Seizure prediction method based on multi-channel graph structure feature mining

Abstract

The application discloses a kind of based on multi-channel atlas structure feature mining's epilepsy attack prediction method, belong to medical image processing technical field.The application includes: signal pre-processing is carried out to electroencephalogram signal, obtains the multi-channel time-frequency domain atlas of electroencephalogram signal based on non-negative Tucker decomposition dimension reduction, contour sampling and contour profile feature extraction are carried out to it to obtain electroencephalogram signal feature, then automatic epilepsy preictal prediction is carried out based on the feature, simultaneously also include automatic identity recognition to the feature that epilepsy preictal is identified.The application small storage overhead, can assist epilepsy patient to be prepared to cope with epilepsy attack;Using non-negative Tucker to reduce the amount of calculation of data dimension reduction;Using the profile of power contour as the multi-channel atlas structure feature of time-frequency domain, the coupling relationship of time, frequency and energy can be well described;Directly using the feature that epilepsy prediction has acquired carries out identity recognition, can simplify system module, accelerate identity recognition response speed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and particularly relates to a seizure prediction method based on multi-channel atlas structure feature mining. BACKGROUND

[0002] Epilepsy is a central nervous system disease caused by abnormal synchronous discharge of a large number of neurons in the brain, and the causes include but are not limited to genetics, drugs, brain injury, metabolic disorders and psychological factors. The common symptoms during the seizure are consciousness disorder, sensory abnormalities and involuntary convulsions. The seizure of epilepsy is unpredictable, rapid and easy to repeat, and the patient may suffer from limb injury or even death due to sudden convulsions and consciousness disorder. In addition, the stigma of epilepsy also brings psychological burden. Most patients can obtain good therapeutic effect through drugs or lesion resection surgery, but about 30% of patients are refractory epilepsy.

[0003] Seizure prediction is of great significance for patients with epilepsy who have not obtained good therapeutic effect. Through seizure prediction, patients can be reminded to take self-protection measures in advance, inform family members or medical staff to provide assistance, and also can link up transcranial magnetic / electric stimulation, brain pacemaker and other devices to inhibit or even avoid the upcoming seizure, so as to minimize the harm caused by epilepsy.

[0004] The seizure of epilepsy can be divided into four periods: pre-seizure period, seizure period, post-seizure period and interictal period. The pre-seizure period is also called the aura period, which generally refers to the tens of minutes before the seizure, and the interictal period refers to the normal period not affected by epilepsy. The essence of seizure prediction is to distinguish the pre-seizure period and the interictal period.

[0005] Compared with intracranial electroencephalogram, magnetic resonance imaging, magnetoencephalogram and positron emission tomography, electroencephalogram (EEG) is most widely used in the field of seizure prediction because of its advantages of low cost, portability, safety and high time resolution. The seizure prediction method based on electroencephalogram can be divided into two categories: traditional machine learning method and neural network based method.

[0006] Traditional machine learning can be divided into two categories according to its designed features: linear feature method and nonlinear feature method. The linear feature method mainly uses time domain, frequency domain, time-frequency domain and spatial domain features of electroencephalogram, and the nonlinear feature method mainly uses Lyapunov exponent or entropy measure. It is worth mentioning that electroencephalogram is a nonlinear multi-source mixed signal, and a single feature often does not contain enough discriminative information required for classification, so these methods generally use multiple features for combination, which leads to redundancy and complexity of the system.

[0007] The neural network method uses convolutional neural network, long short-term memory network, Transformer (transformer network) and graph neural network structure, and can also combine self-encoder and residual connection module to realize end-to-end automatic feature extraction and classification. With strong fitting ability, such method can often obtain excellent performance on existing data, but such method lacks explainability, and overfitting is common, and the inter-individual difference and time-varying of electroencephalogram lead to poor generalization performance; increasing the amount of data is an effective way to improve the generalization performance, but in the medical field such as seizure prediction, data is often scarce and valuable.

[0008] In the seizure prediction or seizure detection, the demand for patient identification has not been concerned in the existing work. Using the same system framework to realize the seizure prediction and the identity recognition can provide personalized services for the patient, for example: based on the identity recognition result, loading the system running parameters optimized for the patient; implementing transcranial magnetic / electric stimulation, brain pacemaker treatment parameters specially set for the patient to block the seizure; notifying the patient's corresponding emergency contact person.

[0009] In addition, if the seizure cannot be prevented, the patient in the seizure state is likely to be unable to effectively communicate with others. Identity recognition can help medical staff quickly obtain the patient's medical history information (such as allergy history, epilepsy sub-type, current treatment plan, etc.), which is crucial for taking the correct first aid measures. SUMMARY

[0010] The present application provides a seizure prediction method based on multi-channel atlas structure features, which uses the contour features of the time-frequency graph contour of the electroencephalogram signal as the classification features to automatically predict the seizure, and further automatically identifies the patient's identity based on the contour features of the contour.

[0011] The technical scheme provided by the present application is as follows:

[0012] The automatic seizure prediction method based on electroencephalogram multi-channel atlas structure feature mining comprises the following steps:

[0013] Step 1: signal preprocessing of the original multi-channel electroencephalogram signal input by the electroencephalogram acquisition device;

[0014] Step 2: continuous wavelet transform is performed on the electroencephalogram signal of each channel after signal preprocessing, and the transformed electroencephalogram data of each channel is obtained, then non-negative Tucker decomposition is used to reduce the dimension of the three-order tensor composed of the electroencephalogram data of all channels, and the reduced multi-channel time-frequency domain atlas is obtained after normalization processing of the reduced time-frequency graph;

[0015] Step 3, extracting a set of contour lines of each channel of the reduced dimension multi-channel time-frequency domain spectrum at non-uniform sampling intervals, extracting the contour features of all contour lines of each channel to obtain a time-frequency map feature vector of the current channel; and obtaining a final electroencephalogram signal feature based on the time-frequency map feature vectors of all channels.

[0016] Step 4, inputting the final electroencephalogram signal feature into the trained bagged tree classifier to automatically identify whether it is in the pre-seizure stage.

[0017] Further, step 4 further comprises: if the identification result of the bagged tree classifier is yes (i.e. predicting that the epilepsy will attack), inputting the corresponding electroencephalogram signal feature into a linear discriminant analysis (LDA) classifier to automatically identify the patient's identity.

[0018] In the present application, for epilepsy prediction, the final electroencephalogram signal feature (denoted as feature vector ) is input into the pre-set bagged tree classifier which has been trained to automatically predict the seizure. The bagged tree is an ensemble learning method, which improves the accuracy and stability of the model by constructing multiple decision trees and combining their prediction results, is not prone to overfitting, has certain robustness and anti-noise ability, and is suitable for solving the problems of low signal-to-noise ratio, low data volume and high time variability in epilepsy prediction; and for patient identity recognition, when the bagged tree predicts that the epilepsy will attack, the same feature vector is input into the pre-set linear discriminant analysis classifier to realize patient identity recognition by feature sharing, avoid the processing cost caused by separately calculating the features for identity recognition, and facilitate the rapid development of downstream tasks relying on identity recognition. Since the electroencephalogram of individuals has great difference, identity recognition does not need the heavy ensemble learning mechanism such as bagged tree, therefore, the present application adopts the simple and efficient linear discriminant analysis which is suitable for multi-classification tasks.

[0019] Further, the signal preprocessing in step 1 comprises: downsampling and filtering to facilitate reducing the data size and reducing signal noise.

[0020] Further, when down-sampling, the original multi-channel electroencephalogram signal is down-sampled to a specified sampling frequency, preferably 200 Hz, allowing a certain range of deviation. Since the electroencephalogram signal is collected on the scalp, the original high-frequency components in the brain will encounter higher impedance in the skull and skin tissue, and the effective frequency range of the collected original multi-channel electroencephalogram signal is approximately 0.5 Hz to 60 Hz. The sampling frequency only needs to ensure that the signal in the effective frequency range will not alias, and a too high sampling frequency will increase the cost of storage and signal processing. Therefore, the present application selects to down-sample the original electroencephalogram signal to 200 Hz. This also has the advantage that when deploying the present application in different electroencephalogram collection hardware, the signal processing module obtains data with a constant sampling frequency at the inlet, which is conducive to the unified design and stable operation of the system.

[0021] Further, when filtering, a band-pass filter is used, preferably a band-pass filter of 0.5 Hz~40 Hz, and the filter is preferably a 4th order Butterworth filter. The wave can suppress noise components outside the effective frequency range, so a band-pass filter of 0.5 Hz~40 Hz is used. The Butterworth filter is a linear phase filter, and it has a smooth response in the passband, which can better avoid distortion of the electroencephalogram signal.

[0022] Further, when performing continuous wavelet transform, a bump wavelet is used as the wavelet basis to obtain good oscillation spike capture capability. In addition, for the time-frequency graph obtained by wavelet transform (all channel electroencephalogram data), the main part with a frequency of 0.5 Hz~40 Hz is intercepted for non-negative Tucker decomposition to further reduce the computational load. The pre-epilepsy electroencephalogram signal may not have significant changes in the time domain, and analysis in the frequency domain is a better perspective. However, due to the time-varying nature of the electroencephalogram signal, its frequency components will change dynamically over time. The present application uses continuous wavelet transform to process each channel of the electroencephalogram signal, and selects the widely used bump wavelet as the wavelet basis to obtain good oscillation spike capture capability.

[0023] The current electroencephalogram acquisition device has multiple channels to record the electrical signals of different brain regions unless it is used in a special scene. The electroencephalogram signals recorded by these channels are not independent: due to the volume conduction effect, the electrode of the current channel collects the superposition of signals generated by multiple independent sources in different brain regions in the skull, which means that there is a lot of correlation and redundancy in the multi-channel electroencephalogram signals collected, so the application adopts non-negative Tucker decomposition to reduce the dimension of the truncated part of the time-frequency diagram obtained by wavelet transform. In the application, the two-dimensional time-frequency matrix of each channel is stacked to naturally form a three-order tensor. Tensor is the generalization of matrix and is the natural representation of multi-dimensional data. Tucker decomposition is a commonly used tensor decomposition method, which can be regarded as a high-dimensional generalization of singular value decomposition, has good data compression and dimension reduction capability, and can extract key features in high-dimensional data. The original Tucker is orthogonal, and the addition of non-negative constraint makes the core tensor and mode matrix obtained by Tucker decomposition non-negative, which is more consistent with the physical reality and has good interpretability. Therefore, the application adopts non-negative Tucker decomposition, uses the core tensor in the decomposition result as the dimension reduction representation of the original data, reduces the time-frequency point number and channel number of the data while retaining the key information. The optimization objective of non-negative Tucker decomposition is:

[0024] ,

[0025] ,

[0026] wherein, is the original electroencephalogram tensor (composed of the truncated part of the time-frequency diagram obtained by wavelet transform, i.e. the electroencephalogram tensor before decomposition), represents the real number field, represents the time point number before decomposition, represents the frequency point number before decomposition, represents the channel number before decomposition, is the core tensor of decomposition, is the mode matrix of different dimensions, n is the dimension identifier, and its value is 1, 2, 3, is the unit matrix, represents the F norm, and are modulo-n multiplication, which is used for multiplication between tensors and matrices, represents the multiplication of the first mode of the core tensor and the mode matrix , represents the multiplication of the second mode of the tensor and the mode matrix Multiplication. The optimization problem can be solved by non-negative alternating least squares, high-order multiplication and many other algorithms. In the present application, the non-negative alternating least squares method is adopted because it has a balanced memory consumption and iteration number, and is suitable for use in scenarios with small data size similar to the present application.

[0027] The rank of the non-negative Tucker decomposition is a key hyperparameter that affects performance. In the present application, the rank is taken as half of the dimension in the corresponding direction of the original tensor. Specifically, the compressed data after dimension reduction is , and the data size will be reduced to one-eighth of the original. Among them, is the Tucker rank.

[0028] Further, in step 3, the contour features include contour proportion features, time domain distribution features and frequency domain distribution features.

[0029] Further, step 3 specifically includes:

[0030] Step 301: A set of contours of the time-frequency domain graph of each channel is extracted in a linearly decreasing manner with a sampling interval, and all coordinate points on each contour are collected to form a two-dimensional coordinate point set , where the subscript is the contour identifier of each channel;

[0031] Step 302: Project the two-dimensional coordinate point set onto the time axis and the frequency axis respectively to obtain a one-dimensional point set of the time axis and a one-dimensional point set of the frequency axis;

[0032] Calculate the contour proportion features from the number of elements of the two-dimensional coordinate point set , and calculate the mean and variance of the one-dimensional point sets and as the time domain distribution features and the frequency domain distribution features, respectively;

[0033] Splice the contour proportion features, the time domain distribution features and the frequency domain distribution features of all contours of each channel to obtain the time-frequency graph feature vector of the current channel; and obtain the final electroencephalogram signal feature based on the time-frequency graph feature vectors of all channels.

[0034] In addition, in step 2, the normalization process specifically includes: normalizing the power of each channel time-frequency graph obtained after performing non-negative Tucker decomposition (i.e., the time-frequency graph of each channel after dimension reduction) to the interval (0, 1] to obtain a normalized time-frequency graph G. This way, the same sampling points can be used when sampling the contours of the time-frequency graphs of different channels.

[0035] Further, step 301 specifically includes:

[0036] In the interval (0, 1], starting from 0, sampling starts at a preset initial interval (preferably 0.05), and after each point is sampled, the sampling interval is attenuated according to a specified value (preferably 0.001), and a plurality of points collected constitute a sampling point set S;

[0037] For any point in the sampling point set S , the height corresponding to the point is defined as x, and the point represents an isohypse of the normalized time-frequency graph G of the current channel with a value of x; all points in the normalized time-frequency graph G with values in the range of are attributed to the isohypse corresponding to the point , and a two-dimensional coordinate point set of the current isohypse is obtained, wherein each two-dimensional coordinate point set corresponds to an isohypse contour, and wherein represents a preset deviation, preferably 0.02; the reason for this is that there is a deviation between the power value calculated by wavelet change and the real power value of the frequency point at the current time, and slightly expanding the value of the isohypse can avoid omission.

[0038] Further, in step 302, the isohypse proportion feature, the time domain distribution feature and the frequency domain distribution feature are respectively:

[0039] The isohypse proportion feature of the isohypse is: , which represents the proportion of the power value corresponding to the current isohypse in all time-frequency points; wherein the total number of points of all isohypses , represents the number of isohypses of the time-frequency domain graph of each channel, represents the number of elements in the two-dimensional coordinate point set of the isohypse ;

[0040] The time domain distribution feature is the mean and variance of the one-dimensional point set ;

[0041] The frequency domain distribution feature is the mean and variance of the one-dimensional point set .

[0042] The technical solution provided by the application at least brings the following beneficial effects:

[0043] (1) The present application can obtain electroencephalogram signals from electroencephalogram acquisition hardware, use non-negative Tucker decomposition dimension reduction, and adopt the contour features of time-frequency graph contour lines as shared features, and then realize automatic prediction of epilepsy and automatic identification of patient identity. Compared with the existing technology, the present application has small storage overhead and can assist epilepsy patients to prepare for seizures;

[0044] (2) The present application uses non-negative Tucker to reduce the dimension of data, which reduces the calculation amount of the subsequent processing flow. Compared with the traditional principal component analysis and other dimension reduction methods, the present application can better explore and retain the complex correlation of the original electroencephalogram data in the time domain, frequency domain and spatial domain;

[0045] (3) The present application uses the contour of power contour as the multi-channel graph structure feature in the time-frequency domain. The electroencephalogram signal has rich discriminative information in the time-frequency domain, and the most important is that high-power frequency components appear at some specific time points. The contour feature of the time-frequency graph contour line can well depict the coupling relationship of time, frequency and energy. By sampling the non-uniform contour line from sparse to dense, the number of contour lines is reduced while ensuring the detailed description of the high-power interval;

[0046] (4) The present application directly uses the features obtained during epilepsy prediction to identify the identity, which can simplify the system module and speed up the response speed of identity recognition. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 is the processing flow chart of the epilepsy seizure prediction method based on multi-channel graph structure feature mining provided by the embodiment of the present application.

[0049] Figure 2 is the processing process schematic diagram of the epilepsy seizure prediction method based on multi-channel graph structure feature mining provided by the embodiment of the present application in application.

[0050] Figure 3 is the contour line section profile graph of the electroencephalogram in the embodiment of the present application when the normalized power is 0.7.

[0051] Figure 4 is the contour line projection schematic diagram of the electroencephalogram time-frequency graph in the embodiment of the present application.

[0052] Figure 5This is the normalized power statistical histogram of the EEG time-frequency graph in this embodiment of the invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0054] It should be noted that the pre-seizure EEG data used in this embodiment of the invention are from patients diagnosed with refractory epilepsy. The EEG acquisition hardware has 18 channels and an original sampling frequency of 500Hz. This embodiment of the invention reads data through a buffer of 3 seconds (1500 time points).

[0055] As one possible implementation, see Figure 1 and Figure 2 The specific implementation steps of the epileptic seizure prediction method based on multi-channel map structure feature mining provided in this embodiment of the invention include:

[0056] Step S1: Perform signal preprocessing on the EEG signal by downsampling and filtering the raw multi-channel EEG signal input from the EEG acquisition device to reduce data size and signal noise.

[0057] Step S2: Obtain the multi-channel time-frequency domain spectrum of the reduced EEG signal, including performing continuous wavelet transform on each channel, using non-negative Tucker decomposition to reduce the dimension of the third-order tensor composed of data from all channels, obtaining the reduced time-frequency image, and then normalizing the reduced time-frequency image.

[0058] The signal preprocessing in step S1 (downsampling and filtering) and step S2 (continuous wavelet transform and non-negative Tucker decomposition) constitute... Figure 1 The data processing stage of the EEG signal shown in the figure;

[0059] Step S3 involves contour line sampling of the time-frequency domain map and contour feature extraction of the contour lines, including extracting contour line groups at non-uniform sampling intervals, calculating the three features of contour line proportion, time domain distribution, and frequency domain distribution, and combining the features of each contour line to form the final EEG signal features.

[0060] Step S4, automatic pre-seizure prediction based on electroencephalogram signal features, inputting the electroencephalogram signal features obtained in step S3 into the trained bagged tree classifier to automatically identify whether the current prediction object is in the pre-seizure stage. Further, step S4 of the embodiment of the present application also includes identity recognition, that is, when it is identified that the electroencephalogram signal features of the current prediction object are in the pre-seizure stage, the same electroencephalogram signal features used for automatic pre-seizure prediction are directly input into the trained linear discriminant analysis classifier to automatically identify the identity of the patient. Then, based on the detection results (automatic pre-seizure prediction results and automatic patient identity recognition results), the downstream related tasks are performed.

[0061] Preferably, step S1 specifically comprises:

[0062] Step S101, downsampling. Since the electroencephalogram signal is collected on the scalp, the original high-frequency components in the brain will encounter high impedance in the skull and skin tissue, and the effective frequency range of the collected signal is approximately 0.5Hz to 60Hz. The sampling frequency only needs to ensure that the signal in the effective frequency range will not alias, and a too high sampling frequency will increase the cost of storage and signal processing. The hardware original sampling frequency of the embodiment is 500Hz, which needs to be downsampled to 200Hz; when the embodiment is deployed in different electroencephalogram collection hardware, the inlet signal acquisition of the deployed signal processing module obtains data with a constant sampling frequency, which is conducive to the unified design and stable operation of the system.

[0063] Step S102, filtering. Filtering can suppress noise components outside the effective frequency range, and the embodiment adopts a 0.5Hz~40Hz band-pass filter, and the filter adopts a 4th order Butterworth filter.

[0064] Preferably, step S2 specifically comprises:

[0065] Step S201, continuous wavelet transform. The embodiment uses continuous wavelet transform to process the electroencephalogram signal of each channel, and preferably uses the widely used bump wavelet as the wavelet basis to obtain good oscillation spike capture capability. For the time-frequency graph obtained by wavelet transform, the main part with a frequency of 0.5Hz ~ 40Hz is intercepted.

[0066] Step S201, non-negative Tucker decomposition is used for dimension reduction, and the non-negative Tucker decomposition optimization target is set to perform dimension reduction processing on the intercepted image of the time-frequency graph obtained by wavelet transform, so as to obtain the time-frequency graph of each channel after dimension reduction.

[0067] Preferably, step 3 specifically comprises:

[0068] Step S301, a set of contour lines are extracted in a linearly decreasing manner of sampling interval, and all coordinate points on each contour line are collected to form an equivalent two-dimensional coordinate point set .

[0069] Step S302, contour feature extraction. The two-dimensional coordinate point set is projected onto two coordinate axes (time axis and frequency axis) respectively to obtain two one-dimensional point sets and . The proportion feature is calculated from the number of elements of the two-dimensional coordinate point set , and the mean and variance of and are calculated as the time domain distribution and frequency domain distribution features, respectively. Finally, the three features of all contour lines are spliced as the final overall feature to obtain the final electroencephalogram signal feature.

[0070] In this embodiment, the power of each channel time-frequency graph after step 201 dimension reduction is normalized to the interval (0, 1], and the normalized time-frequency graph G is obtained. In this way, the same sampling points can be used when sampling the contour lines of the time-frequency graphs of different channels.

[0071] Through the visualization of the time-frequency graph Figure 3 and the power statistical histogram Figure 4 , it can be found that low-power time-frequency points account for the majority and can be regarded as a background. In fact, the difference between different brain wave modes mainly reflects the appearance of high-power time-frequency points at specific time points and frequency points. Given the importance of high-power time-frequency points, it is reasonable to sample the high-power interval at a higher density when determining a set of contour lines. This embodiment includes the following steps when performing step S301 based on the normalized time-frequency graph G:

[0072] In the interval (0, 1], starting from 0, sampling starts with an initial interval of 0.05. After each point is sampled, the sampling interval decays by 0.001, and a total of 27 points are collected to form a sampling point set S. The correspondence between the contour lines and the sampling positions is shown in Table 1:

[0073] Table 1: Sampling position correspondence table of time-frequency graph power contour lines (a total of 27)

[0074]

[0075] The first sampling point in Table 1 is , so the sampling can discard the (0, 0.05) extremely low power interval. From Figure 5The histogram statistics of the time-frequency points in this interval can be seen that the number of time-frequency points in this interval is extremely large, and sampling this interval will greatly increase the amount of data for subsequent processing. More importantly, the extremely low power interval is a kind of background, and there is not much useful information. The interval of the last sampling is 0.023, compared with the initial sampling density, the sampling density of the last segment is approximately doubled, which can realize the non-uniform sampling from sparse to dense in the interval (0, 1].

[0076] For any point in the sampling point set S , the corresponding height is denoted as , that is , which represents the value (height) of the normalized time-frequency diagram G of the channel An isohypse. When determining the points contained in the isohypse, the embodiment considers all points in the normalized time-frequency diagram G whose values are in the range of as belonging to this isohypse. The reason for this is that there is a deviation between the power value calculated by wavelet transform and the real power value of the frequency point at the current time, and slightly expanding the value of the isohypse can avoid omission.

[0077] Through the above processing, the two-dimensional coordinate point set of an isohypse can be determined as , and in the embodiment, .

[0078] Further, step S302 specifically includes:

[0079] For each two-dimensional coordinate point set , there is a contour line corresponding to it. Taking all the horizontal coordinates (time dimension) to form a one-dimensional point set representing the energy distribution characteristics in time; similarly, taking all the vertical coordinates (frequency dimension) to form a one-dimensional point set representing the energy distribution characteristics in frequency. And are one-dimensional sequences, and the embodiment uses the mean and variance of the sequence as the characteristics, and uses , to represent the mean and variance of the one-dimensional point set , respectively, and uses , to represent the mean and variance of the one-dimensional point set , respectively.

[0080] Then the proportion feature of the current isohypse needs to be calculated: let the total number of isohypses be

[0081] ,

[0082] wherein, Contour lines The number of elements represents the number of time-frequency points where the power is the current contour value. (Contour line) Specific gravity characteristics for:

[0083] ,

[0084] in, This represents the percentage of the power value corresponding to the current contour line across all time and frequency points.

[0085] The three features obtained by splicing together are the feature vectors of the current contour lines. v i = [ r i , m T i , v T i , m F i , m F i ] The time-frequency eigenvector of the entire normalized time-frequency graph G is: v ∈ R 1× 135 = [ V 1 , V 2 , … , V 27 ] That is, the number of contour lines included in each channel is defined as follows. Then the eigenvectors of the time-frequency diagram The dimension is .

[0086] The data after non-negative Tucker compression is preserved. Each channel can yield a time-frequency plot feature vector. splicing the time-frequency feature vectors of all channels The final feature vector of this input data is obtained. , i.e., eigenvectors This is the final characteristic of the brainwave signal.

[0087] Furthermore, step S4 specifically involves:

[0088] S401, Epilepsy Prediction. The feature vector... Input the pre-trained bagged tree classifiers to automatically predict epileptic seizures.

[0089] S402 patient identity recognition. When the bagged tree predicts that a seizure will occur, the same feature vector The input pre-set linear discriminant analysis classifier is used for patient identity recognition by means of feature sharing, processing cost caused by calculating features for identity recognition is avoided, and downstream tasks relying on identity recognition can be quickly carried out.

[0090] In the example embodiment, the seizure prediction based on the multi-channel graph structure feature provided by the embodiment of the present application realizes the seizure prediction under the unified feature extraction framework and the patient identity recognition, the related recognition result can be used for the downstream task, the seizure is early warned, the personalized help is provided for the patient, the harm caused by the seizure is reduced to the greatest extent, and the application value is high.

[0091] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0092] The above only describes some embodiments of the present application. Those skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present application. These all belong to the protection scope of the present application.

Claims

1. A method for predicting epileptic seizures based on multi-channel map structural feature mining, characterized in that, Includes the following steps: Step 1: Perform signal preprocessing on the raw multi-channel EEG signals input from the EEG acquisition device; Step 2: Perform continuous wavelet transform on the EEG signal of each channel after signal preprocessing to obtain the transformed EEG data of each channel. Then, use non-negative Tucker decomposition to reduce the dimension of the third-order tensor composed of the EEG data of all channels. Normalize the dimension-reduced time-frequency map to obtain the dimension-reduced multi-channel time-frequency domain map. Step 3: Extract a set of contour lines for each channel of the dimensionality-reduced multi-channel time-frequency domain map using a non-uniform sampling interval, extract the contour features of all contour lines in each channel to obtain the time-frequency map feature vector of the current channel; obtain the final EEG signal features based on the time-frequency map feature vectors of all channels. Step 4: Input the final EEG signal features into the trained bag tree classifier to automatically identify whether the patient is in the pre-seizure stage.

2. The method as described in claim 1, characterized in that, Step 4 also includes: if the identification result of the bag tree classifier is yes, then the corresponding EEG signal features are input into the linear discriminant analysis classifier for automatic identification of the patient's identity.

3. The method as described in claim 1 or 2, characterized in that, The signal preprocessing in step 1 includes downsampling and filtering; during downsampling, the original multi-channel EEG signal is downsampled to around 200Hz; during filtering, a bandpass filter of 0.5Hz to 40Hz is used.

4. The method as described in claim 1 or 2, characterized in that, The continuous wavelet transform uses the Bump wavelet as the wavelet basis, and non-negative Tucker decomposition is performed on the portion of the time-frequency graph obtained by the wavelet transform with frequencies between 0.5 Hz and 40 Hz.

5. The method as described in claim 1 or 2, characterized in that, When performing a nonnegative Tucker decomposition, the rank is half the dimension of the tensor in the corresponding direction before decomposition.

6. The method as described in claim 1 or 2, characterized in that, In step 2, the normalization process is as follows: for the time-frequency graph of each channel obtained after performing non-negative Tucker decomposition, the power is normalized to the interval (0, 1] to obtain the normalized time-frequency graph G.

7. The method as described in claim 6, characterized in that, In step 3, the contour features of contour lines include contour line weight features, time domain distribution features, and frequency domain distribution features.

8. The method as described in claim 7, characterized in that, Step 3 specifically includes: Step 301: Extract a set of contour lines from the time-frequency domain spectrum of each channel using a linearly decreasing sampling interval, and collect all coordinate points on each contour line to form an equivalent two-dimensional coordinate point set. , where subscript Contour identifiers for each channel; Step 302, set the two-dimensional coordinate points Projecting onto the time axis and frequency axis respectively, we obtain a one-dimensional point set on the time axis. and the one-dimensional point set of the frequency axis ; From a set of two-dimensional coordinate points The number of elements is used to calculate the contour line weight characteristics, and the one-dimensional point set is calculated separately. and The mean and variance are used as the time-domain distribution characteristics and frequency-domain distribution characteristics, respectively. The contour weight characteristics, time domain distribution characteristics, and frequency domain distribution characteristics of all contour lines in each channel are combined to obtain the time-frequency map feature vector of the current channel; the final EEG signal characteristics are obtained based on the time-frequency map feature vectors of all channels.

9. The method as described in claim 8, characterized in that, Step 301 specifically involves: In the interval (0, 1], sampling begins from 0 with a preset initial interval. After each point is sampled, the sampling interval is attenuated according to a specified value, and the collected points form a sampling point set S. For any point in the set of sampling points S Let its corresponding height be x, and the point A contour line representing the value of the normalized time-frequency plot G of the current channel as x; all values ​​in the normalized time-frequency plot G are plotted on... All points within the range belong to the point. The corresponding contour lines yield the two-dimensional coordinate point set of the current contour lines. Each set of two-dimensional coordinate points Each corresponds to a contour line outline, among which, This indicates the preset deviation.

10. The method as described in claim 8, characterized in that, In step 302, the contour line weight characteristics, time-domain distribution characteristics, and frequency-domain distribution characteristics are as follows: contour lines The contour line weight characteristics are as follows: The total number of points on all contour lines. N represents the number of contour lines in the time-frequency domain spectrum of each channel. Contour lines Two-dimensional coordinate point set The number of elements; The temporal distribution characteristics are a one-dimensional point set. mean and variance ; The frequency domain distribution is characterized by a one-dimensional point set. mean and variance .

Images