KR102593035B1 - Method and apparatus for diagnosing epilepsy based on deep learning - Google Patents

Abstract

The present invention discloses a deep learning-based epilepsy seizure diagnosis method and device. According to the invention, a processor; and a memory connected to the processor, preprocessing electroencephalogram signals obtained from a plurality of regions from the user, and extracting a first feature vector of the preprocessed electroencephalogram signal through a time-space-frequency network based on a convolutional neural network. Then, through a causal convolution layer, the first feature vectors of previous time points are input from the current time point t at which an epileptic seizure is to be diagnosed, and the second feature vectors of each time period from the previous time point to the current time are extracted, and extended synthesis is performed. Deep extracts a third feature vector at a time after the current point through a product layer, and stores program instructions executed by the processor to determine whether or not an epileptic seizure is present using the extracted third feature vector. A running-based epilepsy seizure diagnosis device is provided.

Description

Deep learning-based epilepsy seizure diagnosis method and apparatus {Method and apparatus for diagnosing epilepsy based on deep learning}

The present invention relates to a deep learning-based epilepsy seizure diagnosis method and device.

Epilepsy is one of the three major chronic brain diseases along with stroke and dementia. Recently, research has been actively conducted to diagnose seizures through analysis of electroencephalogram (EEG) signals.

In particular, the diagnosis of epilepsy seizures through analysis of scalp electroencephalogram signals has the advantage of being able to be detected in real time in everyday life, so it has high technical value.

However, scalp electroencephalogram signals are subject to difficulties such as low performance when applying general classification methods due to noise involved in the measurement process.

Accordingly, research is being conducted to find patterns in electroencephalogram signals using machine learning models, but there is a limitation of generally having a delay of about 4 to 5 seconds before diagnosing a seizure.

Recently, much progress has been made in learning patterns of time-space-frequency characteristics of electroencephalogram signals through deep learning techniques.

Recently, the temporal convolution model, one of the techniques for generating and classifying signals by reflecting the patterns of previous points in time, has shown good results in research dealing with time series data such as audio synthesis.

There are several medical stages before a seizure occurs in an epilepsy patient, so a diagnostic method that takes these into account is needed.

Accordingly, research is needed to reduce the delay in diagnosing seizures in epilepsy patients through causal modeling using a temporal convolution model.

This is a conventional technology that uses machine learning or a probability model to diagnose epileptic seizures using electroencephalogram signals. Korean Patent Registration No. 10-2141185 (title of the invention: electroencephalogram signals using feature extraction along with probability model and machine learning) There is a multi-frequency band coefficient-based epileptic seizure wave detection system. The above-mentioned patent decomposes the frequency band from the electroencephalogram signal into multiple bands, extracts signal coefficients for each band, generates a feature vector from the extracted signal coefficients, and uses this to perform various classifiers (e.g. SVM (Support vector machine), LDA (Linear discriminant analysis). ), etc.), etc.) are proposed to generate and detect seizure conditions, but there is a problem in that the classifier is trained without considering temporal-spatial information, which is one of the important characteristics of EEG signals.

In order to solve the problems of the prior art described above, the present invention uses a deep learning-based causal relationship model rather than a diagnosis system through various machine learning and probability models to extract vectors with characteristics of time, space, and frequency from the electroencephalogram signal. We would like to extract and ultimately propose a deep learning-based epilepsy seizure diagnosis method and device that is more accurate, reduces latency, and is more reliable.

In order to achieve the above-described object, according to an embodiment of the present invention, there is provided a deep learning-based epilepsy seizure diagnosis device, comprising: a processor; and a memory connected to the processor, preprocessing electroencephalogram signals obtained from a plurality of regions from the user, and extracting a first feature vector of the preprocessed electroencephalogram signal through a time-space-frequency network based on a convolutional neural network. Then, through a causal convolution layer, the first feature vectors of previous time points are input from the current time point t at which an epileptic seizure is to be diagnosed, and the second feature vectors of each time period from the previous time point to the current time are extracted, and extended synthesis is performed. Deep extracts a third feature vector at a time after the current point through a product layer, and stores program instructions executed by the processor to determine whether or not an epileptic seizure is present using the extracted third feature vector. A running-based epilepsy seizure diagnosis device is provided.

The program instructions perform band-pass filtering to pass a frequency band of interest among the electroencephalogram signals obtained from the plurality of regions, remove noise from the band-pass filtered signal, and perform data normalization from the noise-removed signal. can do.

The program instructions extract features at a time after the current point through a convolution filter with an expansion ratio of 2 or more, nonlinearize the extracted features through an activation function, and perform the output of the activation function and the causal convolution. The third feature vector can be extracted by adding the residuals between the outputs of the layers.

According to another aspect of the present invention, there is a deep learning-based epilepsy seizure diagnosis device, comprising: a pre-processing unit for pre-processing electroencephalogram signals obtained from a plurality of regions from a user; a feature extraction unit that extracts a first feature vector of the preprocessed electroencephalogram signal through a convolutional neural network-based time-space-frequency network; Through a causal convolution layer, the first feature vectors of previous time points are input from the current time point t at which an epileptic seizure is to be diagnosed, and the second feature vectors of each time period from the previous time point to the current time are extracted, and the extended convolution layer a causal neural network that extracts a third feature vector at a time after the current point through a; and a class classification learning unit that determines whether an epileptic seizure exists using the extracted third feature vector. A deep learning-based epilepsy seizure diagnosis device is provided.

According to another aspect of the present invention, there is a deep learning-based epilepsy seizure diagnosis method including a processor and memory, comprising the steps of preprocessing electroencephalogram signals obtained from a plurality of regions from a user; Extracting a first feature vector of the preprocessed electroencephalogram signal through a convolutional neural network-based time-space-frequency network; Receiving first feature vectors of previous time points from the current time point t at which epileptic seizures are to be diagnosed through a causal convolution layer, and extracting second feature vectors for each time period from the previous time point to the current time point; extracting a third feature vector at a time after the current point through an extended convolution layer; A deep learning-based epilepsy seizure diagnosis method is provided, including the step of determining whether an epileptic seizure exists using the extracted third feature vector.

According to the present invention, feature vectors are extracted based on the time-space-frequency information of the electroencephalogram signal and used to diagnose seizures in epilepsy patients, thereby preventing serious injuries through precise diagnosis and early prediction of seizures in epilepsy patients. and has the advantage of being able to take appropriate action.

Figure 1 is a diagram showing the configuration of a deep learning-based epilepsy seizure diagnosis device according to an embodiment of the present invention.
Figure 2 is a diagram showing the processing process of the preprocessor according to this embodiment.
Figure 3 is a diagram showing the detailed configuration of a causal neural network according to this embodiment.
Figure 4 is a diagram showing the detailed configuration of a causal convolution layer according to this embodiment.
Figure 5 is a diagram to conceptually explain the expansion ratio of the extended convolution layer according to this embodiment.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention.

This embodiment uses a deep learning-based causal model to consider time-space-frequency characteristics in the process of learning the distribution of electroencephalogram signal data to improve seizure diagnosis performance and reduce delay time for epilepsy patients. Extract vector.

The generated feature vector learns a classifier that diagnoses the current condition by considering previous patterns of epileptic seizures in chronological order, ultimately improving classification accuracy and minimizing delay rate.

Figure 1 is a diagram showing the configuration of a deep learning-based epilepsy seizure diagnosis device according to an embodiment of the present invention.

As shown in FIG. 1, the device according to this embodiment may include a preprocessor 100, a feature extraction unit 102, a causal neural network 104, and a class classification learning unit 106.

The preprocessor 100 removes noise from the EEG signal obtained from multiple parts of the user, amplifies the signal, and rectifies the distribution.

Figure 2 is a diagram showing the processing process of the preprocessor according to this embodiment.

As shown in FIG. 2, the preprocessor 100 according to this embodiment performs band-pass filtering to leave the 0.5 to 100 Hz component, which is the frequency band of interest, in the EEG signal (step 200).

Next, the preprocessor 100 performs noise removal filtering to remove noise generated by the device during the EEG signal acquisition process (step 202).

Lastly, in order to match the characteristic range of the acquired EEG signal data, data normalization is performed by calculating the average and variance, adding and dividing them respectively (step 204).

The feature extractor 102 according to this embodiment is composed of a time-space-frequency network, and extracts a first feature vector related to correlation, space, and frequency between a plurality of channels over a short period of time.

Here, the short time may be in the range of several milliseconds.

The feature extraction unit 102 takes into account the characteristics of the EEG signal, considers information in adjacent time periods, and simultaneously extracts a first feature vector regarding the multiple channels and frequencies of the EEG signal obtained from different regions.

The time-space-frequency network according to this embodiment is a convolutional neural network model designed in consideration of the characteristics of the electroencephalogram signal. In order to extract the optimal feature vector from continuous multivariate time series data, a convolution operation is first performed on the time axis. Features are extracted between adjacent times, and a convolution operation is sequentially performed on the channel axis to extract a first feature vector that takes into account the correlation between the plurality of channels of the electroencephalogram signal.

Here, the plurality of channels correspond to different areas for acquiring electroencephalogram signals.

Figure 3 is a diagram showing the detailed configuration of a causal neural network according to this embodiment.

In general, in the case of epileptic seizures, symptoms occur suddenly, but in terms of electroencephalogram signals, aura occurs before the time of seizure, and a causal neural network 104 is used to capture changes in these signals.

Referring to FIG. 3, the causal neural network 104 according to this embodiment may include a causal convolution layer 300, an extended convolution layer 302, an activation function 304, and an adder 306.

Figure 4 is a diagram showing the detailed configuration of a causal convolution layer according to this embodiment.

Referring to FIG. 4, the causal convolution layer 300 simultaneously receives feature vectors of previous time points from time point t (the current time point) at which an epileptic seizure is to be diagnosed, and considers the features of previous time points to determine the second feature of each time period. Extract vector.

The causal convolution layer 300 according to this embodiment is one of the characteristics of epilepsy as a convolutional neural network model designed to satisfy the conditional probability formula that the predicted value in the current time is reflected by the information in the past time. By capturing phenomena that occur before a seizure, it is possible to determine whether a seizure is occurring, thereby reducing the delay in diagnosis.

The extended convolution layer 302 receives the output (second feature vector) of the causal convolution layer 300 and extracts features at a long time after time t through a convolution filter with an expansion ratio of 2 or more.

The activation function 304 non-linearizes the features extracted from the extended convolution layer 302, and the residual between the output of the activation function 304 and the output from the causal convolution layer 300 is sent to the adder 306. By adding, multi-sized features (third feature vector) are extracted.

Figure 5 is a diagram to conceptually explain the expansion ratio of the extended convolution layer according to this embodiment.

The extended convolution layer according to this embodiment extracts features by widening the convolution filter by a preset expansion ratio in order to have a wider receptive field than the existing convolution. Through this, information of a past time longer than EEG signals 1 to t can be efficiently extracted.

Referring again to FIG. 1, the class classification learning unit 106 serves to distinguish which state is currently represented through the generated feature vector. Through this, the patient's current condition is diagnosed and the probability of a seizure occurring is displayed.

The third feature vector extracted through the causal neural network 104 is input to the class classification learning unit 106, and the class classification learning unit 106 diagnoses the current state of the patient through the feature vector. When diagnosing, the probability of a seizure occurring is continuous, and the change in probability from before the seizure occurs until the seizure occurs and ends can be confirmed.

The deep learning-based epilepsy seizure diagnosis device according to this embodiment may include a processor and memory.

Here, the processor may include a central processing unit (CPU) capable of executing a computer program or another virtual machine. Memory may include non-volatile storage devices, such as non-removable hard drives or removable storage devices. Removable storage devices may include compact flash units, USB memory sticks, etc. Memory may also include volatile memory, such as various types of random access memory.

The memory according to this embodiment stores program instructions executed by the processor.

Program instructions according to this embodiment preprocess electroencephalogram signals obtained from a plurality of regions from the user, extract a first feature vector of the preprocessed electroencephalogram signal through a time-space-frequency network based on a convolutional neural network, and Through a causal convolution layer, the first feature vectors of previous time points are input from the current time point t at which an epileptic seizure is to be diagnosed, and the second feature vectors of each time period from the previous time point to the current time are extracted, and the extended convolution layer A third feature vector at a time after the current point is extracted through and the presence or absence of an epileptic seizure is determined using the extracted third feature vector.

Through the above process, precise diagnosis and early prediction of seizures in epilepsy patients are possible.

The above-described embodiments of the present invention have been disclosed for illustrative purposes, and those skilled in the art will be able to make various modifications, changes, and additions within the spirit and scope of the present invention, and such modifications, changes, and additions will be possible. should be regarded as falling within the scope of the patent claims below.

Claims (5)

As a deep learning-based epilepsy seizure diagnosis device,
processor; and
Including a memory connected to the processor,
Preprocess the electroencephalogram signals obtained from multiple areas from the user,
Extracting the first feature vector of the preprocessed electroencephalogram signal through a convolutional neural network-based time-space-frequency network,
Through a causal convolution layer, receive first feature vectors of previous time points from the current time point t at which epileptic seizures are to be diagnosed, and extract second feature vectors for each time period from the previous time point to the current time point,
Extracting a third feature vector at a time after the current point through an extended convolution layer,
To determine whether there is an epilepsy seizure using the extracted third feature vector,
Store program instructions executed by the processor,
The program commands are:
Extract features at a time after the current point through a convolution filter with an expansion ratio of 2 or more, non-linearize the extracted features through an activation function, and between the output of the activation function and the output from the causal convolution layer. A deep learning-based epilepsy seizure diagnosis device that extracts the third feature vector by adding the residuals of . According to paragraph 1,
The program commands are:
Performing band-pass filtering to pass the frequency band of interest among the electroencephalogram signals obtained from the plurality of regions,
Remove noise from the band-pass filtered signal,
A deep learning-based epilepsy seizure diagnosis device that performs data normalization on the noise-removed signal. delete As a deep learning-based epilepsy seizure diagnosis device,
A pre-processing unit that pre-processes electroencephalogram signals obtained from a plurality of regions from the user;
a feature extraction unit that extracts a first feature vector of the preprocessed electroencephalogram signal through a convolutional neural network-based time-space-frequency network;
Through a causal convolution layer, the first feature vectors of previous time points are input from the current time point t at which an epileptic seizure is to be diagnosed, and the second feature vectors of each time period from the previous time point to the current time are extracted, and the extended convolution layer a causal neural network that extracts a third feature vector at a time after the current point through a; and
A class classification learning unit that determines whether an epileptic seizure exists using the extracted third feature vector,
The causal neural network extracts features at a time after the current point through a convolution filter with an expansion ratio of 2 or more, nonlinearizes the extracted features through an activation function, and combines the output of the activation function with the causal convolution. A deep learning-based epilepsy seizure diagnosis device that extracts the third feature vector by adding residuals between outputs in the hierarchy. A method for diagnosing epileptic seizures based on deep learning in a device including a processor and memory, comprising:
Preprocessing electroencephalogram signals obtained from a plurality of regions from the user;
Extracting a first feature vector of the preprocessed electroencephalogram signal through a convolutional neural network-based time-space-frequency network;
Receiving first feature vectors of previous time points from the current time point t at which epileptic seizures are to be diagnosed through a causal convolution layer, and extracting second feature vectors for each time period from the previous time point to the current time point;
extracting a third feature vector at a time after the current point through an extended convolution layer; and
A step of determining whether an epileptic seizure exists using the extracted third feature vector,
The step of extracting the third feature vector is,
Extracting features at a time after the current time through a convolution filter with an expansion ratio of 2 or more;
Non-linearizing the extracted features through an activation function; and
A deep learning-based epilepsy seizure diagnosis method comprising extracting the third feature vector by adding a residual between the output of the activation function and the output of the causal convolution layer.