A lightweight four-classification method for epileptic EEG signals

Abstract

The present application discloses a lightweight four-classification method for epilepsy EEG signals, which relates to the field of EEG signal processing technology. The method collects the user's EEG signals; integrates short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct an EEG signal four-classification model; based on the trained EEG signal four-classification model, the user's EEG signal is used as input, and the probability value of the EEG signal belonging to each category is output; the probability values ​​of the four categories are compared, and the category with the largest probability value is used as the category to which the EEG signal belongs, and a corresponding alarm is sent according to the category. Compared with traditional two-classification and three-classification models, the EEG signal four-classification model of the present application incorporates the characteristics of the late stage of the seizure into the scope of recognition, thereby improving the model accuracy. The lightweight design based on the spectrum reconstruction strategy enables the lightweight four-classification method for epilepsy EEG signals of the present application to be applied to edge wearable device scenarios.

Description

Technical Field

[0001] The present application relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to a lightweight four-classification method for epileptic EEG signals. Background Art

[0002] Epilepsy, a globally widespread neurological disease, has a complex and unpredictable pathogenesis, posing significant challenges to patients' daily lives. In recent years, epileptic seizure monitoring methods based on EEG signals have become a research hotspot, aiming to achieve early warning and accurate classification of seizures through intelligent means. Numerous methods for epilepsy prediction and detection have been proposed. For example, some researchers have explored a stacked one-dimensional convolutional neural network model that combines random selection and data augmentation strategies. Others have achieved automatic prediction of epileptic seizures using a bidirectional long short-term memory network model based on spectral features. Many researchers have also used fast Fourier transforms to extract frequency domain information from patients' EEG signals. However, EEG signals are non-stationary, and fast Fourier transforms have limitations in processing such signals. To address this, some researchers have proposed converting patients' EEG signals from the time domain to the time-frequency domain using methods such as discrete wavelet transforms and short-time Fourier transforms. This not only captures both the temporal and frequency characteristics of the signals but also enhances the suitability of these features for convolutional neural networks.

[0003] However, since the convolution kernel cannot extend to areas beyond the edge, edge pixels can never be located at the center of the kernel function, resulting in significant attenuation of information at the edge of the image in the output feature map. To compensate for this edge effect, the conventional approach is to introduce a zero-padding strategy to expand the effective area that the convolution kernel can reach. However, zero-padding increases the complexity of hardware design, especially for customized accelerators, which require additional consideration of edge filling logic. At the same time, the zero-padding operation increases the amount of computation and inevitably increases hardware energy consumption, which is obviously not friendly to wearable device application scenarios that pursue energy efficiency.

[0004] In actual situations, portable epilepsy monitoring equipment can effectively assist patients in avoiding possible safety threats. Seizure prediction methods can remind patients to move to a safe environment in time. Seizure detection methods can provide patients' relatives or doctors with information about the patient's condition, enabling them to take appropriate rescue measures and therapeutic interventions when the patient has a seizure. However, given that the EEG activity of epilepsy patients exhibits four typical time period characteristics, including pre-ictal, ictal, post-ictal, and interictal periods. The existing three-classification model simply classifies the EEG signals of the post-ictal period as interictal, ignoring the significant differences between the post-ictal and interictal EEGs. This simplified processing will lead to a significant increase in the false alarm rate in practical operations, thereby affecting the reliability of the monitoring system. Summary of the Invention

[0005] The present application aims to solve at least one of the technical problems in the related art to a certain extent. To this end, one purpose of the present application is to propose a lightweight four-classification method for epileptic EEG signals, which improves the reliability of monitoring patients in practical applications.

[0006] One aspect of the present application provides a lightweight four-classification method for epileptic EEG signals, comprising:

[0007] Step S100: collecting the user's EEG signal;

[0008] Step S200: Integrate short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct a four-classification model of EEG signals;

[0009] The method for constructing the four-classification model of EEG signals is as follows:

[0010] Obtaining raw EEG signals, preprocessing the raw EEG signals, and labeling four categories of EEG signal states of epileptic patients, wherein the four categories of EEG signal states of epileptic patients include interictal, preictal, ictal, and postictal, segmenting the four categories in the raw EEG signals using a time t as a slice window length, segmenting the raw EEG signals of each category into m data slices, and constructing a data set containing n data samples using the data slices, wherein each data sample contains data slices of the four categories;

[0011] Using short-time Fourier transform to map each data sample to the time-frequency domain, a time-frequency domain signal of the original EEG signal is obtained, and a spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and the low-frequency part is moved to the center of the spectrum dimension, thereby obtaining an optimized feature representation of the time-frequency domain signal of the original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension;

[0012] The specific implementation method of the spectrum reconstruction strategy is as follows: setting a noise threshold λ, filtering out the noise frequency band greater than or equal to the noise threshold in the [0, A1 Hz] frequency band of the original EEG signal by filtering, retaining the [0, A2 Hz) part, where A2 < λ, cropping the [0, A2 Hz) part to obtain the [0, A3 Hz] part and the (A3 Hz, A2 Hz) part, and moving the [0, A3 Hz] part to the center of the spectrum dimension by translation and splicing;

[0013] Construct a fusion neural network model, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function, and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signals of the optimized original EEG signal in the spectral dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and the softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; the fusion neural network model is trained using the data set to construct a trained four-classification model of EEG signals;

[0014] Step S300: Based on the trained EEG signal four-category model, the user's EEG signal is used as input and the probability value of the EEG signal belonging to each category is output;

[0015] The specific process of taking the user's EEG signal as input and outputting the probability value of the EEG signal belonging to each category based on the trained EEG signal four-classification model is as follows:

[0016] Step S310: Input the user's EEG signal into the trained EEG signal four-classification model, divide the EEG signal into a number of EEG signal segments according to the preset time window length, and use short-time Fourier transform to map each EEG signal segment to the time-frequency domain to obtain a time-frequency domain signal;

[0017] Step S320: Applying a spectrum reconstruction strategy, using a translation and re-splicing method, moves the low-frequency portion of the time-frequency domain signal to the center of the spectrum dimension, and filters the noise band to obtain an optimized time-frequency domain signal;

[0018] Step S330: In the spectrum dimension, a one-dimensional convolutional neural network is used to perform a convolution operation on the optimized time-frequency domain signal to extract frequency domain features;

[0019] Step S340: In the time series dimension, the frequency domain features extracted by the one-dimensional convolutional neural network are modeled using a long short-term memory network to obtain the time domain features of the frequency domain features in the time series;

[0020] Step S350: Input the frequency domain features and time domain features into the fully connected layer and the softmax activation function, classify the EEG signal segments, and output the probability values ​​of the EEG signal segments belonging to the four categories;

[0021] Step S400: comparing the probability values ​​of the four categories, taking the category with the largest probability value as the category to which the EEG signal belongs, and sending a corresponding alarm based on the category; wherein, the category to which the EEG signal belongs is determined to implement the EEG signal four-classification model for epileptic seizure prediction and seizure detection; when the category belongs to the pre-ictal stage, epileptic seizure prediction is implemented; when the category belongs to the ictal stage, post-ictal stage, or interictal stage, epileptic seizure detection is implemented;

[0022] The specific method of comparing the probability values ​​of the four categories, taking the category with the largest probability value as the category to which the EEG signal belongs, and sending the corresponding alarm according to the category is: classifying through a fully connected layer and a softmax function, outputting the probability value of the EEG signal segment belonging to each category, comparing the sizes of the four probability values, and taking the category with the largest probability value as the category to which the EEG signal segment belongs; if the category is pre-ictal, the user is warned, and if the category is ictal, post-ictal or interictal, the doctor is notified.

[0023] One aspect of the present application provides a lightweight four-classification system for epileptic EEG signals, comprising:

[0024] A sensor module for collecting the user's EEG signals;

[0025] The processing unit module is used to integrate short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct a four-category model of EEG signals; the method for constructing the four-category model of EEG signals is as follows: obtaining the original EEG signal, preprocessing the original EEG signal, marking the four categories of EEG signal states of epileptic patients, the four categories of EEG signal states of epileptic patients include interictal period, preictal period, ictal period and postictal period, dividing the four categories in the original EEG signal with a time length t as the slice window length, dividing the original EEG signal of each category into m data slices, and using the data slices to construct a data set containing n data samples, each data sample contains data slices of four categories; using short-time Fourier transform to map each data sample to the time-frequency domain to obtain the time-frequency domain signal of the original EEG signal, and applying the spectrum reconstruction strategy to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and moving the low-frequency part to The central position of the spectrum dimension is used to obtain the characteristic representation of the time-frequency domain signal of the optimized original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension; a fusion neural network model is constructed, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signal of the optimized original EEG signal in the spectrum dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and the softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; the fusion neural network model is trained using the data set to construct a trained four-classification model of EEG signals;

[0026] The output interface module is used to take the user's EEG signal as input based on the trained EEG signal four-category model and output the probability value of the EEG signal belonging to each category;

[0027] The communication module is used to compare the probability values ​​of the four categories, take the category with the largest probability value as the category to which the EEG signal belongs, and send a corresponding alarm according to the category to which it belongs; wherein, the category to which the EEG signal belongs is determined to realize the EEG signal four-classification model for the onset prediction and onset detection of epileptic patients. When the category to which it belongs is the pre-ictal stage, the onset prediction of the epileptic patient's condition is realized; when the category to which it belongs is the ictal period, post-ictal period or interictal period, the onset detection of the epileptic patient's condition is realized.

[0028] One aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in a lightweight four-classification method for epileptic EEG signals are implemented.

[0029] One aspect of the present application provides a readable storage medium storing a computer program, wherein the computer program is suitable for being loaded by a processor to execute steps in a lightweight four-classification method for epileptic EEG signals.

[0030] The lightweight four-classification method for epileptic EEG signals proposed in this application has the following advantages over the existing technology:

[0031] This application provides a lightweight four-classification method for epilepsy EEG signals that fully covers all stages of epilepsy, while meeting the functional requirements of epilepsy onset prediction and detection. Incorporating feature recognition of the late seizure phase into model prediction and detection eliminates the false positive problem caused by traditional two-classification and three-classification methods that fail to consider the late seizure phase and are easily misidentified as other phases. This improves the recognition accuracy of the model and ensures high reliability in monitoring patients in practical applications.

[0032] The EEG signal four-classification model proposed in this application integrates short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network. The spectrum reconstruction strategy reduces the influence of noise, so that the low-frequency part is concentrated in the central area of ​​the spectrum dimension, allowing the one-dimensional convolutional neural network to directly perform convolution operations on these key frequency parts without using padding operations. The omission of padding operations significantly reduces the complexity of hardware design, making the model proposed in this application more suitable for deployment on hardware devices;

[0033] The spectrum reconstruction strategy optimizes the time-frequency domain signal through two steps: noise band filtering and low-frequency shifting. Noise band filtering suppresses background noise and improves signal quality; low-frequency shifting enhances spectrum concentration, facilitating feature extraction using a one-dimensional convolutional neural network. These two steps work together to enhance the optimization of the time-frequency domain feature representation, laying the foundation for one-dimensional convolutional neural network feature extraction and four-classification of epileptic EEG signals. Furthermore, the spectrum reconstruction strategy has low computational complexity and is easy to implement, which helps improve the overall system's operational efficiency and achieve a lightweight model design.

[0034] The combined use and collaboration of a one-dimensional convolutional neural network and a long short-term memory network ensures that both the time and frequency domains of the EEG signal are fully captured. By leveraging the characteristics of the one-dimensional convolutional neural network and the long short-term memory network, which are suitable for recognizing the frequency and time domains, respectively, each network module only recognizes the information it is suitable for, achieving excellent recognition accuracy with a low parameter count. This design effectively reduces the total computational complexity and parameter count, making the proposed lightweight four-classification method for epilepsy EEG signals highly lightweight and extremely low in hardware overhead.

[0035] Compared with traditional classification models, the lightweight four-classification method of epilepsy EEG signals provided in this application can simultaneously realize seizure prediction and seizure detection in epilepsy patients. When the category result output by the four-classification model of the EEG signal is pre-ictal, the prediction of the patient's disease attack is realized, so that the patient can take precautions in advance based on the early warning. When the category result output by the four-classification model of the EEG signal is ictal period, post-ictal period or interictal period, it means that the patient is already in the process of disease attack, and the specific attack stage can be detected to facilitate medical staff to carry out corresponding treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of a lightweight four-classification method for epileptic EEG signals provided in this application;

[0037] Figure 2 A schematic diagram of a method for preprocessing EEG signals for a lightweight four-classification method of epileptic EEG signals provided in this application;

[0038] Figure 3 A schematic diagram of a short-time Fourier transform method for a lightweight four-classification method of epileptic EEG signals provided in this application;

[0039] Figure 4 A schematic diagram of a spectrum reconstruction strategy for a lightweight four-classification method of epileptic EEG signals provided in this application;

[0040] Figure 5 Schematic diagram of a one-dimensional convolutional neural network and a long short-term memory network for a lightweight four-classification method of epileptic EEG signals provided in this application;

[0041] Figure 6 Schematic diagram of the experimental results provided for this application;

[0042] Figure 7 Schematic diagram of the four-class confusion matrix of the experimental results provided for this application;

[0043] Figure 8 Schematic diagram of the effect of the postictal phase on specificity provided for this application;

[0044] Figure 9 This is a functional module diagram of a lightweight four-classification system for epileptic EEG signals provided in this application;

[0045] Figure 10 A schematic diagram of the structure of an electronic device provided in this application;

[0046] Figure 11 This is a schematic diagram of the structure of a readable storage medium provided by this application. DETAILED DESCRIPTION

[0047] To better understand the present application, various aspects of the present application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely descriptions of exemplary embodiments of the present application and are not intended to limit the scope of the present application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0048] In the accompanying drawings, the size, dimensions, and shapes of the elements have been slightly adjusted for ease of illustration. The accompanying drawings are for illustration only and are not drawn strictly to scale. As used herein, the terms "substantially," "approximately," and similar terms are used to indicate approximate values, not degrees, and are intended to illustrate inherent deviations in measurements or calculations that would be recognized by a person of ordinary skill in the art. In addition, in this application, the order in which the steps are described does not necessarily represent the order in which these steps would occur in actual operation, unless otherwise specified or inferred from the context.

[0049] It should also be understood that expressions such as "including", "comprising", "having", "containing" and / or "comprising" are open rather than closed expressions in this specification, which indicate the presence of the stated features, elements and / or components, but do not exclude the presence of one or more other features, elements, components and / or combinations thereof. In addition, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features rather than just the individual elements in the list. In addition, when describing embodiments of the present application, "may" is used to mean "one or more embodiments of the present application". And, the term "exemplary" is intended to refer to an example or illustration.

[0050] Unless otherwise defined, all words used herein (including engineering terms and scientific and technological terms) have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that, unless otherwise specified in this application, words defined in commonly used dictionaries should be interpreted as having the same meaning as they do in the context of the relevant technology, and should not be interpreted in an idealized or overly formal sense.

[0051] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0052] Example 1

[0053] like Figure 1 As shown in FIG, a lightweight four-classification method for epileptic EEG signals provided by this application includes:

[0054] Step S100: collecting the user's EEG signal;

[0055] The EEG signal is a non-stationary signal that indicates the temporal changes in the electrical activity of neurons in the cerebral cortex and contains relevant information about epileptic seizures.

[0056] Step S200: Integrate short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct a four-classification model of EEG signals;

[0057] The method for constructing the four-classification model of EEG signals is as follows:

[0058] Obtaining raw EEG signals, preprocessing the raw EEG signals, and labeling four categories of EEG signal states of epileptic patients, wherein the four categories of EEG signal states of epileptic patients include interictal, preictal, ictal, and postictal, segmenting the four categories in the raw EEG signals using a time t as a slice window length, segmenting the raw EEG signals of each category into m data slices, and constructing a data set containing n data samples using the data slices, wherein each data sample contains data slices of the four categories;

[0059] Using short-time Fourier transform to map each data sample to the time-frequency domain, a time-frequency domain signal of the original EEG signal is obtained, and a spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and the low-frequency part is moved to the center of the spectrum dimension, thereby obtaining an optimized feature representation of the time-frequency domain signal of the original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension;

[0060] The time-frequency domain signal is a two-dimensional representation obtained by performing short-time Fourier transform on the EEG signal segment;

[0061] The low-frequency part refers to the part of the time-frequency domain signal with a frequency of 0 to bHz; the value of b is set by those skilled in the art according to actual needs;

[0062] Construct a fusion neural network model, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function, and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signals of the optimized original EEG signal in the spectral dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and the softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; the fusion neural network model is trained using the data set to construct a trained four-classification model of EEG signals;

[0063] The specific implementation method of the spectrum reconstruction strategy is as follows: setting a noise threshold λ, filtering out the noise frequency band greater than or equal to the noise threshold in the [0, A1 Hz] frequency band of the original EEG signal by filtering, retaining the [0, A2 Hz) part, where A2 < λ, cropping the [0, A2 Hz) part to obtain the [0, A3 Hz] part and the (A3 Hz, A2 Hz) part, and moving the [0, A3 Hz] part to the center of the spectrum dimension by translation and splicing;

[0064] The values ​​of the noise threshold λ and A3 are set by those skilled in the art based on actual needs and experience;

[0065] The spectrum reconstruction strategy is used to optimize the feature representation of the time-frequency domain signals of the electroencephalogram (EEG) signals of epileptic patients. This strategy effectively suppresses background noise and enhances the spectral concentration of the signal through two main steps: noise band filtering and low-frequency part shifting, thereby improving the accuracy of subsequent feature extraction and classification.

[0066] Figure 2 A schematic diagram of a method for preprocessing EEG signals of a lightweight four-classification method for epilepsy EEG signals provided in this application; Figure 2 As shown in the figure, the preprocessing of the original EEG signal includes data labeling, data slicing and data set division. In the data labeling stage, the original EEG signal is labeled with the four categories of the EEG signal state of epileptic patients, and the parts that do not belong to the four categories are discarded; each category of the labeled original EEG signal is divided into m data slices with a time length t as the slice window length, t is 10s, and a data set is constructed. Each data sample in the data set contains data slices of the interictal period, preictal period, ictal period and postictal period;

[0067] In step S200, the EEG signal states of epileptic patients are divided into interictal, preictal, ictal, and postictal phases, innovatively incorporating the postictal phase into the scope of model recognition, which greatly improves the accuracy of the model.

[0068] Short-time Fourier transform is used to extract the time-frequency domain signals of the EEG signal. The time-frequency domain signals facilitate feature extraction in the subsequent EEG signal four-classification model. A spectrum reconstruction strategy is used to filter the noise frequency band in the time-frequency domain signals, and the low-frequency part is moved to the center of the spectrum dimension. A one-dimensional convolutional neural network is then used to extract frequency domain features. The advantage of a one-dimensional convolutional neural network is that it has a small number of parameters, which makes the EEG signal four-classification model lightweight, so that it can be applied in edge wearable device scenarios that are sensitive to power consumption.

[0069] Step S300: Based on the trained EEG signal four-category model, the user's EEG signal is used as input and the probability value of the EEG signal belonging to each category is output;

[0070] The specific process of taking the user's EEG signal as input and outputting the probability value of the EEG signal belonging to each category based on the trained EEG signal four-classification model is as follows:

[0071] Step S310: Input the user's EEG signal into the trained EEG signal four-classification model, divide the EEG signal into a number of EEG signal segments according to the preset time window length, and use short-time Fourier transform to map each EEG signal segment to the time-frequency domain to obtain a time-frequency domain signal;

[0072] The length of the time window is set by those skilled in the art according to actual needs;

[0073] Figure 3 A schematic diagram of the short-time Fourier transform method for a lightweight four-classification method of epileptic EEG signals provided in this application is provided. The EEG signal segments with a time window length of 10s are short-time Fourier transformed to obtain the time characteristics and frequency characteristics of the EEG signal segments on each electrode channel.

[0074] The spectrum dimension represents the amplitude of different frequency components, and the time series dimension represents different time windows in the time series;

[0075] The time-frequency domain signal reveals the changes in the frequency components of the EEG signal at different times;

[0076] Step S320: Applying a spectrum reconstruction strategy, using a translation and re-splicing method, moves the low-frequency portion of the time-frequency domain signal to the center of the spectrum dimension, and filters the noise band to obtain an optimized time-frequency domain signal;

[0077] Figure 4A schematic diagram of a spectrum reconstruction strategy for a lightweight four-classification method of epileptic EEG signals provided in this application. The noise frequency band is filtered out in the range of 0 to 120 Hz, the 0 to 59 Hz part is retained, the 0 to 59 Hz part is cropped to obtain the 0 to 45 Hz part and the 46 to 59 Hz part, and the 0 to 45 Hz part is moved to the center of the spectrum dimension by translation and splicing. The spectrum reconstruction strategy aims to selectively emphasize the key frequency bands and place the key frequency bands in a prominent position that is easily recognized by the one-dimensional convolutional neural network, thereby improving the significance of the key frequency bands and reducing the impact of noise.

[0078] Step S330: In the spectrum dimension, a one-dimensional convolutional neural network is used to perform a convolution operation on the optimized time-frequency domain signal to extract frequency domain features;

[0079] Step S340: In the time series dimension, the frequency domain features extracted by the one-dimensional convolutional neural network are modeled using a long short-term memory network to obtain the time domain features of the frequency domain features in the time series;

[0080] Figure 5 This application provides a schematic diagram of a one-dimensional convolutional neural network and a long short-term memory network for a lightweight four-classification method of epileptic EEG signals. The one-dimensional convolutional neural network and long short-term memory network of this application only need to recognize the information that they are suitable for recognition, and are still sufficient to complete the recognition task at a smaller network size. This theoretical calculation amount and parameter optimization design make it particularly suitable for resource-constrained edge computing environments, such as wearable devices.

[0081] Step S350: Input the frequency domain features and time domain features into the fully connected layer and the softmax activation function, classify the EEG signal segments, and output the probability values ​​of the EEG signal segments belonging to the four categories;

[0082] Step S400: comparing the probability values ​​of the four categories, taking the category with the largest probability value as the category to which the EEG signal belongs, and sending a corresponding alarm based on the category; wherein, the category to which the EEG signal belongs is determined to implement the EEG signal four-classification model for epileptic seizure prediction and seizure detection; when the category belongs to the pre-ictal stage, epileptic seizure prediction is implemented; when the category belongs to the ictal stage, post-ictal stage, or interictal stage, epileptic seizure detection is implemented;

[0083] The specific method of comparing the probability values ​​of the four categories, taking the category with the largest probability value as the category to which the EEG signal belongs, and sending the corresponding alarm according to the category is: classifying through a fully connected layer and a softmax function, outputting the probability value of the EEG signal segment belonging to each category, comparing the sizes of the four probability values, and taking the category with the largest probability value as the category to which the EEG signal segment belongs; if the category is pre-ictal, the user is warned, and if the category is ictal, post-ictal or interictal, the doctor is notified.

[0084] The warning threshold is set by those skilled in the art and is set and adjusted according to actual needs and classification results.

[0085] Example 2

[0086] In order to evaluate the generalization performance of the model, this application uses a 5-fold cross-validation technique to randomly divide the data set into five equal parts. In each iteration, one of the subsets is used as the test set, and the remaining four subsets are used as training sets. The number of training rounds for each experiment is 300. In order to evaluate the classification performance of the model, this application uses total accuracy, specificity, and sensitivity as evaluation indicators. Sensitivity measures the accuracy of the method provided by this application in identifying positive samples to ensure that no cases are missed; specificity measures the ability of the method provided by this application to correctly identify negative samples to avoid misdiagnosis.

[0087] The four-classification network model used in this application is capable of handling both epileptic seizure detection and prediction tasks. To evaluate the model's epileptic seizure prediction performance, this application designates the ictal, interictal, and postictal stages as negative examples, while distinguishing the preictal stage as a positive example, thereby facilitating a detailed assessment of the model's specificity and sensitivity in predicting epileptic seizures. When evaluating the model's epileptic seizure detection performance, the preictal, interictal, and postictal stages are considered negative examples, while the ictal stage is set as a positive example.

[0088] Figure 6 The experimental results provided for this application show the model accuracy of the four-category classification accuracy for seizure prediction sensitivity (SENS-P), seizure detection sensitivity (SENS-D), seizure prediction specificity (SPEC-P), seizure detection specificity (SPEC-D), and seizure classification accuracy for 10 cases selected from the CHB-MIT dataset: chb01, chb02, chb03, chb04, chb05, chb07, chb08, chb09, chb10, and chb22. The EEG four-classification model proposed in this application achieved excellent classification results in the majority of patients, with an accuracy exceeding 99% in cases such as chb03, chb04, chb07, chb08, and chb09. Figure 7The diagram of the four-class confusion matrix for the experimental results provided for this application shows the confusion matrix for the four-class classification of epileptic EEG signals. The average accuracy rates for SENS-P, SENS-D, SPEC-P, SPEC-D, and the four-class classification were 97.90%, 99.89%, 99.28%, 100%, and 98.44%, respectively.

[0089] The EEG signal four-classification model proposed in this application not only performs both epileptic seizure prediction and detection tasks, but also achieves highly competitive classification accuracy with a relatively low number of parameters. Using only 3.7K parameters, this model achieves a significant parameter reduction compared to more recent models, with the reduction ranging from 1.7 to 60.7 times.

[0090] Traditional two-class and three-class models cannot cover the late stage of seizures during training. In order to evaluate the impact of the late stage of seizures on recognition under real-world conditions, this application conducts two sets of experiments on the two-class and three-class models: one set is under ideal conditions and excludes late stage samples from the test set, while the other set includes these samples to simulate real-world scenarios. The experiments show the seizure prediction and detection performance of the three-class model. The two-class model does not have the seizure detection function, so only the seizure prediction performance is obtained. Figure 8 The figure below shows the effect of the post-ictal phase on specificity provided in this application. After incorporating post-ictal samples, the specificity of both the two-category and three-category seizure prediction models decreased by more than 20%, while the specificity of the three-category seizure detection model decreased relatively little, only about 5%. An important reason for the significant decrease in seizure prediction specificity is that the characteristics of the post-ictal phase are more similar to those of the pre-ictal phase, making it easier for the model to misclassify post-ictal samples as pre-ictal samples.

[0091] The model proposed in this application performs well even in real-world scenarios involving postictal samples, maintaining a specificity of 99.28% for epileptic seizure prediction and 100% for detection. This finding highlights the significant advantage of the four-classification model in improving the specificity of epileptic seizure prediction and detection tasks in real-world scenarios, demonstrating its effectiveness compared to traditional methods.

[0092] Example 3

[0093] like Figure 9 As shown in FIG, a lightweight four-classification system for epilepsy EEG signals provided by this application includes:

[0094] A sensor module for collecting the user's EEG signals;

[0095] The processing unit module is used to integrate short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct a four-category model of EEG signals; the method for constructing the four-category model of EEG signals is as follows: obtaining the original EEG signal, preprocessing the original EEG signal, marking the four categories of EEG signal states of epileptic patients, the four categories of EEG signal states of epileptic patients include interictal period, preictal period, ictal period and postictal period, dividing the four categories in the original EEG signal with a time length t as the slice window length, dividing the original EEG signal of each category into m data slices, and using the data slices to construct a data set containing n data samples, each data sample contains data slices of four categories; using short-time Fourier transform to map each data sample to the time-frequency domain to obtain the time-frequency domain signal of the original EEG signal, and applying the spectrum reconstruction strategy to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and moving the low-frequency part to The central position of the spectrum dimension is used to obtain the characteristic representation of the time-frequency domain signal of the optimized original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension; a fusion neural network model is constructed, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signal of the optimized original EEG signal in the spectrum dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and the softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; the fusion neural network model is trained using the data set to construct a trained four-classification model of EEG signals;

[0096] The output interface module is used to take the user's EEG signal as input based on the trained EEG signal four-category model and output the probability value of the EEG signal belonging to each category;

[0097] A communication module is used to compare the probability values ​​of the four categories, take the category with the largest probability value as the category to which the EEG signal belongs, and send a corresponding alarm based on the category. The category to which the EEG signal belongs is determined to implement the EEG signal four-classification model for epileptic seizure prediction and seizure detection. When the category belongs to the pre-ictal stage, the epileptic seizure prediction is achieved; when the category belongs to the ictal stage, post-ictal stage, or interictal stage, the epileptic seizure detection is achieved.

[0098] Optionally, the communication module further includes a wireless communication means for ensuring timely communication even when the user is in a remote or inconvenient operating environment;

[0099] Optionally, the lightweight epileptic EEG signal four-classification system is designed to be portable or wearable, integrated with low-power components, and suitable for long-term continuous monitoring of the EEG signal status of epileptic patients;

[0100] Optionally, the lightweight four-classification system for epilepsy EEG signals is connected to a remote monitoring platform via wireless communication technology to achieve real-time data transmission and remote medical support functions.

[0101] Example 4

[0102] Figure 10 This is a schematic diagram of the electronic device structure provided by an embodiment of the present application. Figure 10 According to another aspect of the present application, an electronic device is provided. The electronic device may include one or more processors and one or more memories. The memories may store computer-readable code that, when executed by the one or more processors, may execute the lightweight four-classification method for epileptic EEG signals as described above.

[0103] The method or system according to the embodiment of the present application can also be used by Figure 10 The electronic device architecture shown in FIG. Figure 10As shown, the electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, an input / output component, a hard disk, etc. The storage device in the electronic device, such as a ROM or a hard disk, can store a lightweight four-classification method for epilepsy EEG signals provided in this application. A lightweight four-classification method for epilepsy EEG signals may, for example, include: collecting the user's EEG signals; integrating short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct an EEG signal four-classification model; the method for constructing the EEG signal four-classification model is: obtaining the original EEG signal, pre-processing the original EEG signal, and marking the four categories of the EEG signal state of the epilepsy patient, the four categories of the EEG signal state of the epilepsy patient include interictal period, preictal period, ictal period and postictal period, and dividing the four categories in the original EEG signal with a slice window length of time t, and dividing the original EEG signal of each category into The signal is divided into m data slices, and the data slices are used to construct a data set containing n data samples, each data sample contains four categories of data slices; each data sample is mapped to the time-frequency domain using short-time Fourier transform to obtain the time-frequency domain signal of the original EEG signal, and the spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and the low-frequency part is moved to the central position of the spectrum dimension to obtain the feature representation of the optimized time-frequency domain signal of the original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension; a fusion neural network model is constructed, including an input layer, a one-dimensional convolutional neural network, a long-short-term convolutional neural network, and a time series dimension. The LSTM network is composed of a time-domain memory network, a fully connected layer, a softmax activation function, and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signals of the optimized original EEG signal in the spectrum dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and the softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; the fusion neural network model is trained using the data set to construct A trained four-category model of EEG signals is constructed; based on the trained four-category model of EEG signals, the user's EEG signals are used as input, and the probability value of the EEG signals belonging to each category is output; the probability values ​​of the four categories are compared, and the category with the largest probability value is used as the category to which the EEG signal belongs, and a corresponding alarm is sent according to the category to which it belongs; wherein, the category to which the EEG signal belongs is determined to realize the EEG signal four-category model for the prediction and detection of epileptic seizures, and when the category to which it belongs is the pre-ictal stage, the prediction of the epileptic seizure is realized, and when the category to which it belongs is the ictal stage, the post-ictal stage or the interictal stage, the detection of the epileptic seizure is realized. Furthermore, the electronic device may also include a user interface. Of course, Figure 10 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 10 One or more components of an electronic device are shown.

[0104] Example 5

[0105] Figure 11 This is a schematic diagram of the structure of a readable storage medium provided by an embodiment of the present application. Figure 11 As shown, it is a readable storage medium according to an embodiment of the present application. Computer-readable instructions are stored on the computer-readable storage medium. When the computer-readable instructions are executed by the processor, a lightweight four-classification method for epilepsy EEG signals according to the embodiment of the present application described with reference to the above figures can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache). Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0106] In addition, according to the implementation mode of the present application, the process described in the above reference flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be run by a processor to execute instructions corresponding to the method steps provided in the present application, such as: collecting the user's EEG signal; integrating short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to construct a four-category EEG signal model; the method for constructing the four-category EEG signal model is: obtaining the original EEG signal, pre-processing the original EEG signal, and marking the four categories of the EEG signal state of the epileptic patient, the four categories of the EEG signal state of the epileptic patient include interictal period, preictal period, During the seizure period and post-seizure period, the original EEG signal is divided into four categories with a slice window length of t, and the original EEG signal of each category is divided into m data slices. The data slices are used to construct a data set containing n data samples, and each data sample contains data slices of four categories; each data sample is mapped to the time-frequency domain using short-time Fourier transform to obtain the time-frequency domain signal of the original EEG signal, and the spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and move the low-frequency part to the central position of the spectrum dimension to obtain the feature representation of the optimized time-frequency domain signal of the original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension degree; construct a fusion neural network model, including input layer, one-dimensional convolutional neural network, long short-term memory network, fully connected layer, softmax activation function and output layer; the one-dimensional convolutional neural network is used to perform convolution operation on the time-frequency domain signal of the optimized original EEG signal in the spectrum dimension to extract the frequency domain features of the original EEG signal; the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension to obtain the time domain features of the frequency domain features of the original EEG signal in the time series; the fully connected layer and softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample; use data The fusion neural network model is trained to construct a trained EEG signal four-classification model; based on the trained EEG signal four-classification model, the user's EEG signal is used as input, and the probability value of the EEG signal belonging to each category is output; the probability values ​​of the four categories are compared, and the category with the largest probability value is used as the category to which the EEG signal belongs, and a corresponding alarm is sent according to the category to which it belongs; wherein, the category to which the EEG signal belongs is determined to realize the EEG signal four-classification model for epileptic seizure prediction and seizure detection. When the category to which it belongs is the pre-ictal stage, the epileptic patient's condition is predicted; when the category to which it belongs is the ictal period, the post-ictal period or the interictal period, the epileptic patient's condition is detected.When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are performed.

[0107] The methods, apparatus, and devices of the present application may be implemented in many ways. For example, the methods, apparatus, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the methods according to the present application. Therefore, the present application also covers recording media that store programs for executing the methods according to the present application.

[0108] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0109] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A lightweight four-classification method for epileptic EEG signals, characterized by: include: Collect the user's EEG signals; Integrate short-time Fourier transform, spectrum reconstruction strategy, one-dimensional convolutional neural network and long short-term memory network to build a four-classification model of EEG signals; The method for constructing the four-classification model of the electroencephalogram (EEG) signal is as follows: obtaining the original EEG signal, preprocessing the original EEG signal, marking the four categories of the EEG signal state of the epileptic patient, wherein the four categories of the EEG signal state of the epileptic patient include interictal period, preictal period, ictal period, and postictal period, segmenting the four categories in the original EEG signal with a time length t as the slice window length, segmenting the original EEG signal of each category into m data slices, and constructing a data set containing n data samples using the data slices, wherein each data sample contains data slices of the four categories; Using short-time Fourier transform to map each data sample to the time-frequency domain, a time-frequency domain signal of the original EEG signal is obtained, and a spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and the low-frequency part is moved to the center of the spectrum dimension, thereby obtaining an optimized feature representation of the time-frequency domain signal of the original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension; Construct a fusion neural network model, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function, and an output layer. The one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signals of the optimized original EEG signal in the spectral dimension to extract the frequency domain features of the original EEG signal. The long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension, obtaining the time domain features of the frequency domain features of the original EEG signal in the time series. The fully connected layer and softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample. The fusion neural network model is trained using the dataset to construct a trained four-category EEG signal model. Based on the trained EEG signal four-classification model, the user's EEG signal is taken as input and the probability value of the EEG signal belonging to each category is output; Compare the probability values ​​of the four categories, take the category with the largest probability value as the category to which the EEG signal belongs, and send a corresponding alarm according to the category to which it belongs; wherein, the category to which the EEG signal belongs is determined to realize the EEG signal four-classification model for epileptic seizure prediction and seizure detection of epileptic patients; when the category belongs to the pre-ictal stage, the epileptic seizure prediction of the patient's condition is realized; when the category belongs to the ictal stage, post-ictal stage or interictal stage, the epileptic seizure detection of the patient's condition is realized; The specific implementation method of the spectrum reconstruction strategy is as follows: setting a noise threshold λ, filtering the noise frequency band greater than or equal to the noise threshold in the [0, A1Hz] frequency band of the original EEG signal by filtering, retaining the [0, A2Hz) part, where A2 < λ, cropping the [0, A2Hz) part to obtain the [0, A3Hz] part and the (A3Hz, A2Hz) part, and moving the [0, A3Hz] part to the center of the spectrum dimension by translation and splicing.

2. A lightweight four-classification method for epileptic EEG signals according to claim 1, characterized in that: The specific process of taking the user's EEG signal as input and outputting the probability value of the EEG signal belonging to each category based on the trained EEG signal four-classification model is as follows: Input the user's EEG signal into the trained EEG signal four-classification model, divide the EEG signal into a EEG signal segments according to the preset time window length, and use short-time Fourier transform to map each EEG signal segment to the time-frequency domain to obtain the time-frequency domain signal; Applying the spectrum reconstruction strategy, using the translation and re-splicing method, the low-frequency part of the time-frequency domain signal is moved to the center of the spectrum dimension, and the noise band is filtered to obtain the optimized time-frequency domain signal; In the spectrum dimension, a one-dimensional convolutional neural network is used to perform convolution operations on the optimized time-frequency domain signal to extract frequency domain features; In the time series dimension, the long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network to obtain the time domain features of the frequency domain features in the time series; The frequency domain features and time domain features are input into the fully connected layer and the softmax activation function to classify the EEG signal segments and output the probability values ​​of the EEG signal segments belonging to the four categories.

3. A lightweight four-classification method for epileptic EEG signals according to claim 2, characterized in that: The specific method of comparing the probability values ​​of the four categories, taking the category with the largest probability value as the category to which the EEG signal belongs, and sending the corresponding alarm according to the category is: classifying through a fully connected layer and a softmax function, outputting the probability value of the EEG signal segment belonging to each category, comparing the sizes of the four probability values, and taking the category with the largest probability value as the category to which the EEG signal segment belongs; if the category is pre-ictal, the user is warned, and if the category is ictal, post-ictal or interictal, the doctor is notified.

4. A lightweight four-classification system for epileptic EEG signals, which is implemented based on a lightweight four-classification method for epileptic EEG signals according to any one of claims 1 to 3, characterized in that: include: A sensor module for collecting the user's EEG signals; A processing unit module is used to integrate a short-time Fourier transform, a spectrum reconstruction strategy, a one-dimensional convolutional neural network, and a long short-term memory network to construct a four-category model for EEG signals; the method for constructing the four-category model for EEG signals is as follows: obtaining original EEG signals, preprocessing the original EEG signals, marking four categories of EEG signal states of epileptic patients, the four categories of EEG signal states of epileptic patients including interictal, preictal, ictal, and postictal, segmenting the original EEG signals into four categories using a time length t as a slice window length, segmenting the original EEG signals of each category into m data slices, and using the data slices to construct a data set containing n data samples, each data sample containing data slices of the four categories; Using short-time Fourier transform to map each data sample to the time-frequency domain, a time-frequency domain signal of the original EEG signal is obtained, and a spectrum reconstruction strategy is applied to filter the noise frequency band in the time-frequency domain signal of the original EEG signal, and the low-frequency part is moved to the center of the spectrum dimension, thereby obtaining a feature representation of the time-frequency domain signal of the optimized original EEG signal; the dimensions of the time-frequency domain signal include the spectrum dimension and the time series dimension; a fusion neural network model is constructed, including an input layer, a one-dimensional convolutional neural network, a long short-term memory network, a fully connected layer, a softmax activation function, and an output layer; the one-dimensional convolutional neural network is used to perform a convolution operation on the time-frequency domain signal of the optimized original EEG signal in the spectrum dimension to extract the frequency domain features of the original EEG signal; The long short-term memory network is used to model the frequency domain features extracted by the one-dimensional convolutional neural network in the time series dimension, obtaining the time domain features of the frequency domain features of the original EEG signal in the time series. The fully connected layer and softmax activation function are used to calculate the probability value of the data sample belonging to each category, and the output layer is used to output the classification of the data sample. The fusion neural network model is trained using the dataset to construct a trained four-category EEG signal model. The output interface module is used to take the user's EEG signal as input based on the trained EEG signal four-category model and output the probability value of the EEG signal belonging to each category; The communication module is used to compare the probability values ​​of the four categories, take the category with the largest probability value as the category to which the EEG signal belongs, and send a corresponding alarm according to the category to which it belongs; wherein, the category to which the EEG signal belongs is determined to realize the EEG signal four-classification model for the onset prediction and onset detection of epileptic patients. When the category to which it belongs is the pre-ictal stage, the onset prediction of the epileptic patient's condition is realized; when the category to which it belongs is the ictal period, post-ictal period or interictal period, the onset detection of the epileptic patient's condition is realized.

5. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the lightweight four-classification method for epileptic EEG signals as claimed in any one of claims 1 to 3 are implemented.

6. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the steps of the lightweight four-classification method for epileptic EEG signals according to any one of claims 1 to 3.

Images