Spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation

Abstract

The present invention discloses a spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation. First, a plurality of EEG waveform features are used to perform sample screening on each EEG channel data based on a threshold method, and finally a multi-channel spike wave candidate sample is obtained through data segmentation; data analysis is performed on the obtained multi-channel spike wave candidate sample, and a plurality of weight calculations are performed based on the optimal channel. A weighted algorithm is used to generate single-channel candidate sample data from the multi-channel spike wave candidate sample; finally, a single-channel spike wave classification algorithm is used to complete the classification detection of the generated data. The present invention can not only extract effective information from each channel data and improve spike wave recognition performance through a weighted generation algorithm for multi-channel data, but also provide a process for screening multi-channel spike wave candidate samples, and provide a standard for subsequent establishment of data sets for other similar algorithms. In addition, by further analyzing the weights used in weighted data generation, the specific channel location where the spike wave is generated can also be detected.

Description

Technical Field

[0001] The present invention belongs to the field of electroencephalogram (EEG) signal processing and relates to a spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation. Background Art

[0002] Epilepsy is a common chronic neurological disease that poses a serious threat to the health and life of children and adults. Spike waves and their complex waveforms are the pathological basis of epileptic seizures. Spike wave discharge duration and location are of great significance, and spike wave detection is the first step in determining these parameters.

[0003] Existing spike wave detection methods can be divided into two types based on the input EEG data dimension: single-channel EEG data and multi-channel EEG data. The single-channel data-based method has a small amount of data, simpler feature extraction and classification model building methods, and faster detection speed; the multi-channel data-based method takes into account the physical model of electrical signal conduction and improves detection accuracy through features between multiple channels. These detection methods have the following disadvantages:

[0004] 1. Methods based on single-channel data lack information about other EEG channels at the same moment, and are less effective in identifying artifacts such as spikes.

[0005] 2. Compared with the former, the method based on multi-channel data has high algorithm complexity and is time-consuming, and non-spike channel data may also interfere with recognition.

[0006] Based on the characteristics of spike waves in single-channel and multi-channel data, the present invention proposes a spike wave discharge detection algorithm. By screening out multi-channel sample data that meets the spike wave characteristics from the EEG signal, a single-channel data sample is calculated through an intelligent weighted data generation algorithm. Combined with the existing single-channel detection algorithm based on time series features and LSTM network architecture (patent number: 202011504990.4), a spike wave detection effect with higher accuracy, stronger anti-interference ability and better generalization ability is achieved. Summary of the Invention

[0007] This invention uses multi-channel spike wave weighted fusion to propose an efficient spike wave discharge detection algorithm. Its core innovations are: (1) the threshold method of waveform characteristics is used to quickly screen out sample data that meets the spike wave characteristics from the multi-channel EEG signal, and the secondary screening of samples is completed based on the characteristics of spike wave discharge under the bipolar lead EEG data. While achieving fast and efficient screening of spike wave candidate samples, it also removes some artifact interference with similar waveforms, effectively improving the actual detection efficiency; (2) based on the intelligent weighting algorithm of multi-angle weights, single-channel samples are generated, while retaining the EEG multi-channel information, feature dimensionality reduction and high-recognition feature extraction are achieved, greatly reducing the computational complexity. The proposed data screening and sample generation method is combined with the existing single-channel detection algorithm based on time series features and LSTM network architecture to achieve high-precision spike wave classification detection.

[0008] The spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation includes the following steps:

[0009] Step 1: Utilize multiple EEG waveform features to perform sample screening on each EEG channel data based on the threshold method, and obtain multi-channel spike wave candidate sample information by summarizing the screening results. Then, artifacts are removed based on the multi-channel characteristics of spike wave discharges, and finally multi-channel spike wave candidate samples are obtained through data segmentation.

[0010] The multi-channel spike wave candidate sample information includes the discharge time and channel distribution information of possible spike wave discharge events; the artifact removal is based on the "tit-for-tat" phenomenon unique to spike wave discharges under bipolar lead electroencephalogram to eliminate interference information in the multi-channel spike wave candidate sample information that does not meet the above phenomenon; the data segmentation refers to segmenting the input EEG signal based on the obtained multi-channel spike wave candidate sample information to obtain multi-channel spike wave candidate samples.

[0011] Step 2: Perform data analysis on the obtained multi-channel spike candidate samples, perform multiple weight calculations based on the optimal channel, and finally generate single-channel candidate sample data from the multi-channel spike candidate samples through a weighted algorithm.

[0012] The data analysis refers to determining the channel in each channel of the candidate sample that is most likely to have spike discharges based on amplitude and waveform similarity; the weights include three weights based on amplitude, waveform similarity, and distance from the optimal channel; the weighted algorithm refers to performing a weighted sum of the channel data of the multi-channel candidate samples based on the three weights obtained above, and finally obtaining the single-channel sample data.

[0013] Step 3: Use the single-channel spike classification algorithm to complete the classification detection of the generated data.

[0014] The single-channel spike classification algorithm refers to a single-channel detection algorithm based on time series features and LSTM network architecture. The detection process includes sample time series feature extraction, feature fusion and stacked Bi-LSTM neural network classification.

[0015] The specific implementation of step 1 is as follows:

[0016] 1-1. The original input multi-channel EEG data was filtered in the basic frequency domain using a 1-70 Hz bandpass filter and a 50 Hz notch filter. The processed data was used for subsequent data analysis.

[0017] 1-2. The discharge time of the initial spike candidate samples on each EEG channel was determined by minima. Five waveform features of each sample were further calculated, including amplitude, duration, left half-wave slope, right half-wave slope, and overall slope. Samples were then screened using a thresholding method. Only samples that passed the thresholding method met all threshold conditions, ultimately obtaining single-channel spike candidate samples suspected of having spike discharges on all channels. The threshold hyperparameters used in the thresholding method were determined by the upper and lower limits of the 99.9% confidence intervals of the five feature databases calculated using the Monte Carlo method. The dataset used for statistical analysis was a database of labeled spike samples.

[0018] 1-3. After obtaining all single-channel spike candidate sample information, the label information of all channels is summarized to obtain multi-channel candidate spike sample information, including candidate spike discharge time and channel distribution information.

[0019] The channel labeling information is summarized as follows: when the discharge time difference between candidate spikes in different channels is less than 0.02 seconds, based on the spike discharge generation and conduction model, it is determined that these discharge phenomena detected on the electrodes of different channels are caused by the same discharge source inside the brain. In this case, information is summarized, and the candidate spike discharge time in the obtained multi-channel spike candidate sample information is the mean of the discharge time of all single-channel samples. The channel distribution information records all channel numbers appearing in the sample group.

[0020] 1-4. When locating and analyzing waveforms with significant polarity changes using bipolar EEG data, there are more pronounced features compared to data using reference leads. Clinically, this characteristic phenomenon caused by spike discharges is referred to as "tit-for-tat," meaning that spike discharges can only occur at all electrodes between the first bipolar lead channel with a positive wave and the last bipolar lead channel with a negative wave. Based on this phenomenon, artifact removal is performed on multi-channel candidate spike sample information using the following steps:

[0021] (1) Obtain bipolar lead EEG data related to the channel where candidate spike wave discharge phenomena are recorded.

[0022] (2) Traverse the bipolar lead data corresponding to each multi-channel spike wave candidate sample information in turn, and find the bipolar lead channel where the first positive peak and the last negative peak are located.

[0023] (3) According to the corresponding principle, the interference channels that do not meet the "tit-for-tat" condition in the multi-channel spike candidate sample information are eliminated.

[0024] 1-5. Based on the multi-channel spike candidate sample information after artifact removal, multi-channel spike candidate samples are generated through data segmentation. Each sample contains multi-channel data centered at the discharge time point and lasting 0.25 seconds.

[0025] The specific operations of step 2 are as follows:

[0026] 2-1. Determine the first-class weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the waveform amplitude amplitude , the steps are as follows:

[0027] (1) The amplitude feature quantity amplitude is recorded as the amplitude of the waveform at the center of each channel data.

[0028] (2) The amplitude weight matrix obtained by normalizing the amplitude feature values ​​of all channels is recorded as the weight matrix W amplitude .

[0029] 2-2. Calculate the standard spike data set, which is the cluster center data obtained by the K-means clustering algorithm based on the existing labeled spike sample database.

[0030] 2-3. Determine the second type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the similarity with the standard spike data set shape , the steps are as follows:

[0031] (1) Calculate the waveform feature shape of each channel data of the sample based on the minimum Euclidean distance Edist between each channel data and the standard spike data set, and further map the Euclidean distance data to the similarity space in the range of 0 to 1 through the Gaussian function. The calculation formula is as follows:

[0032]

[0033] (2) The data obtained after the mean normalization of all channel waveform weights is recorded as the waveform weight matrix W shape .

[0034] 2-4. Determine the optimal channel based on the amplitude weights and waveform weights of each channel of the multi-channel spike candidate sample. The optimal channel is defined as the channel with the largest product of the amplitude weight and the waveform weight within the multi-channel sample.

[0035] 2-5. Determine the third type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the distance from the optimal channel distance , the steps are as follows:

[0036] (1) Calculate the distance feature of each channel data of the sample. If the current channel and the optimal channel are in the same brain hemisphere, the source distance feature is defined as:

[0037] distance=0.75 Mdist

[0038] Otherwise defined as:

[0039] distance = 0.5 × 0.75 Mdist

[0040] Where Mdist is the Manhattan distance between the current channel and the optimal channel.

[0041] (2) The data matrix obtained by normalizing the distance weights of all channels is recorded as the distance weight matrix W distance .

[0042] 2-6. Based on the three types of weight matrices obtained in the previous steps and the multi-channel data multiChannelData, weighted synthesis of single-channel data samples singleChannelData is performed. The calculation formula is as follows:

[0043]

[0044] Where n is the number of channels contained in the multi-channel data.

[0045] The specific operations of step 3 are as follows:

[0046] 3-1. Extract two types of time series features from single-channel data, namely the nonlinear energy features calculated based on the smooth nonlinear energy operator and the morphological features obtained based on the mathematical morphology nonlinear filter.

[0047] 3-2. The EEG data, the extracted nonlinear energy, and the morphological features are sequentially merged into a 1×3-dimensional single-step feature vector according to their corresponding relationship on the time axis as the input of the pre-trained stacked Bi-LSTM neural network classifier network. The classifier finally outputs the classification result of the sample as spike or non-spike.

[0048] The beneficial effects of the present invention are as follows:

[0049] This spike detection method, based on intelligent multi-channel EEG screening and weighted sample generation, not only extracts effective information from each channel's data through a weighted generation algorithm for multi-channel data, improving spike recognition performance, but also provides a process for screening candidate multi-channel spike samples, setting a standard for establishing datasets for subsequent similar algorithms. Furthermore, further analysis of the weights used in weighted data generation can also detect the specific channel location where spikes originate. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic structural diagram of an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the "tit-for-tat" phenomenon of spike waves in multiple channels;

[0052] Figure 3 This is a spike wave recognition effect diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, the implementation steps of the spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation have been described in detail in the invention content, that is, the technical solution of the present invention mainly includes the following steps:

[0055] Step 1: Utilize multiple EEG waveform features to perform sample screening on each EEG channel data based on the threshold method, and obtain multi-channel spike wave candidate sample information by summarizing the screening results. Then, artifacts are removed based on the multi-channel characteristics of spike wave discharges, and finally multi-channel spike wave candidate samples are obtained through data segmentation.

[0056] The multi-channel spike wave candidate sample information includes the discharge time and channel distribution information of possible spike wave discharge events; the artifact removal is based on the "tit-for-tat" phenomenon unique to spike wave discharges under bipolar lead electroencephalogram to eliminate interference information in the multi-channel spike wave candidate sample information that does not meet the above phenomenon; the data segmentation refers to segmenting the input EEG signal based on the obtained multi-channel spike wave candidate sample information to obtain multi-channel spike wave candidate samples.

[0057] Step 2: Perform data analysis on the obtained multi-channel spike candidate samples, perform multiple weight calculations based on the optimal channel, and finally generate single-channel candidate sample data from the multi-channel spike candidate samples through a weighted algorithm.

[0058] The data analysis refers to determining the channel in each channel of the candidate sample that is most likely to have spike discharges based on amplitude and waveform similarity; the weights include three weights based on amplitude, waveform similarity, and distance from the optimal channel; the weighted algorithm refers to performing a weighted sum of the channel data of the multi-channel candidate samples based on the three weights obtained above, and finally obtaining the single-channel sample data.

[0059] Step 3: Use the single-channel spike classification algorithm to complete the classification detection of the generated data.

[0060] The single-channel spike classification algorithm refers to a single-channel detection algorithm based on time series features and LSTM network architecture. The detection process includes sample time series feature extraction, feature fusion and stacked Bi-LSTM neural network classification.

[0061] The specific implementation of step 1 is as follows:

[0062] 1-1. The original input multi-channel EEG data was filtered in the basic frequency domain using a 1-70 Hz bandpass filter and a 50 Hz notch filter. The processed data was used for subsequent data analysis.

[0063] 1-2. The discharge time of the initial spike candidate samples on each EEG channel was determined by minima. Five waveform features of each sample were further calculated, including amplitude, duration, left half-wave slope, right half-wave slope, and overall slope. Samples were then screened using a thresholding method. Only samples that passed the thresholding method met all threshold conditions, ultimately obtaining single-channel spike candidate samples suspected of having spike discharges on all channels. The threshold hyperparameters used in the thresholding method were determined by the upper and lower limits of the 99.9% confidence intervals of the five feature databases calculated using the Monte Carlo method. The dataset used for statistical analysis was a database of labeled spike samples.

[0064] 1-3. After obtaining all single-channel spike candidate sample information, the label information of all channels is summarized to obtain multi-channel candidate spike sample information, including candidate spike discharge time and channel distribution information.

[0065] The channel labeling information is summarized as follows: when the discharge time difference between candidate spikes in different channels is less than 0.02 seconds, based on the spike discharge generation and conduction model, it is determined that these discharge phenomena detected on the electrodes of different channels are caused by the same discharge source inside the brain. In this case, information is summarized, and the candidate spike discharge time in the obtained multi-channel spike candidate sample information is the mean of the discharge time of all single-channel samples. The channel distribution information records all channel numbers appearing in the sample group.

[0066] 1-4. When locating and analyzing waveforms with significant polarity changes using bipolar lead data, there are more obvious characteristics compared to data using reference lead data. Clinically, the characteristic phenomenon caused by spike discharges is called "tit-for-tat," which describes the possibility of spike discharges occurring at all electrodes between the first bipolar lead channel with a positive wave and the last bipolar lead channel with a negative wave. For example, Figure 2 Based on this phenomenon, artifact removal is performed on the multi-channel candidate spike wave sample information. The steps are as follows:

[0067] (1) Obtain bipolar lead EEG data related to the channel where candidate spike wave discharge phenomena are recorded.

[0068] (2) Traverse the bipolar lead data corresponding to each multi-channel spike wave candidate sample information in turn, and find the bipolar lead channel where the first positive peak and the last negative peak are located.

[0069] (3) According to the corresponding principle, the interference channels that do not meet the "tit-for-tat" condition in the multi-channel spike candidate sample information are eliminated.

[0070] 1-5. Based on the multi-channel spike candidate sample information after artifact removal, multi-channel spike candidate samples are generated through data segmentation. Each sample contains multi-channel data centered at the discharge time point and lasting 0.25 seconds.

[0071] The specific operations of step 2 are as follows:

[0072] 2-1. Determine the first-class weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the waveform amplitude amplitude , the steps are as follows:

[0073] (1) The amplitude feature quantity amplitude is recorded as the amplitude of the waveform at the center of each channel data.

[0074] (2) The amplitude weight matrix obtained by normalizing the amplitude feature values ​​of all channels is recorded as the weight matrix W amplitude .

[0075] 2-2. Calculate the standard spike data set, which is the cluster center data obtained by the K-means clustering algorithm based on the existing labeled spike sample database.

[0076] 2-3. Determine the second type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the similarity with the standard spike data set shape , the steps are as follows:

[0077] (1) Calculate the waveform feature shape of each channel data of the sample based on the minimum Euclidean distance Edist between each channel data and the standard spike data set, and further map the Euclidean distance data to the similarity space in the range of 0 to 1 through the Gaussian function. The calculation formula is as follows:

[0078]

[0079] (2) The data obtained after the mean normalization of all channel waveform weights is recorded as the waveform weight matrix W shape .

[0080] 2-4. Determine the optimal channel based on the amplitude weights and waveform weights of each channel of the multi-channel spike candidate sample. The optimal channel is defined as the channel with the largest product of the amplitude weight and the waveform weight within the multi-channel sample.

[0081] 2-5. Determine the third type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the distance from the optimal channel distance , the steps are as follows:

[0082] (1) Calculate the distance feature of each channel data of the sample. If the current channel and the optimal channel are in the same brain hemisphere, the source distance feature is defined as:

[0083] distance=0.75 Mdist

[0084] Otherwise defined as:

[0085] distance = 0.5 × 0.75 Mdist

[0086] Where Mdist is the Manhattan distance between the current channel and the optimal channel.

[0087] (2) The data matrix obtained by normalizing the distance weights of all channels is recorded as the distance weight matrix W distance .

[0088] 2-6. Based on the three types of weight matrices obtained in the previous steps and the multi-channel data multiChannelData, weighted synthesis of single-channel data samples singleChannelData is performed. The calculation formula is as follows:

[0089]

[0090] Where n is the number of channels contained in the multi-channel data.

[0091] The specific operations of step 3 are as follows:

[0092] 3-1. Extract two types of time series features from single-channel data, namely the nonlinear energy features calculated based on the smooth nonlinear energy operator and the morphological features obtained based on the mathematical morphology nonlinear filter.

[0093] 3-2. The EEG data, the extracted nonlinear energy, and the morphological features are sequentially merged into a 1×3-dimensional single-step feature vector according to their corresponding relationship on the time axis as the input of the pre-trained stacked Bi-LSTM neural network classifier network. The classifier finally outputs the classification result of the sample as spike or non-spike.

[0094] Figure 3 This is a diagram showing the effect of spike wave recognition on multi-channel measured data according to an embodiment of the present invention, wherein the top circle marks the position of the identified and located spike wave discharge.

[0095] In order to achieve better detection results for spike discharges, the following will introduce the selection and design of parameters in practical applications, which can serve as a reference for other applications of this invention:

[0096] In steps 1-3, 0.25 seconds is selected as the duration of a frame because the duration of spike discharge is 0.02-0.07s and the duration of sharp wave discharge is 0.07-0.2s. Although the definitions of spikes and sharp waves are different in clinical practice, in the field of automatic identification of spikes and sharp waves, the two are collectively referred to as epileptic transients or spikes. The spikes mentioned in this patent refer to the collection of spikes and sharp waves in the medical sense.

[0097] In steps 2-3, the hyperparameters μ and σ of the Gaussian function are set to 0 and 7 respectively. After actual measurement, a more reasonable weight based on the Euclidean distance can be calculated.

[0098] The spike wave detection method proposed in the present invention, based on multi-channel EEG intelligent screening and weighted sample generation, obtains multi-channel sample data through screening in the input EEG, and then converts the multi-channel sample data into single-channel sample data through a weighted data generation algorithm. Finally, a high-performance single-channel spike wave detection algorithm is used to identify the sample as a spike wave or non-spike wave to complete the detection. Subsequently, the spike wave discharge channel location and related statistics can be further provided to effectively assist doctors in diagnosis and treatment. Due to the high complexity and susceptibility of EEG signals to interference, and the presence of normal physiological electrical signals with high waveform feature similarity but non-spike signals, traditional single-channel detection algorithms have poor anti-interference capabilities. Many interferences in EEG signals are misjudged as spike waves. Multi-channel detection algorithms are highly complex and time-consuming. This algorithm comprehensively considers the advantages and disadvantages of single-channel and multi-channel recognition algorithms, performs sample screening on multi-channel data, and synthesizes multi-channel data into single-channel data through an intelligent weighted data generation algorithm, significantly improving the algorithm's recognition ability and anti-interference ability on the basis of lower computational complexity. In summary, the algorithm proposed in the present invention for data screening and intelligent weighting of multi-channel data to generate single-channel data can effectively extract information from multi-channel data and generate meaningful single-channel data samples. In combination with the existing high-performance single-channel spike wave detection algorithm, it can achieve a high-performance spike wave detection function.

Claims

1. A spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation, characterized in that: The steps include: Step 1: Filter samples of each EEG channel data based on the threshold method using multiple EEG waveform features, and obtain multi-channel spike candidate sample information by summarizing the screening results. Then, artifacts are removed based on the multi-channel characteristics of spike discharges, and finally multi-channel spike candidate samples are obtained through data segmentation; The multi-channel spike wave candidate sample information includes the discharge time and channel distribution information of possible spike wave discharge events; the artifact removal is based on the "tit-for-tat" phenomenon unique to spike wave discharges under bipolar lead electroencephalogram to eliminate interference information in the multi-channel spike wave candidate sample information that does not meet the above phenomenon; the data segmentation refers to segmenting the input EEG signal based on the obtained multi-channel spike wave candidate sample information to obtain multi-channel spike wave candidate samples; Step 2: Analyze the data of the obtained multi-channel spike candidate samples, perform multiple weight calculations based on the optimal channel, and finally generate single-channel candidate sample data from the multi-channel spike candidate samples through a weighted algorithm; The data analysis refers to determining the channel in each candidate sample that is most likely to have spike discharges based on amplitude and waveform similarity; the weights include three weights based on amplitude, waveform similarity, and distance from the optimal channel; the weighting algorithm refers to performing a weighted sum of the channel data of the multi-channel candidate samples based on the three weights obtained above, ultimately obtaining single-channel sample data; The specific operations of step 2 are as follows: 2-1. Determine the first-class weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the waveform amplitude amplitude , the steps are as follows: (1) The amplitude characteristic quantity amplitude is recorded as the amplitude of the waveform at the center of each channel data; (2) The amplitude weight matrix obtained by normalizing the amplitude feature values ​​of all channels is recorded as the weight matrix W amplitude ; 2-2. Calculating a standard spike data set, which is the data of each cluster center obtained by the K-means clustering algorithm based on an existing labeled spike sample database; 2-3. Determine the second type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the similarity with the standard spike data set shape , the steps are as follows: (1) Calculate the waveform feature shape of each channel data of the sample based on the minimum Euclidean distance Edist between each channel data and the standard spike data set, and further map the Euclidean distance data to the similarity space in the range of 0 to 1 through the Gaussian function. The calculation formula is as follows: (2) The data obtained after the mean normalization of all channel waveform weights is recorded as the waveform weight matrix W shape ; 2-4. Determine the optimal channel based on the amplitude weights and waveform weights of each channel in the multi-channel spike candidate sample. The optimal channel is defined as the channel with the largest product of the amplitude weight and the waveform weight within the multi-channel sample. 2-5. Determine the third type of weight matrix for each channel of the multi-channel spike candidate sample, that is, the weight matrix W based on the distance from the optimal channel distance , the steps are as follows: (1) Calculate the distance feature of each channel data of the sample. If the current channel and the optimal channel are in the same brain hemisphere, the source distance feature is defined as: distance=0.75 Mdist Otherwise defined as: distance=0.5×0.75 Mdist Where Mdist is the Manhattan distance between the current channel and the optimal channel; (2) The data matrix obtained by normalizing the distance weights of all channels is recorded as the distance weight matrix W distance ; 2-6. Based on the three types of weight matrices obtained in the previous steps and the multi-channel data multiChannelData, weighted synthesis of single-channel data samples singleChannelData is performed. The calculation formula is as follows: Where n is the number of channels contained in the multi-channel data; Step 3: Use a single-channel spike classification algorithm to complete the classification detection of the generated data; The single-channel spike classification algorithm refers to a single-channel detection algorithm based on time series features and LSTM network architecture. The detection process includes sample time series feature extraction, feature fusion and stacked Bi-LSTM neural network classification.

2. The spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation according to claim 1, characterized in that: The specific implementation of step 1 is as follows: 1-1. Filter each channel of the original input multi-channel EEG data through a 1-70 Hz bandpass filter and a 50 Hz notch filter in the basic frequency domain, and the processed data is used for subsequent data analysis; 1-2. The discharge time of the initial spike candidate samples on each EEG channel was determined by the minimum point. Five waveform features of each sample were further calculated, including amplitude, duration, left half-wave slope, right half-wave slope, and overall slope. Samples were screened for each feature using a threshold method. The samples that passed the screening were required to meet all threshold conditions. Ultimately, single-channel spike candidate samples suspected of having spike discharges on all channels were obtained. The threshold hyperparameters used in the threshold method were determined by the upper and lower limits of the 99.9% confidence intervals of the five feature databases calculated using the Monte Carlo method. The dataset used for statistical analysis was from a database of labeled spike samples. 1-3. After obtaining all single-channel spike candidate sample information, the label information of all channels is summarized to obtain multi-channel candidate spike sample information, including candidate spike discharge time and channel distribution information; The channel labeling information is summarized as follows: when the discharge time difference between candidate spikes in different channels is less than 0.02 seconds, based on the spike discharge generation and conduction model, it is determined that these discharge phenomena detected on the electrodes of different channels are caused by the same discharge source in the brain. In this case, the information is summarized, and the candidate spike discharge time in the resulting multi-channel spike candidate sample information is the mean of the discharge time of all single-channel samples. The channel distribution information records all channel numbers appearing in the sample group; 1-4. When locating and analyzing waveforms with significant polarity changes using bipolar EEG data, there are more pronounced features compared to data using reference leads. Clinically, this characteristic phenomenon caused by spike discharges is referred to as "tit-for-tat," describing the possibility of spike discharges occurring at all electrodes between the first bipolar lead channel with a positive wave and the last bipolar lead channel with a negative wave. Based on this phenomenon, artifact removal is performed on multi-channel candidate spike sample information using the following steps: (1) Obtain bipolar lead EEG data related to the channel where candidate spike wave discharge phenomena are recorded; (2) traverse the bipolar lead data corresponding to each multi-channel spike wave candidate sample information in turn, and find the bipolar lead channel where the first positive peak and the last negative peak are located; (3) According to the corresponding principle, the interference channels that do not meet the "tit-for-tat" condition are eliminated from the multi-channel spike candidate sample information; 1-5. Based on the multi-channel spike candidate sample information after artifact removal, multi-channel spike candidate samples are generated through data segmentation. Each sample contains multi-channel data centered at the discharge time point and lasting 0.25 seconds.

3. The spike wave detection method based on multi-channel EEG intelligent screening and weighted sample generation according to claim 2, characterized in that: The specific operations of step 3 are as follows: 3-1. Extract two types of time series features from single-channel data: nonlinear energy features calculated based on a smooth nonlinear energy operator and morphological features obtained based on a mathematical morphology nonlinear filter; 3-2. The EEG data, the extracted nonlinear energy, and the morphological features are sequentially merged into a 1×3-dimensional single-step feature vector according to their corresponding relationship on the time axis as the input of the pre-trained stacked Bi-LSTM neural network classifier network. The classifier finally outputs the classification result of the sample as spike or non-spike.

Images