A single channel seizure detection device

Abstract

The application relates to a single-channel seizure detection device, comprising: a data acquisition unit configured to acquire single-channel electroencephalogram data to be detected and to pre-process the single-channel electroencephalogram data to be detected; an electroencephalogram feature extraction unit configured to extract time-frequency features of the pre-processed single-channel electroencephalogram data and to perform standardization processing on the extracted time-frequency features; and a seizure detection unit configured to sequentially pass the standardization-processed time-frequency features through an input layer, an Embedding layer, a feature generalization module, a global pooling layer, a 1x1 convolution layer and a Linear output layer to obtain a detection result; the application combines the light convolution calculation advantages of the partial convolution idea and the advantages of the Style-domain generalization algorithm, and further divides and processes frequency domain information to reduce the increase of calculation amount, so that the detection model enhances the cross-patient detection effect of single-channel seizure electroencephalogram and makes the network lighter, and the generalization and real-time performance of the detection model are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing technology, and in particular relates to a single-channel epileptic seizure detection device. Background Technology

[0002] Electroencephalography (EEG) is an important clinical technique for the diagnosis and treatment of childhood epilepsy. Currently, clinical practice mainly relies on doctors visually observing characteristic waveforms in the EEG and visually assessing abnormal discharges in the child's brain to make manual judgments. Multi-lead EEG recordings for children with epilepsy typically take one to several days. Doctors induce seizures under safe conditions and simultaneously assess and label the interictal, preictal, and ictal signals of the EEG. However, the prolonged observation, assessment, and care of the child are extremely time-consuming and energy-draining for pediatric medical staff, potentially impacting their physical and mental health. Furthermore, manual interpretation of the images is highly subjective, easily leading to a higher rate of misjudgment. New intelligent seizure detection technology for childhood epilepsy is primarily based on machine learning methods. It extracts features such as fuzzy entropy, sample entropy, and HIGUCHI fractal dimension from the seizure signals and trains a classifier for detection, thus reducing the workload of pediatric medical staff to some extent. However, research in this field lacks non-induced, real-time acquisition of electroencephalograms (EEGs) of children with epilepsy, and the amount of data is too small, making it difficult to develop real-time monitoring technology for children with epilepsy and hindering subsequent research, thus limiting the development of automatic detection technology for childhood epilepsy seizures.

[0003] With the increasing intersection of deep learning and neuroscience, automated detection technology for real-time EEG seizures in children with epilepsy has received considerable attention. Currently, deep learning-based automated detection methods for childhood epilepsy seizures offer high accuracy, but most models suffer from high redundancy, poor real-time performance, reliance on individual child EEG data, and weak generalization across patients based on single-channel EEG data. Therefore, current deep learning-based automated detection technology for real-time EEG seizures in children with epilepsy is insufficient to meet the needs of clinical monitoring and real-time EEG acquisition for children with epilepsy. Summary of the Invention

[0004] To address the problems existing in the background art, the present invention provides a single-channel epileptic seizure detection device, comprising:

[0005] S1: Data acquisition unit, configured to acquire single-channel EEG signal data to be detected and preprocess the single-channel EEG signal data to be detected;

[0006] S2: EEG feature extraction unit, configured to extract time-frequency features from preprocessed single-channel EEG signal data and standardize the extracted time-frequency features;

[0007] S3: Seizure detection unit, configured to sequentially pass the standardized time-frequency features through the input layer, embedding layer, feature generalization module, global pooling layer, 1×1 convolutional layer and linear output layer to obtain the detection result; the feature generalization module includes four feature generalization units connected in series, each feature generalization unit including: multiple generalization layers connected in series and a PatchMerging layer; the generalization layer is used to generalize the input features.

[0008] Preferably, the preprocessing of the acquired single-channel EEG signal data includes:

[0009] The single-channel EEG signal data to be detected is input into a Butterworth bandpass filter for filtering to obtain intermediate EEG data. The intermediate EEG data is then denoised using the Independent Component Analysis (ICA) algorithm to obtain preprocessed single-channel EEG signal data.

[0010] Preferably, the extraction of time-frequency features from the preprocessed single-channel EEG signal data includes:

[0011]

[0012] Where X(f,τ) represents the extracted time-frequency features; X(t) represents the preprocessed single-channel EEG signal data; π represents pi; e represents the natural base; i represents the imaginary part of the complex number; R represents the real number; τ represents the time variable; f represents the frequency variable; and t represents time.

[0013] Preferably, the standardization process for the extracted time-frequency features includes:

[0014]

[0015] Among them, X Norm (f,τ) represents the time-frequency feature after standardization, X(f,τ) represents the extracted time-frequency feature; μ represents the mean of feature X(f,τ); δ represents the variance of feature X(f,τ).

[0016] Preferably, the input layer is configured to input features X Norm (f,τ) expands by one dimension, causing it to expand to X. Norm (1,f,τ); The Embedding layer is configured to process the input features X Norm (1,f,τ) is projected as a four-dimensional feature X Embed (1,c,f,τ), where c is the number of channels.

[0017] Preferably, the Generalization layer is configured to divide the input features into high-frequency components, low-frequency components, and original input features. The high-frequency components are input into the generalization convolution branch A for processing to obtain output feature a, and the low-frequency components are input into the partial convolution branch B for processing to obtain output feature b. Output features a and b are concatenated in the frequency domain to obtain feature F. Feature F is added to the original input feature c to obtain the output feature of the Generalization layer. The generalization convolution branch A includes a SCNorm module for generalizing the input high-frequency components. The partial convolution branch B includes a 3×3 Pconv layer, a 1×1 dilation layer, a BatchNorm layer, a ReLU activation layer, and a 1×1 bottleneck layer, sequentially connected to perform one 3×3 32-channel partial convolution operation, one 1×1 64-channel convolution operation, one BatchNorm operation, one ReLU activation, and one 1×1 32-channel convolution operation on the input low-frequency components.

[0018] Preferably, the SCNorm module is configured to perform Softmax processing on the input features to obtain fuzzy weights. and The input features are processed by inputting them into branch 1 and branch 2 respectively; branch 1 includes a BatchNorm layer and a first FFT layer connected in series; branch 2 includes a second FFT layer; based on the fuzzy weights... and The amplitude of the output from the first FFT layer and the amplitude of the output from the second FFT layer are weighted to obtain the intermediate amplitude. The intermediate amplitude and the phase of the output from the second FFT layer are then input into the inverse FFT layer for processing to obtain the output characteristics of the SCNorm module.

[0019] Preferably, the Pconv layer is configured to extract c of the input features. p The feature of the number of channels is used for convolution operation. Where, n v This represents the set threshold parameter, c. k This indicates the number of channels for the input feature.

[0020] The present invention has at least the following beneficial effects

[0021] This invention combines the lightweight convolutional computation advantages of partial convolution with the advantages of Style-domain generalization algorithms. While enhancing the cross-patient detection function of single-channel epileptic seizure EEG, it also makes the network more lightweight, greatly improving the generalization and real-time performance of the detection model. This effectively promotes the development of pediatric epilepsy detection technology and drives epilepsy seizure detection and prediction research towards the big data stage. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0023] Figure 2 This is a schematic diagram of the detection model of the present invention;

[0024] Figure 3 This is a schematic diagram of the Generalization layer of the present invention;

[0025] Figure 4 This is a schematic diagram of the SCNorm module of the present invention;

[0026] Figure 5 This is a schematic diagram of the structure of a single-channel epileptic seizure detection device in an embodiment of the present invention. Detailed Implementation

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0029] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0030] Please see Figure 1 The present invention provides a single-channel epileptic seizure detection device, comprising:

[0031] S1: Data acquisition unit, configured to acquire single-channel EEG signal data to be detected and preprocess the single-channel EEG signal data to be detected;

[0032] Preferably, the preprocessing of the acquired single-channel EEG signal data includes:

[0033] The single-channel EEG signal data to be detected is input into a Butterworth bandpass filter for filtering to obtain intermediate EEG data. The intermediate EEG data is then denoised using the Independent Component Analysis (ICA) algorithm to obtain preprocessed single-channel EEG signal data. The ICA algorithm removes artifacts such as electrooculography (EOG) and electromyography (EMG) in the first intermediate EEG data by comparing different components to obtain the second intermediate EEG data.

[0034] In this embodiment, considering the sampling rate, information content, and data label matching degree of children's EEG data, the public dataset of children's epileptic seizure EEG from Boston Children's Hospital and the dataset of children's epileptic seizure EEG from the Children's Hospital Affiliated to Chongqing Medical University are used as the training set for the detection model; the training set has been labeled with epileptic seizure segments by relevant hospital personnel. As one embodiment, when training the detection model, the BinaryCrossEntropy loss function is used as the loss function of the intelligent detection model of children's epileptic seizure signals, and the stochastic gradient descent optimization algorithm is used as the parameter optimization algorithm for the model.

[0035] In another embodiment of the present invention, a single-channel EEG acquisition device is used to acquire single-channel EEG signal data in real time on the child's head. The single-channel EEG acquisition device includes a TGAMN EEG acquisition module and an OpenBCI series EEG acquisition module.

[0036] S2: EEG feature extraction unit, configured to extract time-frequency features from preprocessed single-channel EEG signal data and standardize the extracted time-frequency features;

[0037] Preferably, the extraction of time-frequency features from the preprocessed single-channel EEG signal data includes:

[0038]

[0039] Where X(f,τ) represents the extracted time-frequency features; X(t) represents the preprocessed single-channel EEG signal data; π represents pi; e represents the natural base; i represents the imaginary part of the complex number; R represents the real number; τ represents the time variable; f represents the frequency variable; and t represents time.

[0040] Preferably, the standardization process for the extracted time-frequency features includes:

[0041]

[0042] Among them, X Norm (f,τ) represents the standardized time-frequency feature, X(f,τ) represents the extracted time-frequency feature; μ represents the mean of feature X(f,τ); δ represents the variance of feature X(f,τ). Standardizing the extracted time-frequency feature can reduce gradient explosion and gradient vanishing problems during model training and improve the generalization performance of the model.

[0043] Please see Figure 2 S3: Seizure Detection Unit, configured to sequentially pass the standardized time-frequency features through an input layer, an embedding layer, a feature generalization module, a global pooling layer, a 1×1 convolutional layer (one 1×1 convolutional operation with 128 channels), and a linear output layer (one linear layer and one sigmoid layer) to obtain the detection result; the feature generalization module includes four feature generalization units connected in series, each feature generalization unit including: multiple generalization layers connected in series and a patchmerging layer; the generalization layer is used to generalize the input features.

[0044] In this embodiment, the number of generalization layers in the four feature generalization units are L1, L2, L3, and L4, respectively, where L1 <L2<L3<L4;

[0045] Preferably, the input layer is configured to input features X Norm (f,τ) expands by one dimension, causing it to expand to X. Norm (1,f,τ); The Embedding layer is configured to process the input features X Norm (1,f,τ) is projected as a four-dimensional feature X Emb (1,c,f,τ), where c is the number of channels.

[0046] Please see Figure 3Preferably, the Generalization layer is configured to divide the input features into high-frequency components, low-frequency components, and original input features. The high-frequency components are input into the generalization convolution branch A for processing to obtain output feature a. The low-frequency components are input into the partial convolution branch B for processing to obtain output feature b. Output features a and b are concatenated in the frequency domain to obtain feature F. Feature F is added to the original input feature c to obtain the output feature of the Generalization layer. The generalization convolution branch A includes a SCNorm module for generalizing the input high-frequency components. The partial convolution branch B includes a 3×3 Pconv layer, a 1×1 dilation layer, a BatchNorm layer, a ReLU activation layer, and a 1×1 bottleneck layer, sequentially connected to perform one 3×3 32-channel partial convolution operation, one 1×1 64-channel convolution operation, one BatchNorm operation, one ReLU activation, and one 1×1 32-channel convolution operation on the input low-frequency components.

[0047] In this embodiment, the Generalization layer is configured to divide the input features into high-frequency components, low-frequency components, and the original input features are represented as follows: For example, for input feature X Embed (1,c,f,τ) is divided into high-frequency components. Low frequency components With the original input X Embed (1, c, f, τ). The high-frequency components are input to the SCNorm module for generalization processing, and the low-frequency components are input to the Pconv layer for partial convolution processing to reduce the model's convolution computation and memory throughput.

[0048] Please see Figure 4 Preferably, the SCUORM module is configured to perform Softmax processing on the input features to obtain fuzzy weights. and The input features are processed by inputting them into branch 1 and branch 2 respectively; branch 1 includes a BatchNorm layer and a first FFT layer connected in series; branch 2 includes a second FFT layer; based on the fuzzy weights... and The amplitude of the output from the first FFT layer and the amplitude of the output from the second FFT layer are weighted to obtain the intermediate amplitude. The intermediate amplitude and the phase of the output from the second FFT layer are input into the inverse FFT layer for processing to obtain the output characteristics of the SCNorm module.

[0049] Preferably, the Pconv layer is configured to extract c of the input features. p The feature of the number of channels is used for convolution operation. Where, nv This represents the set threshold parameter, c. k This indicates the number of channels for the input feature.

[0050] Please see Figure 5 The present invention provides a specific embodiment of a single-channel epileptic seizure detection device, comprising: a single-channel head-mounted electroencephalogram (EEG) acquisition module, a pediatric epileptic seizure EEG detection platform, and a data storage module;

[0051] The single-channel head-mounted EEG acquisition module is used to acquire single-channel EEG signal data from the child's head, and includes:

[0052] The system includes a single-channel EEG acquisition chip, acquisition electrodes, a Bluetooth module, a power supply module, and a headband. The single-channel EEG acquisition chip is a TGAM single-channel EEG acquisition chip used to perform analog-to-digital conversion of EEG signals.

[0053] The Bluetooth module uses a highly integrated BK3432 chip with a 2Mbps data rate to establish communication with the host computer.

[0054] The power supply module is used to independently power the single-channel EEG chip and the Bluetooth module;

[0055] The headband is equipped with a movable buckle to secure the dry electrodes, which are used to place the dry electrodes close to the child's forehead for scalp EEG acquisition.

[0056] The pediatric epilepsy seizure EEG detection platform includes: a data preprocessing module, an epilepsy seizure detection module, a visualization module, and a user information management module;

[0057] The data preprocessing module is equipped with Butterworth filter, wavelet transform and ICA algorithm, which are used to filter and denoise the raw EEG of the children that have been collected.

[0058] The epilepsy seizure detection module uses the pediatric epilepsy seizure detection method provided by this invention to perform real-time detection of the collected EEG data of the child.

[0059] The visualization module serves as the main interface for displaying the child's original EEG data, preprocessed data, or detected epileptic seizure data collected by the EEG acquisition module.

[0060] The user information management module and the data storage module establish a connection using the TCP protocol. The module sends, receives, and processes HTTP messages to send the real-time collected raw EEG of the child or the detected and labeled EEG to the data storage module for storage.

[0061] As one embodiment, the data storage module is written in C++ and combined with MySQL to build a server for storing user information and user data;

[0062] After the acquisition module acquires the child's EEG signal, the single-channel EEG signal data to be tested is input to the detection platform via Bluetooth module. After the data is processed by the detection model on the detection platform, the detection results are output and automatically marked and segmented.

[0063] The pediatric epilepsy seizure EEG detection platform is available on both Windows and Android platforms. The Android platform only supports static visualization of user data, while connecting with the Windows platform to monitor the child's condition in real time. In one embodiment, the detection system acquires the child's EEG using the single-channel acquisition module as a lower-level machine, and then transmits the single-channel EEG signal data to the EEG detection platform via Bluetooth. After unpacking the raw data, the child's original EEG is obtained. This raw EEG can be visualized in real time, statically, filtered, denoised, labeled, and automatically detected, labeled, segmented, and stored on the EEG detection platform.

[0064] Existing technologies require medical personnel to undergo extensive training to manually identify and determine if a seizure has occurred in children, then manually label the data before further research. This process is time-consuming, labor-intensive, produces small datasets, lacks portability, and is not standardized; varying skill levels among medical personnel can lead to different results. This invention, however, can automatically segment, label, and store detected seizure signals while assisting medical personnel in detecting childhood epilepsy. This invention not only reduces the workload of medical personnel and the cost per test but also helps expand the overall pediatric epilepsy EEG database. The entire process is automated and standardized, enabling automatic detection of childhood epilepsy seizures. It represents a significant advancement in the health monitoring and EEG data collection for children with epilepsy.

[0065] In summary, this invention combines the lightweight convolutional computation advantages of partial convolution with the advantages of Style-domain generalization algorithms. While enhancing the cross-patient detection capabilities of single-channel epileptic seizure EEG, it also makes the network more lightweight, greatly improving the generalization and real-time performance of the detection model. This effectively promotes the development of pediatric epilepsy detection technology and drives epilepsy seizure detection and prediction research towards the big data stage.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A single-channel epileptic seizure detection device, characterized in that, include: S1: Data acquisition unit, configured to acquire single-channel EEG signal data to be detected and preprocess the single-channel EEG signal data to be detected; S2: EEG feature extraction unit, configured to extract time-frequency features from preprocessed single-channel EEG signal data and standardize the extracted time-frequency features; S3: Seizure Detection Unit, configured to sequentially pass standardized time-frequency features through an input layer, an embedding layer, a feature generalization module, a global pooling layer, a 1×1 convolutional layer, and a linear output layer to obtain detection results; the feature generalization module includes four feature generalization units connected in series, each feature generalization unit including: multiple generalization layers connected in series and a PatchMerging layer; the generalization layer is used to generalize the input features; The Generalization layer is configured to divide the input features into high-frequency components, low-frequency components, and the original input features. The high-frequency components are processed by the generalization convolutional branch A to obtain output feature a, and the low-frequency components are processed by the partial convolutional branch B to obtain output feature b. Output features a and b are concatenated in the frequency domain to obtain feature F. Feature F is then added to the original input feature c to obtain the output feature of the Generalization layer. The generalization convolutional branch A includes a SCNorm module for generalizing the input high-frequency components. The partial convolutional branch B includes a 3×3Pconv layer, a 1×1 dilation layer, a BatchNorm layer, a ReLU activation layer, and a 1×1 bottleneck layer, sequentially connected in series, for processing the input low-frequency components once. 32-channel partial convolution operation, 1 time 64-channel convolution operation, 1 BatchNorm operation, 1 ReLU activation, 1 32-channel convolution operation; The generalization layer is configured to divide the input features into high-frequency components. Low-frequency components and original input features ;in, Represents frequency variables; Represents a time variable; Number of channels; The Pconv layer is configured to extract input features. The feature of the number of channels is used for convolution operation. ,in, This indicates the set threshold parameter. This indicates the number of channels for the input feature.

2. The single-channel epileptic seizure detection device according to claim 1, characterized in that, The preprocessing of the single-channel EEG signal data to be detected includes: The single-channel EEG signal data to be detected is input into a Butterworth bandpass filter for filtering to obtain intermediate EEG data. The intermediate EEG data is then denoised using the Independent Component Analysis (ICA) algorithm to obtain preprocessed single-channel EEG signal data.

3. The single-channel epileptic seizure detection device according to claim 1, characterized in that, The time-frequency features of the extracted preprocessed single-channel EEG signal data include: in, This represents the extracted time-frequency features; This represents preprocessed single-channel EEG signal data; Represents pi; Represents the natural base; Represent the imaginary part of a complex number; Represent real numbers; Represents a time variable. Represents frequency variables; Indicates time.

4. The single-channel epileptic seizure detection device according to claim 1, characterized in that, The standardization process for the extracted time-frequency features includes: in, This represents the time-frequency characteristics after standardization. This represents the extracted time-frequency features; Representation of features The mean; Representation of features The variance.

5. The single-channel epileptic seizure detection device according to claim 1, characterized in that, The input layer is configured to process the input features Expand one dimension, causing it to expand to The Embedding layer is configured to process the input features. Projection as four-dimensional features ,in, This represents the number of channels.

6. The single-channel epileptic seizure detection device according to claim 1, characterized in that, The SCNorm module is configured to perform Softmax processing on the input features to obtain fuzzy weights. and The input features are processed by inputting them into branch 1 and branch 2 respectively; branch 1 includes a BatchNorm layer and a first FFT layer connected in series; branch 2 includes a second FFT layer; based on the fuzzy weights and The amplitude of the output from the first FFT layer and the amplitude of the output from the second FFT layer are weighted to obtain the intermediate amplitude. The intermediate amplitude and the phase of the output from the second FFT layer are then input into the inverse FFT layer for processing to obtain the output characteristics of the SCNorm module.

Images