Artificial intelligence-based device and method for predicting cardiac disease risk

Abstract

An artificial intelligence-based cardiac disease risk prediction device and method according to one embodiment are disclosed. The artificial intelligence-based cardiac disease risk prediction device according to one embodiment comprises at least one processor and at least one memory for storing commands executed by the at least one processor, wherein the at least one processor inputs at least one of first input data and second input data of a subject to a learning model and predicts the subject's cardiac disease risk based on the results output by the learning model.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to an artificial intelligence-based heart disease risk prediction device and method. More specifically, it relates to a technology for informing the risk of occurrence of paroxysmal atrial fibrillation and the risk of new atrial fibrillation occurring in the near future using measurement results such as electrocardiograms and / or chest X-rays.

[0002] [Cross-reference to related applications] This application claims priority based on Korean Patent Application No. 10-2023-0041988, filed on March 30, 2023, and the entire specification thereof is incorporated into this application.

Background Art

[0003] Deep learning models based on artificial neural network structures can be trained to evaluate the risks of various heart diseases by analyzing digital time-series data obtained by digitizing electrocardiogram waveforms and image data obtained by imaging them. Therefore, in recent years, technologies for applying such artificial intelligence systems to various areas of heart diseases have been developed.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Heart diseases, such as paroxysmal atrial fibrillation, are diseases in which atrial fibrillation does not occur normally but occurs intermittently. Therefore, there is a possibility of missing paroxysmal atrial fibrillation unless it is monitored with an electrocardiogram device over a long period of time. However, such monitoring work is costly, and patients also need to wear a monitoring device for a long time, which is very inconvenient.

[0005] In recent years, many wearable, compact electrocardiogram devices have been developed and sold to alleviate these inconveniences. However, these devices are also expensive, require prolonged wear, and can occasionally cause severe skin allergies, making consistent application to all patients difficult. Therefore, a separate high-risk group cleaning method is needed to selectively apply such monitoring only to high-risk patients.

[0006] Embodiments of this disclosure provide an artificial intelligence-based cardiac disease risk prediction device and method that can detect high-risk groups for cardiac diseases such as paroxysmal atrial fibrillation by utilizing input information that can be obtained in a short time (less than 1 minute), such as chest X-ray or electrocardiogram (single, non-monitoring), without performing long-term electrocardiogram monitoring. [Means for solving the problem]

[0007] An artificial intelligence-based cardiac disease risk prediction device according to one embodiment comprises at least one processor and at least one memory for storing commands executed by the at least one processor, wherein the at least one processor inputs at least one of a subject's first input data and second input data into a learning model and predicts the subject's cardiac disease risk based on the results output by the learning model.

[0008] The learning model includes an encoder unit configured to convert input data, including at least one of the first input data and the second input data, into a vector, and a task unit configured to assign a score to the input data to predict the probability that the subject has heart disease.

[0009] The task unit reflects the additional information about the subject in the raw score of the input data. The system may be configured to calculate a corrected score and predict the probability that the subject has heart disease based on the corrected score.

[0010] The additional information may be information entered by the subject, or it may be generated by preprocessing recorded information that has been stored in advance by the encoder unit.

[0011] The task unit may include a first task unit configured to calculate a first individual score which is corrected by reflecting the additional information in a first vector for the first input data, and a second task unit configured to calculate a second individual score which is corrected by reflecting the additional information in a second vector for the second input data.

[0012] The task unit may be configured to calculate an integrated score that is corrected by reflecting the additional information in the entire vector generated based on the first vector for the first input data and the second vector for the second input data.

[0013] The learning model is trained based on constructed training data, which may be constructed by labeling acquired input data with a first label, a second label, or a pseudo-label.

[0014] The aforementioned training data may be constructed by assigning a first label to the input data if a heart disease has already been diagnosed at a first point in time in the patient's medical history that forms the basis of the acquired input data, and by assigning the first label to the input data if abnormal signs corresponding to the heart disease are detected at a second point in time in the acquired input data.

[0015] The aforementioned training data may be constructed by assigning pseudo-labels to the input data if the acquired input data is interrupted and measured before a second time point, and no abnormal signs are detected in the input data up to a third time point before the second time point.

[0016] The training data may be constructed by assigning a second label to the input data if the acquired input data is measured from a second time point up to a fourth time point, and no abnormal signs are detected in the input data up to the fourth time point.

[0017] The first input data is electrocardiogram (ECG) examination information, the second input data is chest X-ray examination information, and the cardiac disease may include atrial fibrillation.

[0018] According to one embodiment, an artificial intelligence-based cardiac disease risk prediction method is performed by an artificial intelligence-based cardiac disease risk prediction device comprising at least one processor and at least one memory for storing commands executed by the at least one processor, and comprises the steps of inputting at least one of first input data and second input data of a subject into a learning model, and predicting the cardiac disease risk of the subject based on the results output by the learning model.

[0019] The learning model includes an encoder unit configured to convert input data, including at least one of the first input data and the second input data, into a vector, and a task unit configured to assign a score to the input data to predict the probability that the subject has heart disease.

[0020] The task unit may be configured to adjust the raw score of the input data to reflect additional information about the subject, calculate a corrected score, and predict the probability that the subject has heart disease based on the corrected score.

[0021] The additional information may be information entered by the subject, or it may be generated by preprocessing recorded information that has been stored in advance by the encoder unit.

[0022] The task unit may include a first task unit configured to calculate a first individual score obtained by reflecting the additional information in a first vector for the first input data, and a second task unit configured to calculate a second individual score obtained by reflecting the additional information in a second vector for the second input data.

[0023] The task unit may be configured to calculate an integrated score obtained by reflecting the additional information in the entire vector generated based on the first vector for the first input data and the second vector for the second input data.

[0024] The learning model is learned based on the constructed learning data, and the learning data may be constructed by labeling the acquired input data with any one of a first label, a second label, and a pseudo label.

[0025] The learning data is constructed by assigning a first label to the input data when a heart disease has already been diagnosed at a first time point in the medical history of the patient from whom the acquired input data is derived, and may also be constructed by assigning a first label to the input data when an abnormal sign corresponding to a heart disease is detected at a second time point in the acquired input data.

[0026] The learning data may be constructed by assigning a pseudo label to the input data when the acquired input data is measured after being interrupted before a second time point and no abnormal sign is detected in the input data until a third time point before the second time point.

[0027] The learning data may be constructed by assigning a second label to the input data when the acquired input data is measured until a fourth time point after the second time point and no abnormal sign is detected in the input data until the fourth time point.

[0028] The first input data is electrocardiogram examination information, the second input data is chest X-ray examination information, and the heart disease may include atrial fibrillation.

Advantages of the Invention

[0029] When selecting a high-risk group of heart diseases such as paroxysmal atrial fibrillation according to the embodiments of the present disclosure, electrocardiogram monitoring can be selectively applied, which can reduce medical costs and improve patient convenience.

[0030] In addition, according to the embodiments of the present disclosure, insufficient medical data can be constructed as learning data by pseudo-labeling or the like, so that the learning model can be efficiently learned even in an environment where the input data is limited.

Brief Description of the Drawings

[0031] [Figure 1] It is a block diagram for explaining an apparatus for predicting a heart disease risk based on artificial intelligence according to an embodiment. [Figure 2] It is a diagram for explaining the architecture of a learning model according to an embodiment. [Figure 3] It is a diagram for explaining an example of an inference process executed by the learning model of FIG. 2. [Figure 4] It is a diagram for explaining another example of an inference process executed by the learning model of FIG. 2. [Figure 5] It is an exemplary diagram for explaining the architecture and inference process of an additional learning model according to an embodiment. [Figure 6] It is an exemplary diagram for explaining the architecture and inference process of an additional learning model according to an embodiment. [Figure 7] It is an exemplary diagram for explaining the architecture and inference process of an additional learning model according to an embodiment. [Figure 8] It is a diagram for explaining an algorithm for constructing learning data according to an embodiment. [Figure 9]This is a flowchart illustrating an artificial intelligence-based method for predicting the risk of cardiac disease according to one embodiment. [Modes for carrying out the invention]

[0032] The following description is provided to aid in a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, this is merely illustrative and the invention is not limited thereto.

[0033] In describing one embodiment, if a specific description of prior art related to the present invention is deemed to obscure the gist of the embodiment, such detailed description will be omitted.

[0034] The terms described below are defined in consideration of the function of the present invention, and may differ depending on the intent or conventions of the user or operator. Therefore, their definitions should be based on the content throughout this specification. The terms used in the detailed description of the invention are merely for the purpose of describing one embodiment and should not be used to limit it.

[0035] Unless otherwise specified, singular expressions imply the meaning of the plural form. Terms such as "first," "second," etc., are used to distinguish various components and are not necessarily limited by their meaning.

[0036] In this specification, the concept of cardiac disease includes atrial fibrillation.

[0037] In this specification, "learning model" refers to a learning model that predicts the risk of cardiac diseases such as atrial fibrillation.

[0038] In this specification, atrial fibrillation means atrial fibrillation or atrial fibrillation, and atrial flutter.

[0039] On the other hand, the apparatus of the present invention may be entirely hardware, or it may have aspects that are partly hardware and partly software. For example, a unit may refer to an apparatus for transmitting and receiving data of a specific format and content using an electronic communication method, and the software associated therewith.

[0040] In this specification, terms such as “part,” “module,” “server,” “system,” “device,” or “terminal” (e.g., encoder part, task part) refer to a combination of hardware and software driven by that hardware. For example, here, hardware may be a data processing device including a CPU or other processor. Software driven by hardware may be a running process, object, executable file, or thread of execution. It can also refer to a ution, a program, etc.

[0041] In this specification, the encoder unit may be referred to as an encoder or encoding execution processor. Similarly, the task unit may be referred to as a task execution processor.

[0042] In this specification, "before" may include both the present and the past.

[0043] Figure 1 is a block diagram illustrating an artificial intelligence-based cardiac disease risk prediction device 10 according to one embodiment.

[0044] Referring to Figure 1, the artificial intelligence-based cardiac disease risk prediction device 10 includes a processor 100 and memory 200.

[0045] The processor 100 inputs at least one of the subject's first input data and second input data into the learning model 210.

[0046] Here, the first input data and the second input data may be information indicating the electrical activity of the heart or information indicating the anatomical structure, respectively. For example, the first input data and the second input data may be electrocardiogram (ECG) examination information and chest X-ray examination information, respectively.

[0047] As a non-limiting example, the electrocardiogram (ECG) test information may be multi-channel ECG test information, for example, ECG test information measured from multiple leads (e.g., a 12-lead ECG), or multi-channel one-dimensional time-series data. Alternatively, the ECG test information may be an image data array obtained by converting each time-series data into an image on a single two-dimensional plane. Or, the ECG signal may be image array data obtained by capturing or photographing the image data array. In other words, the ECG test information is intended to have any format, regardless of the data format, as long as it is information that records the electrical activity of the heart.

[0048] As a non-limiting example, the second input data may be chest X-ray examination information generated from an X-ray imaging device. The corresponding chest X-ray examination information may be X-ray images taken from various directions, and preferably original images or data arrays having a standard format. For example, the chest X-ray examination information may be medical digital images and communication standard (DICOM) image original files or image data arrays taken from X-rays emitted from the rear to front (PA), front to back (AP), or one to the other (LAT) of the subject, or image data arrays captured or taken from chest X-ray film or corresponding DICOM images. In other words, the chest X-ray examination information is intended to have a free format regardless of the data format, as long as it is information that records the anatomical structure of the heart.

[0049] The processor 100 predicts the risk of cardiac disease in the subject based on the risk prediction results output by the learning model 210.

[0050] Here, a typical example of cardiac disease is paroxysmal atrial fibrillation. Specifically, cardiac disease may include paroxysmal atrial fibrillation and persistent atrial fibrillation as types of atrial fibrillation classified according to their mode of onset, duration, and whether or not they resolve spontaneously. Furthermore, as mentioned above, it may also include atrial flutter.

[0051] The processor 100 analyzes the first input data and at least one of the second input data and calculates a score based on this score, and then analyzes the subject's heart condition, such as paroxysmal atrial fibrillation. Disease risk can be predicted. Specifically, the processor 100 can predict the risk of cardiac disease in a subject based on the results of the learning model 210's analysis of the first input data and the second input data, or the overall information obtained by integrating the first input data and the second input data.

[0052] The processor 100 can predict the risk of cardiac disease in a subject based on a correction score calculated by the learning model 210 by adding additional information to at least one of the first and second input data. For example, the processor 100 can predict the risk of cardiac disease in a subject based on a correction score calculated by the learning model 210 by adding additional information to each of the first and second input data, or to the combined information obtained by integrating the first and second input data.

[0053] Here, additional information refers to information that describes the health status of the subject, and may include demographic information, symptoms, signs, medical history, and other test results.

[0054] As a specific example, the additional information may include values ​​relating to at least one of the following: the subject's age, sex, symptoms such as palpitations, underlying diseases such as heart failure, vital signs, echocardiogram values, and Holter electrocardiogram measurements. The additional information may also be a numerical vector extracted by transforming the aforementioned information using normalization or by embedding it using a separate artificial intelligence encoder (preprocessing of additional information).

[0055] Memory 200 stores at least one command to be executed by processor 100. Memory 200 may also store the learning model 210 that processor 100 accesses. The architecture and inference process of the learning model 210 will be described later.

[0056] In Figure 1, the learning model 210 is shown to be stored in memory 200, but it may be stored in an external storage device other than the example shown, and is not necessarily limited to the example shown.

[0057] Figure 2 is a diagram illustrating the architecture of the learning model 210.

[0058] Referring to Figure 2, the exemplary learning model 210 includes an encoder unit 211 and a task unit 212.

[0059] The encoder unit 211 may be configured to generate vectors corresponding to the input data of the learning model 210. Here, the input data may include first input data and second input data.

[0060] For example, the encoder unit 211 can extract features from input data in image format and convert the image into a vector.

[0061] As another example, the encoder unit 211 can extract features from multi-channel time-series input data and convert the time-series data into a vector.

[0062] As another example, the encoder unit 211 can embed words or documents into text-formatted input data and convert words or sentences into vectors.

[0063] The encoder unit 211 includes a Convolutional Neural Network, a Recurrent Neural Network, and a Transformer. r) and may consist of at least one multi-layer perceptron. In this case, it is preferable, but not limited, that the neural network of the encoder unit 211 utilizes one that has already been optimized or trained for other purposes.

[0064] The task unit 212 may be configured to assign scores to input data and predict the probability that a subject has heart disease.

[0065] Here, the task unit 212 may be configured to predict the probability of heart disease using a vector corresponding to the input data output by the encoder unit 211 (i.e., the output vector of the encoder unit) as an input value. The task unit 212 may also include a numerical vector corresponding to the additional information as an additional input value.

[0066] The task unit 212 may consist of a classifier including a fully connected layer and an activation function. For example, by passing the final output through a sigmoid function, a value between 0 and 1 can be output as a probability.

[0067] On the other hand, the encoder unit 211 and the task unit 212 may be trained using an optimizer (e.g., Stochastic Gradient Descent, Nesterov Accelerated Gradient, RMSprop, Adam, Adam-W Optimizer) and a loss function (e.g., Cross-Entropy Loss with / without label 0smoothing).

[0068] Figure 3 is a diagram illustrating an example of the inference process performed by the learning model 210.

[0069] The encoder unit 211 of the learning model 210 can map input data to a low-dimensional space and generate corresponding vectors. For example, when the encoder unit 211 receives first input data and / or second input data as input data, it generates corresponding first vectors and / or second vectors. That is, when the encoder unit 211 receives only first input data, it outputs only first vectors, and when it receives both first and second input data, it outputs first and second vectors corresponding to each.

[0070] At that time, the task unit 212 may add additional information to the first and second vectors to calculate a corrected score for having a cardiac disease such as paroxysmal atrial fibrillation.

[0071] Figure 4 below illustrates the inference process of another example of the learning model 210.

[0072] Figure 4 illustrates the inference process of another example performed by the learning model 210. For the sake of clarity, redundant explanations have been omitted.

[0073] The task unit 212 of the learning model 210 may calculate a corrected score related to heart disease based on the first vector, the second vector, and the first additional information.

[0074] For example, the task unit 212 may calculate a raw score related to heart disease based on the first vector and the second vector, and then calculate a corrected integrated score related to heart disease by reflecting the first additional information in the raw score.

[0075] Subsequently, the task unit 212 may calculate a corrected final integrated score by further incorporating the second additional information into the modified integrated score. At that time, the first and second additional information may be provided in a pre-processed format so that the learning model 210 can process it.

[0076] Figure 5 is an illustrative diagram illustrating the architecture and inference process of the additional learning model 210. For the sake of clarity, redundant explanations have been omitted.

[0077] Referring to Figure 5, the additional learning model 210 may include a first encoder unit 211-1, a second encoder unit 211-2, and a task unit 212. The first encoder unit 211-1 and the second encoder unit 211-2 may be used separately to effectively process different types of input data.

[0078] The first encoder unit 211-1 may be used to convert time-series data into a vector. For example, the first encoder unit 211-1 is configured to receive electrocardiogram data as first input data and generate a corresponding vector.

[0079] The second encoder unit 211-2 may be used to convert image data into a vector. For example, the second encoder unit 211-2 is configured to receive a chest X-ray as second input data and generate a corresponding vector.

[0080] Note that the first and second are mentioned to identify them as different entities and are not limited to the above example. That is, the first encoder unit 211-1 and the second encoder unit 211-2 may perform opposite functions, unlike in the above example.

[0081] The task unit 212 may receive the first vector and the second vector output from the first encoder unit 211-1 and the second encoder unit 211-2, and further calculate a corrected integrated score using the received additional information. In this case, the task unit 212 may calculate an integrated score based on the entire vector obtained by integrating the first vector and the second vector, and then calculate a corrected integrated score by reflecting the additional information.

[0082] In Figure 5, the learning model 210 is designed to process different types of input data, so it is preferable to use it when both the first and second input data are acquired.

[0083] Figure 6 is an illustrative diagram illustrating the architecture and inference process of the additional learning model 210. For the sake of clarity, redundant explanations have been omitted.

[0084] The additional learning model 210 in Figure 6 may include the first learning model 201 and the second learning model 202.

[0085] Here, the first learning model 201 may be a model configured to calculate a score corresponding to the first input data. The second learning model 202 may be a model configured to calculate a score corresponding to the second input data.

[0086] Specifically, the first learning model 201 may generate a first vector by encoding the first input data in the first encoder unit 211-1, and then calculate a corrected first individual score by receiving the first vector and additional information as input in the first task unit 212-1.

[0087] Similarly, the second learning model 202 may calculate a corrected second individual score by encoding the second input data in the second encoder unit 211-2 to generate a second vector, and by receiving the second vector and additional information as input in the second task unit 212-2.

[0088] Subsequently, the learning model 210 may calculate a corrected combined score based on the corrected first individual score and the corrected second individual score output by the first learning model 201 and the second learning model 202, respectively.

[0089] Specifically, the learning model 210 may calculate a corrected integrated score based on the arithmetic mean or weighted mean of the corrected first individual score and the corrected second individual score.

[0090] The learning model 210 may calculate a corrected integrated score as an output result obtained when applying regression analysis, machine learning algorithms, and artificial neural network models using the corrected first individual score and the corrected second individual score as input data.

[0091] In Figure 6, the learning model 210 is designed to process different types of input data, so it is preferable to use it when both the first and second input data are acquired.

[0092] Figure 7 is an illustrative diagram illustrating the architecture and inference process of the additional learning model 210.

[0093] The additional learning model 210 in Figure 7 may include the first learning model 201 and the second learning model 202, similar to Figure 6.

[0094] Here, the first additional information may be used to correct individual scores together with the first and second vectors in the second encoder unit 211-2, the first task unit 212-1, and the second task unit 212-2, respectively. The second additional information may be used to correct the integrated score.

[0095] In Figure 7, the learning model 210 is designed to process different types of input data, so it is preferable to use it when both the first and second input data are acquired.

[0096] As illustrated and explained in Figures 2 to 7, the learning model 210 may have an architecture in which various numbers and combinations of encoder units 211 and task units 212 are linked together, and additional information is input to the input terminal of the task unit 212 as a numerical vector or added to the output value of the task unit 212.

[0097] Furthermore, the encoder unit 211, task unit 212, first encoder unit 211-1, second encoder unit 211-2, first task unit 212-1, and second task unit 212-2, as illustrated and described in Figures 2 to 7, may be individually trained depending on the type of input data to be processed. For example, it is preferable that the encoder unit 211 and task unit 212 are trained based on a pair of first and second input data, and that the first encoder unit 211-1, second encoder unit 211-2, first task unit 212-1, and second task unit 212-2 are trained based on either first or second input data, respectively.

[0098] Figure 8 is a diagram illustrating the algorithm for constructing training data.

[0099] Referring to Figure 8, the labeling and classification system has a hierarchical structure configured to label data using medical outcomes at a specific point in time.

[0100] Firstly, if it is confirmed that the patient from whom the input data is based had a heart condition prior to the first time point, the input data may be assigned the first label. Here, the first time point may be the time when the input data was acquired, or a point in time prior to that.

[0101] For example, if the presence of heart disease is recorded in the patient's medical records, pre-measured biosignal information, and pre-analyzed diagnostic results, a first label is assigned to the input data, and it is constructed as training data.

[0102] Here, the first label may mean positive.

[0103] Furthermore, the medical record may include information such as the presence or absence of atrial fibrillation, the time when atrial fibrillation was diagnosed, and the time of input. Pre-measured biosignal information may include the presence or absence of atrial fibrillation and the time of the examination, or it may be calculated by a separate algorithm. Pre-analyzed diagnostic results may include the presence or absence and time of atrial fibrillation detected during Holter monitoring, loop recorder, or monitoring with wearable medical devices.

[0104] Secondly, if it is confirmed that the input data contains abnormal signs of heart disease prior to the second time point, the first label may be assigned to the input data. Here, the second time point is a time point after the first time point, and may be, for example, a predetermined time point at which monitoring of the input data is terminated.

[0105] As a concrete example, input data from patients without heart disease is obtained, and if abnormal signs of heart disease are detected in the input data, a first label is assigned to the input data and it is constructed as training data.

[0106] At that time, abnormal signs can be detected using a separate learning model.

[0107] Thirdly, if the acquired input data is interrupted and incompletely measured before the second time point, and no abnormal signs are detected in the input data up to the third time point prior to the second time point, a third label may be assigned to the input data.

[0108] Here, "interruption" may mean that observation is stopped before the initially planned observation period because no abnormal signs of heart rate were detected.

[0109] Here, the third label may refer to a pseudo-label. The pseudo-label is a value predicted by the learning model 210 for the corresponding data, and may represent, for example, the probability of having a heart condition such as paroxysmal atrial fibrillation.

[0110] In that case, the third point in time may be after the first point in time when the measurement of the input data began, and before the second point in time which corresponds to the predetermined end time of the measurement of the input data.

[0111] On the other hand, the acquired input data is not interrupted before the second time point; in other words, it is measured completely from the second time point to the fourth time point, and there are no abnormalities in the input data up to the fourth time point. If no signs are detected, a second label may be assigned to the relevant input data.

[0112] Here, the second label may mean negative.

[0113] On the other hand, the measurement of input data and the assessment of abnormal signs may be initiated or determined by medical tests such as electrocardiograms, Holter monitors, and loop recorders.

[0114] It goes without saying that the learning model 210 may be semi-supervised learning using the learning data constructed as shown in Figure 8, or supervised learning using other pre-constructed learning data.

[0115] On the other hand, the training data may be constructed by the processor 100 or an external entity through the classification system shown in Figure 8.

[0116] Figure 9 is a flowchart illustrating a method for predicting the risk of heart disease based on artificial intelligence according to one embodiment.

[0117] Referring to Figure 9, the method shown in Figure 9 is performed by an artificial intelligence-based cardiac disease risk prediction device 10 according to one embodiment of Figure 1.

[0118] First, the artificial intelligence-based cardiac disease risk prediction device 10 according to one embodiment inputs at least one of the subject's first input data and second input data into a pre-trained learning model 210.

[0119] Subsequently, the artificial intelligence-based cardiac disease risk prediction device 10 according to one embodiment predicts the degree of cardiac disease risk of the subject based on the results output by the learning model 210.

[0120] Although Figure 9 illustrates the method in multiple steps, at least some of the steps may be performed in a different order, combined with other steps, omitted, divided into more detailed steps, or at least one additional step not shown may be added.

[0121] Although representative embodiments of the present invention have been described in detail above, a person with ordinary skill in the art to which the present invention pertains will understand that various modifications can be made to the above embodiments, as long as they do not deviate from the scope of the present invention. Accordingly, the scope of the present invention is not limited to the embodiments described above, but must be defined not only by the claims described later, but also by equivalent claims. [Industrial applicability]

[0122] An artificial intelligence-based cardiac disease risk prediction device and method according to one embodiment can predict cardiac disease risk using a learned model and is therefore usable in the digital medical industry.

Claims

1. At least one processor, An artificial intelligence-based cardiac disease risk prediction device comprising at least one memory for storing commands executed by the at least one processor, The aforementioned at least one processor is At least one of the subject's first input data and second input data is input into the learning model. An artificial intelligence-based cardiac disease risk prediction device that predicts the cardiac disease risk of the subject based on the results output by the learning model.

2. The learning model includes an encoder unit configured to convert input data, which includes at least one of the first input data and the second input data, into a vector, The artificial intelligence-based cardiac disease risk prediction device according to claim 1, comprising a task unit configured to assign a score to the input data and predict the probability that the subject has cardiac disease.

3. The task unit calculates a corrected score by reflecting the additional information about the subject in the raw score of the input data. The artificial intelligence-based cardiac disease risk prediction device according to claim 2, configured to predict the probability that the subject has cardiac disease based on the correction score.

4. The artificial intelligence-based cardiac disease risk prediction device according to claim 3, wherein the additional information is either information entered by the subject, or generated by preprocessing recorded information stored in advance by the encoder unit.

5. The task unit includes a first task unit configured to calculate a first individual score which is corrected by reflecting the additional information in a first vector for the first input data, The artificial intelligence-based cardiac disease risk prediction device according to claim 3, further comprising: a second task unit configured to calculate a second individual score corrected by reflecting the additional information in a second vector for the second input data.

6. The artificial intelligence-based cardiac disease risk prediction device according to claim 3, wherein the task unit is configured to calculate an integrated score that is corrected by reflecting the additional information in the entire vector generated based on the first vector for the first input data and the second vector for the second input data.

7. The aforementioned learning model is trained based on the constructed training data. The artificial intelligence-based cardiac disease risk prediction device according to claim 1, wherein the learning data is constructed by labeling acquired input data with a first label, a second label, or a pseudo-label.

8. The artificial intelligence-based cardiac disease risk prediction device according to claim 7, constructed by assigning a first label to the input data if the learning data is based on a first point in time in the patient's medical history which is the source of the acquired input data, and assigning the first label to the input data if abnormal signs corresponding to cardiac disease are detected in the acquired input data at a second point in time.

9. If the acquired input data is interrupted and measured before the second time point, and no abnormal signs are detected in the input data up to the third time point before the second time point, then the learning data is obtained. A cardiac disease risk prediction device based on artificial intelligence according to claim 7, constructed by assigning pseudo-labels to the input data.

10. The artificial intelligence-based cardiac disease risk prediction device according to claim 7, wherein the learning data is constructed by measuring the acquired input data from a second time point up to a fourth time point, and if no abnormal signs are detected in the input data up to the fourth time point, a second label is assigned to the input data.

11. The first input data is electrocardiogram examination information, The second input data is chest X-ray examination information, An artificial intelligence-based cardiac disease risk prediction device according to any one of claims 1 to 10, wherein the cardiac disease includes atrial fibrillation.

12. At least one processor, A method performed by an artificial intelligence-based cardiac disease risk prediction device comprising at least one memory for storing commands executed by at least one processor, A step of inputting at least one of the subject's first input data and second input data into a learning model, An artificial intelligence-based method for predicting the risk of cardiac disease, comprising the step of predicting the risk of cardiac disease of the subject based on the results output by the learning model.

13. The learning model includes an encoder unit configured to convert input data, which includes at least one of the first input data and the second input data, into a vector, A method for predicting the risk of heart disease based on artificial intelligence according to claim 12, comprising a task unit configured to assign a score to the input data and predict the probability that the subject has heart disease.

14. The task unit adjusts the raw score of the input data to reflect additional information about the subject, and calculates a corrected score. The artificial intelligence-based method for predicting the risk of heart disease according to claim 13, which is configured to predict the probability that the subject has heart disease based on the correction score.

15. The method for predicting cardiac disease risk based on artificial intelligence according to claim 14, wherein the additional information is either information entered by the subject or generated by preprocessing recorded information stored in advance by the encoder unit.

16. The task unit includes a first task unit configured to calculate a first individual score which is corrected by reflecting the additional information in a first vector for the first input data, The artificial intelligence-based method for predicting cardiac disease risk according to claim 14, further comprising: a second task unit configured to calculate a second individual score corrected by reflecting the additional information in a second vector for the second input data.

17. The artificial intelligence-based method for predicting cardiac disease risk according to claim 14, wherein the task unit is configured to calculate an integrated score that is corrected by reflecting the additional information in the entire vector generated based on the first vector for the first input data and the second vector for the second input data.

18. The aforementioned learning model is trained based on the constructed training data. The aforementioned training data assigns a first label, a second label, and a pseudo-label to the acquired input data. A method for predicting cardiac disease risk based on artificial intelligence according to claim 12, constructed by labeling one of the bells.

19. The aforementioned training data, If a heart disease has already been diagnosed at the first point in the patient's medical history that forms the basis of the acquired input data, a first label is assigned to that input data to construct the data. A method for predicting cardiac disease risk based on artificial intelligence according to claim 18, which is constructed by assigning a first label to the input data if an abnormal sign corresponding to cardiac disease is detected in the acquired input data at a second time point.

20. The artificial intelligence-based method for predicting cardiac disease risk according to claim 18, wherein the learning data is constructed by assigning a pseudo-label to the input data if the acquired input data is interrupted and measured before a second time point, and no abnormal signs are detected in the input data up to a third time point before the second time point.

21. The artificial intelligence-based method for predicting cardiac disease risk according to claim 18, wherein the learning data is constructed by assigning a second label to the input data if the acquired input data is measured from a second time point up to a fourth time point, and no abnormal signs are detected in the input data up to the fourth time point.

22. The first input data is electrocardiogram examination information, The second input data is chest X-ray examination information, A method for predicting the risk of cardiac disease based on artificial intelligence according to any one of claims 12 to 21, wherein the cardiac disease includes atrial fibrillation.

Images