second learning dataset are output as multiple latent vectors located close to each other in the latent space.

[0165] Thereafter, in step S3300 , the computing device 100 may perform clustering on the plurality of latent vectors using a clustering model.

[0166] According to an embodiment, the computing device 100 may use a clustering model to perform clustering on the plurality of latent vectors.

[0167] As an example, the clustering model may be a model trained to perform clustering using a plurality of latent vectors extracted by the embedding model, and may be a K-Nearest Neighbors (KNN) model.

[0168] In one embodiment, the clustering model may perform clustering based on similarity distances between a plurality of latent vectors corresponding to each ROI signal.

[0169] More specifically, during the clustering process, the distance or similarity between potential vectors is calculated to determine the proximity or similarity of these vectors. Typically, the distance is calculated using a measurement method such as Euclidean distance or cosine similarity.

[0170] For example, the distance between latent vector A extracted from a particular region of interest signal and latent vector B extracted from other regions of interest signals is calculated to evaluate the similarity between the two vectors. When the distance between A and B is close, the two vectors may represent ECG signals containing the same or similar patterns. The clustering model groups multiple latent vectors into clusters based on this distance information.

[0171] In the case of the KNN (K-Nearest Neighbor) model, clustering is performed by finding the K nearest neighboring vectors of a particular vector and including the corresponding vector in the cluster to which these vectors belong. For example, when K = 3, the three nearest neighboring vectors of a latent vector A are found and assigned to the cluster containing these three neighboring vectors. This process is repeated for all latent vectors, and eventually, vectors sharing similar characteristics are grouped into the same cluster.

[0172] This process is useful for classifying multiple patterns in electrocardiogram data and analyzing how each pattern correlates with a specific type or risk level of heart disease. For example, if a particular cluster contains information related to irregularities in heart rate, that cluster can be used as an indicator of a condition such as arrhythmia.

[0173] As a specific example, when clustering is performed on each latent vector using a clustering model, four clusters may be formed. In the above clustering process, multiple vectors are clustered based on the similarity between the latent vectors.

[0174] Assuming there are four clusters, namely, a first cluster, a second cluster, a third cluster, and a fourth cluster, these clusters can each represent a different ECG pattern. For example, the first cluster may include signals with a normal heart rate pattern, while the second cluster may include signals with a specific abnormality such as an arrhythmia. The third and fourth clusters may each include signals reflecting other types of heart diseases or conditions. While the above description describes the formation of four clusters (or groups), this is merely an example; the actual number of clusters may vary depending on the characteristics of the data and the settings of the clustering model.

[0175] This clustering automatically identifies specific patterns or anomalies in ECG data and analyzes the characteristics of each cluster to provide information useful for early diagnosis and management of heart disease. The clustering results provide valuable data for medical professionals to comprehensively evaluate patient conditions and develop customized treatment plans.

[0176] Furthermore, in step S3400 , the computing device 100 may construct a vector database based on information related to the plurality of latent vectors and the clusters corresponding to the latent vectors.

[0177] The vector database of the present invention may be a NoSQL database server for storing and managing information related to embedded vectors, their cluster information, and distances between vectors.

[0178] The vector database of the present invention is optimized for analyzing and managing multiple potential vectors extracted from multiple electrocardiogram data. Thus, characteristic patterns of electrocardiogram signals can be systematically stored and searched for when needed.

[0179] As an example, the vector database of the present invention can support risk assessment and condition monitoring based on ECG data. When newly collected ECG data from a specific patient is received, the latent vector extracted from the data can be compared with existing vectors stored in the vector database to evaluate similarity. This similarity evaluation can determine which cluster the patient's ECG data belongs to and which previously observed patterns the corresponding cluster resembles.

[0180] Furthermore, in an embodiment, the computing device 100 may extract representative vectors associated with the plurality of clusters corresponding to the clustering result, i.e., principal vectors. Such principal vectors are vectors representing each cluster, which may represent the center of the plurality of latent vectors within the cluster.

[0181] More specifically, when the data stored in a vector database is massive, for example, searching through 300,000 pieces of ECG data, it can be very time-consuming. Typically, when new ECG data for a specific patient is input, it is segmented into multiple region-of-interest signals during preprocessing, and 80 to 120 embedding vectors are generated for each region-of-interest signal. For example, individually comparing the new ECG data against all the vectors stored in the database is computationally intensive and inefficient. This process is time-consuming and unsuitable for real-time analysis.

[0182] To shorten search time and improve efficiency, multiple latent vectors generated from the patient's ECG data can be compared with the principal vector of each cluster to quickly identify similar clusters. During this process, the inner product between each embedding vector and the principal vector is calculated to evaluate similarity, and this information is used to determine which cluster the corresponding ECG data belongs to.

[0183] Rather than comparing the entire data one by one, this method compares the principal vectors representing each cluster, allowing for the rapid identification of relevant clusters. This results in a significant contribution to the rapid analysis and effective management of ECG data, the rapid assessment of heart disease risk, and real-time monitoring of patient status.

[0184] Furthermore, in the embodiment, heart disease prediction information related to electrocardiogram data of a prediction target may be generated through the following process.

[0185] In an embodiment, a method for generating heart disease prediction information related to electrocardiogram data of a prediction object may include: acquiring the electrocardiogram data of the prediction object; performing preprocessing on the electrocardiogram data of the prediction object; using an embedding model to generate multiple latent vectors corresponding to the preprocessed electrocardiogram data; respectively comparing each latent vector with a representative vector corresponding to each cluster to classify each latent vector into one of each of the above-mentioned clusters; and generating hierarchical information related to the risk level of heart disease based on the classification results of each latent vector into multiple clusters.

[0186] In an embodiment, the subject of the prediction may be an individual receiving an electrocardiogram (ECG) for the purpose of assessing and managing the risk of heart disease. This may be a person undergoing a general health checkup, a person with a family history of heart disease, or a patient previously diagnosed with heart disease or exhibiting related symptoms. The subject of the prediction may continuously assess their heart status through regular ECG monitoring and utilize the computing device 100 of the present invention for the early detection and prevention of heart disease.

[0187] More specifically, after acquiring the ECG data of the prediction target, the data is preprocessed to remove noise and purify the signal. As described above, preprocessing may include the process of removing unnecessary signal components and emphasizing important signal features in order to improve the quality of the ECG data, and may also include the process of dividing each ECG signal into appropriate units (predefined window sizes) for processing. Typically, the predefined window size can be defined as an interval containing important information in the ECG data, for example, a length of about 30 seconds to about 1 minute. In this process, the ECG data can be divided into specified time intervals, and each interval can be analyzed independently.

[0188] The preprocessed ECG data is input into an embedding model (e.g., an encoder extracted from an autonomous regenerative model pre-learned using a first learning dataset), thereby generating a latent vector (or embedding vector) for each data point. The generated multiple latent vectors are compared with existing vectors stored in a vector database, and each vector is classified into a specific cluster based on similarity. Each cluster has a specific pattern that reflects the heart condition, so the risk of heart disease is assessed based on which cluster the latent vector belongs to.

[0189] Furthermore, computing device 100 can generate hierarchical information related to the risk level of heart disease based on the classification results of multiple latent vectors into multiple clusters. More specifically, during preprocessing, multiple region-of-interest signals are segmented and acquired at specific time units for the predicted electrocardiogram data. As each region-of-interest signal is input into the embedding model, multiple latent vectors associated with each region-of-interest signal are generated. In this case, predictive information related to the risk level of heart disease can be generated based on the cluster into which each latent vector is classified and the number of clusters.

[0190] As a specific example, if the ECG data for the prediction target consists of ten 30-second region-of-interest signals, the corresponding 10 region-of-interest signals can be transformed into ten latent vectors using the embedding model. These 10 latent vectors can then be classified into one of four predefined clusters. Each predefined cluster represents a specific cardiac state, with the first cluster representing the highest risk level, and the fourth cluster representing the lowest risk level.

[0191] The computing device 100 can predict the overall risk level based on how the 10 input signals are classified into clusters.

[0192] For example, if eight signals belong to the first cluster and two to the second, the patient's risk of heart disease can be assessed as high. Conversely, if five signals belong to the second cluster, three to the third cluster, and two to the fourth cluster, the risk can be assessed as moderate. If two signals belong to the third cluster and eight to the fourth cluster, the risk can be assessed as low. If all ten signals belong to the fourth cluster, the patient's risk of heart disease can be assessed as extremely low.

[0193] Specifically, the computing device 100 of the present invention can use an embedding model to analyze signals in regions of interest (ROIs) within electrocardiogram (ECG) data to obtain multiple latent vectors. Each latent vector is evaluated based on the learning dataset to determine which of multiple clusters predefined by the clustering model it belongs to, thereby enabling a hierarchical prediction of a patient's risk of heart disease. In other words, the computing device 100 compares the multiple latent vectors generated by the embedding model with predefined clusters to determine which cluster each vector belongs to, and uses this information to perform a multidimensional assessment of the patient's heart condition.

[0194] This stratified risk level prediction method improves the accuracy of ECG data analysis and facilitates a comprehensive understanding of a patient's heart condition. Thus, the calculation device of the present invention systematically and clearly stratifies and provides a patient's heart disease risk level, thereby providing crucial information for medical professionals to take appropriate preventive measures or establish treatment plans. This enables early detection and appropriate treatment of heart disease, thereby contributing to patient health management.

[0195] Figure 14 The flowchart is an exemplary flowchart for illustrating a process of generating a heart disease risk prediction model through ensemble learning of multiple tree series models according to an embodiment of the present invention.

[0196] Reference Figure 14 The steps of generating a heart disease risk prediction model may include: step S3500, obtaining user meta-information corresponding to multiple electrocardiogram data respectively; step S3600, obtaining multiple latent vectors and cluster information corresponding to the multiple latent vectors from a vector database; and step S3700, training multiple tree models based on heart rate variability characteristic information, user meta-information, multiple latent vectors and cluster information corresponding to the latent vectors, and performing integrated learning to merge the outputs of each tree model to generate a heart disease risk prediction model.

[0197] More specifically, in step S3500, obtaining user meta-information may include collecting personal health information such as the patient's age, gender, medical history, and lifestyle. This information reflects each patient's health status and characteristics and significantly impacts the interpretation of ECG data. In particular, this meta-information is a crucial variable for more accurately assessing the risk of heart disease. This allows for customized predictions and assessments based on individual patient health factors.

[0198] In step S3600, multiple latent vectors and the cluster information corresponding to each vector are retrieved from a vector database. A latent vector summarizes the characteristics of the ECG data, while the cluster information represents the specific cardiac condition or disease pattern to which the corresponding vector belongs. This allows identification of the cluster to which the patient's ECG data belongs and assessment of the risk of cardiac disease indicated by the cluster.

[0199] In step S3700, multiple tree models are trained based on the heart rate variability characteristic information, user meta-information, latent vectors, and cluster information. Each tree model uses a variety of tree-based machine learning algorithms such as decision trees, random forests, and gradient boosting trees to predict the risk of heart disease.

[0200] Cluster information can be obtained from a vector database and may include information about the cluster to which each vector belongs and information related to the distances between vectors. For example, cluster information provides information about which cluster each potential vector belongs to and distance information indicating the similarities or differences between multiple vectors within a cluster. This information plays an important role in assessing the cluster to which a specific ECG pattern belongs and the similarity of that pattern to other ECG data. This allows for analysis of the characteristics of ECG data and accurately determines whether the corresponding data is associated with a specific disease or condition.

[0201] More specifically, the process of training multiple tree series models based on heart rate variability characteristic information, user meta information, latent vectors and cluster information is as follows.

[0202] First, heart rate variability characteristic information, an index of heart rate variability extracted from electrocardiogram data, reflects the activity of the cardiac autonomic nervous system. User meta-information represents a patient's individual health status, including age, gender, medical history, and lifestyle. Latent vectors represent the characteristics of electrocardiogram data in a low-dimensional space, and cluster information indicates which cardiac condition or disease the vector is associated with.

[0203] This diverse information serves as input for training multiple tree-based models. For example, the random forest model uses this input to construct multiple decision trees, combining the results of each tree to provide a final prediction. Gradient boosted trees use sequential model learning to correct for errors in previous models at each step, aiming for more accurate predictions.

[0204] Cluster information is particularly important in this context. The distance between each latent vector's cluster and its neighbors provides a deeper understanding of the characteristics of ECG data and can be used to assess the risk of heart disease. This information plays a crucial role in decision-making in various tree models, which use it to learn data patterns.

[0205] In the final ensemble, the predictions of the individual tree models are combined to derive a single, comprehensive forecast. During this process, the weights of the individual models are adjusted to improve the reliability and accuracy of the final forecast. The ensemble process mitigates the weaknesses of individual models and combines their strengths to provide a more robust and consistent forecast.

[0206] As a result, this comprehensive learning and integration process can reflect the multidimensional characteristics of ECG data to accurately and precisely predict the risk level of heart disease.

[0207] Afterward, ensemble learning is performed to combine the outputs of each tree model. This combines the predictions of the individual models to improve the accuracy and reliability of the final prediction. Ensemble learning allows for more precise predictions of the risk of heart disease, enabling comprehensive assessments of a patient's health status and the development of early diagnosis and preventive treatment plans.

[0208] Specifically, the present invention combines the advantages of multiple models through ensemble learning of multiple tree models, achieving higher predictive performance than a single model and improving robustness to data noise or variation. Thus, the model can provide consistent predictions across multiple patient data sets and reduce prediction uncertainty.

[0209] In summary, the computing device of the present invention utilizes three neural network models to more accurately predict the risk of heart disease. First, the embedding model transforms electrocardiogram (ECG) data into low-dimensional latent vectors to summarize important signal features. Second, the clustering model clusters multiple latent vectors to determine the cluster to which each vector belongs. Finally, the heart disease risk prediction model combines heart rate variability characteristic information, user meta-information, latent vectors, and cluster information to predict the risk of heart disease. Through this combined access method, the computing device can more accurately and consistently predict the risk of heart disease.

[0210] Specifically, the final heart disease risk prediction process utilizes multiple tree models based on four variables: heart rate variability characteristics, user meta-information, latent vectors, and cluster information. Each model generates its own output. These outputs are combined through ensemble learning to generate the final output: the predicted heart disease risk level. This structure improves the accuracy of heart disease risk prediction, achieving a multidimensional and comprehensive assessment by combining multiple variables. This plays a crucial role in more accurately reflecting individual patients' health status, enabling healthcare professionals to clearly understand their risk level and develop optimal treatment plans.

[0211] The heart disease risk prediction model generated through final learning can stratify patients based on their ECG data and provide predictive information related to their future heart disease risk. This stratified information plays a key role in multidimensionally evaluating a patient's heart health and developing personalized treatment plans. This allows medical professionals to more clearly understand a patient's risk level, enabling more precise treatment and preventive measures, thereby facilitating early detection and management of heart disease.

[0212] Specifically, the system comprehensively analyzes patient meta-information from electrocardiogram (ECG) data to categorize the likelihood of heart disease into multiple risk levels. For example, the system stratifies risk levels into low, medium, and high, allowing medical staff to clearly understand which patients need to focus their attention. This increases the likelihood that patients in high-risk groups will receive early diagnosis and treatment, while patients in low-risk groups can reduce unnecessary testing and receive effective health management. Consequently, this system plays a significant role in achieving efficient allocation of medical resources, preventing complications caused by heart disease, and improving patients' quality of life.

[0213] The embodiments described above can be implemented using hardware structural elements, software structural elements, and / or a combination of hardware structural elements and software structural elements. For example, the devices, methods, and structural elements described in various embodiments can be implemented using a conventional computer or special-purpose computer such as a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor (DSP), a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device that executes instructions and can respond. The processor device can run an operating system (OS) and software applications executed on the operating system. Furthermore, the processing device can access, store, operate, process, and generate data in response to the execution of the software. For ease of understanding, there are descriptions of using a single processing device. Those skilled in the art will understand that the processing device includes multiple processing elements and / or multiple types of processing elements. For example, the processing device may include multiple processors or a processor and a controller. Furthermore, other processing configurations such as parallel processors may also be employed.

[0214] Software may include a computer program, code, instructions, or a combination of at least one of these, which configures a processing device or, independently or collectively, instructs a processing device to perform the desired operation. Software and / or data may be stored on any type of machine, component, physical device, virtual equipment, computer storage medium, or device for interpretation by a processing device or for providing instructions or data to a processing device. Software may be distributed across computer systems connected via a network and stored or executed using a distributed method. Software and data may be stored on a computer-readable recording medium.

[0215] The method of the embodiment is implemented in the form of program instructions that can be run by a variety of computer units and recorded on a computer-readable medium. The computer-readable medium can store program instructions, data files, data structures, etc. alone or in combination, and the program instructions recorded on the medium can be specially designed or constructed for the embodiment or commonly used by ordinary technicians in the field of computer software. As an example of a computer-readable recording medium, it includes magnetic media (magnetic media) such as hard disks, floppy disks, and tapes, optical recording media (optical media) such as CD-ROMs and DVDs, magneto-optical media (magneto-optical media) such as floppy disks (floptical disks), and read-only memories (ROMs), random access memories (RAMs), flash memories, etc., which are specially constructed hardware devices for storing and running program instructions. As an example of program instructions, it includes mechanical codes formed by compilers and high-level language codes that are run by computers using interpreters, etc.

[0216] The hardware devices described above may be operated as one or more software modules to perform the tasks of the embodiments, and vice versa.

[0217] As described above, although the embodiments are described with limited reference to the accompanying drawings, a person skilled in the art may employ various technical modifications and variations based on the embodiments. For example, the described techniques may be performed in a different order than described, and / or the described systems, structures, devices, circuits, and other structural elements may be combined or combined in a different manner than described, and even if replaced or substituted with other structural elements or equivalent technical solutions, appropriate results may be achieved.

[0218] Therefore, these examples, other embodiments and equivalents to the scope of protection of the invention are within the scope of protection of the invention described later.

Claims

1. A deep learning model learning method for predicting heart disease, executed by a processor included in a computing device, characterized in that: The steps include: The processor obtains a plurality of discontinuous individual heart rates from a plurality of electrocardiogram data, wherein the plurality of electrocardiogram data is non-standardized data obtained from a plurality of patients and constituted by continuous signals; as well as As the plurality of acquired individual heart rates are input into a deep learning model for predicting a patient's heart disease, the processor trains the deep learning model to predict a heart disease category for each of the plurality of acquired individual heart rates.

2. The deep learning model learning method for predicting heart disease according to claim 1, characterized in that: The step of training the deep learning model includes the following steps: using the first individual heart rate obtained from at least one electrocardiogram data among the multiple individual heart rates obtained to train the deep learning model.

3. The deep learning model learning method for predicting heart disease according to claim 1, characterized in that: The step of obtaining the plurality of discontinuous individual heart rates comprises the following steps: Filtering the plurality of electrocardiogram data respectively to remove noise; and A peak detection algorithm is applied to each of the plurality of electrocardiogram data from which the noise is removed to subdivide the plurality of electrocardiogram data into a plurality of individual heart rates.

4. The deep learning model learning method for predicting heart disease according to claim 1, characterized in that: The step of training the deep learning model includes the following steps: using binary cross entropy and an optimizer to optimize the parameters of the deep learning model.

5. The deep learning model learning method for predicting heart disease according to claim 1, characterized in that: The method further includes the step of applying, by the processor, a second individual heart rate obtained from at least one electrocardiogram data among the obtained plurality of individual heart rates to the trained deep learning model to determine a threshold value for distinguishing the heart disease category.

6. The deep learning model learning method for predicting heart disease according to claim 5, characterized in that: The step of determining a threshold value for distinguishing the heart disease categories comprises the following steps: collecting, for each of the second individual heart rates, a probability score associated with each individual heart rate isolated from the same electrocardiogram data; averaging the probability scores of each collected second individual heart rate to derive a final probability score; as well as A threshold value for distinguishing the cardiac disease categories is fine-tuned based on the derived final probability score.

7. The deep learning model learning method for predicting heart disease according to claim 5, characterized in that: The determined threshold is stored together with the weighted value of the deep learning model.

8. A method for predicting heart disease using a deep learning model, executed by a processor included in a computing device, characterized in that: The steps include: The processor receives electrocardiogram data, wherein the electrocardiogram data is non-standardized data consisting of continuous signals for a patient; acquiring a plurality of discrete individual heart rates from the acquired electrocardiogram data; deriving, for each of the plurality of individual heart rates obtained as a result of inputting the plurality of individual heart rates obtained into a deep learning model, a probability value of being predicted as belonging to a specific heart disease category; as well as predicting the patient's heart disease using the derived plurality of probability values, The deep learning model is a model that performs learning using a plurality of individual heart rates obtained from a plurality of electrocardiogram data of a plurality of patients.

9. The heart disease prediction method using a deep learning model according to claim 8, characterized in that: The step of predicting the patient's heart disease comprises the steps of comparing the average of the derived probability values ​​with a threshold value to distinguish the patient's heart disease category, The threshold is determined by applying a plurality of individual heart rates obtained from at least one electrocardiogram data among a plurality of electrocardiogram data of the plurality of patients to the deep learning model, and the threshold is a value for distinguishing the heart disease category.

10. The heart disease prediction method using a deep learning model according to claim 8, characterized in that: The step of obtaining the plurality of discontinuous individual heart rates comprises the following steps: filtering the received electrocardiogram data to remove noise; and A peak detection algorithm is applied to the electrocardiogram data from which the noise is removed to subdivide the electrocardiogram data into a plurality of individual heart rates.

11. The heart disease prediction method using a deep learning model according to claim 8, characterized in that: The deep learning model is in a state of optimizing parameters based on binary cross entropy and optimizer.

12. A computing device, characterized in that: include: processor; Network interface; Memory; and A computer program, loaded into the memory and executed by the processor, The processor performs the method according to claim 1 or 8 by executing at least one instruction included in the computer program.

13. A computer program, characterized in that In combination with a computing device, the method according to claim 1 or 8 is stored in a computer readable recording medium.

Images