Individualized epileptic seizure type classification method and system based on super-dimensional calculation

Abstract

The invention provides an individualized epileptic seizure type classification method and system based on super-dimensional calculation, and the method comprises the steps: firstly obtaining a multi-channel electroencephalogram signal of a to-be-classified patient during epileptic seizure, and constructing a super-dimensional vector representing the to-be-classified patient; wherein the super-dimensional vector is formed by combining a channel base vector and a feature vector of each channel in the multi-channel electroencephalogram signal, the channel base vector is a binary super-dimensional vector, and the similarity between the channel base vectors represents the physical relative position distance between the corresponding channels; calculating the similarity between the super-dimensional vector representing the to-be-classified patient and the prototype vector of each epileptic seizure type, and judging the epileptic seizure type of the prototype vector with the highest similarity with the super-dimensional vector representing the to-be-classified patient as the seizure type of the to-be-classified patient during epileptic seizure; according to the method, the super-dimensional vector based on channel embedding is designed, the physiological position characteristics are embedded into the super-dimensional vector in combination with the physiological position of an actual channel, and the classification accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of epileptic seizure type classification, and in particular to a method and system for classifying individualized patient epileptic seizure types based on hyperdimensional computing. Background Art

[0002] Epilepsy is a chronic neurological disease with recurrent seizures. Epileptic seizures are caused by abnormal discharges of neurons in the brain, which makes it impossible for the central nervous system to maintain the balance between excitation and inhibition of brain neurons, resulting in transient functional disorders in the brain. In order to ensure that patients receive the best treatment, in addition to whether epilepsy is detected, it is crucial to correctly classify the type of epilepsy for each patient. This can help doctors predict the course of epileptic seizures, locate epileptic foci, and thus make the most accurate epilepsy treatment. Due to the huge individual differences between epilepsy patients, it is very urgent to classify individualized epileptic seizure types. The existing classification of epileptic seizure types mostly ignores the individual differences of epilepsy, and often uses deep learning methods to classify epileptic seizures. However, for individual patients, the amount of data is not enough to support deep learning, so the accuracy of classification will also decrease, which is not conducive to doctors making customized treatments for patients' specific conditions.

[0003] In recent years, automatic epilepsy detection algorithms based on electroencephalograms have emerged in an endless stream. Deep learning-based convolutional neural networks, recurrent neural networks, etc. have achieved good results in epilepsy detection. With the application of attention mechanisms, the detection results have also improved. However, epilepsy detection algorithms based on deep learning often have high power consumption and are difficult to apply to real-time epilepsy detection on wearable devices. A major challenge in real-time epilepsy detection on wearable devices is to design an efficient framework that can improve the reliability of recognition while meeting the strict computing, memory, and power constraints of embedded platforms that work in real time. In addition, epilepsy detection is difficult because epileptic seizures are greatly affected by the individuality of patients and have a variety of types. Therefore, multi-seizure type classification is crucial in epilepsy research, which makes personalized treatment plans possible, thereby better controlling epileptic seizures. Grouping patients according to seizure type helps to select appropriate treatment methods, helps to understand the mechanisms and characteristics of different seizure types, and promotes the development of targeted intervention measures. Summary of the invention

[0004] In order to solve the problem of epileptic seizure type classification, especially the personalized patient epileptic seizure type classification, that is, how to use the patient's own existing partial seizure data training model to determine the type of subsequent seizures of the patient, the present invention proposes an individualized epileptic seizure type classification method and system based on hyperdimensional computing.

[0005] The technical solution adopted by the present invention is as follows, comprising the following steps:

[0006] A method for individualized epileptic seizure type classification based on hyperdimensional computing, comprising:

[0007] Acquire a multi-channel EEG signal of a patient to be classified when the patient has an epileptic seizure, and construct a hyperdimensional vector representing the patient to be classified; wherein the hyperdimensional vector is composed of channel basis vectors and feature vectors of some channels in the multi-channel EEG signal, and the similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels;

[0008] Calculating the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vectors of various epileptic seizure types, and determining the epileptic seizure type of the prototype vector corresponding to the highest similarity with the hyperdimensional vector representing the patient to be classified as the seizure type of the patient to be classified when the patient has an epileptic seizure;

[0009] Among them, the prototype vectors of various types of epileptic seizures are constructed and obtained using a trainer based on a data set; the data set contains a hyperdimensional vector corresponding to a multi-channel electroencephalogram signal during each historical epileptic seizure.

[0010] Furthermore, the channel basis vector is a binary super-dimensional vector.

[0011] Furthermore, the channel basis vector is constructed by the following method:

[0012] Combining the two-dimensional matrix, the propagation characteristics of EEG signals and the hemispheric characteristics of the brain, all channels used in multi-channel EEG signal acquisition are grouped, where the signal has high channel similarity within a group and low channel similarity between groups;

[0013] For the first channel of each group, a randomly initialized binary superdimensional vector is constructed as the corresponding channel basis vector. Then, each group flips numflips digits based on the channel basis vector of the previous channel in turn according to the distance to construct all the channel basis vectors.

[0014] Further, the grouping is divided into the following 5 groups:

[0015] (1) left group from left frontal pole Fp1 to left occipital pole O1: ‘FP1-F7’, ‘F7-T3’, ‘T3-T5’, ‘T5-O1’;

[0016] (2) right group from right frontal pole Fp to right occipital O2: ‘FP2-F8’, ‘F8-T4’, ‘T4-T6’, ‘T6-O2’;

[0017] (3) left central group from left frontal pole Fp1 to left occipital O1: ‘FP1-F3’, ‘F3-C3’, ‘C3-P3’, ‘P3-O1’;

[0018] (4) right central group from right frontal pole Fp2 to right occipital O2: ‘FP2-F4’, ‘F4-C4’, ‘C4-P4’, ‘P4-O2’;

[0019] (5) The central group passing through left middle temporal T3, left central C3, central point Cz, right central C4, and right middle temporal T4: ‘T3-C3’, ‘C3-CZ’, ‘CZ-C4’, and ‘C4-T4’.

[0020] Furthermore, the channel basis vector is constructed by the following method:

[0021] Combining the two-dimensional matrix, the propagation characteristics of EEG signals and the hemispheric characteristics of the brain, all channels used in multi-channel EEG signal acquisition are divided into 5 groups:

[0022] (1) left group from left frontal pole Fp1 to left occipital pole O1: ‘FP1-F7’, ‘F7-T3’, ‘T3-T5’, ‘T5-O1’;

[0023] (2) right group from right frontal pole Fp to right occipital O2: ‘FP2-F8’, ‘F8-T4’, ‘T4-T6’, ‘T6-O2’;

[0024] (3) left central group from left frontal pole Fp1 to left occipital O1: ‘FP1-F3’, ‘F3-C3’, ‘C3-P3’, ‘P3-O1’;

[0025] (4) right central group from right frontal pole Fp2 to right occipital O2: ‘FP2-F4’, ‘F4-C4’, ‘C4-P4’, ‘P4-O2’;

[0026] (5) the central group passing through left middle temporal T3, left central C3, central point Cz, right central C4, and right middle temporal T4: ‘T3-C3’, ‘C3-CZ’, ‘CZ-C4’, and ‘C4-T4’;

[0027] The first channels of the left group, the right group and the central group are randomly initialized as the first vectors of their corresponding groups; the first channels of the left central group and the right central group are respectively obtained by adding the channel basis vectors of the m channels closest to them, and then normalizing the sum vector through the majority voting threshold method to obtain the corresponding channel basis vector; if m is an even number, an additional hyperdimensional vector is randomly generated and added to the channel basis vectors of the m channels as the final sum vector.

[0028] Furthermore, the numflips formula is as follows:

[0029] numflips=(HD_dim / α / (numch / group-1))

[0030] α means that this group only flips the 1 / α digits of the channel basis vector of the first channel, HD_dim refers to the dimension of the hyperdimensional vector, numch refers to the number of all channels, and group refers to the number of channel groups.

[0031] Furthermore, the feature vector is composed of a feature position vector and an eigenvalue vector, wherein the feature position vector is used to encode the position information of the corresponding feature in the feature vector, and the eigenvalue vector is a hyperdimensional vector constructed based on the eigenvalue of the EEG signal of the corresponding channel.

[0032] Furthermore, the feature vector is constructed based on the number of feature peaks of each channel of the multi-channel EEG signal.

[0033] Furthermore, before extracting the peak number of the EEG signal of each channel, it also includes performing a line thinning process on the channel EEG signal.

[0034] Furthermore, the trainer is composed of an accumulator and a judge. The accumulator accumulates the super-dimensional vectors of data samples belonging to the same class into their corresponding class vectors according to different attack types. After all the data are accumulated, the judge normalizes the total class vectors obtained for each type, and the final super-dimensional vector obtained is the prototype vector of each type.

[0035] A personalized epileptic seizure type classification system based on hyperdimensional computing, used to implement the personalized epileptic seizure type classification method based on hyperdimensional computing, comprising:

[0036] A data acquisition and preprocessing module is used to acquire multi-channel EEG signals of patients to be classified during epileptic seizures, and construct a hyperdimensional vector representing the patients to be classified; wherein the hyperdimensional vector is composed of a channel basis vector and a feature vector of each channel in the multi-channel EEG signal, the channel basis vector is a binary hyperdimensional vector, and the similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels;

[0037] A classification module calculates the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vectors of each type of epileptic seizure, and determines the epileptic seizure type corresponding to the prototype vector with the highest similarity to the hyperdimensional vector representing the patient to be classified as the seizure type of the patient to be classified; wherein the prototype vectors of each type of epileptic seizure are constructed based on a data set using a trainer; and the data set contains a number of hyperdimensional vectors corresponding to epileptic seizure types.

[0038] A storage medium containing computer executable instructions, which, when executed by a computer processor, implement the individualized epileptic seizure type classification method based on hyperdimensional computing.

[0039] A computer program product comprises a computer program / instruction, which, when executed by a processor, implements the steps of a method for individualized epileptic seizure type classification based on hyperdimensional computing.

[0040] Compared with the prior art, the advantages and effects of the present invention are:

[0041] (1) A personalized epileptic seizure type classification method based on hyperdimensional computing was proposed. For the first time, the use of epileptic seizure classification for individual patients was studied, and hyperdimensional computing, which is highly applicable on embedded devices, was used to classify the epileptic seizure types of personalized patients, which can help doctors provide customized treatment for patients.

[0042] (2) A new feature for classifying epileptic seizure types was designed: the DPP feature. In addition to significantly improving the classification accuracy in the hyperdimensional computing model, the newly proposed feature can also improve the classification accuracy of common machine learning classifiers such as support vector machines and lightgbm.

[0043] (3) An initialization module Chembed based on the channel-embedded hyperdimensional basis vector was designed. It combined the physiological position of the actual channel and embedded the physiological position characteristics into the hyperdimensional vector, effectively improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of an epilepsy classification method according to an embodiment of the present invention;

[0045] Figure 2 This is an EEG segment image after preprocessing in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of an EEG topological mapping method according to an embodiment of the present invention;

[0047] Figure 4 Schematic diagram of a method for generating channel basis vectors according to an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the DPP feature extraction process according to an embodiment of the present invention; DETAILED DESCRIPTION

[0049] Hyperdimensional computing (HD) is a computational framework based on random distribution representation inspired by neuroscience. It enhances the symbolic and neural nature of computation by representing data in high-dimensional space. HD uses hyperdimensional vectors to represent data, and the dimensions are usually more than 10,000, which not only improves the robustness and flexibility of data representation, but also can efficiently calculate the similarity between data. In recent years, HD has aroused great interest in many fields, especially in the biomedical field. HD has been used in electrocardiogram, electroencephalogram error-related point detection, gesture recognition, electromyography, etc., and is an effective alternative to machine learning.

[0050] Although hyperdimensional computing has achieved remarkable results in epileptic seizure detection, existing research has mainly focused on distinguishing epileptic seizures from non-seizure states, but has failed to fully consider the diversity and individual characteristics of epileptic seizure types. In particular, how to use only a small portion of each patient's data for training and effectively identify multiple types of epileptic seizures. The present invention aims to fill this gap and designs a hyperdimensional vector initialization module (Chembed) based on channel embedding. This module successfully maps the physiological structure of the brain and the physical distance of the electrodes into similarity representations in hyperdimensional space through a sophisticated channel basis vector initialization strategy, taking into account the three-dimensional ellipsoid shape of the brain and the physical layout of the electrodes, allowing the characteristics of brain signal propagation and the inherent properties of the cerebral hemispheres to be captured. At the same time, combined with a carefully designed group flipping strategy, the model is able to learn more detailed and individualized epileptic seizure characteristics on a limited data set. This innovation provides a new perspective for the development of customized treatment strategies. It embeds novel elements in the existing technology system, while promoting the exploration of the field of epilepsy research and providing patients with more diverse treatment options.

[0051] The present invention is further described below in conjunction with the accompanying drawings, including the basic principles, main features and advantages of the present invention. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. The present invention may also have various changes and improvements, which all fall within the scope of protection of the present invention.

[0052] A specific embodiment of the present invention discloses an individualized epileptic seizure type classification method based on hyperdimensional computing. The present invention constructs a channel basis vector that combines the physiological position of the actual channel, embeds the physiological position characteristics into the hyperdimensional vector representing the patient to be classified, and then judges the epileptic seizure type by combining the similarity with various prototype vectors, such as Figure 1 As shown in the flowchart, it mainly includes two parts: the construction of various prototype vectors and the actual judgment of individualized epileptic seizure types combined with various prototype vectors, as follows:

[0053] The construction of various prototype vectors includes the following steps:

[0054] 1. Data acquisition, constructing a data set, wherein the data set contains a number of hyperdimensional vectors corresponding to epileptic seizure types; specifically, the following sub-steps are included:

[0055] 1.1 Obtain sample set:

[0056] The currently publicly available dataset containing annotations of epileptic seizure types is the Temple University Epilepsy Seizure Data Corpus (TUSZ). TUSZ is the only open source EEG dataset containing annotations of multiple epileptic seizure types, which contains more than 30,000 clinical scalp EEG records collected from outpatient care units, intensive care units, emergency management units, emergency rooms, and various other hospital centers since 2002. The original EEG signal is a multi-channel recording, where the number of channels ranges from 20 to 128. The data is digitized using a 16-bit A / D converter, and the frequency range of the samples is from 250Hz to 1024Hz. The annotation team classified multiple epileptic seizure types based on the EEG report and preliminary analysis. 8 types of epileptic seizures are annotated, and the list of seizure type labels is shown in Table 1. Due to the lack of myoclonic seizure MYSZ seizure type in the dataset, this embodiment excludes it when selecting seizure types. There are 55 patients with multiple seizure types in the data, and each patient has 2 to 5 seizure types. The present invention selects this publicly available data set to obtain the EEG signals and corresponding epileptic seizure labels of each epileptic seizure from these 55 patients, and uses this as a sample set.

[0057] Table 1 List of seizure type labels

[0058]

[0059]

[0060] 1.2 Sample set data preprocessing, including:

[0061] (1) Data resampling: The sampling frequencies of all EEG data in the TUSZ dataset are not the same, and include different sampling frequencies from 250 Hz to 1028 Hz. The data are uniformly resampled to 250 Hz.

[0062] (2) Use of montage: Montage refers to the rational and orderly arrangement of EEG channels. These channels are created to display the activity of the entire head and provide information on lateralization and localization. Common montages include unipolar montage and bipolar montage. The TUSZ dataset uses the 10-20 international standardized electrode placement method. The 10-20 system electrode placement method is the standard electrode placement method specified by the International Electroencephalography Association. The distance from the midpoint of the frontal pole to the root of the nose and the distance from the occipital point to the external occipital protuberance each account for 10% of the total length of this line, and the remaining points are separated by 20% of the total length of this line. Therefore, it is named the 10-20 system. This dataset adopts two different montage recording methods. Montage is a differential view of the data, providing the difference voltage between the electrode and the reference node. Average reference and linked ear reference are two different unipolar montages used in EEG recording. Average reference uses the average value of a certain number of electrodes as a reference, and linked ear reference uses a leading adapter to link the left and right ears. However, in the actual selection process, bipolar montage is often used because bipolar montage can reduce noise. Experiments have found that TCP is the most effective montage. Therefore, in the invention, the following is selected: Figure 3 The 20-lead bipolar montages shown are: 'FP1-F7', 'F7-T3', 'T3-T5', 'T5-O1', 'FP2-F8', 'F8-T4', 'T4-T6', 'T6-O2', 'T3-C3', 'C3-CZ', 'CZ-C4', 'C4-T4', 'FP1-F3', 'F3-C3', 'C3-P3', 'P3-O1', 'FP2-F4', 'F4-C4', 'C4-P4', 'P4-O2'. Among them, FP1, F7, T3, etc. represent standard EEG electrodes on the 10-20 system, and the form similar to 'FP1-F7' represents the subtraction of the EEG data of two electrodes connected by '-', and the EEG electrodes obtained by subtraction are used as the bipolar montage EEG electrodes (channels).

[0063] (3) Data slicing: Clipping training, also known as the window method, is an important data preprocessing method that aims to generate more samples and improve the classification performance of the classifier. In the present invention, each seizure segment in the acquired sample set is sliced ​​into 2s non-overlapping data segments. The 2s data size is selected based on the commonly used selection method in historical literature. The preprocessed EEG data is as follows Figure 2 shown.

[0064] 1.3 Training set and test set division

[0065] The present invention is directed to personalized epileptic seizure type classification. Therefore, the division of training sets and test sets is based on individual patients. For individualized patients, hyperdimensional computing can capture the characteristics of signals through a small number of calculations. Therefore, only one seizure type from each patient needs to be selected as training. Specifically, hyperdimensional computing can obtain EEG seizure type data features from a small number of learnings, and randomly select one piece of each seizure type data of each patient, slice the data as a training set, and slice the remaining seizure data segments as test set data. This is to prevent the data in the training set from appearing in the test set and affecting the experimental classification results.

[0066] 1.4 Constructing a corresponding hyperdimensional vector based on each episode data segment slice, specifically including the following sub-steps:

[0067] 1.4.1 Constructing channel basis vectors

[0068] This is the highlight of the present invention. The brain is a three-dimensional ellipsoid. Although the 10-20 electrodes clearly define the physical distance between electrodes, it is difficult to quantify it for analysis. Therefore, the present invention constructs a channel basis vector for each channel. The similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels, thereby mapping the brain topology to a two-dimensional matrix and approximately mapping the distance between channels to a two-dimensional space. In one embodiment, the method for constructing the above-mentioned channel basis vector is as follows: combining the two-dimensional matrix, the propagation characteristics of the EEG signal and the hemispheric characteristics of the brain, all channels are divided into several groups, and the group stipulates: the signal has high channel similarity within the group and low channel similarity between groups. For each group, the initialization vector of the first channel in the group is first determined as the first vector, that is, the channel basis vector of the first channel. The dimension of the initialization vector is set to 10,000 in the experiment, and the generated channel basis vector is a binary super-dimensional vector (composed of 0 and 1). Based on this, it is preliminarily determined that the channel basis vectors between each group have a smaller similarity, which conforms to the physical relative position distance between the corresponding channels between the groups. A channel can be randomly selected as the first channel. In a preferred embodiment, they can be sorted from small to large according to their physical position size, for example, according to the position of the head from front to back, from left to right, or from back to front, from right to left, and the first one in the sorting is selected as the first channel to facilitate the subsequent construction of the channel basis vector. Then, the first vector is flipped in a specified number of dimensions (from 0 to 1 or from 1 to 0). The number is determined based on the distance between the remaining channels in the group and the first channel. Based on this, it is further determined that the channel basis vectors in each group have a greater similarity, and the channel basis vectors between different groups have a smaller similarity, which is consistent with the physical relative position distance between the channels. Through this method, the physiological position characteristics are embedded into the hyperdimensional vector in combination with the actual channel's physiological position. In a more specific embodiment, as shown in Figure 3Taking the 20-lead bipolar montage shown in the figure as an example, the specific steps are as follows: a. First, the grouping strategy, such as Figure 3 As shown, there are 5 groups in total: (1) The left group from the left frontal pole (Fp1) to the left occipital (O1): 'FP1-F7', 'F7-T3', 'T3-T5', 'T5-O1'. (2) The right group from the right frontal pole (Fp2) to the right occipital (O2): 'FP2-F8', 'F8-T4', 'T4-T6', 'T6-O2'. (3) The left central group from the left frontal pole (Fp1) to the left occipital (O1): 'FP1-F3', 'F3-C3', 'C3-P3', 'P3-O1'. (4) The right central group from the right frontal pole (Fp2) to the right occipital (O2): 'FP2-F4', 'F4-C4', 'C4-P4', 'P4-O2'. (5) Central group passing through left middle temporal (T3), left central (C3), central point (Cz), right central (C4), right middle temporal (T4): 'T3-C3', 'C3-CZ', 'CZ-C4', 'C4-T4'. According to literature reports, the FPz-Fz-Cz-Pz-Oz group is used less frequently, and there are already left groups, left central groups, right groups, and right central groups that can reflect the longitudinal propagation of EEG in the left and right hemispheres, so this group is not included in the grouping range. b. Flip strategy: First, the first channel of each group is randomly initialized as the first vector of their corresponding group. According to the characteristics of the hyperdimensional vector, the Hamming distance of the randomly initialized hyperdimensional vector is about 5000, which is approximately orthogonal, indicating that they are not similar. The remaining three channels after each group are based on the super-dimensional vector of the first channel. According to the distance from the first vector, they flip numflips digits on the basis of the previous channel vector in turn (the original 0 is flipped to 1, and the original 1 is flipped to 0). Specifically, according to the distance from the first vector, the second vector in the group is flipped numflips digits on the basis of the first vector, the third vector in the group is flipped numflips digits on the basis of the second vector in the group, and so on. Figure 4 As shown. Among them, the numflips formula is as follows:

[0069] numflips=(HD_dim / α / (numch / group-1))

[0070] Here, the number α means that the group only flips the 1 / α digits of the first vector to ensure that the similarity within the group is as large as possible. During the experiment, the experiments were set to 2, 4, 8, and 16, and it was found that the setting of 4 was the best. HD_dim refers to the dimension of the hyperdimensional vector, which was set to 10000 in the experiment. numch refers to the number of all channels, which was set to 20 in the experiment. group refers to the number of channel groups, which was set to 5 in this embodiment. Table 2 shows the Hamming distance between the generated channel basis vector and the first channel vector, which is in line with expectations.

[0071] In another specific implementation scheme, the first channels of the left group, the right group and the central group are first randomly initialized as the first vectors of their corresponding groups. Since the source channels of the left central group and the right central group are the same as those of the left group and the right group, respectively, and are very close to the central point, a (1, 10000) super-dimensional vector is randomly generated according to the binding characteristics of the super-dimensional vector, and it is added to the FP1-Fz and T3-C3 channel basis vectors (randomly generated) that are closer to the FP1-F3 channel basis vector, and the majority voting threshold method is set to normalize the sum vector, and the super-dimensional vector of FP1-F3 can be obtained as the channel basis vector of the first channel of the left central group, which will be similar to the first three. Similarly, a (1, 10000) super-dimensional vector is randomly generated, and it is added to the FP1-Fz and T3-C3 that are closer to the FP2-F4, and the majority voting threshold method is set to normalize the sum vector, and the super-dimensional vector of FP2-F4 can be obtained as the channel basis vector of the first channel of the right central group. The reason for generating a random vector is that for the addition of an odd number of super-dimensional vectors, the calculation results can be normalized by majority voting. However, this cannot be achieved for an even number of vectors. It is known that as a binding operation, for addition, the final generated vector will be similar to all input vectors, so introducing a random vector can help reduce the deviation caused by the addition of an even number of super-vectors. After the first channel of each group is initialized, the remaining three channels after each group are based on the super-dimensional vector of the first channel, and according to the distance from the first vector, they flip numflips digits on the basis of the previous channel vector in turn to obtain the channel basis vectors of all channels.

[0072] 1.4.2 Construct feature vector;

[0073] The present invention is applicable to a combination of one or more features currently widely used in the binary classification of epileptic seizures, including average absolute value, average energy, standard deviation, ratio of absolute average values ​​of adjacent coefficients, skewness, kurtosis, etc.; in a specific embodiment, the feature vector of the present invention is composed of a feature position vector and a feature value vector, and the feature position vector is a randomly generated N super-dimensional vector IDi (1≤i≤N) to encode the position information of the corresponding feature in the feature vector, and N represents the feature type. Different feature types have different feature positions, and there is no specific relationship between them, so the feature position vector is randomly initialized. The feature value vector refers to a vector encoded by a specific feature value. According to the level of feature quantization, the similarity between the feature value vectors is set, and the first level quantization and the last level quantization are the least similar. The intermediate quantization level flips the dimension value of the first level quantization super-dimensional vector according to the distance, and the first level quantization super-dimensional vector is randomly initialized. In this embodiment, the feature quantization level is set to the same size as the super-dimensional vector level, that is, 10000. Feature position vectors, eigenvalue vectors and channel basis vectors. These basis vectors are stored in the model's item memory (IM) to facilitate feature encoding and can be regarded as a symbol table or dictionary of vectors defined in the system.

[0074] Table 2: Hamming distance between all channel basis vectors and the first channel basis vector

[0075]

[0076] In a specific implementation scheme, taking the number of peaks based on data slices as a feature as an example, the method for constructing a feature vector is as follows:

[0077] Step 1: DPP feature extraction: In order to simulate the process of doctors identifying epileptic seizures in EEG signals, the present invention extracts the number of signal peaks from the linearly simplified signal as the characteristic value of the signal segment, and names this new feature DPP. Specifically: the number of peaks of each channel of each data slice is obtained, and it is further quantified into a characteristic value vector; in a more specific implementation scheme, the Douglas-Peucker polyline thinning algorithm can be performed on the data before feature extraction to optimize the data. The original sequence is represented as a piecewise linear representation, small fluctuations are ignored, and data points reflecting the main characteristics of the signal are retained. Simulate the signal observed by the human eye. The Douglas-Peucker algorithm is an algorithm for polyline thinning, which simplifies the polyline by reducing the number of points on the polyline while keeping the overall shape of the polyline basically unchanged. The basic idea of ​​the algorithm is to connect the first and last points of the polyline, and then calculate the distance from all points on the polyline to the straight line to find the maximum distance value dmax. If dmax is less than a given threshold DDD, all the middle points on the broken line are discarded; if dmax is greater than or equal to DDD, the point corresponding to dmax is retained, and the broken line is divided into two parts with this point as the boundary, and the method is repeated for these two parts. This process will continue until all dmax are less than DDD, at which time the thinning of the broken line is completed. The advantages of this algorithm include translation and rotation invariance, that is, the thinning result of the broken line does not depend on the specific position and direction of the broken line, but only on the given threshold. In this embodiment, the threshold is set to 10. After the Douglas-Peucker algorithm, the EEG signal is linearly simplified, and the most prominent part of the original data signal can be retained.

[0078] In this embodiment, the data has 20 channels in total, and each multi-dimensional EEG signal finally obtains a feature matrix of (20, 1). Figure 5 The upper left picture represents the thinning process of the line, the star-shaped points represent the points selected by the thinning, the upper right picture is the original picture, the lower right picture is the picture after linear simplification, and the lower left picture is the peak point selection.

[0079] Step 2: Using the proposed arithmetic operation, the channel basis vector and the eigenvector are initialized to obtain a vector representing the data sample, which is called record-based encoding. The eigenvector uses two types of hyperparameters: the feature position vector and the eigenvalue vector, which represent the feature position and the eigenvalue, respectively. The feature position vector encodes the feature position information by randomly generating N hyperdimensional vectors IDi (1≤i≤N). For the eigenvalue vector, its value is first normalized and discretized into a fixed number of boxes, that is, it is quantized into p levels. The eigenvalue vector L generates p hyperdimensional vectors based on the eigenvalue, also known as the level vector Li, i represents the i-th level vector, and there is a correlation between the level vectors. The Hamming distance between the first level vector and the last level vector is the largest and the similarity is the lowest. This can be achieved by predefining the number of flipping bits of the vector to ensure the correlation of the quantization vector level. In this embodiment, p is selected as 10000. Then, each discrete feature position vector is bound to its eigenvalue vector to generate a feature vector, and the binding formula is shown in the following formula.

[0080]

[0081] Here H represents the generated feature vector, L represents the quantized feature value vector, and ID represents the feature position vector. Represents the eigenvalue vector at the feature position.

[0082] After the binding is completed, the bound feature vector is bound to its corresponding channel basis vector. The formula is as follows.

[0083]

[0084] E i ∈{E1,E2,...,E M},where 1≤i≤M

[0085] Here S represents the generated data sample vector, H represents the feature vector on each channel, E represents the channel basis vector, and M is the number of channels.

[0086] Finally, the total hyperdimensional vector is normalized in the same way as when the channel basis vector is initialized, that is, a majority voting method is used. If the value of the hyperdimensional vector dimension is greater than half of the number of channels, it is assigned a value of 1, otherwise it is assigned a value of 0.

[0087] 1.5 Model training, constructing prototype vectors for various types of epileptic seizures:

[0088] After the encoding is completed, the data is sent to the trainer for training. In this embodiment, the trainer is the same as the hyperdimensional calculation for binary classification, and is composed of an accumulator and a judge. The accumulator accumulates the data sample vectors belonging to the same class into the corresponding class vector according to different attack types. After all the training data are accumulated, the judge normalizes the total class vector obtained for each type. The normalization method is the same as that during encoding, using the majority voting method. The final hyperdimensional vector is the prototype vector of each type, and is also the standard for classifying test data later. During training, the output vector of the encoder is stored in the associative memory (AM).

[0089] 2. Practical application: Determine the epileptic seizure type of the patient to be classified by combining the similarity with various prototype vectors

[0090] Obtain multi-channel EEG signals of the patient to be classified during epileptic seizure, and use the same encoding method, that is, after feature extraction, the feature encoding method constructs a hyperdimensional vector representing the patient to be classified;

[0091] Calculate the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vector of each type of epileptic seizure, and determine the epileptic seizure type of the prototype vector corresponding to the prototype vector with the highest similarity to the hyperdimensional vector representing the patient to be classified as the seizure type of the epileptic seizure of the patient to be classified; in a specific implementation scheme, cosine similarity or Hamming distance can be used to measure the similarity between it and the generated prototype vector, and the seizure type corresponding to the model vector with the highest similarity to the hyperdimensional vector of the data to be tested is determined to be the seizure type of the data to be tested.

[0092] In this embodiment, the Hamming distance is used. When the Hamming distance between two vectors is close to 0, it means that they are similar, and close to 0.5 means they are orthogonal or dissimilar. The normalized Hamming distance formula of two vectors A and B is shown in the following formula. d represents the vector dimension.

[0093]

[0094] Due to the imbalance in the number of data of various types of seizures, the present invention chooses to use the weightedF1 index as an evaluation index, and statistics the accuracy and precision of the classification results. As shown in Table 3, different features are classified according to the method of the present invention (hyperdimensional computing model, hyper), and compared with the commonly used machine learning models SVM and LIGHTGBM models. It can be found that from a single feature point of view, the DPP feature is higher than FFT, and the F1 score obtained by amptitude on each classifier is higher, and the result of a single DPP feature is close to the Mav6 feature result of 6 commonly used time series features combined, which proves the effectiveness of the DPP feature proposed in the present invention for the classification of epileptic seizure types. And the F1 effect of the proposed hyperdimensional computing model is better than that of SVM and LIGHTGBM, indicating that the hyperdimensional computing of the present invention has great potential for classification in terms of epileptic seizure types. Among them, the Mav6 feature is 6 features widely used in the binary classification of epileptic seizures, including the average absolute value, average energy, standard deviation, the ratio of the absolute mean values ​​of adjacent coefficients, skewness, and kurtosis. amptitude and FFT represent a time domain feature and a frequency domain feature, respectively.

[0095] Table 3: Comparison of the effects of the features used in different models with other features

[0096]

[0097] The initialization methods of 5 commonly used channel basis vectors are compared with the initialization method based on channel topology structure embedding chembed proposed by the present invention, as shown in Table 4. It can be seen that the F1 of chembed, an initialization method based on channel topology structure embedding, has obtained the highest score, with an average increase of about 2 percentage points, proving that this initialization method embeds the physiological position characteristics into the hyperdimensional vector representing the patient to be classified, effectively improving the accuracy of judging the type of epileptic seizure. The five commonly used initialization methods are specifically introduced as follows: random represents random initialization; sandiwich represents randomly initializing a hyperdimensional vector for the first channel according to the input EEG data, and then the hyperdimensional vector randomly flips a specified number of vector bits for each channel; scalenorand represents randomizing a hyperdimensional vector for the first channel, and then sequentially flips a specified number of vector bits according to the specified EEG order; groupscrand represents grouping EEG, using the same basis vector for the EEG in the same group, and randomly generating hyperdimensional vectors for EEG in different groups.

[0098] Table 4: Different channel basis vector initialization methods for DPP features

[0099]

[0100] Corresponding to the aforementioned embodiment of an individualized epileptic seizure type classification method based on hyperdimensional computing, the present invention also provides an embodiment of an individualized epileptic seizure type classification system based on hyperdimensional computing.

[0101] An embodiment of the present invention provides an individualized epileptic seizure type classification system based on hyperdimensional computing, comprising:

[0102] A data acquisition and preprocessing module is used to acquire multi-channel EEG signals of patients to be classified during epileptic seizures, and construct a hyperdimensional vector representing the patients to be classified; wherein the hyperdimensional vector is composed of a channel basis vector and a feature vector of each channel in the multi-channel EEG signal, the channel basis vector is a binary hyperdimensional vector, and the similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels;

[0103] A classification module is used to calculate the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vector of each type of epileptic seizure, and determine the epileptic seizure type corresponding to the prototype vector with the highest similarity to the hyperdimensional vector representing the patient to be classified as the seizure type of the patient to be classified; wherein the prototype vector of each type of epileptic seizure is constructed based on a data set using a trainer; and the data set contains a number of hyperdimensional vectors corresponding to epileptic seizure types.

[0104] Furthermore, the data acquisition and preprocessing module includes:

[0105] d1: Vector initialization module Chembed: This is the highlight of the present invention. In the initialization of the super-dimensional basis vector, the present invention proposes an initialization module for the super-dimensional vector based on channel embedding, which combines the physiological position of the actual channel, embeds the physiological position characteristics into the super-dimensional vector, and converts the physical distance between channels into the similarity between basis vectors. According to the size of the similarity, the initialization vector is designed. Since there is no relationship between the features of the feature position vector, random initialization is adopted. The eigenvalue vector is initialized according to the feature quantization level.

[0106] d2: EEG feature extraction module: A new feature based on linear simplification, Douglas-Peucker Peak (DPP), is proposed. The processed EEG signal is linearly simplified, and the classic Douglas-Peucker thinning algorithm is used: the original sequence is represented as a piecewise linear representation, small fluctuations are ignored, and data points reflecting the main characteristics of the signal are retained to simulate the signal observed by the naked eye. Afterwards, the number of peaks of the simplified signal is extracted as the feature of the signal. Linear simplification is mainly based on the frequency information of the signal, and the number of peaks is extracted based on the signal amplitude. The newly extracted DPP feature retains the time-frequency information of the signal, which is conducive to signal classification.

[0107] d3: Feature encoding module: discretize the features to different levels to form a feature value vector, bind it with the generated feature position vector and channel basis vector, and generate a hyperdimensional vector of the EEG signal.

[0108] An embodiment of an individualized epileptic seizure type classification system based on hyperdimensional computing of the present invention can be applied to any device with data processing capability, and the device with data processing capability can be a device or apparatus such as a computer.

[0109] The device embodiment can be implemented by software, hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capability in which it is located to read the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, it includes a processor, memory, network interface, and non-volatile memory. In addition, any device with data processing capability in which the device in the embodiment is located can also include other hardware according to the actual function of the device with data processing capability, which will not be described in detail.

[0110] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0111] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0112] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, an individualized epileptic seizure type classification method based on hyperdimensional computing in the above embodiment is implemented.

[0113] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0114] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for individualized epileptic seizure type classification based on hyperdimensional computing, characterized in that: include: Acquire a multi-channel EEG signal of a patient to be classified when the patient has an epileptic seizure, and construct a hyperdimensional vector representing the patient to be classified; wherein the hyperdimensional vector is composed of channel basis vectors and feature vectors of some channels in the multi-channel EEG signal, and the similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels; Calculating the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vectors of various epileptic seizure types, and determining the epileptic seizure type corresponding to the prototype vector having the highest similarity to the hyperdimensional vector representing the patient to be classified as the seizure type of the patient to be classified; Among them, the prototype vectors of various types of epileptic seizures are constructed and obtained using a trainer based on a data set; the data set contains a hyperdimensional vector corresponding to a multi-channel electroencephalogram signal during each historical epileptic seizure.

2. The method according to claim 1, characterized in that The channel basis vector is a binary super-dimensional vector; the channel basis vector is constructed by the following method: Combining the two-dimensional matrix, the propagation characteristics of EEG signals and the hemispheric characteristics of the brain, all channels used in multi-channel EEG signal acquisition are grouped, where the signal has high channel similarity within a group and low channel similarity between groups; For the first channel of each group, a randomly initialized binary superdimensional vector is constructed as the corresponding channel basis vector. Then, each group flips numflips digits based on the channel basis vector of the previous channel in turn according to the distance to construct all the channel basis vectors.

3. The method according to claim 2, characterized in that All channels used in the multi-channel EEG signal acquisition are grouped into the following 5 groups: (1) left group from left frontal pole Fp1 to left occipital pole O1: ‘FP1-F7’, ‘F7-T3’, ‘T3-T5’, ‘T5-O1’; (2) right group from right frontal pole Fp to right occipital O2: ‘FP2-F8’, ‘F8-T4’, ‘T4-T6’, ‘T6-O2’; (3) left central group from left frontal pole Fp1 to left occipital O1: ‘FP1-F3’, ‘F3-C3’, ‘C3-P3’, ‘P3-O1’; (4) right central group from right frontal pole Fp2 to right occipital O2: ‘FP2-F4’, ‘F4-C4’, ‘C4-P4’, ‘P4-O2’; (5) The central group passing through left middle temporal T3, left central C3, central point Cz, right central C4, and right middle temporal T4: ‘T3-C3’, ‘C3-CZ’, ‘CZ-C4’, and ‘C4-T4’.

4. The method according to claim 3, characterized in that The channel basis vector is constructed by the following method: The first channels of the left group, the right group and the central group are randomly initialized as their channel basis vectors and the first vectors of the corresponding groups; the first channels of the left central group and the right central group are respectively obtained by adding the channel basis vectors of the m channels closest to them, and then normalizing the sum vector through the majority voting threshold method to obtain the corresponding channel basis vector; if m is an even number, an additional super-dimensional vector is randomly generated and added to the channel basis vectors of the m channels as the final sum vector.

5. The method according to claim 2, characterized in that: The numflips formula is as follows: numflips=(HD_dim / α / (numch / group-1)) α means that this group only flips the 1 / α digits of the channel basis vector of the first channel, HD_dim refers to the dimension of the hyperdimensional vector, numch refers to the number of all channels, and group refers to the number of channel groups.

6. The method according to claim 1, characterized in that The feature vector is composed of a feature position vector and a feature value vector, wherein the feature position vector is used to encode the position information of the corresponding feature in the feature vector, and the feature value vector is a hyperdimensional vector constructed based on the eigenvalue of the EEG signal of the corresponding channel.

7. The method according to claim 1, characterized in that The feature vector is constructed based on the number of feature peaks of each channel of the multi-channel EEG signal.

8. An individualized epileptic seizure type classification system based on hyperdimensional computing, characterized in that: A method for implementing an individualized epileptic seizure type classification method based on hyperdimensional computing as described in any one of claims 1 to 7, comprising: A data acquisition and preprocessing module is used to acquire multi-channel EEG signals of patients to be classified during epileptic seizures, and construct a hyperdimensional vector representing the patients to be classified; wherein the hyperdimensional vector is composed of a channel basis vector and a feature vector of each channel in the multi-channel EEG signal, the channel basis vector is a binary hyperdimensional vector, and the similarity between the channel basis vectors represents the physical relative position distance between the corresponding channels; A classification module calculates the similarity between the hyperdimensional vector representing the patient to be classified and the prototype vectors of each type of epileptic seizure, and determines the epileptic seizure type corresponding to the prototype vector with the highest similarity to the hyperdimensional vector representing the patient to be classified as the seizure type of the patient to be classified; wherein the prototype vectors of each type of epileptic seizure are constructed based on a data set using a trainer; and the data set contains a number of hyperdimensional vectors corresponding to epileptic seizure types.

9. A storage medium comprising computer executable instructions, which, when executed by a computer processor, implements a method for individualized epileptic seizure type classification based on hyperdimensional computing as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method for individualized epileptic seizure type classification based on hyperdimensional computing as described in any one of claims 1 to 7 are implemented.

Images