KR20230036487A - System for diagnosing and explaining cardiac disorder based on explainable artificial intelligence deep learning - Google Patents

Abstract

According to the present invention, disclosed is an XAI deep learning-based heart disease diagnosing and interpreting system, which can diagnose a heart disease and provide a reason for diagnosis using an explainable neural network model. The XAI deep learning-based heart disease diagnosing and interpreting system, comprises: an XAI deep learning model unit (110) constructed to use a heart disease and electrocardiogram data determining the heart disease, as learning data, having one or more corresponding interpretation modules for interpreting one or more characteristics determining a diagnosis of the heart disease, and deriving the diagnosis of the heart disease and determination characteristics of the diagnosed heart disease; an input unit (120) for receiving the electrocardiogram data to provide the same for the XAI deep learning model unit (110); and an output unit (130) for outputting the diagnosis of the heart disease and the determination characteristics of the diagnosed heart disease by using the electrocardiogram data from the input unit (120) as an input, by the XAI deep learning model unit (110).

Description

XAI deep learning based heart disease diagnosis and interpretation system {SYSTEM FOR DIAGNOSING AND EXPLAINING CARDIAC DISORDER BASED ON EXPLAINABLE ARTIFICIAL INTELLIGENCE DEEP LEARNING}

The present invention relates to a heart disease diagnosis and interpretation system based on XAI deep learning, which can be applied to clinical practice by increasing the accuracy and interpretability of diagnosis by providing reasons for diagnosis and diagnosis of heart disease by an explainable neural network model.

As is well known, since the development of electrocardiography, knowledge related to electrocardiography has expanded exponentially, and information on the electrical function of the heart can be obtained from electrocardiography, and various heart diseases such as arrhythmia, coronary artery disease, and myocardial disease can be diagnosed.

Meanwhile, research on AI algorithms for electrocardiograms is being actively conducted, and AI algorithms detect heart failure, predict atrial fibrillation among arrhythmia rhythms, or determine gender.

In this way, by overcoming human limitations, AI algorithms can detect subtle changes in electrocardiogram waveforms, and furthermore, electrocardiogram interpretation can be improved.

Atrial fibrillation is a major cause of rising morbidity and mortality, and timely diagnosis and treatment are important. AI algorithms are used to improve the performance of diagnostic algorithms using biosignals such as electrocardiograms and oxygen saturation curves, but AI algorithms are not understandable by humans. As a black box that does not explain predictions in a possible way, it has a problem of low reliability due to poor transparency and interpretability.

In addition, in the conventional deep learning structure, since the neural network is a unit, there is a limit in confirming only the contents of the judgment such as the presence or absence of heart disease, the type of heart disease, or the degree of heart disease.

Therefore, since the deep learning model cannot be trusted due to lack of transparency and interpretability, a technology capable of constructing an explainable artificial intelligence model with improved accuracy and interpretability that can present the diagnosis and reason for heart disease together is required. .

Korean Patent Registration No. 10-2156087 (Method for predicting paroxysmal atrial fibrillation in normal sinus rhythm electrocardiogram condition using deep learning, 2020.09.15) Korean Patent Publication No. 10-2021-0066322 (Atrial Fibrillation Prediction System and Prediction Method, 2021.06.07)

The technical problem to be achieved by the spirit of the present invention is to provide diagnosis and reasons for diagnosis of heart disease by an explainable neural network model, thereby increasing the accuracy and interpretability of diagnosis, which can be applied to clinical practice, XAI deep learning-based heart disease diagnosis. and an interpretation system.

In order to achieve the above object, an embodiment of the present invention uses one or more corresponding analysis modules that use heart disease and electrocardiogram data for determining heart disease as learning data and analyze one or more characteristics for determining the diagnosis of heart disease. An XAI deep learning model unit configured to derive the diagnosis of heart disease and the judgment characteristics of the diagnosed heart disease; an input unit receiving electrocardiogram data and providing the data to the XAI deep learning model unit; And an output unit for outputting the diagnosis of heart disease and the characteristics of the diagnosed heart disease by taking the electrocardiogram data from the input unit as input by the XAI deep learning model unit; XAI deep learning-based heart disease diagnosis and analysis system.

Here, the heart disease is atrial fibrillation, and the analysis module may analyze the judgment characteristics of the atrial fibrillation.

In addition, the analysis module combines an irregularity analysis module for analyzing heartbeat irregularities from the electrocardiogram data, a P wave analysis module for analyzing the absence of a P wave, and an output value of the irregularity analysis module and the P wave analysis module. It may be composed of an ensemble module that determines atrial fibrillation.

In addition, the irregularity analysis module or the P-wave analysis module includes first to fourth residual blocks, a first FC layer and a second FC layer, and the first to fourth residual blocks include two convolutions. layer and two batch normalization layers are each repeated, and the fourth residual block is connected to a platen layer, the platen layer is connected to the first FC layer, and the second FC layer is connected to a single output node. The single output node outputs a first probability of heartbeat irregularity and a second probability of absence of P wave, respectively, by a sigmoid activation function, and the ensemble module outputs the first probability and the second probability. By combining the probabilities and outputting the probability of existence of atrial fibrillation, the existence of atrial fibrillation can be predicted, and the predicted judgment characteristics of atrial fibrillation can be analyzed.

Meanwhile, the ECG data may be single-lead, 6-lead, or 12-lead ECG data.

In addition, the input unit may include a pre-processing module for removing noise and normalizing the electrocardiogram data.

In addition, the verification data of the XAI deep learning model unit may be composed of internal verification data extracted from the same diagnostic group as the training data at a certain rate and external verification data extracted from a diagnostic group different from the training data.

In addition, the output unit may provide a treatment corresponding to the characteristics of the diagnosed heart disease.

In addition, the XAI deep learning model unit may be a NBDT model.

According to the present invention, diagnosis of heart disease and reasons for diagnosis are provided by an explainable neural network model to increase the accuracy and interpretability of diagnosis so that it can be applied to clinical practice, and the accuracy of the model can be increased by utilizing internal and external verification data. There is an effect of providing a corresponding treatment according to a diagnosis reason, and diagnosing various heart diseases other than atrial fibrillation using electrocardiogram data.

1 shows a schematic configuration diagram of a heart disease diagnosis and interpretation system based on XAI deep learning according to an embodiment of the present invention.
FIG. 2 illustrates an XAI deep learning model unit of the XAI deep learning based heart disease diagnosis and interpretation system of FIG. 1 .
FIG. 3 illustrates a dataset table of electrocardiogram data for the construction of the XAI deep learning model unit of FIG. 2 .
FIG. 4 is a flowchart of a diagnosis and analysis process by the XAI deep learning-based heart disease diagnosis and interpretation system of FIG. 1 .

Hereinafter, embodiments of the present invention having the above-described characteristics will be described in more detail with reference to the accompanying drawings.

The XAI deep learning-based heart disease diagnosis and interpretation system according to an embodiment of the present invention uses heart disease and electrocardiogram data for determining heart disease as learning data, and analyzes one or more characteristics for determining the diagnosis of heart disease. The XAI deep learning model unit 110, constructed to derive the diagnosis of heart disease and the judgment characteristics of the diagnosed heart disease, with the above analysis module, receives electrocardiogram data and provides it to the XAI deep learning model unit 110 An output unit 130 that outputs the diagnosis of heart disease and the judgment characteristics of the diagnosed heart disease by taking the electrocardiogram data from the input unit 120 as input by the input unit 120 and the XAI deep learning model unit 110 Including, it is a gist to provide diagnosis and diagnosis reasons of heart disease by an explainable neural network model.

Hereinafter, referring to FIGS. 1 to 4, the XAI deep learning-based heart disease diagnosis and interpretation system of the above configuration is specifically described as follows.

First of all, the XAI deep learning model unit 110 uses digitally stored electrocardiogram (ECG) data among heart disease and biosignals for determining heart disease as training data, and uses the diagnosis of heart disease as training data. It is constructed to derive the diagnosis of heart disease and the judgment characteristics of the diagnosed heart disease by having one or more corresponding analysis modules for interpreting one or more characteristics to be judged, and thus determining not only the diagnosis name but also the judgment criteria, reason for judgment, or judgment of the diagnosis of the heart disease. You may be asked to provide evidence.

Here, the XAI deep learning model unit 110 uses single lead I and six leads (I, II, III, VL, VR) as training data for model construction or validation data for validating the model. ,VF) or 12-lead ECG data, etc., can be used in various formats.

In addition, the verification data of the XAI deep learning model unit 110 is composed of internal verification data extracted from the same diagnostic group as the training data at a certain rate and external verification data extracted from a diagnostic group different from the training data, and external verification The overfitting problem can be solved through the data to verify the accuracy of the model.

For example, referring to <Table 1> in FIG. 3, a PTB-XL ECG data set including 21,837 12-lead ECGs with a sampling rate of 400 Hz and 10,605 12-lead ECGs with a sampling rate of 500 Hz are included. The Chapman ECG data set with a sample rate of 300 Hz, the PhysioNet ECG data set containing 8,528 single-lead ECGs with a sampling rate of 300 Hz, and the Sejong ECG data set containing 128,399 12-lead ECGs with a sampling rate of 500 Hz are exemplified. When constructing the model in the example, training data and internal verification data are selected by randomly splitting at a ratio of 9:1 from the population of the Sejong ECG data set, and PTB-XL ECG data set and Chapman ECG data are selected as external verification data. Sets and PhysioNet ECG data sets can be selected.

In this embodiment, the XAI deep learning model unit 110 is a NBDT (a neural network and a decision tree) combined to maintain a high level of interpretability while approaching the accuracy of the neural network. Neural-Backed Decision Tree) model.

Meanwhile, the heart disease is atrial fibrillation, and the analysis module is a module for classifying the characteristics of atrial fibrillation, and can analyze criteria, reasons, or grounds for determining atrial fibrillation.

For reference, atrial fibrillation is an arrhythmia disease that causes irregular pulses in the form of fast pulses due to loss of regular contraction of the atria and irregular tremors. A normal P wave does not appear, only irregular lines appear.

Accordingly, in order to identify the irregularity of heartbeat and the absence of P wave, which are the characteristics of atrial fibrillation, the analysis module, referring to FIG. 2, analyzes the irregularity of heartbeat from electrocardiogram data. A P wave analysis module 112 for analyzing the absence of, and an ensemble module 113 for determining the existence probability of atrial fibrillation by combining the output values of the irregularity analysis module 111 and the P wave analysis module 112. may consist of

Specifically, the irregularity analysis module 111 or the P-wave analysis module 112, respectively, the first residual block to the fourth residual block to make learning easier, convolution and max pooling ( A fully connected layer and a second FC layer that classify electrocardiogram data images into defined labels by taking results of Max Pooling process.

Here, as shown enlarged in FIG. 2, the first to fourth residual blocks have two convolution layers (CONV) and the possibility of performing learning faster or falling into a local optimum problem. Two batch normalization layers (BN) to reduce , may be configured by repeating each.

In addition, the fourth residual block is connected to a flatten layer that converts into a single vector, the flatten layer is connected to the first FC layer, the second FC layer is connected to a single output node, and the single output node is connected to, The first probability of heartbeat irregularity and the second probability of absence of P wave can be output by a sigmoid activation function that determines activation according to the size of the weighted sum for the previous layer. .

Finally, the ensemble module 113 may combine the first probability and the second probability to output a probability of existence of atrial fibrillation, thereby predicting the existence of atrial fibrillation and interpreting characteristics of the predicted atrial fibrillation.

On the other hand, the XAI deep learning model unit 110 is not limited to the aforementioned atrial fibrillation, but is applied to various heart diseases such as cardiomyopathy, congenital heart disease, valve disease, heart failure, pericardial disease, hypertension, arteriosclerosis or coronary artery disease Thus, by extracting and analyzing the shape of electrocardiogram data providing body signals such as 2D diagnostic data, 3D diagnostic data or waveforms, N characteristics for diagnosing heart disease are extracted by an auto encoder, and the characteristics The corresponding diagnosis of heart disease according to the combination may be learned to derive a diagnosis as a prediction result according to electrocardiogram data as a predictor variable.

Next, the input unit 120 receives the electrocardiogram data to predict the diagnosis name and judgment characteristics of heart disease and provides it to the XAI deep learning model unit 110 .

Meanwhile, the input unit 120 may include a pre-processing module 121 that removes noise and normalizes ECG data.

For example, the pre-processing module 121 may remove artifacts by removing 1 second each from the beginning and the end of 12-lead ECG data stored digitally for 10 seconds to secure ECG data of higher quality.

Next, the output unit 130 receives the electrocardiogram data, which is a predictor variable from the input unit 120, as input by the XAI deep learning model unit 110, and determines the diagnosis of heart disease as a prediction result and the diagnosed heart disease By outputting the characteristics, it is possible to confirm the possibility of a diagnosis and the criterion, reason or basis for determining the diagnosis.

That is, the output unit 130 presents whether the heartbeat is regular or the P wave exists together with the diagnosis result of not having atrial fibrillation, or presents the P wave because the heartbeat is irregular along with the diagnosis result of having atrial fibrillation. It can be suggested because it is absent.

In addition, the output unit 130 may provide a treatment corresponding to the judgment characteristic of the diagnosed heart disease, for example, irregular heartbeat or absence of a P wave.

Figure 4 shows a flow chart of the diagnosis and interpretation process by the XAI deep learning-based heart disease diagnosis and interpretation system of Figure 1, with reference to this, briefly described in detail as follows.

First, in the XAI deep learning model unit construction step (S110), digitally stored electrocardiogram data among biosignals that determine heart disease and heart disease are used as learning data, and one or more characteristics for determining the diagnosis of heart disease are analyzed. XAI deep learning model unit 110 is built to derive the diagnosis of heart disease and the characteristics of the diagnosed heart disease by providing one or more analysis modules that Or, the basis for judgment may be presented together.

Here, various types of data such as single lead (lead I), six leads (I, II, III, VL, VR, VF) or 12 lead ECG data are used as learning data for model construction or verification data for validating the model. ECG data are available.

In addition, the verification data is composed of internal verification data extracted at a certain ratio from the same diagnostic group as the learning data and external verification data extracted from a diagnostic group different from the training data, thereby solving the overfitting problem through the external verification data. It can be used to verify the accuracy of the model.

On the other hand, the heart disease is atrial fibrillation, and the analysis module is a module for classifying the characteristics of atrial fibrillation, and it is possible to analyze the criteria, reasons, or judgment characteristics of the basis for the diagnosis of atrial fibrillation.

For reference, atrial fibrillation is an arrhythmia disease that causes irregular pulses in the form of fast pulses due to loss of regular contraction of the atria and irregular tremors. A normal P wave does not appear, only irregular lines appear.

Accordingly, in order to identify the irregularity of heartbeat and the absence of P wave, which are the characteristics of atrial fibrillation, the analysis module, referring to FIG. 2, analyzes the irregularity of heartbeat from electrocardiogram data. It may be composed of a P wave analysis module 112 for analyzing the absence of, and an ensemble module 113 for determining the existence probability of atrial fibrillation by combining the output values of the irregularity analysis module 111 and the P wave analysis module 112. there is.

Specifically, the irregularity analysis module 111 or the P wave analysis module 112 takes the results of the first to fourth residual blocks and the convolution and max pooling processes, respectively, which make learning easier, to obtain an electrocardiogram data image It may include a 1st FC layer and a 2nd FC layer that classify into defined labels.

Here, as shown enlarged in FIG. 2, each of the first to fourth residual blocks includes two convolutional layers and two batch normalization layers to perform learning faster or to reduce the possibility of falling into a local optimization problem. It can be configured repeatedly.

In addition, the fourth residual block is connected to a platen layer that converts into a single vector, the platen layer is connected to the first FC layer, the second FC layer is connected to a single output node, and the single output node is connected to the previous layer. A first probability of heartbeat irregularity and a second probability of absence of a P wave may be respectively output by a sigmoid activation function that determines whether or not to activate according to the magnitude of the weighted sum.

Finally, the ensemble module 113 may combine the first probability and the second probability to output a probability of existence of atrial fibrillation, thereby predicting the existence of atrial fibrillation and interpreting characteristics of the predicted atrial fibrillation.

On the other hand, the XAI deep learning model unit 110 is not limited to the aforementioned atrial fibrillation, but is applied to various heart diseases such as cardiomyopathy, congenital heart disease, valve disease, heart failure, pericardial disease, hypertension, arteriosclerosis or coronary artery disease Thus, by analyzing the electrocardiogram data, N characteristics for diagnosing heart disease are extracted, and the corresponding diagnosis of heart disease according to the combination of characteristics is learned, and the prediction result according to the electrocardiogram data, which is a predictor variable, can be derived. there is.

Thereafter, in the test data input step (S120), electrocardiogram data, which is a predictor variable to predict the diagnosis and judgment characteristics of heart disease, is received and provided to the XAI deep learning model unit 110.

Meanwhile, the test data input step ( S120 ) may further include a preprocessing step ( S121 ) of normalizing and removing noise of the electrocardiogram data by the preprocessing module 121 .

For example, in the preprocessing step (S121), artifacts may be removed by removing 1 second each of the start and end parts of the 12-lead ECG data stored digitally for 10 seconds to secure ECG data of higher quality.

Thereafter, in the prediction result output step (S130), the XAI deep learning model unit 110 uses the electrocardiogram data, which is a predictor variable from the input unit 120, as an input to determine the irregularity of the heartbeat and the absence of a P wave. , the prediction result of the diagnosis of heart disease and the judgment characteristics of the diagnosed heart disease are output to confirm the possibility of the diagnosis and the criterion, reason or basis for determining the diagnosis.

That is, in the prediction result output step (S130), whether the heartbeat is regular or the P wave is presented together with the diagnosis result of no atrial fibrillation, or whether the heartbeat is irregular together with the diagnosis result of atrial fibrillation. It can be suggested that the P wave is absent.

Thereafter, in the treatment providing step (S140), through the output unit 130, a treatment corresponding to the judgment characteristics of the diagnosed heart disease such as atrial fibrillation, for example, irregular heartbeat or absence of a P wave, may be provided. .

In addition, it is possible to provide computer software stored in a computer-readable recording medium that performs the diagnosis and interpretation method by the above-mentioned XAI deep learning-based heart disease diagnosis and interpretation system.

Therefore, by the configuration of the heart disease diagnosis and interpretation system based on XAI deep learning as described above, diagnosis of heart disease and diagnosis reasons are provided by an explainable neural network model, thereby increasing the accuracy and interpretability of diagnosis and applying it to clinical practice. It is possible to increase the accuracy of the model by utilizing internal and external verification data, to provide a corresponding treatment according to the reason for diagnosis, and to diagnose various heart diseases other than atrial fibrillation using electrocardiogram data.

The embodiments described in this specification and the configurations shown in the drawings are only one of the most preferred embodiments of the present invention, and do not represent all of the technical spirit of the present invention, so various equivalents that can replace them at the time of this application. It should be understood that there may be waters and variations.

110: XAI deep learning model unit 111: irregularity analysis module
112: P wave analysis module 113: ensemble module
120: input unit 121: preprocessing module
130: output unit
S110: XAI deep learning model building step
S120: Test data input step
S121: pre-processing step
S130: Prediction result output step
S140: Treatment provision step

Claims (9)

Heart disease and electrocardiogram data for determining heart disease are used as learning data, and one or more corresponding analysis modules are provided to interpret one or more characteristics for determining the diagnosis of heart disease, thereby determining the diagnosis of heart disease and the diagnosed heart disease. Built to derive the characteristics, XAI deep learning model unit;
an input unit receiving electrocardiogram data and providing the data to the XAI deep learning model unit; and
An output unit for outputting a diagnosis name of a heart disease and a judgment characteristic of the diagnosed heart disease by taking the electrocardiogram data from the input unit as an input by the XAI deep learning model unit;
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the heart disease is atrial fibrillation, and the analysis module analyzes the judgment characteristics of the atrial fibrillation,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 2,
The analysis module combines an irregularity analysis module for analyzing heartbeat irregularities from the electrocardiogram data, a P wave analysis module for analyzing the absence of a P wave, and output values of the irregularity analysis module and the P wave analysis module to determine atrial fibrillation. Characterized in that it consists of an ensemble module that determines
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 3,
The irregularity analysis module or the P-wave analysis module includes a first residual block to a fourth residual block, a first FC layer and a second FC layer,
Two convolutional layers and two batch normalization layers are repeated in the first to fourth residual blocks, respectively, and the fourth residual block is connected to a platen layer, and the platen layer is connected to the first FC layer. , and the second FC layer is connected to a single output node, and the single output node outputs a first probability of heartbeat irregularity and a second probability of absence of P wave, respectively, by a sigmoid activation function. do,
The ensemble module combines the first probability and the second probability to output the existence probability of atrial fibrillation,
Characterized in that predicting the presence of the atrial fibrillation and interpreting the judgment characteristics of the predicted atrial fibrillation,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the electrocardiogram data is single-lead, 6-lead or 12-lead ECG data,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the input unit includes a pre-processing module for removing noise and normalizing the electrocardiogram data,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the verification data of the XAI deep learning model unit consists of internal verification data extracted at a certain rate from the same diagnostic group as the training data and external verification data extracted from a diagnostic group different from the training data,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the output unit provides a treatment corresponding to the judgment characteristics of the diagnosed heart disease,
XAI deep learning based heart disease diagnosis and interpretation system.
According to claim 1,
Characterized in that the XAI deep learning model unit is an NBDT model,
XAI deep learning based heart disease diagnosis and interpretation system.

Images