Disease diagnosis method and device using electrocardiogram data

Abstract

A disease diagnosis method using electrocardiogram data according to some embodiments of the present invention may comprises the steps of: performing first learning for a first encoder using first learning data including electrocardiogram data for a first number of leads; performing second learning for a second encoder initialized with information about the first encoder trained through the first learning using second learning data including electrocardiogram data for a second number of leads smaller than the first number; and providing input data including electrocardiogram data for the second number of leads to the second encoder trained through the second learning to classify a disease class associated with the input data.

Description

Disease diagnosis method and device using electrocardiogram data

[0001] The present invention relates to a disease diagnosis method and device using electrocardiogram data, and more specifically, to a disease diagnosis method and device using electrocardiogram data that reduces the amount of learning by using transfer learning while increasing diagnostic accuracy.

[0002] The content described in this section merely provides background information for the present embodiment and does not constitute prior art.

[0003] Electrocardiography (ECG) data records the electrical activity of the heart, specifically the electrical signals generated when the heart contracts and relaxes. ECG data can contain important information about the heart's health, rhythm, and structural abnormalities. ECG data can be collected through electrodes attached to the patient's chest, arms, and legs.

[0004] Electrocardiogram data can be collected using various leads (or channels). Leads are used to record the heart's electrical activity from various angles, and by measuring electrical signals from electrodes attached to the patient's skin, they help to graphically represent the heart's electrical activity.

[0005] Although electrocardiograms are relatively simple and widely used diagnostic tools, they are complex to interpret because accurate interpretation of the complex electrical signals generated in each part of the heart requires specialized knowledge, and interpretations can vary depending on the interpreter's perspective.

[0006] Moreover, since various heart diseases can exhibit similar electrocardiogram patterns, it is possible that subtle differences in electrocardiogram data could lead to misdiagnosis of different heart diseases.

[0007] Recently, various studies have been conducted to utilize artificial intelligence to more objectively interpret electrocardiogram data. However, the following limitations pose challenges to AI-based diagnosis of heart disease.

[0008] First, the most commonly used electrocardiogram (ECG) is the 12-lead ECG, which measures the heart's electrical activity in various directions and dimensions, providing comprehensive information about the heart's condition. Because it is the most commonly used ECG, the 12-lead ECG can be an easy way to collect training data. While the availability of a large amount of training data can be a significant advantage for AI training, it can also be prohibitively time-consuming and costly to generate accurate labels for each training data point.

[0009] Furthermore, due to the complexity of the 12-lead ECG test, it can be difficult to perform on wearable or simplified portable devices. Therefore, more simplified testing methods such as 3-lead ECG, 5-lead ECG, and single-lead ECG have recently been utilized. However, these simplified ECG tests can consist of various combinations of leads depending on the testing equipment, and the disadvantage is that it is difficult to collect sufficient training data for each lead combination.

[0010] In order to solve the problems described above, the present specification provides a disease diagnosis method and device that performs pre-learning using a relatively large number of generalized learning data and performs precision learning using a relatively small number of simplified learning data, thereby providing a faster and more accurate learning method.

[0011] The purpose of the present invention is to provide a disease diagnosis method and device using electrocardiogram data with a fast learning speed.

[0012] Another object of the present invention is to provide a disease diagnosis method and device using electrocardiogram data with high diagnostic accuracy.

[0013] The objectives of the present invention are not limited to those mentioned above. Other objectives and advantages of the present invention not mentioned above can be understood through the following description and will be more clearly understood through the embodiments of the present invention. Furthermore, it will be readily apparent that the objectives and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims.

[0014] A disease diagnosis method using electrocardiogram data according to some embodiments of the present invention for solving the above problem may include a step of performing first learning on a first encoder using first learning data including electrocardiogram data for a first number of leads, a step of performing second learning on a second encoder initialized with information on the first encoder learned through the first learning using second learning data including electrocardiogram data for a second number of leads smaller than the first number, and a step of providing input data including electrocardiogram data for the second number of leads to the second encoder learned through the second learning to classify a disease class associated with the input data.

[0015] In some embodiments, the first training data may not include labeling data, and the second training data may include electrocardiogram data for the second number of leads and labeling data associated with the electrocardiogram data for the second number of leads.

[0016] In some embodiments, the first learning may include generating first learning patch data using the first learning data, aligning the first learning patch data to generate first learning preprocessing data, performing positional embedding on the first learning preprocessing data to generate first learning embedding data, randomly masking the first learning embedding data to generate first learning masking data, and training the first encoder using the first learning masking data and the first learning embedding data.

[0017] In some embodiments, the second learning may include initializing a second encoder using information about the first encoder learned through the first learning, generating second learning patch data using the second learning data, aligning the second learning patch data to generate second learning preprocessing data, performing positional embedding on the second learning preprocessing data to generate second learning embedding data, and training the second encoder using the second learning embedding data and labeling data associated with the second learning embedding data.

[0018] In some embodiments, the step of generating the first learning preprocessing data may include arranging the first learning patch data according to predetermined conditions.

[0019] In some embodiments, the first size of the first training data set for the first training data may be larger than the second size of the second training data set for the second training data.

[0020] In some embodiments, the first learning may be self-supervised learning and the second learning may be supervised learning.

[0021] According to some embodiments of the present invention for solving the above problem, a disease diagnosis device using electrocardiogram data includes an input module that inputs input data and preprocesses the input data to generate preprocessed data, a learning module that receives the preprocessed data from the input module and generates output data using the preprocessed data, and an output module that receives output data from the learning module and generates and outputs classified data using the output data, wherein the learning module includes a first learning unit that is pre-learned using electrocardiogram data for a first number of leads, and a second learning unit that is precisely learned using electrocardiogram data for a second number of leads smaller than the first number, wherein the second learning unit can be initialized using the pre-learned first learning unit.

[0022] In some embodiments, the first learning data for pre-training the first learning unit may not include labeling data, and the second learning data for precision training the second learning unit may include electrocardiogram data for the second number of leads and labeling data associated with the electrocardiogram data for the second number of leads.

[0023] In some embodiments, the input module may divide electrocardiogram data included in the input data to generate a plurality of patches, and place the plurality of patches according to predetermined conditions to generate the preprocessed data.

[0024] The disease diagnosis method and device using electrocardiogram data according to some embodiments of the present invention have the advantage of enabling fast and accurate learning of an artificial intelligence model by performing precise learning on an encoder pre-trained using first learning data and using second learning data.

[0025] The disease diagnosis method and device using electrocardiogram data according to some embodiments of the present invention have the advantage of reducing the labeling time by performing self-supervised learning on a relatively large number of first learning data, and ensuring the accuracy of an artificial intelligence model by performing supervised learning on a relatively small number of second learning data.

[0026] In addition to the above-described contents, the specific effects of the present invention are described together with the specific matters for carrying out the invention below.

[0027] FIG. 1 is a drawing for explaining the configuration of a disease diagnosis device using electrocardiogram data according to some embodiments of the present invention.

[0028] FIG. 2 is a drawing for explaining the configuration of an input module according to some embodiments of the present invention.

[0029] FIG. 3 is a drawing for exemplarily explaining the operation of an input module according to some embodiments of the present invention.

[0030] FIG. 4 is a diagram illustrating the configuration of a learning module according to some embodiments of the present invention.

[0031] FIG. 5 is a diagram illustrating an overall learning method of a learning module according to some embodiments of the present invention.

[0032] FIG. 6 is a diagram illustrating the configuration of a first learning unit included in a learning module according to some embodiments of the present invention.

[0033] FIG. 7 is a diagram for explaining a first learning method for a first learning unit according to some embodiments of the present invention.

[0034] FIG. 8 is a drawing for exemplarily explaining a process in which a first learning method is performed for a first learning unit according to some embodiments of the present invention.

[0035] FIG. 9 is a diagram illustrating the configuration of a second learning unit included in a learning module according to some embodiments of the present invention.

[0036] FIG. 10 is a diagram illustrating a second learning method for a second learning unit according to some embodiments of the present invention.

[0037] FIG. 11 is a drawing for explaining a method of diagnosing a disease using a disease diagnosis device according to some embodiments of the present invention in which learning has been completed.

[0038] FIG. 12 is a diagram for explaining the hardware configuration of an electronic device constituting a disease diagnosis device using electrocardiogram data according to some embodiments of the present invention.

[0039] The terms and words used in this specification and claims should not be interpreted based on their general or dictionary meanings. In accordance with the principle that inventors can define the concepts of terms and words to best describe their inventions, they should be interpreted in a way that is consistent with the technical concept of the present invention. Furthermore, the embodiments described in this specification and the configurations depicted in the drawings are merely examples of how the present invention can be realized and do not fully represent the technical concept of the present invention. Therefore, it should be understood that various equivalents, modifications, and applicable examples may exist as of the time of filing.

[0040] The terms first, second, A, B, etc. used in this specification and claims may be used to describe various components, but the components should not be limited by these terms. These terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term "and / or" includes any combination of a plurality of related listed items or any item among a plurality of related listed items.

[0041] The terminology used in this specification and claims is for the purpose of describing specific embodiments only and is not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. It should be understood that terms such as "comprise" or "have" in this application do not preclude the presence or addition of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification.

[0042] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0043] Terms defined in commonly used dictionaries should be interpreted to have a meaning consistent with their meaning in the context of the relevant technology, and will not be interpreted in an idealized or overly formal sense unless expressly defined in this application.

[0044] In addition, each configuration, process, procedure or method included in each embodiment of the present invention may be shared within a scope that is not technically inconsistent with each other.

[0045]

[0046] FIG. 1 is a drawing for explaining the configuration of a disease diagnosis device using electrocardiogram data according to some embodiments of the present invention.

[0047] Referring to FIG. 1, a disease diagnosis device (10) may include an input module (100), a learning module (200), and an output module (300).

[0048] The input module (100) can receive input data (D_in). The input data (D_in) may be electrocardiogram data measured over a specific period of time. The input module (100) can preprocess the received input data (D_in) to generate preprocessed data (D_pre). The input module (100) can generate preprocessed data (D_pre) by, for example, patching the input data (D_in) into a predetermined size and aligning the patched input data. A detailed description thereof will be provided later. The input module (100) can provide the generated preprocessed data (D_pre) to the learning module (200).

[0049] The learning module (200) can be provided with preprocessed data (D_pre). The learning module (200) can perform a calculation using the preprocessed data (D_pre) and generate output data (D_out) as a result thereof. The learning module (200) can be pre-trained through deep learning or machine learning. The learning module (200) can provide the generated output data (D_out) to the output module (300). A specific method for training the learning module (200) will be described below. The learning module (200) can provide the generated output data (D_out) to the output module (300).

[0050] The output module (300) can receive output data (D_out). The output module (300) can generate classification data (D_clsf) using the output data (D_out). The classification data (D_clsf) can mean the classification result of the disease contained in the input data (D_in), i.e., the diagnosis result. In other words, the classification data (D_clsf) can include the diagnosis result for one or more diseases.

[0051] That is, a disease diagnosis device (10) according to some embodiments of the present invention may include an input module (100), a learning module (200), and an output module (300), and for a more specific description of the input module (100), reference will be made to FIGS. 2 and 3.

[0052]

[0053] FIG. 2 is a diagram illustrating the configuration of an input module according to some embodiments of the present invention. FIG. 3 is a diagram exemplarily illustrating the operation of an input module according to some embodiments of the present invention.

[0054]

[0055] Referring to FIGS. 2 and 3, the input module (100) may include a patch generation unit (110) and a patch sorting unit (120). Input data (D_in) may be provided to the patch generation unit (110). The patch generation unit (110) may divide the received input data (D_in) to generate a plurality of patches. In other words, the patch generation unit (110) may divide the received input data (D_in) into a predetermined size to generate a plurality of patches, thereby generating patch data (D_pct). Here, the patch data (D_pct) may mean a set of a plurality of patches generated by the patch generation unit (110).

[0056] For example, if the input data (D_in) is 12-lead data measured at 500 Hz for 10 seconds, the input data (D_in) can include 500*10*12 = 60,000 values. At this time, if the size of each patch generated by the patch generation unit (110) is 50, the patch generation unit (110) can generate 1,200 patches. That is, the patch generation unit (110) can generate patch data (D_pct) by dividing 60,000 input data (D_in) into sizes of 50 and generating 1,200 patches. That is, if the input data (D_in) is assumed to be electrocardiogram data collected for a first period of time, each patch can mean electrocardiogram data collected for a second period of time that is shorter than the first period of time. However, this is merely a simple example to help understanding the explanation, and the embodiments are not limited thereto.

[0057] The patch data (D_pct) generated by the patch generation unit (110) may be provided to the patch sorting unit (120). The patch sorting unit (120) may receive the patch data (D_pct) and generate preprocessing data (D_pre) by sorting the order of the patches included in the patch data (D_pct). The patch sorting unit (120) may generate preprocessing data (D_pre) by arranging the patches included in the patch data (D_pct) in a row, for example, as illustrated in FIG. 3. The patch sorting unit (120) may generate preprocessing data (D_pre) by sequentially arranging the patches included in the patch data (D_pct) in a row according to the order of each channel (e.g., lead). However, this is merely an exemplary description and the embodiments are not limited thereto. For example, the patch sorting unit (120) may generate preprocessing data (D_pre) by randomly arranging patches included in the patch data (D_pct) or according to other specific rules. The preprocessing data (D_pre) generated by the patch sorting unit (120) may be provided to the learning module (200). The learning module (200) may perform first learning using first learning data and second learning using second learning data in advance, and for a more specific description of such learning methods, reference is made to FIGS. 4 to 10.

[0058]

[0059] FIG. 4 is a diagram for explaining the configuration of a learning module according to some embodiments of the present invention. FIG. 5 is a diagram for explaining the overall learning method of a learning module according to some embodiments of the present invention. FIG. 6 is a diagram for explaining the configuration of a first learning unit included in a learning module according to some embodiments of the present invention. FIG. 7 is a diagram for explaining a first learning method for a first learning unit according to some embodiments of the present invention. FIG. 8 is a diagram for exemplarily explaining a process of performing a first learning method for a first learning unit according to some embodiments of the present invention. FIG. 9 is a diagram for explaining the configuration of a second learning unit included in a learning module according to some embodiments of the present invention. FIG. 10 is a diagram for explaining a second learning method for a second learning unit according to some embodiments of the present invention.

[0060]

[0061] Referring to FIGS. 4 and 5, the learning module (200) may include a first learning unit (210) and a second learning unit (220). The first learning unit (210) may perform first learning, which is pre-learning, and the second learning unit (220) may perform second learning, which is precision learning.

[0062] That is, the learning module (200) can pre-learn the first learning unit (210) using the first learning data (S100), and can precisely learn the second learning unit (220) using at least a part of the first learning unit (210) learned through step S100 and the second learning data (S200).

[0063] The first training data may be electrocardiogram data having a first number of channels (e.g., leads). For example, the first training data may be 12-lead electrocardiogram data, but this is merely an example and the embodiments are not limited thereto. Furthermore, the first training data may not include label data.

[0064] Meanwhile, the second learning data may include electrocardiogram data having a second number of channels. The second number may be smaller than the first number. For example, the second learning data may be 3-lead electrocardiogram data, but this is merely an example and the embodiments are not limited thereto. Furthermore, the second learning data may include label data for the electrocardiogram data.

[0065] In other words, the first learning unit (210) that learns using the first learning data can learn through self-supervised learning that does not require a label, and the second learning unit (220) that learns using the second learning data can learn through supervised learning.

[0066] Additionally, in some embodiments, the first size of the first training data set, which includes the first training data, may be larger than the second size of the second training data set, which includes the second training data. This may mean that the first training data may be more easily collected, while the second training data may be relatively more difficult to collect. However, the embodiments are not necessarily limited thereto.

[0067] According to some embodiments, the first learning unit (210) may receive first learning preprocessing data (Dlrn_pre1) and use the same to generate first learning feedback data (Dlrn_fd1). The first learning feedback data (Dlrn_fd1) may be backpropagated to the first learning unit (210) and used to train the first learning unit (210). Similarly, the second learning unit (220) may receive second learning preprocessing data (Dlrn_pre2) and use the same to generate second learning feedback data (Dlrn_fd2). The second learning feedback data (Dlrn_fd2) may be backpropagated to the second learning unit (220) and used to train the second learning unit (220). The first learning preprocessing data (Dlrn_pre1) may refer to the first learning data that has passed through the input module (100), and the second learning preprocessing data (Dlrn_pre2) may refer to the second learning data that has passed through the input module (100).

[0068] Meanwhile, the second learning unit (220) can receive information (info_enc) about the first encoder of the first learning unit (210) for which learning has been completed, and the information (info_enc) about the first encoder can be used to initialize the second encoder of the second learning unit (220). To explain the first learning method for the first learning unit (210), reference is made further to FIGS. 6 to 8.

[0069]

[0070] Referring to FIGS. 1, 2, 4, 6, and 7, the input module (100) can receive first learning data. As described above, the first learning data may not include a label. The patch generation unit (110) of the input module (100) can generate first learning patch data using the first learning data (S110). In other words, the patch generation unit (110) can generate the first learning patch data by dividing the first learning data and generating a plurality of patches. The first learning patch data can be provided to the patch alignment unit (120). The patch alignment unit (120) of the input module (100) can generate first learning preprocessing data (Dlrn_pre1) by aligning the first learning patch data (S120). The first learning preprocessing data (Dlrn_pre1) can be provided to the first learning unit (210).

[0071] The first learning unit (210) may include a first embedding layer (211), a masking layer (213), a first encoder (215), a first decoder (217), and a first output layer (219).

[0072] The first embedding layer (211) can receive the first learning preprocessing data (Dlrn_pre1). The first embedding layer (211) can perform positional embedding on the first learning preprocessing data (Dlrn_pre1) to generate the first learning embedding data (Dlrn_emb1) (S130). Positional embedding is a technique used in sequence data processing and can be used to transmit the order or position information of data to the learning model. That is, the first embedding layer (211) can perform an embedding process of converting each patch included in the first learning preprocessing data (Dlrn_pre1) into a vector of a predetermined dimension. At this time, a unique positional embedding can be added according to the position of each patch so that the learning model (e.g., the first encoder (215) and the first decoder (217)) can consider the position of each patch. In other words, as illustrated in FIG. 8, the first learning embedding data (Dlrn_emb1) may include an embedding vector that takes into account the location of each patch included in the first learning preprocessing data (Dlrn_pre1). The first learning embedding data (Dlrn_emb1) generated in the first embedding layer (211) may be provided to the masking layer (213) and the first output layer (219).

[0073] The masking layer (213) can receive the first learning embedding data (Dlrn_emb1). The masking layer (213) can perform random masking on the first learning embedding data (Dlrn_emb1) to generate the first learning masking data (Dlrn_msk1) (S140). For example, the masking layer (213) can generate the first learning masking data (Dlrn_msk1) by removing information on some of the embedding vectors among the plurality of embedding vectors included in the first learning embedding data (Dlrn_emb1). The first learning masking data (Dlrn_msk1) can be provided to the first encoder (215).

[0074] The first encoder (215) can receive the first learning masking data (Dlrn_msk1). The first encoder (215) can first encode the first learning masking data (Dlrn_msk1) to generate the first learning encoding data (Dlrn_enc1). That is, the first encoder (215) can generate the first learning encoding data (Dlrn_enc1) by extracting features for the first learning masking data (Dlrn_msk1). In other words, the first learning encoding data (Dlrn_enc1) can include information about features that are vectorized from important portions of the first learning embedding data (Dlrn_emb1). The first learning encoding data (Dlrn_enc1) can be provided to the first decoder (217).

[0075] The first decoder (217) can receive the first learning encoding data (Dlrn_enc1). The first decoder (217) can perform a first decoding of the first learning encoding data (Dlrn_enc1) to generate the first learning decoding data (Dlrn_dec1) (S150). That is, the first decoder (217) can generate the first learning decoding data (Dlrn_dec1) that restores the first learning embedding data (Dlrn_emb1) using the first learning encoding data (Dlrn_enc1). In other words, the first encoder (215) performs the role of extracting the main features of the embedding vectors included in the first learning masking data (Dlrn_msk1), and the first decoder (217) can generate the first learning decoding data (Dlrn_dec1) by restoring the first learning embedding data (Dlrn_emb1) using the main features extracted by the first encoder (215). The first learning decoding data (Dlrn_dec1) can be provided to the first output layer (219).

[0076] The first output layer (219) can receive the first learning embedding data (Dlrn_emb1) and the first learning decoding data (Dlrn_dec1). The first output layer (219) can generate the first learning feedback data (Dlrn_fd1) using the cost function of the first learning embedding data (Dlrn_emb1) and the first learning decoding data (Dlrn_dec1). The first learning feedback data (Dlrn_fd1) generated in the first output layer (219) can be backpropagated to the first encoder (215) and the first decoder (217) and used to perform the first learning for the first encoder (215) and the first decoder (217) (S160). According to some embodiments, after pre-training the first learning unit (210) through steps S110 to S160, the second learning unit (220) can be precisely trained using at least a portion of the trained first learning unit (210). To explain the second learning, which is the precise learning of the second learning unit (220), further reference is made to FIGS. 9 and 10 .

[0077]

[0078] Referring to FIGS. 1, 2, 4, 9, and 10, the second learning unit (220) can receive information (info_enc) about the first encoder included in the pre-learned first learning unit (210) (S210). In other words, the second encoder (223) of the second learning unit (220) can be initialized using the information (info_enc) about the pre-learned first encoder.

[0079] The input module (100) can receive second learning data. As described above, the second learning data can include labels. In other words, the second learning data can include electrocardiogram data and labels for disease information associated with each electrocardiogram data. The patch generation unit (110) of the input module (100) can generate second learning patch data using the second learning data (S220). That is, the patch generation unit (110) can generate second learning patch data by dividing the electrocardiogram data included in the second learning data and generating a plurality of patches. The second learning patch data can be provided to the patch alignment unit (120). The patch alignment unit (120) of the input module (100) can align the second learning patch data to generate second learning preprocessing data (Dlrn_pre2) (S230). The second learning preprocessing data (Dlrn_pre2) can be provided to the second learning unit (220).

[0080] The second learning unit (220) may include a second embedding layer (221), a second encoder (223), and a second output layer (225). As described above, the second encoder (223) may be initialized using information (info_enc) about the pre-trained first encoder. For example, the second learning unit (220) may utilize the structure of a visual transformer (ViT). The visual transformer may be an architecture that processes an image as a sequence and learns relationships between various parts within the image. A detailed description of the visual transformer will be omitted.

[0081] The second embedding layer (221) can receive the second learning preprocessing data (Dlrn_pre2). The second embedding layer (221) can perform positional embedding on the second learning preprocessing data (Dlrn_pre2) to generate the second learning embedding data (Dlrn_emb2) (S240). That is, the second embedding layer (221) can perform an embedding process for converting each patch included in the second learning preprocessing data (Dlrn_pre2) into a vector of a predetermined dimension. The second learning embedding data (Dlrn_emb2) generated by the second embedding layer (221) can be provided to the second encoder (223).

[0082] The second encoder (223) may perform a second encoding on the second learning embedding data (Dlrn_emb2) to generate the second learning encoding data (Dlrn_enc2) (S250). For example, the second encoder (223) may further receive class tokens for effectively classifying the second learning embedding data (Dlrn_emb2), but the embodiments are not limited thereto. The second learning encoding data (Dlrn_enc2) may be provided to the second output layer (225).

[0083] The second output layer (225) can receive the second learning encoding data (Dlrn_enc2) and the corresponding labeling data (Dlrn_lbd). The second output layer (225) can perform second learning for the second encoder (223) using the second learning encoding data (Dlrn_enc2) and the corresponding labeling data (Dlrn_lbd) (S260). For example, the second output layer (225) can perform an operation for disease diagnosis by applying a linear layer and a softmax function to the second learning encoding data (Dlrn_enc2) to calculate the probability that the second learning encoding data (Dlrn_enc2) belongs to each disease class. Subsequently, the second output layer (225) can generate the second learning feedback data (Dlrn_fd2) by calculating a cost function between the operation result and the labeling data (Dlrn_lbd). The second learning feedback data (Dlrn_fd2) can be backpropagated to the second encoder (223) and used to perform second learning for the second encoder (223). Referring further to FIG. 11 to explain a method of performing disease diagnosis using a disease diagnosis device that has undergone learning through the first and second learning.

[0084]

[0085] Figure 11 is a diagram illustrating a method for diagnosing a disease using a disease diagnosis device according to some embodiments of the present invention, after learning has been completed. For convenience of explanation, details identical or similar to those described above are omitted or briefly explained.

[0086] Referring to FIGS. 1, 5, and 11, input data (D_in) may be provided to an input module (100) of a disease diagnosis device (10). A patch generation unit (110) of the input module (100) may divide the input data (D_in) into a plurality of patches to generate patch data (D_pct). The patch data (D_pct) may be provided to a patch alignment unit (120). The patch alignment unit (120) may align the patch data (D_pct) to generate preprocessed data (D_pre). The preprocessed data (D_pre) may be provided to a learning module (200). At this time, the learning module (200) may be learned in advance through the first learning and second learning described above.

[0087] For disease diagnosis, the second learning unit (220) of the learning module (200) may be utilized. In other words, the first learning unit (210) is utilized for pre-learning, and for disease diagnosis, the second learning unit (220) that has undergone precise learning may be utilized. The second embedding layer (221) of the second learning unit (220) may positionally embed preprocessed data (D_pre) to generate embedded data (D_emb). The embedded data (D_emb) may be provided to the pre-learned second encoder (223).

[0088] The second encoder (223) can perform a second encoding on the embedded data (D_emb) to generate encoded data (D_enc). The encoded data (D_enc) can be provided to the second output layer (225).

[0089] The second output layer (225) can diagnose a disease contained in the input data (D_in) using the encoding data (D_enc). For example, the second output layer (225) can calculate the probability for each disease class contained in the input data (D_in) and generate the output data (D_out). The output data (D_out) can be provided to the output module (300).

[0090] The output module (300) can generate classification data (D_clsf) using the output data (D_out). For example, the classification data (D_clsf) may include the disease with the highest probability among the probabilities for the disease classes included in the output data (D_out). However, this is only one embodiment, and the embodiments are not limited thereto. For example, the classification data (D_clsf) may include all information about disease classes exceeding a specific probability threshold, or may include information about disease classes according to various other conditions.

[0091]

[0092] In some embodiments, the relatively large size of the first training data set facilitates securing a sufficient number of training data sets for the first encoder (215) to extract features for the embedding vectors. Furthermore, since the first encoder (215) is pre-trained through self-supervised learning, the time and cost required for labeling the first training data can be reduced.

[0093] Meanwhile, even if the size of the second learning data set including labeling information is relatively small, the second encoder (223) is precisely trained using information about the pre-trained first encoder (215) to extract features for the embedding vector, so that an artificial intelligence model with sufficiently high accuracy can be trained with only a smaller number of epochs.

[0094] In other words, the disease diagnosis method and device using electrocardiogram data according to some embodiments of the present invention can perform self-supervised learning using first learning data, which is commonly used and relatively easy to collect learning data from, and then perform supervised learning using second learning data for more precise calculations. Through this, the disease diagnosis method and device according to some embodiments of the present invention have the advantage of being able to learn an artificial intelligence model with higher accuracy with a smaller learning amount.

[0095]

[0096] The results of comparing the accuracy of a disease diagnosis device that performed first and second learning according to some embodiments of the present invention and a disease diagnosis device that performed only second learning are shown in Table 1, and it can be confirmed that the accuracy of a disease diagnosis device according to some embodiments of the present invention is higher.

[0097] Study MethodAvg. AUROCPTB-XLCPSC 2018Performed only the second study0.8770.899Performed the first and second studies0.9040.970

[0098] FIG. 12 is a diagram for explaining the hardware configuration of an electronic device constituting a disease diagnosis device using electrocardiogram data according to some embodiments of the present invention.

[0099] Referring to FIG. 12, an electronic device (1000) constituting a disease diagnosis device according to some embodiments of the present invention may include a controller (1010), an input / output device (1020), a memory (1030), an interface (1040), a storage (1050), and a bus (1060). The controller (1010), the input / output device (1020), the memory (1030), the interface (1040), and / or the storage (1050) may be coupled to each other via a bus (1060). The bus (1060) corresponds to a path through which data is transferred.

[0100] Specifically, the controller (1010) may include at least one of a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphics Processing Unit), an NPU (Neural Processing Unit), a digital signal processor, a microcontroller, an application processor (AP), and logic elements capable of performing functions similar thereto.

[0101] The input / output device (1020) may include at least one of a keypad, a keyboard, a touchscreen, and a display device.

[0102] The memory (1030) can load data and / or programs, etc. For example, the memory (1030) can load a program composed of instructions for performing a disease diagnosis method according to some embodiments of the present invention. At this time, the memory (1030) is an operating memory for improving the operation of the controller (1010) and may include high-speed DRAM and / or SRAM. The memory (1030) may include one or more volatile memory devices such as DDR SDRAM (Double Data Rate Static DRAM) and SDR SDRAM (Single Data Rate SDRAM) and / or one or more non-volatile memory devices such as EEPROM (Electrically Erasable Programmable ROM) and flash memory.

[0103] The interface (1040) may perform a function of transmitting data to or receiving data from a communication network. The interface (1040) may be wired or wireless. For example, the interface (1040) may include an antenna or a wired or wireless transceiver.

[0104] Storage (1050) can store and preserve data and / or programs. For example, storage (1050) can store a program composed of instructions for performing a disease diagnosis method according to some embodiments of the present invention. In this specification, memory (1030) and storage (1050) are used separately, but the embodiments are not limited thereto. Memory (1030) and storage (1050) can be used to substantially store or load data or programs, and thus the two terms can be used interchangeably. Storage (1050) can include one or more non-volatile memory devices such as a solid state drive (SSD), a hard drive, and flash memory.

[0105] Alternatively, the electronic device (1000) according to embodiments of the present invention may be a system formed by connecting multiple electronic devices via a network. In this case, each module or combination of modules may be implemented as the electronic device (1000). However, the present embodiment is not limited thereto.

[0106] Additionally, the electronic device (1000) can transmit data through a network. This may include a network based on wireless Internet technology and short-range communication technology.

[0107] The wireless Internet technology may include, for example, at least one of Wireless LAN (WLAN), Digital Living Network Alliance (DMNA), Wireless Broadband (Wibro), World Interoperability for Microwave Access (Wimax), High Speed ​​Downlink Packet Access (HSDPA), High Speed ​​Uplink Packet Access (HSUPA), IEEE 802.16, Long Term Evolution (LTE), Long Term Evolution-Advanced (LTE-A), Wireless Mobile Broadband Service (WMBS), and 5G NR (New Radio) technologies. However, the present embodiment is not limited thereto.

[0108] Short-range communication technologies may include, for example, at least one of Bluetooth, Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra-Wideband (UWB), ZigBee, Near Field Communication (NFC), Ultra Sound Communication (USC), Visible Light Communication (VLC), Wi-Fi, Wi-Fi Direct, and 5G NR (New Radio). However, the present embodiment is not limited thereto.

[0109] An electronic device (1000) communicating through a network may comply with technical standards and standard communication methods for mobile communication. For example, the standard communication method may include at least one of GSM (Global System for Mobile communication), CDMA (Code Division Multi Access), CDMA2000 (Code Division Multi Access 2000), EV-DO (Enhanced Voice-Data Optimized or Enhanced Voice-Data Only), WCDMA (Wideband CDMA), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTEA (Long Term Evolution-Advanced), and 5G NR (New Radio). However, the present embodiment is not limited thereto.

[0110]

[0111] The above description is merely an example of the technical idea of ​​the present embodiment, and those skilled in the art will appreciate that various modifications and variations can be made without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of ​​the present embodiment, but rather to explain it, and the scope of the technical idea of ​​the present embodiment is not limited by these embodiments. The scope of protection of the present embodiment should be interpreted by the claims below, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of rights of the present embodiment.

Claims

1. A disease diagnosis method performed in a disease diagnosis device composed of an electronic device, A step of performing first learning for a first encoder using first learning data including electrocardiogram data for a first number of leads; A step of performing second learning on a second encoder initialized with information on a first encoder learned through the first learning, using second learning data including electrocardiogram data for a second number of leads smaller than the first number; and A step of providing input data including electrocardiogram data for the second number of leads to a second encoder learned through the second learning, thereby classifying a disease class associated with the input data. A method for diagnosing diseases using electrocardiogram data.

2. In paragraph 1, The above first learning data does not include labeling data, The second learning data includes electrocardiogram data for the second number of leads and labeling data associated with the electrocardiogram data for the second number of leads. A method for diagnosing diseases using electrocardiogram data.

3. In paragraph 1, The above first learning is, A step of generating first learning patch data using the first learning data; A step of generating first learning preprocessing data by aligning the first learning patch data; A step of generating first learning embedding data by performing positional embedding on the first learning preprocessing data; A step of generating first learning masking data by randomly masking the first learning embedding data; and A step of training the first encoder using the first learning masking data and the first learning embedding data, A method for diagnosing diseases using electrocardiogram data.

4. In paragraph 3, The second learning above is, A step of initializing a second encoder using information about the first encoder learned through the first learning; A step of generating second learning patch data using the second learning data; A step of generating second learning preprocessing data by aligning the second learning patch data; A step of generating second learning embedding data by performing positional embedding on the second learning preprocessing data; and A step of training the second encoder using the second learning embedding data and labeling data associated with the second learning embedding data, A method for diagnosing diseases using electrocardiogram data.

5. In paragraph 3, The step of generating the above first learning preprocessing data is: Including arranging the first learning patch data according to predetermined conditions, A method for diagnosing diseases using electrocardiogram data.

6. In paragraph 1, The first size of the first learning data set for the first learning data is larger than the second size of the second learning data set for the second learning data. A method for diagnosing diseases using electrocardiogram data.

7. In paragraph 1, The above first learning is self-supervised learning, and the above second learning is supervised learning. A method for diagnosing diseases using electrocardiogram data.

8. An input module that inputs input data and preprocesses the input data to generate preprocessed data; A learning module that receives the preprocessing data from the input module and generates output data using the preprocessing data; and An output module that receives output data from the learning module and generates and outputs classification data using the output data, The learning module includes a first learning unit that is pre-learned using electrocardiogram data for a first number of leads, and a second learning unit that is precisely learned using electrocardiogram data for a second number of leads smaller than the first number, wherein the second learning unit is initialized using the pre-learned first learning unit. A disease diagnosis device using electrocardiogram data.

9. In paragraph 8, The first learning data for pre-learning the above first learning unit does not include labeling data, The second learning data for precisely learning the second learning unit includes electrocardiogram data for the second number of leads and labeling data associated with the electrocardiogram data for the second number of leads. A disease diagnosis device using electrocardiogram data.

10. In paragraph 8, The above input module, Splitting the electrocardiogram data included in the input data to create multiple patches, and arranging the multiple patches according to predetermined conditions to create the preprocessing data. A disease diagnosis device using electrocardiogram data.

Images