Method and analysis device for classifying heart disease of subject by using electrocardiogram data

Abstract

This method for classifying a heart disease of a subject by using ECG data comprises steps in which an analysis device: receives ECG data about the subject; and inputs the ECG data into a trained deep learning model and classifies the heart disease of the subject on the basis of values that are output by the deep learning model. The deep learning model is a model trained using training data in which the amount of training data for each class is imbalanced with respect to multiple heart diseases, and is trained using a cross-entropy loss to which class weight is applied.

Description

Method and analysis device for classifying a subject's heart disease using electrocardiogram data

[0001] The technology described below relates to a technique for diagnosing or classifying heart disease using electrocardiogram data.

[0002] An electrocardiogram (ECG) represents information about the heart's electrical signals. Traditionally, ECG data has been used to diagnose cardiovascular diseases and conditions. Meanwhile, various techniques for classifying ECG data using learning models have been studied.

[0003] Because biosignals like ECGs have individual characteristics, standardizing them is challenging. Furthermore, collecting sufficient training data for deep learning model training is challenging. Deep learning models can use ECG data to classify various types of heart disease, but the number of training data can vary significantly across classes. Therefore, deep learning models interpreting ECG data struggle to derive accurate results for specific heart diseases by analyzing data from different subjects.

[0004] The technology described below aims to provide a deep learning-based technique for classifying multiple heart diseases based on ECG data.

[0005] A method for classifying a subject's heart disease includes a step of an analysis device receiving the subject's ECG data, a step of the analysis device inputting the ECG data into a trained deep learning model, and a step of classifying the subject's heart disease based on a value output by the deep learning model.

[0006] An analysis device for classifying a subject's heart disease includes an interface device for receiving the subject's ECG data, a storage device for storing a deep learning model for receiving the ECG data and classifying the heart disease, and a calculation device for inputting the received ECG data into the deep learning model and classifying the subject's heart disease based on a value output by the deep learning model.

[0007] The technique described below can accurately classify heart disease even when the number of training data for different classes is imbalanced.

[0008] Figure 1 is an example of a system for classifying heart disease using ECG data.

[0009] Figure 2 is an example of the learning process of a deep learning model.

[0010] Figure 3 is an example of the structure of a deep learning model that classifies heart disease using ECG data.

[0011] Figure 4 is an example of an analysis device for classifying heart disease.

[0012] The technology described below is susceptible to various modifications and embodiments. Specific embodiments are illustrated and described in detail in the drawings. However, this does not limit the technology described below to specific embodiments, and it should be understood that all modifications, equivalents, and alternatives fall within the spirit and scope of the technology described below.

[0013] Terms such as first, second, A, and B may be used to describe various components, but these components are not limited by these terms and are used solely to distinguish one component from another. For example, without departing from the scope of the technology described below, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. The term "and / or" includes a combination of a plurality of related described items or any of a plurality of related described items.

[0014] As used herein, the singular expressions should be understood to include the plural expressions unless the context clearly dictates otherwise, and the term "comprises" and the like should be understood to mean the presence of a described feature, number, step, operation, component, part, or combination thereof, but not to exclude the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0015] Before going into a detailed description of the drawings, it should be made clear that the division of components in this specification is merely a division based on the main function of each component. In other words, two or more components described below may be combined into a single component, or a single component may be further subdivided into two or more components with more detailed functions. In addition to its own main function, each component described below may additionally perform some or all of the functions of other components, and of course, some of the main functions of each component may be exclusively performed by other components.

[0016] Additionally, in performing a method or method of operation, each process constituting the method may occur in a different order than the stated order, unless the context clearly indicates a specific order. That is, each process may occur in the same order as the stated order, may be performed substantially simultaneously, or may be performed in the opposite order.

[0017] The technology described below is a technique for diagnosing a subject's heart disease using ECG (electrocardiogram) data.

[0018] The following explains how the analysis device uses a learning model to classify the presence or type of heart disease based on ECG data. The analysis device can be implemented using a variety of data processing devices. For example, the analysis device can be implemented as a PC, a networked server, a smart device, or a chipset embedded with a dedicated program.

[0019] Machine learning models include decision trees, random forests, K-nearest neighbors (KNN), naive Bayes, support vector machines (SVMs), and artificial neural networks (ANNs). The technology described below uses a deep learning model among ANNs to classify and diagnose heart disease from ECG data.

[0020] Figure 1 is an example of a system (100) for classifying heart disease using ECG data. In Figure 1, the analysis device is illustrated as an example of a computer terminal (130) and a server (140).

[0021] The electrocardiogram device (110) generates ECG data for a subject (patient). The electrocardiogram device (110) can store the generated ECG data in an EMR (Electronic Medical Record, 120) or a separate database (DB).

[0022] The learning device (50) analyzes ECG data to build a deep learning model that classifies heart disease. The learning device (50) refers to a computer device capable of preprocessing ECG data and training a deep learning model. The learning device (50) can build a deep learning model using certain training data. The learning process and model structure will be described later. The computer terminal (130) and server (140) acquire the built deep learning model and use it for subject analysis.

[0023] In Fig. 1, a user (A) can analyze ECG data using a computer terminal (130). The computer terminal (130) can receive ECG data from an electrocardiogram device (110) or an EMR (120) via a wired or wireless network. In some cases, the computer terminal (130) may be a device physically connected to the electrocardiogram device (110). The computer terminal (130) can preprocess ECG data in a predetermined manner. The computer terminal (130) can receive ECG data and classify or diagnose heart disease based on the values ​​output by a deep learning model. User A can check the analysis results on the computer terminal (130).

[0024] The server (140) can receive ECG data from an electrocardiogram device (110) or an EMR (120). The server (140) can preprocess the ECG data. The server (140) can receive ECG data and classify or diagnose heart disease based on the values ​​output by the deep learning model. The server (140) can transmit the analysis results to the terminal of user A.

[0025] The computer terminal (130) and / or server (140) may also store the analysis results in the EMR (120).

[0026] Below, we describe a deep learning model for classifying heart disease and the training process of the deep learning model.

[0027] The researchers used data from 7,355 patients collected at Asan Medical Center in Seoul. The data comprised data on eight types of heart disease, as shown in Table 1 below.

[0028] Number of disease names and data (%) 1. Atrial flutter (atrial flutter) 655 (8.9) 2. 1st or 2nd degree AV block (Mobits type 1) (1st / 2nd degree AV block) 278 (3.8) 3. Irregular narrow QRS tachyarrhythmia (QRS tachycardia) 1,451 (19.7) 4. Supraventricular tachycardia (PSVT) 307 (4.2) 5. High degree or complete AV block (high / complete AV block) 299 (4.0) 6. Sinus node dysfunction (sinus node dysfunction) 850 (11.6) 7. Sinus tachyarrhythmia (sinus tachycardia) 1,132 (15.4) 8. Ventricular premature contraction (VPC) (Premature ventricular contractions) 2,382 (32.4) Total 7,355 (100)

[0029] Looking at Table 1 above, the number of data points varies significantly and is imbalanced across diseases. This imbalance is defined as a difference in the number of data points between two different classes exceeding a certain threshold.

[0030] Figure 2 illustrates an example of the training process (200) of a deep learning model. The deep learning model classifies heart disease using ECG data. Therefore, the deep learning model can also be called a classification model. The database consists of data with varying numbers of data for each heart disease class, as shown in Table 1.

[0031] The learning device can preprocess the raw data of the ECG in a certain manner (210).

[0032] The learning device can remove low-frequency data of 0.05 to 1 Hz, which corresponds to noise caused by breathing or movement, from the patients' data using the Savitzky-Golay filter.

[0033] Additionally, the learning device can normalize the size of the ECG signal to a constant. For example, the learning device can normalize the size of the data to the range [-1, 1] by applying min-max normalization. The learning device can normalize the data using a formula such as Equation 1 below. In Equation 1, min(data) is the smallest value among all data, max(data) is the largest value among all data, data is the value of a specific data, and scale is the normalized value of the specific data.

[0034]

[0035] The learning device divides the collected data into training data and validation data (220). The researcher divided the 7,355 data sets into 4,707 training data, 1,471 test data, and 1,177 validation data.

[0036] The learning device performs learning of a classification model using learning data (220).

[0037] The learning device extracts a single input data point from the training data and performs the learning process. The learning device compares the output value (the degree of heart disease classification) from the classification module that received the input data with the correct answer, thereby updating the model's parameters. The learning device repeats the learning process using data from the training data point.

[0038] Once the learning process is complete, the learning device performs validation on the learned classification model (230). The learning device extracts one of the validation data sets and performs the validation process. The learning device inputs the selected data into the classification model, classifies heart disease, and compares the classification results with the correct answer value to perform validation.

[0039] Inception-V3 is a CNN (Convolutional Neural Network)-based model that minimizes information loss in pooling while reducing the number of parameters in a limited memory environment (see C. Szegedy et al., 'Rethinking the inception architecture for computer vision,' in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, pp. 2818-2826.).

[0040] Figure 3 illustrates the structure of a deep learning model that classifies heart disease using ECG data. The left side of Figure 3 shows the overall structure of the deep learning model, while the right side shows the structure of the modules that make up the feature extraction layer.

[0041] A deep learning model can include an input layer, a feature extraction layer, and an output layer. The structure of each layer in a deep learning model can utilize any of a variety of models or modules.

[0042] The input layer may include a stem layer. The stem layer reduces the size (dimension) of the input data. The stem layer may sequentially include three convolutional layers, a max pooling layer, two convolutional layers, and a max pooling layer.

[0043] In Figure 3, the dotted block corresponds to the feature extraction layer. The feature extraction layer includes an inception module and a grid reduction module. The inception module is structured to reduce the amount of computation while extracting significant features. The inception module uses inception module A, inception module B, and inception module C, which have different structures. Inception module A uses two 3 × 3 filters instead of a 5 × 5 filter. The inception module performs factorization and asymmetric convolution. The factorization technique reduces the amount of computation of the convolution by factoring the filter size of the convolution layer. The grid reduction module is placed between the inception modules. The grid reduction module reduces the dimensionality while reducing the loss of expressive power. The grid reduction module concatenates the outputs of the pooling layer and the convolution layer.

[0044] The output layer can output classification values ​​(probability values ​​for classes) by performing an activation function (sigmoid) and global average pooling on the feature values ​​output from the feature extraction layer. The output layer can also be configured as a fully connected layer.

[0045] In the training process of a deep learning model, cross-entropy can be used as a loss. Cross-entropy (CE) loss is expressed as shown in Equation 2 below.

[0046]

[0047] p target represents the probability of the target. Typically, the weights of the loss function are the same regardless of class. In this case, if the number of training data is imbalanced across multiple classes, the constructed deep learning model may exhibit poor performance for some classes.

[0048] To address the problem of imbalanced training data across classes, researchers assigned class weights to the loss function. The class weights are given in Equation 2 below.

[0049]

[0050] Class weights are parameters that weight the loss as in Equation 2. The numerator (the number of data) in Equation 3 is the total number of training data, class number is the number of classes, and class data is the number of training data for each class.

[0051] For the training data in Table 1, the class weights are as shown in Table 2 below. That is, the class weights have low values ​​when the number of data is large, and high values ​​when the number of data is small.

[0052] Class Class Weight Data Count Atrial flutter 1.27 16 55 1st or 2nd degree AV block (Mobits type 1) 2.98 52 79 Irregular narrow QRS tachyarrhythmia 0.57 41 45 1 PSVT 2.7 13 30 7 High degree or complete AV block 2.7 8 52 9 Sinus node dysfunction 0.98 85 0 Sinus tachyarrhythmia 0.73 61 132 VPC 0.35 2 38 2

[0053] Additionally, the researchers used focal loss instead of cross-entropy loss. The focal loss is given by Equation 4 below.

[0054]

[0055] α target and (1 - p target ) γ determines the weights of the loss function. 1 - p target ) γThe items reduce the weight of the loss according to γ ​​α target and γ can be determined by grid search.

[0056] The researcher verified the performance of the constructed classification model. The researcher used a classification model constructed using imbalanced training data and a classification model constructed using balanced training data. The balanced training data was data extracted from 279 classes in Table 1. The performance of the models is shown in Table 3 below. In Table 3, 'Imbalance' is a model that uses binary cross-entropy loss on imbalanced data, 'Focal loss' is a model that uses focal loss on imbalanced data, 'Class weight' is a model that uses binary cross-entropy loss with class weights on imbalanced data, and 'Balance' is a model that uses binary cross-entropy loss on balanced training data.

[0057] Model TypeAccuracyPrecisionRecallF1 ScoreImbalance0.920.920.920.92Focal Loss0.960.960.960.96Class Weight0.950.940.950.95Balance0.850.860.860.86

[0058] In Table 3, "Balance" showed relatively low performance due to the small number of training data. When imbalanced data was used, models using focal loss (Focal loss) and binary cross-entropy loss with class weights (Class weight) showed higher performance.

[0059] Table 4 below shows the individual performance (F1 score) of the classification modules for multiple heart diseases.

[0060] ClassImbalance datasetFocal lossClass weightBalance datasetAtrial flutter0.770.870.860.851st or 2nd degree AV block(Mobits type 1)0.850.940.910.68PSVT0.780.830.810.92High degree or complete AV block0.750.880.880.85Irregular narrow QRS tachyarrhythmia0.920.980.970.65Sinus node dysfunction0.970.980.980.98Sinus tachyarrhythmia0.930.950.960.96VPC0.990.990.990.99

[0061] As shown in Table 4, models applying the standard binary cross-entropy loss to imbalanced training data exhibited poor performance for classes with limited data. Models using balanced training data also exhibited poor performance for some classes, as they could not utilize the entire dataset. When using imbalanced data, models using the focal loss (Focal loss) and the class-weighted binary cross-entropy loss (Class weight) showed relatively high performance across all classes.

[0062] Figure 4 illustrates an example of an analysis device (300) for classifying heart disease. The analysis device (300) corresponds to the aforementioned analysis devices (130 and 140 of Figure 1). The analysis device (300) may be physically implemented in various forms. For example, the analysis device (300) may take the form of a computer device such as a PC, a network server, a data processing chipset, or the like.

[0063] The analysis device (300) may include a storage device (310), a memory (320), a computing device (330), an interface device (340), a communication device (350), and an output device (360).

[0064] The storage device (310) can store the patient's ECG data.

[0065] The storage device (310) can store a deep learning model that classifies heart disease. In this case, the storage device (310) can also store multiple models for different types of heart disease.

[0066] In this case, the deep learning model is a model trained with imbalanced training data. The deep learning model may be trained using focal loss. Alternatively, the deep learning model may be trained using cross-entropy loss weighted by the class weights in Equation 3.

[0067] The storage device (310) can store analysis results (images, text, etc.).

[0068] The memory (320) can store data and information generated during the process of classifying heart disease using ECG data by the analysis device (300).

[0069] The interface device (340) is a device that receives certain commands and data from the outside.

[0070] The interface device (340) can receive the subject's ECG data from a physically connected input device or an external storage device.

[0071] The interface device (340) can also transmit the classification results of heart disease to an external object.

[0072] A communication device (350) refers to a configuration that receives and transmits certain information through a wired or wireless network.

[0073] The communication device (350) can receive the subject's ECG data from an external object. The communication device (350) can receive clinical information.

[0074] Alternatively, the communication device (350) may transmit the classification results of heart disease to an external object such as a user terminal.

[0075] Meanwhile, the interface device (340) is meant to include an interface that transmits data or information received from the communication device (350) into the analysis device (300).

[0076] An output device (360) is a device that outputs certain information. The output device (360) can output interfaces, analysis results, etc. required for the data processing process.

[0077] The computational device (330) may preprocess the subject's initial ECG data in a predetermined manner. The preprocessing may include filtering noise information and normalizing the ECG signal size, as described above.

[0078] The computational device (330) inputs (preprocessed) ECG data into a pre-trained deep learning model, and can classify the subject's heart disease based on the classification result (probability value) output by the deep learning model.

[0079] At this time, the deep learning model may be a separate model for each of the multiple heart diseases. In this case, the computing device (330) can input the subject's ECG data into multiple deep learning models to obtain classification results for each heart disease. The computing device (330) can select the heart disease with the highest probability value among multiple heart diseases as the subject's heart disease.

[0080] The computing device (330) may be a device such as a processor, AP, or chip embedded with a program that processes data and performs certain operations.

[0081] Additionally, the method for classifying heart disease using ECG data as described above can be implemented as a program (or application) including an executable algorithm that can be executed on a computer. The program can be stored and provided on a temporary or non-transitory computer-readable medium.

[0082] A non-transitory readable medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the various applications or programs described above may be stored and provided on a non-transitory readable medium, such as a CD, DVD, hard disk, Blu-ray disk, USB, memory card, ROM (read-only memory), PROM (programmable read only memory), EPROM (Erasable PROM, EPROM), EEPROM (Electrically EPROM), or flash memory.

[0083] Temporarily readable media refers to various types of RAM, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous DRAM (Synclink DRAM, SLDRAM), and Direct Rambus RAM (DRRAM).

[0084] The present embodiment and the drawings attached to the present specification only clearly illustrate a part of the technical idea included in the above-described technology, and it is obvious that all modified examples and specific embodiments that can be easily inferred by a person skilled in the art within the scope of the technical idea included in the specification and drawings of the above-described technology are included in the scope of the rights of the above-described technology.

Claims

A step in which the analysis device receives the subject's ECG (electrocardiogram) data; and The above analysis device includes a step of inputting the ECG data into a trained deep learning model and classifying the heart disease of the subject based on the value output by the deep learning model. The above deep learning model is a method for classifying a subject's heart disease using ECG data, which is a model trained using cross-entropy loss with class weights applied, while using training data with an imbalanced number of training data per class for multiple heart diseases. In the first paragraph, The above deep learning model includes multiple classification models for atrial flutter, first / second degree atrioventricular block, QRS tachycardia, supraventricular tachycardia, high / complete atrioventricular block, sinusitis, sinus tachycardia, and premature ventricular contraction, respectively. The above analysis device inputs the ECG data into each of the plurality of classification models, and a method for classifying the subject's heart disease using the ECG data to classify the subject's heart disease. In the first paragraph, The above class weights are a method for classifying a subject's heart disease using ECG data determined by the formula below. (Class weight is the class weight, the number of data is the total number of training data, class number is the number of classes, and class data is the number of training data for each class) In the first paragraph, A method for classifying a subject's heart disease using ECG data, wherein the above deep learning model extracts features of the ECG data using a plurality of inception modules and a grid reduction module. An interface device for receiving the subject's ECG (electrocardiogram) data; A storage device that stores a deep learning model that receives ECG data and classifies heart disease; and Including a calculation device that inputs the ECG data received above into the deep learning model and classifies the heart disease of the subject based on the value output by the deep learning model, The above deep learning model is an analysis device that classifies a subject's heart disease by learning using learning data with an imbalanced number of learning data per class for multiple heart diseases, and using a cross-entropy loss to which class weights are applied. In paragraph 5, The above deep learning model includes multiple classification models for atrial flutter, first / second degree atrioventricular block, QRS tachycardia, supraventricular tachycardia, high / complete atrioventricular block, sinusitis, sinus tachycardia, and premature ventricular contraction, respectively. The above analysis device is an analysis device that classifies the subject's heart disease by inputting the ECG data into each of the plurality of classification models. In paragraph 5, The above class weights are an analysis device that classifies the subject's heart disease determined by the formula below. (Class weight is the class weight, the number of data is the total number of training data, class number is the number of classes, and class data is the number of training data for each class) In paragraph 5, The above deep learning model is an analysis device that classifies a subject's heart disease by extracting features of the ECG data using a plurality of inception modules and a grid reduction module.

Images