Automatic disease classification method using wavelet function-applied convolutional neural network

Abstract

The present invention relates to an automatic disease classification method using a wavelet function-applied convolutional neural network. According to the present invention, the automatic disease classification method comprises the steps: acquiring a bio-signal from a sensor attached to a part of the body of a patient; acquiring a wavelet-transformed bio-signal by inputting the acquired bio-signal into a first learning model based on wavelet transform; extracting a feature value, by inputting the acquired wavelet-transformed bio-signal into a second learning model to analyze a morphological feature included in the bio-signal; and classifying a type of a disease by inputting the extracted feature value into a third learning model based on a decision-making algorithm.

Description

Automatic disease classification method using a convolutional neural network with wavelet functions

[0001] The present invention relates to an automatic disease classification method using a convolutional neural network applying a wavelet function, and more particularly, to an automatic disease classification method for extracting a wavelet-transformed biosignal by applying a compressed or expanded wavelet function to a biosignal input through a wavelet transform-based convolutional neural network, and inputting the wavelet-transformed biosignal into a convolutional neural network to classify a disease type.

[0002] Arrhythmias are abnormal heart rhythms, irregular heartbeats (tachycardia or bradycardia), or insufficient cardiac output, resulting in abnormal conduction of cardiac action potentials along abnormal pathways or sequences. Because these conditions can lead to acute cardiovascular disease or chronic complications, appropriate management and monitoring are essential.

[0003] Analysis of the presence and type of arrhythmia is possible by analyzing the form of the electrocardiogram (ECG), and in particular, with the recent development of deep neural networks (DNN), automatic arrhythmia classification techniques utilizing them have been proposed.

[0004] Among these, convolutional neural networks (CNNs) are primarily utilized, demonstrating superior performance due to their advantages in analyzing patterns in time-series inputs. However, these DNN / CNN-based arrhythmia classification techniques suffer from a "black-box" issue, limiting their practical application in clinical settings.

[0005] The technology underlying the present invention is disclosed in Republic of Korea Patent No. 10-2163217 (announced on October 8, 2020).

[0006] The technical problem to be achieved by the present invention is to provide an automatic disease classification method for extracting a wavelet-transformed biosignal by applying a compressed or expanded wavelet function to a biosignal input through a wavelet-transform-based convolutional neural network, and classifying a disease type by inputting the wavelet-transformed biosignal into the convolutional neural network.

[0007] According to an embodiment of the present invention for achieving such a technical task, an automatic disease classification method includes a step of obtaining a biosignal from a sensor attached to a part of a patient's body, a step of inputting the obtained biosignal into a first learning model based on wavelet transform to obtain a wavelet-transformed biosignal, a step of inputting the obtained wavelet-transformed biosignal into a second learning model to analyze morphological features included in the biosignal and extract feature values, and a step of inputting the extracted feature values ​​into a third learning model based on a decision-making algorithm to classify a type of disease.

[0008] The above biosignal may include at least one of electrocardiogram (ECG) data, photoplethysmogram (PPG) data, respiratory signal, electromyogram (EMG) data, electroencephalogram (EEG) data, and electrogastrogram (EGG) data.

[0009] The above first learning model can generate various wavelet functions by applying a predefined mother wavelet to an input biosignal to compress or expand it, and perform a cross-correlation operation between the generated wavelet function and the input biosignal to output a wavelet-transformed biosignal.

[0010] The above first learning model is composed of wavelet kernels each having a different scale, and the wavelet kernels can be determined through the following mathematical formula.

[0011]

[0012] Here, Ψ is the mother wavelet. is defined within, where l represents the length of the kernel, and a i represents the scale factor, which is a unitless parameter.

[0013] The above first learning model can output a wavelet-transformed biosignal using the following mathematical formula.

[0014]

[0015] Here, x represents the input data.

[0016] In this way, according to the present invention, by providing a disease classification model based on wavelet (WavelNet) transformation, it is possible to resolve the black-box issue, which is a chronic problem of conventional deep neural network (DNN)-based solutions, thereby enabling clinical application.

[0017] In addition, the disease classification model based on wavelet (WavelNet) transform does not require a separate preprocessing process, is not restricted by the frequency domain of the input signal or the sampling frequency of the sensor, and has variable time-frequency resolution and flexibility in selecting the mother wavelet, so it can be widely used in the analysis of not only electrocardiogram (ECG), but also photoplethysmogram (PPG), respiratory signal, electromyogram (EMG), electroencephalogram (EEG), and electrogastrogram (EGG).

[0018] FIG. 1 is a schematic diagram illustrating an automatic disease classification device according to an embodiment of the present invention.

[0019] FIG. 2 is a flowchart for explaining an automatic disease classification method using an automatic disease classification device according to an embodiment of the present invention.

[0020] Figure 3 is a drawing for explaining the automatic disease classification method illustrated in Figure 2.

[0021] Figure 4 is a flowchart for explaining step S220 illustrated in Figure 2.

[0022] Figure 5 is an exemplary diagram for explaining step S222 shown in Figure 4.

[0023] FIG. 6 is a diagram showing the results of a performance comparison between a wavelet-based arrhythmia classification model applied to an embodiment of the present invention and a conventional CNN-based arrhythmia classification model.

[0024] Below, with reference to the attached drawings, embodiments of the present invention are described in detail so that those skilled in the art can easily implement them. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in the drawings, parts irrelevant to the description are omitted for clarity of description, and similar parts are designated with similar reference numerals throughout the specification.

[0025] Throughout the specification, whenever a part is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise stated.

[0026] Then, with reference to the attached drawings, an embodiment of the present invention will be described in detail so that a person having ordinary knowledge in the technical field to which the present invention pertains can easily carry out the present invention.

[0027] Hereinafter, an automatic disease classification device according to an embodiment of the present invention will be described in more detail using FIG. 1.

[0028] FIG. 1 is a schematic diagram illustrating an automatic disease classification device according to an embodiment of the present invention.

[0029] As illustrated in FIG. 1, an automatic disease classification device (100) according to an embodiment of the present invention includes a data receiving unit (110), a wavelet transform unit (120), a feature value extraction unit (130), and a disease classification unit (140).

[0030] First, the data receiving unit (110) receives a biosignal. Here, the biosignal includes at least one of electrocardiogram (ECG) data, photoplethysmogram (PPG) data, respiratory signal, electromyogram (EMG) data, electroencephalogram (EEG) data, and electrogastrogram (EGG) data.

[0031] Next, the wavelet transform unit (120) constructs a first learning model based on wavelet transform, inputs the received biosignal into the constructed first learning model, and outputs a biosignal to which wavelet transform has been applied.

[0032] The feature value extraction unit (130) inputs the biosignal to which the wavelet transformation has been applied into the second learning model. Then, the second learning model outputs feature values ​​from the input biosignal.

[0033] Finally, the disease classification unit (140) inputs the output feature values ​​into a third learning model based on a decision-making algorithm to classify the type of disease.

[0034] Hereinafter, an automatic disease classification method using an automatic disease classification device (100) according to an embodiment of the present invention will be described in more detail using FIGS. 2 to 5. However, in order to increase understanding of the embodiment of the present invention, an example will be described of a method of setting electrocardiogram (ECG) data as a biosignal among a plurality of biosignals and classifying arrhythmia according to the input electrocardiogram data. However, the present invention is not necessarily limited thereto, and may be applied to automatically classify various diseases using a biosignal including at least one of electrocardiogram (ECG) data, photoplethysmogram (PPG) data, respiratory signal, electromyogram (EMG) data, electroencephalogram (EEG) data, and electrogastrogram (EGG) data.

[0035] FIG. 2 is a flowchart for explaining an automatic disease classification method using an automatic disease classification device according to an embodiment of the present invention, and FIG. 3 is a drawing for explaining the automatic disease classification method illustrated in FIG. 2.

[0036] Below, as shown in FIGS. 2 and 3, the automatic disease classification device (100) according to an embodiment of the present invention receives electrocardiogram data (S210).

[0037] The examiner attaches multiple electrodes to the patient's arms, legs, and chest, and then measures the magnitude and direction of the current. Specifically, the examiner attaches electrodes to the patient's body in response to three standard limb leads, three amplified limb leads, and six precordial leads, acquiring electrocardiogram waveforms measured in 12 leads. The electrocardiogram waveform indicates the direction and amplitude of the QRS rhythm.

[0038] However, as the smart health device market grows, electrocardiogram (ECG) measurements have become possible through a minimal number of channels, in addition to standard ECG tests. Therefore, the present invention utilizes two or more channels to receive ECG data from the examiner or the relevant patient.

[0039] Next, the wavelet transform unit (120) inputs the received electrocardiogram data into the first learning model to obtain wavelet-transformed electrocardiogram data (S220).

[0040] Figure 4 is a flowchart for explaining step S220 illustrated in Figure 2.

[0041] As shown in Fig. 4, the wavelet transform unit (120) generates a dataset (S221).

[0042] For example, the wavelet transform unit (120) classifies electrocardiogram data obtained from multiple subjects into five grades as described in Table 1 below.

[0043]

[0044] The wavelet transform unit (120) classifies a plurality of electrocardiogram data classified into five classes into training, verification, and test sets, respectively.

[0045] Next, the wavelet transform unit (120) constructs a first learning model based on wavelets and trains the first learning model using the generated training data set (S222).

[0046] Figure 5 is an exemplary diagram for explaining step S222 shown in Figure 4.

[0047] As illustrated in FIG. 5, the wavelet transform unit (120) adopts various mother wavelets using three models of the Daubechies wavelet family (Db2, Db4, Db6) and three models of the Symlet wavelet family (Sym2, Sym4, Sym6), and then constructs a first learning model based on the adopted mother wavelets.

[0048] Meanwhile, in the embodiment of the present invention, a mother wavelet extended by the length of the kernel is adopted, with the values ​​limited to 0 to 1. At this time, the kernel size is calculated using the following mathematical expression 1.

[0049]

[0050] Here, Ψ is the mother wavelet. is defined within, where l represents the length of the kernel, and a i represents the scale factor, which is a unitless parameter.

[0051] Meanwhile, when the mother wavelet is expanded, i.e., when the scale factor is greater than 1, the scaled wavelet can be trimmed. On the other hand, when the mother wavelet is compressed, the scaled wavelet can retain the characteristics of the mother wavelet. Therefore, the scale factor is limited to values ​​between 0 and 1 for compressing the mother wavelet.

[0052] Once the kernel size is determined, the wavelet transform unit (120) causes the first learning model to calculate the cross-correlation between the input value and the kernel.

[0053] That is, the first learning model based on wavelet transform is trained using a training data set generated to output wavelet-transformed electrocardiogram data using the following mathematical expression 2.

[0054]

[0055] Here, x represents the input data.

[0056] After step S222 is completed, the wavelet transform unit (120) inputs the electrocardiogram data received in step S210 into the first learning model to obtain wavelet-transformed electrocardiogram data (S223).

[0057] The wavelet transform unit (120) inputs the received electrocardiogram data into the first learning model. Then, the first learning model applies a predefined mother wavelet to the input biosignal to generate various wavelet functions that are compressed or expanded, and performs a cross-correlation operation between the generated wavelet function and the input biosignal to output wavelet-transformed electrocardiogram data.

[0058] Next, the feature value extraction unit (130) extracts feature values ​​from the wavelet-transformed electrocardiogram data (S230).

[0059] The feature value extraction unit (130) constructs a second learning model to further analyze the wavelet-transformed electrocardiogram data, and inputs the wavelet-transformed electrocardiogram data into the constructed second learning model to extract feature values ​​from the wavelet-transformed electrocardiogram data.

[0060] To elaborate, the waveform in electrocardiogram data largely includes the P wave, QRS complex, and T wave, and the intervals are composed of the PR interval and the QT interval. Arrhythmia refers to a condition in which the pulse is not regular or the frequency deviates from the normal range, and this can be output in the form of irregular output of the P wave, QRS complex, and T wave, or shortened or lengthened PR or QT interval.

[0061] Therefore, the second learning model extracts feature values ​​by comparing wavelet-transformed electrocardiogram data with standard electrocardiogram data.

[0062] Finally, the disease classification unit (140) classifies the type of arrhythmia using the extracted feature values ​​(S240).

[0063] The disease classification unit (140) constructs a third learning model based on a decision-making algorithm and inputs feature values ​​into the constructed third learning model. The third learning model then classifies the presence or absence of arrhythmia and the type of arrhythmia based on the input feature values ​​and outputs the results.

[0064] FIG. 6 is a diagram showing the results of a performance comparison between a wavelet-based arrhythmia classification model applied to an embodiment of the present invention and a conventional CNN-based arrhythmia classification model.

[0065] Figure 6 compares the performance of six different wavelet (WavelNet)-based arrhythmia classification techniques developed using Daubechies-2,4,6, and Symlet-2,4,6 with the existing CNN-based arrhythmia classification technique. Most of the wavelet (WavelNet)-based arrhythmia classification techniques outperformed the existing CNN-based arrhythmia classification techniques, and among them, the technique utilizing the Symlet-4 mother wavelet showed the best performance. Specifically, the sensitivities of the WavelNet-based model (adopting Symlet-4) for non-ectopic beat, supraventricular ectopic beat, and ventricular ectopic beat classification were 91.4%, 49.3%, and 91.4%, respectively, which significantly improved the performance of the CNN-based model (87.3%, 14.9%, and 83.3%, respectively). The positive rate and false positive rate also showed improved performance. The overall accuracy also showed excellent performance, ranging from 83.7% (CNN model) to 89.1% (WavelNet model).

[0066] In this way, the automatic disease classification method according to the present invention can show superior performance compared to conventional CNN-based solutions by providing a disease classification model based on wavelet (WavelNet) transform.

[0067] Furthermore, by observing the learned parameter, scale factor AI, we can interpret which frequency ranges the wavelet transform-based disease classification model (WavelNet) has learned to primarily analyze for accurate disease classification. This approach addresses the "black-box" issue, a persistent problem with conventional deep neural network (DNN)-based solutions, thereby enabling clinical application.

[0068] In addition, the automatic disease classification method according to the present invention is a disease classification model based on wavelet (WavelNet) transform that does not require a separate preprocessing process, is not restricted by the frequency domain of the input signal or the sampling frequency of the sensor, and has variable time-frequency resolution and flexibility in selecting a mother wavelet, so that it can be widely used in the analysis of not only electrocardiogram (ECG), but also photoplethysmogram (PPG), respiratory signal, electromyogram (EMG), electroencephalogram (EEG), electrogastogram (EGG), etc.

[0069]

[0070] While the present invention has been described with reference to the embodiments illustrated in the drawings, these are merely exemplary, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible. Therefore, the true scope of technical protection of the present invention should be determined by the technical spirit of the appended claims.

[0071] <Explanation of symbols>

[0072] 100: Automatic disease classification device

[0073] 110: Data receiving unit

[0074] 120: Wavelet transform section

[0075] 130: Feature value extraction unit

[0076] 140: Disease Classification Department

Claims

1. A step of acquiring biosignals from a sensor attached to a part of the patient's body; A step of inputting the acquired biosignal into a first learning model based on wavelet transform to obtain a wavelet-transformed biosignal, A step of inputting the acquired wavelet-transformed biosignal into a second learning model, analyzing the morphological features included in the biosignal, and extracting feature values; and An automatic disease classification method including a step of classifying the type of disease by inputting the extracted feature values ​​into a third learning model based on a decision-making algorithm.

2. In paragraph 1, The above biosignals are, An automatic disease classification method comprising at least one of electrocardiogram (ECG) data, photoplethysmogram (PPG) data, respiratory signal, electromyogram (EMG) data, electroencephalogram (EEG) data, and electrogastric (EGG) data.

3. In paragraph 1, The above first learning model is, Generate various wavelet functions by compressing or expanding the input biosignal by applying a predefined mother wavelet, An automatic disease classification method for performing a cross-correlation operation between the generated wavelet function and an input biosignal to output a wavelet-transformed biosignal.

4. In paragraph 3, The above first learning model is, It consists of wavelet kernels, each with a different scale. The above wavelet kernel is an automatic disease classification method determined by the following mathematical formula: Here, Ψ is the mother wavelet. is defined within, where l represents the length of the kernel, and a i represents the scale factor, which is a unitless parameter.

5. In paragraph 4, The above first learning model is, An automatic disease classification method that outputs a wavelet-transformed biosignal using the following mathematical formula: Here, x represents the input data.

Images