Electroencephalogram signal data classification processing method and system based on multi-dimensional feature representation

Abstract

This invention provides a method and system for classifying and processing electroencephalogram (EEG) signal data based on multi-dimensional feature representation. It relates to the fields of EEG signal processing, feature representation, pattern recognition, and machine learning. The method includes: S1, acquiring multi-channel EEG signal data and performing preprocessing; S2, extracting multi-dimensional EEG signal features to characterize EEG signal classification; S3, using the Crowned Porcupine optimization algorithm to optimize the machine learning model and output EEG signal classification evaluation indicators; S4, determining dual thresholds and completing three-level classification of EEG signals. This invention can simultaneously integrate EEG signal features to characterize EEG signal classification anomalies, improving evaluation sensitivity and stability; constructing continuous EEG signal classification evaluation indicators to perform three-level signal hierarchical recognition and complete EEG signal classification judgment; and using the Crowned Porcupine optimization algorithm to optimize the machine learning model to improve the accuracy, robustness, and generalization performance of the three-class machine learning model.

Description

Technical Field

[0001] This invention relates to the fields of electroencephalogram (EEG) signal processing, feature representation, pattern recognition, and machine learning, specifically to a method and system for classifying and processing EEG signal data based on multi-dimensional feature representation. Background Technology

[0002] Electroencephalogram (EEG) signals are bioelectrical signals continuously acquired over time through multiple acquisition channels. They are characterized by high temporal resolution, numerous acquisition channels, complex frequency components, and significant non-stationarity. During EEG signal classification, different categories of EEG samples typically exhibit differences in frequency domain energy distribution, nonlinear complexity, microstate dynamics, and inter-channel functional connectivity topology. Existing EEG signal classification methods often focus on single-band power, a limited number of channel features, or single network indicators, making it difficult to simultaneously characterize changes in EEG signals at the spectral scale, nonlinear dynamic scale, sub-second microstate scale, and multi-channel network scale. Furthermore, differences in the dimensions, stability, and redundancy of different feature types can lead to feature redundancy, model parameter dependence on empirical settings, and insufficient cross-sample batch generalization performance if these features are directly concatenated into the classification model. Therefore, it is necessary to propose a classification and processing method and system for EEG signal data based on multi-dimensional feature representation. This method involves standardizing preprocessing of multi-channel EEG signals, extracting spectral features, nonlinear complexity features, microstate features, and functional connectivity network features, and combining these with a feature-selected and optimized machine learning model to achieve stable classification of EEG signal sample categories. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention aims to provide a method and system for classifying and processing EEG signal data based on multi-dimensional feature representation. This method simultaneously integrates spectral features, nonlinear complexity features, microstate features, and EEG signal classification network features, characterizing EEG signal classification anomalies at multiple scales and improving evaluation sensitivity and stability. During the machine learning modeling stage, the CPO (Corrugated Porcine Optimization) algorithm is introduced to globally optimize key hyperparameters of the model. Compared to traditional empirical parameter tuning or local search methods, this method more effectively obtains model parameter combinations suitable for the current EEG signal feature distribution, improving the accuracy, robustness, and generalization performance of the machine learning three-class classification model. Furthermore, a continuous EEG signal classification evaluation index is constructed, and two thresholds are set for three-level signal hierarchical recognition. Therefore, in addition to completing the classification judgment, it can also reflect the continuous evolution trend of EEG signals.

[0004] Specifically, on the one hand, the present invention provides a method for classifying and processing electroencephalogram (EEG) signal data based on multi-dimensional feature representation, which includes the following steps: S1: Preprocess the multi-channel EEG signal data to obtain preprocessed EEG signal data. Classify EEG signal features based on pre-labeled EEG signal feature patterns; S2: Extract multidimensional EEG signal features that can characterize the classification of EEG signals; set the power spectral density as... Determine the frequency band of EEG signal data Internal bandwidth power To obtain the relative power The spectral features of the EEG signal were extracted; the intrinsic mode components were subjected to Hilbert transform to obtain the marginal spectrum. Construct intrinsic mode component weights The weighted feature marginal spectrum is obtained. Determine the Hilbert yellow state entropy Hilbert's Yellow Response Entropy The nonlinear complexity features of EEG signal data are extracted; based on global field power and topological stability, microstate peak evaluation parameters are constructed. ; Filter the set of candidate microstate peak points Micro-state features of EEG signal data are extracted; the preprocessed EEG signals are divided into time windows, and the stability coefficient of each time window is determined. The functional connection strength is then corrected to obtain the functional connection strength. The functional connectivity matrices corrected for each frequency band are fused to obtain the final EEG signal network matrix. Extract features from the EEG signal classification network; S3: Based on the multidimensional EEG signal features extracted in step S2, construct EEG signal feature vectors, perform standardization and screening, use the Crowned Porcupine optimization algorithm to optimize the machine learning model, output three classes of posterior probabilities, and construct a continuous EEG signal classification evaluation index. ; S4: Determine the first threshold based on the continuous evaluation index of the sample. Second threshold Based on the classification and evaluation indicators of continuous EEG signals determined in step S3 With the first threshold Second threshold The system performs three-level classification of EEG signals, controls associated medical devices at different levels, and stores the classified EEG signals in a database for display on the associated medical device display terminal.

[0005] Preferably, step S2 specifically includes: S21: Based on the power spectral density is Determine the bandwidth and power of EEG signals in different frequency bands. To obtain the relative power ; S22: Perform Hilbert transform on the intrinsic mode components; construct component weights. The weighted feature marginal spectrum is obtained. We obtain Hilbert's yellow state entropy. Hilbert's Yellow Response Entropy ; S23: Construct microstate peak evaluation parameters based on global field power and topological stability. ; Filter the set of candidate microstate peak points Extract average duration, average occurrence rate, coverage, and average global field power; S24: Divide the preprocessed EEG signal into time windows and determine the stability coefficient of the time windows. The functional connection strength is then corrected to obtain the functional connection strength. By fusing the functional connectivity matrix, the final EEG signal network matrix is ​​obtained. .

[0006] Preferably, in step S22, the intrinsic mode component weights are constructed. The weighted feature marginal spectrum is obtained. Specifically: ; ; in, For the first The weighted feature marginal spectrum of each channel; For the first The first channel The weights of each intrinsic mode component; For the first The first channel Each intrinsic mode component at frequency Marginal spectrum at the location; The frequency of the brainwave signal; This represents the total number of intrinsic modal components. Number the frequency bands for EEG signal data; For the first The first channel The energy percentage of each intrinsic mode component; For the first The first channel Modal stability coefficients of each intrinsic modal component; For the first The first channel Modal stability coefficients of each intrinsic modal component; Index for summing intrinsic modal components; For the first The energy of each intrinsic mode component; This is the parameter for correcting the denominator.

[0007] Preferably, in step S23, microstate peak evaluation parameters are constructed based on global field power and topological stability. ; Filter the set of candidate microstate peak points Specifically: ; ; in, For evaluating the peak value of the microstate; This is the topological stability adjustment coefficient; For the first Topological stability at each sampling point; This is the set of candidate microstate peak points; For microstate peak evaluation parameters The Quantile threshold; Quantile index for peak evaluation parameters of microstate; ; This refers to the time sampling point number in the characteristic time spectrum.

[0008] Preferably, the EEG signal network matrix in step S24 The method for obtaining it is as follows: ; ; ; in, This is the stability coefficient for the time window; For the first The first channel and the first The first channel in the The average connection strength across multiple time windows in a frequency band; For the first The first channel and the first The first channel in the Standard deviation of connection strength across multiple time windows in a frequency band; For the first The first channel and the first The first channel in the Stability correction function connection strength on each frequency band; For the first The first channel and the first The strength of EEG signal connectivity between channels; For the first Weights of each frequency band; Total number of frequency bands; This is a frequency band index.

[0009] Preferably, the classification of EEG signal features in step S1 is specifically as follows: first EEG signal feature category C0, second EEG signal feature category C1 and third EEG signal feature category C2, corresponding to the feature control group HC, the first feature classification group noMHE and the second feature classification group MHE identified by the machine learning three-classification model.

[0010] Preferably, step S3 specifically includes: S31: Concatenate the features extracted in step S2 to form a multidimensional EEG signal feature vector. Standardization processing is performed to form a multidimensional EEG signal feature vector. ; S32: Train the machine learning model optimized using the Crowned Porcupine optimization algorithm; input the subset of core features of the EEG signal obtained in step S31 into the machine learning three-class classification model to construct the machine learning three-class classification model in the parameter space. Finding the optimal set of parameters This ensures that the machine learning three-class classification model achieves optimal validation performance. S33: Apply the optimized machine learning model from step S32 to EEG signal classification, and output the posterior probability of the feature control group. Posterior probability of the first feature classification group Posterior probability of the second feature classification group .

[0011] Preferably, the performance evaluation function of the machine learning three-class classification model in step S32 is: , ; ; in, This is a performance evaluation function for a machine learning three-class classification model. The optimal set of parameters; The parameters that allow the objective function to reach its maximum value; In the model with parameters Real-time recognition performance; The classification accuracy is given by the model parameter θ. To identify performance weighting coefficients; This is the weighting coefficient for classification accuracy.

[0012] Preferably, the three-level classification of EEG signals in step S4 specifically involves: When continuous EEG signal classification assessment index When the risk is low, the first EEG signal feature category C0 in step S1 is identified, the category label is feature control group HC, and a standard working mode instruction is sent to the associated medical device. When the classification assessment index of continuous EEG signals meets When the risk is determined to be medium, corresponding to the second EEG signal feature category C1 in step S1, the category label is the first feature classification group noMHE, and an enhanced acquisition mode command is sent to the associated medical device. When continuous EEG signal classification assessment index If the risk is high, it corresponds to the third EEG signal feature category C2 in step S1, with the category label being the second feature classification group MHE, and a deep analysis mode command is sent to the associated medical device.

[0013] On the other hand, the present invention provides an EEG signal data classification and processing system based on a multi-dimensional feature representation method, which includes: an EEG signal acquisition and processing module, a multi-dimensional feature extraction module, a feature fusion and screening module, a model optimization and training module, a continuous index construction module, a signal hierarchical output module, and a data management module. The EEG signal acquisition and processing module is used to acquire multi-channel EEG signal data and perform quality control and preprocessing on the raw EEG signals. Its function is to improve the quality of EEG signals and reduce the impact of noise and artifacts on multi-dimensional feature extraction. The multi-dimensional feature extraction module is used to extract spectral features, nonlinear complexity features, microstate features, and EEG signal network features from the preprocessed EEG signals. Its role is to characterize the frequency domain changes, nonlinear complexity changes, sub-second topological dynamic changes, and brain region connectivity changes of EEG signals from multiple dimensions. The feature fusion and filtering module is used to concatenate, standardize and filter various features output by the multi-dimensional feature extraction module. Its function is to reduce feature redundancy and improve model training efficiency and generalization performance. The model optimization and training module is used to build an EEG signal classification model and to globally optimize the model hyperparameters through optimization algorithms. Its purpose is to reduce the dependence on manual parameter tuning and improve the model's classification performance, robustness and cross-sample generalization performance. The continuous index construction module is used to construct continuous EEG signal classification evaluation indexes based on the three types of posterior probabilities output by the optimized machine learning model. Its function is to convert discrete classification probabilities into continuous numerical indexes, thereby improving the interpretability of EEG signal classification results. The signal stratification output module is used to output signal stratification results based on continuous EEG signal classification evaluation indicators and dual thresholds. Its function is to convert the model calculation results into easily understandable engineering output results. The data management module is used to store and manage EEG signal data, feature data, model parameters and output results. Its function is to enable the system's calculation process to be traced and verified, and to facilitate subsequent model updates and engineering deployment.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on multi-channel EEG signal data, this invention integrates spectral features, nonlinear complexity features, microstate features and EEG signal classification network features, which can characterize EEG signal classification abnormalities from multiple scales and improve assessment sensitivity and stability.

[0015] (2) In the machine learning modeling stage, the present invention introduces the Crowned Porcupine Optimization Algorithm (CPO) to perform global optimization of key hyperparameters of the model. Compared with traditional empirical parameter tuning or local search methods, it can more effectively obtain the model parameter combination suitable for the current EEG signal feature distribution, thereby improving the accuracy, robustness and generalization performance of the machine learning three-class classification model.

[0016] (3) This invention does not only output a single discrete classification label, but constructs a continuous EEG signal classification evaluation index and further sets two thresholds to identify the three-level signals of the feature control group HC, the first feature classification group noMHE and the second feature classification group MHE. Therefore, while completing the classification judgment, it can also reflect the continuous evolution trend of EEG signals. Attached Figure Description

[0017] Figure 1 This is a control block diagram of the EEG signal data classification and processing method based on multi-dimensional feature representation of the present invention; Figure 2 This is a flowchart of the machine learning EEG signal classification and evaluation method of the present invention; Figure 3 This is a flowchart of the multi-channel EEG signal data acquisition and preprocessing process in this invention; Figure 4 This is a flowchart of the multidimensional EEG signal feature extraction process in this invention; Figure 5 This is a flowchart of the machine learning model training process based on the CPO optimization algorithm of the hog cactus in this invention. Figure 6 This is a schematic diagram of the EEG signal classification and evaluation indicators and the dual-threshold three-level signal layering in this invention; Figure 7 This is a schematic diagram of the single-sample processing and output effect of the EEG signal data classification and processing system of the present invention; Figure 8 This is a diagram illustrating the usage effect of the electroencephalogram (EEG) signal data classification and processing system of the present invention. Detailed Implementation

[0018] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0019] This invention proposes a method for classifying and processing electroencephalogram (EEG) signal data based on multi-dimensional feature representation, such as... Figure 1As shown, multi-channel EEG signal data were collected and preprocessed; multi-dimensional EEG signal features characterizing EEG signal classification were extracted; the Crowned Porcupine optimization algorithm was used to optimize the machine learning model to output EEG signal classification evaluation indicators; dual thresholds were determined and three-level classification of EEG signals was completed. Figure 2 The diagram illustrates the EEG signal classification and processing flow. First, multi-channel resting-state EEG signals are acquired and preprocessed, including filtering, downsampling, bad lead processing, artifact removal, rereference, and effective data segmentation. Second, spectral features, nonlinear complexity features, microstate features, and functional connectivity network features are extracted from the preprocessed EEG signals to construct a multi-dimensional EEG signal feature vector. Third, the multi-dimensional EEG signal feature vector is standardized and feature-filtered to obtain a core feature subset. This core feature subset is then input into a machine learning three-class classification model optimized by the CPO algorithm, which outputs the posterior probabilities that the EEG signal to be processed belongs to the preset first EEG signal feature category C0 (control group HC), the second EEG signal feature category C1 (noMHE group), and the third EEG signal feature category C2 (MHE group). Finally, a continuous EEG signal classification index is constructed based on the three posterior probabilities, and this index is divided into three category intervals using a dual threshold to obtain the EEG signal classification results. The specific steps include: S1: Acquire multi-channel EEG signal data and perform preprocessing. Acquire multi-channel EEG signal data from the EEG acquisition device and classify the samples into three categories—C0, C1, and C2—based on pre-labeled EEG signal feature patterns. These category labels are used only to represent the category attributes of the EEG signal samples in the feature space and serve as supervision labels for training and validating machine learning models.

[0020] In this embodiment of the invention, the number of EEG acquisition channels is 32; the preferred sampling frequency is 250Hz to 1000Hz. The acquisition of multi-channel EEG signal data is as follows: ; in, Multichannel EEG signal data; For the first Each channel at time EEG signal data; Number the EEG acquisition channels; This refers to the number of EEG acquisition channels; This is a time parameter.

[0021] Multi-channel EEG signal data is preprocessed to obtain preprocessed EEG signal data. The preprocessing in this embodiment includes power frequency notch filtering, bandpass filtering, downsampling, bad lead identification and interpolation, independent component analysis for artifact removal, EEG and EMG artifact removal, rereference processing, and effective data segmentation. The preprocessed EEG signal data is as follows: ; in, This refers to preprocessed EEG signal data; For the first Each channel at time Preprocessed EEG signal data.

[0022] like Figure 3 As shown, in the multi-channel EEG signal data acquisition and preprocessing stage, resting-state EEG signal data is first acquired to obtain multi-channel EEG signal data. Then, power frequency notch filtering, bandpass filtering, downsampling, bad lead identification and interpolation, independent component analysis for artifact removal, EEG and EMG artifact removal, rereference processing, and effective data segmentation are performed sequentially to obtain high-quality preprocessed EEG signal data. Through the above preprocessing steps, the impact of power frequency interference, artifacts, and bad leads on subsequent analysis results is effectively reduced, improving the quality of EEG signal data and the stability of feature extraction, laying the foundation for subsequent multi-dimensional EEG signal feature analysis.

[0023] S2: Based on preprocessed EEG signal data Extracting multidimensional EEG signal features that can characterize the classification of EEG signals; specifically including the following steps: S21: Extract the spectral characteristics of the EEG signal. Perform power spectrum estimation on the EEG signals of each channel, and extract the absolute and relative power of the Delta, Theta, Alpha, Beta, and Gamma bands. Specifically, the Delta band is the low-frequency band of the delta rhythm, preferably 1Hz–4Hz; the Theta band is the theta rhythm band, preferably 4Hz–8Hz; the Alpha band is the alpha rhythm band, preferably 8Hz–13Hz; the Beta band is the beta rhythm band, preferably 13Hz–30Hz; and the Gamma band is the high-frequency band of the gamma rhythm, preferably 30Hz–45Hz, with the specific upper limit determined based on the sampling frequency and filtering settings.

[0024] Let the first Each channel at frequency The power spectral density at is Then the frequency band of EEG signal data The bandwidth power within is: ; in, For the first Each channel is in the EEG signal data frequency band Within the bandwidth power; For the first Each channel at frequency Power spectral density at; The frequency of the brainwave signal; This refers to the frequency band for EEG signal data. This represents the lower limit of the frequency band for EEG signal data. This represents the upper limit of the frequency band for EEG signal data.

[0025] No. Each channel is in the EEG signal data frequency band The relative power within is set as follows: ; in, For the first Each channel is in the EEG signal data frequency band Relative power within; For the first One frequency band of EEG signal data; This represents the total number of frequency bands for EEG signal data. The frequency band number is assigned to the EEG signal data.

[0026] S22: Extracting nonlinear complexity features from EEG signal data. This invention constructs Hilbert-Huang state entropy (HHSE) and Hilbert-Huang response entropy (HHRE) based on preprocessed EEG signals to characterize the complexity changes of EEG signals under non-stationary conditions. For the first... Preprocessed EEG signals from each channel Perform empirical mode decomposition to obtain One intrinsic mode IMF component and one residual term, specifically: ; in, For the first Preprocessed EEG signals from each channel; For the first The first channel Each intrinsic mode IMF component; The total number of intrinsic mode IMF components; This is a residual term.

[0027] Perform a Hilbert transform on each intrinsic mode (IMF) component to obtain the corresponding characteristic time spectrum. And further obtained the first The marginal spectrum of each intrinsic mode IMF component is as follows: ; in, For the first The first channel Each intrinsic mode IMF component at frequency Marginal spectrum at the location; For the first Characteristic temporal spectrum of each intrinsic mode IMF component; The time sampling point number in the characteristic time spectrum; The total number of sampling points in the time dimension of the characteristic time spectrum.

[0028] The characteristic time spectrum in the above formula is specifically the Hilbert-Huang time spectrum, which is an adaptive time-frequency representation method for nonlinear and non-stationary signals constructed based on Empirical Mode Decomposition (EMD) and Hilbert Transform (Hilbert Transform). Its core lies in the joint characterization of instantaneous frequency and instantaneous amplitude, breaking through the dependence of traditional Fourier analysis on stationarity.

[0029] No. Energy of each intrinsic mode IMF component for: ; in, For the first The energy of each intrinsic mode IMF component.

[0030] Get the first Energy percentage of each intrinsic mode IMF component for: ; in, For the first The first channel Energy percentage of each intrinsic mode IMF component; Index for summation of intrinsic mode IMF components; For the first Energy of each intrinsic mode IMF component; This is the parameter for correcting the denominator.

[0031] Furthermore, the preprocessed EEG signals were divided into several time periods, and the first time period was calculated for each time period. The energy of each intrinsic mode IMF component in each time period is used to obtain the average energy of that intrinsic mode IMF component. and energy standard deviation . No. Modal stability coefficients of individual intrinsic modal IMF components for: ; in, For the first The first channel Modal stability coefficients of each intrinsic modal IMF component; To obtain the energy mean of the intrinsic mode IMF components, data from multiple time periods needs to be collected. To determine the energy standard deviation of the intrinsic mode IMF components, data from multiple time periods needs to be collected.

[0032] When the energy fluctuation of a certain intrinsic mode IMF component is small over multiple time periods, then Smaller A larger value indicates higher stability of the intrinsic mode IMF component; when the energy fluctuation of a certain intrinsic mode IMF component is large, then... A smaller value indicates lower stability of the intrinsic mode IMF component. The energy proportion and modal stability coefficient of the intrinsic mode IMF component are used to construct the weights of the intrinsic mode IMF component. for: ; in, For the first The first channel The weights of each intrinsic mode IMF component.

[0033] This weight is used to enhance the intrinsic mode factor (IMF) components with higher energy proportions and better stability when constructing the characteristic marginal spectrum, while reducing the influence of IMF components with lower energy or poorer stability. The characteristic marginal spectrum used in this embodiment is the Hilbert-Huang marginal spectrum, which is a global statistical representation of the signal's frequency domain energy distribution in the Hilbert-Huang transform (HHT). It is the product of integrating the Hilbert spectrum along the time axis, directly reflecting the probability and cumulative energy intensity of each frequency component throughout the signal's duration. It is a core tool for identifying the dominant frequencies of non-stationary signals.

[0034] Based on the weights of the intrinsic mode IMF components, a weighted feature marginal spectrum is constructed: ; in, For the first The weighted feature marginal spectrum of each channel.

[0035] Within the target frequency range of state entropy Within this, the weighted feature marginal spectrum is normalized to obtain the state entropy frequency probability. Specifically: ; in, For the state entropy target frequency range, the first The normalized probability of each frequency point; The number of frequency points within the target frequency range of the state entropy; For the first One frequency point; For the first One frequency point; This is the index of frequency points within the target frequency range of the state entropy.

[0036] Determine the Hilbert yellow state entropy for: ; in, For the first Hilbert yellow state entropy for each channel.

[0037] Within the target frequency range of response entropy Within this, the weighted feature marginal spectrum is normalized to obtain the response entropy frequency probability. for: ; in, For the response entropy target frequency range, the first The normalized probability of each frequency point; The number of frequency points within the target frequency range for response entropy.

[0038] Hilbert Yellow Response Entropy for: ; in, For the first Hilbert yellow response entropy of each channel.

[0039] Through the above processing, Hilbert's Huang state entropy Hilbert's Yellow Response Entropy It can reflect the energy contribution and stability changes of the intrinsic mode (IMF) components of each EEG signal, and serve as input for the nonlinear complexity feature of the EEG signal into the subsequent classification model.

[0040] like Figure 4As shown, in the multidimensional EEG signal feature extraction stage, this invention characterizes EEG signal data from multiple levels. First, it extracts absolute and relative power features in frequency bands such as Delta, Theta, Alpha, Beta, and Gamma to reflect the frequency domain energy distribution changes of EEG signal data. Second, it extracts nonlinear complexity features such as state entropy, response entropy, Hilbert-Huang state entropy, and Hilbert-Huang response entropy to characterize the complexity and non-stationary dynamic changes of EEG signal data. Third, it extracts features such as the average duration, average occurrence rate, coverage, and average global field power of microstates to characterize the dynamic organization characteristics of EEG signal data on a sub-second timescale. Fourth, it constructs a functional connectivity matrix by removing biased weighted phase lag parameters and further extracts EEG signal classification network features such as node clustering parameters, node efficiency, global efficiency, modularity, small-world property, and club parameters to reflect the local aggregation performance, information transmission efficiency, and overall topological organization pattern of the EEG signal classification network. Through the above multidimensional feature joint characterization, EEG signal classification anomalies can be identified more comprehensively.

[0041] After feature extraction, spectral features, nonlinear complexity features, microstate features, and EEG signal classification network features are concatenated to form a multidimensional EEG signal feature vector corresponding to the EEG signal. Subsequently, the feature vector is standardized to eliminate the dimensional differences between features. Recursive feature elimination and other methods are used to filter features to obtain a core feature subset for model training, reduce redundant feature interference, and improve model training efficiency and generalization performance.

[0042] S23: Extracting microstate features from EEG signal data. The global field power of the preprocessed EEG signal data from S1 is calculated, and the scalp potential topology map corresponding to the peak of the global field power is extracted. Cluster analysis is performed based on the potential topology map of the EEG signal data to obtain microstate templates for several classes of EEG signal data. These microstate templates are then backfitted to continuous EEG signal data to obtain microstate label sequences for each time step. Based on the microstate label sequences of the EEG signal data, the average duration, average occurrence rate, coverage, and average global field power features of each type of microstate are extracted to characterize the dynamic organization pattern of the EEG signal data on a sub-second timescale.

[0043] This invention determines candidate peak points of microstates based on global field power and topological stability, and performs microstate template clustering and microstate statistical feature calculation based on the candidate peak points.

[0044] No. Average amplitude of the channel at each sampling point for: ; in, For the first The average amplitude of the channel at each sampling point.

[0045] No. Global field power at each sampling point for: ; in, For the first Global field power at each sampling point.

[0046] The first The topological vector of scalp potential at each sampling point is denoted as: ; in, For the first Topological vector of scalp potential at each sampling point; This is the transpose symbol.

[0047] Get the first Topological stability at each sampling point for: ; in, For the first Topological stability at each sampling point; This is the function for calculating the correlation coefficient. To take the absolute value.

[0048] Topological stability is used to represent the degree of consistency between the spatial distribution of scalp potentials at the current sampling point and the spatial distribution at adjacent time points. The higher the topological stability, the more stable the spatial topology corresponding to that sampling point.

[0049] Based on global field power and topological stability, construct microstate peak evaluation parameters. for: ; in, For evaluating the peak value of the microstate; This is the topological stability adjustment coefficient.

[0050] When two sampling points have similar global field powers, the sampling point with higher topological stability has higher [potential]. These are more suitable as samples for microstate clustering. Based on the microstate peak evaluation parameters... Screening candidate microstate peak points from local peak points of fluorescently labeled GFP. for: ; in, This is the set of candidate microstate peak points; For microstate peak evaluation parameters The Quantile threshold; The quantile index for the microstate peak evaluation parameter is 0.70 to 0.90 in this example.

[0051] For the set of candidate microstate peak points Clustering is performed on the topological vectors corresponding to each candidate peak point to obtain... Micro-state template. Set to 4, 5, or 6, with a value of 4. Clustering uses a spatial correlation clustering method. The microstate templates obtained from clustering are then backfitted to continuous EEG signals to obtain the microstate label sequence corresponding to each sampling point. For the first Microstate-like, extract average duration Average occurrence rate Coverage and average global field power . (To obtain the first) Average duration of microstates for: ; in, For the first The average duration of microstates; For the first The number of consecutive segments of microstate-like states; For the first Microstate The duration of a continuous segment.

[0052] No. Average occurrence rate of microstates for: ; in, The total effective EEG recording time.

[0053] No. Coverage of micro-states for: ; in, For the first Coverage of micro-states; To be assigned as the first The number of sampling points for micro-states.

[0054] No. Average global field power of microstate for: ; in, For the first The average global field power of the microstate.

[0055] Through the above processing, microstate features can characterize the topological organization of EEG signals on a short time scale and improve the representativeness and stability of microstate templates.

[0056] S24: Extract features from the EEG signal classification network. The functional connectivity strength between channels is calculated based on multi-channel EEG signal data, and a functional connectivity matrix is ​​constructed. The functional connectivity index is the debiased weighted phase lag parameter. The functional connectivity matrix is ​​thresholded to form the EEG signal classification network. Graph theory features such as network node clustering parameters, node efficiency, global efficiency, modularity, small-world property, and club parameters are extracted to characterize the local clustering performance, information transmission efficiency, and overall topological organization of the EEG signal classification network.

[0057] The preprocessed EEG signals are divided into multiple time windows. Let the first window be... Within the first time window, the first The first channel and the first The first channel in the The cross spectrum of each frequency band is The debiased weighted phase lag exponent is then expressed as: ; in, For the first The first channel and the first The first channel in the Debiased weighted phase lag index in each frequency band ; For mutual spectrum; For operations involving the imaginary part; Total number of time windows; Index for time windows.

[0058] To further reflect the stability of the connection within each time window, the connection strength within each time window was calculated separately. And further calculate its mean over multiple time windows. and standard deviation for: ; in, For the first The first channel and the first The first channel in the The average connection strength across multiple time windows in a frequency band; For the first The first channel and the first The first channel in the The standard deviation of connection strength across multiple time windows in a frequency band.

[0059] No. The first channel and the first The first channel in the Time window stability coefficient in each frequency band for: ; in, This represents the stability coefficient for the time window.

[0060] When a connection is relatively stable over multiple time windows Smaller Larger; when a connection is enhanced only occasionally within a few time windows, Larger The value is relatively small. Based on the time window stability coefficient, the functional connectivity strength is corrected to obtain the stability-corrected functional connectivity strength. for: ; in, For the first The first channel and the first The first channel in the Stability correction function connection strength on each frequency band.

[0061] For multiple frequency bands, the corrected functional connectivity matrices of each frequency band are fused to obtain the final EEG signal network matrix. for: ; in, For the first The first channel and the first The strength of EEG signal connectivity between channels; For the first Weight of each frequency band It can be set to equal weights, or it can be determined based on the classification contribution of each frequency band in the training set; This represents the total number of frequency bands.

[0062] EEG signal network matrix By performing sparsification or thresholding, the adjacency matrix of the EEG signal network is obtained. for: ; in, The first element in the adjacency matrix of the EEG signal network Line number Column elements; This represents the threshold for brain network connectivity.

[0063] Based on the adjacency matrix of the brainwave signal network or weighted matrix Extract graph theory features such as node clustering coefficient, node efficiency, global efficiency, modularity, small-world property, and rich club coefficient.

[0064] Node clustering coefficients characterize the degree of clustering in the neighborhood of a local node; node efficiency characterizes the information transmission performance between a single node and other nodes; global efficiency characterizes the overall information transmission efficiency of the entire brain network; modularity characterizes the modular organization level of the brain network; small-world property characterizes whether the brain network simultaneously possesses high local clustering and short global paths; and rich club coefficient characterizes whether tightly connected cores are formed between highly connected nodes. Finally, the node clustering parameters, node efficiency, global efficiency, modularity, small-world property, and club parameters of the EEG signal classification network are statistically summarized to form a feature set for subsequent EEG signal classification network applications.

[0065] S3: Construct EEG signal feature vectors and perform standardization and screening. Use the Crowned Porcupine Optimization Algorithm (CPO) to optimize the machine learning model, output three types of posterior probabilities, and construct a classification evaluation index for continuous EEG signals.

[0066] S31: Constructing and standardizing the EEG signal feature vector. The spectral features, nonlinear complexity features, microstate features, and EEG signal classification network features extracted in S2 are concatenated to form a multidimensional EEG signal feature vector: ; in, This is a multidimensional EEG signal feature vector; For the first Each EEG signal feature vector; The feature dimension; This is the matrix transpose symbol.

[0067] To eliminate differences in the dimensions of various features, the feature vectors of multidimensional EEG signals are standardized, specifically as follows: ; in, For the first The results of standardized processing of multidimensional EEG signal feature vectors; For the first Each EEG signal feature vector; For the first The mean of each feature in the training set; For the first The standard deviation of each feature in the training set.

[0068] The standardized feature vector that makes up the multidimensional EEG signal feature vector is as follows: ; in, The standardized feature vectors are the feature vectors of the multidimensional EEG signal.

[0069] A recursive feature elimination method is employed to filter standardized feature vectors of multi-dimensional EEG signals, obtaining a subset of core EEG features. Specifically, using standardized feature vectors from the training set as input, a support vector machine (SVM), extreme gradient boosting (AGES), or random forest (RFS) model is first selected as the base evaluator. Then, the base evaluator is trained on the current feature set, and each feature is ranked according to model weights, feature gains, and feature importance scores. Next, the features with the lowest importance are removed according to a preset elimination ratio, forming an updated candidate feature set. This training, ranking, and elimination process is repeated until the number of candidate features reaches a preset lower limit. For each round of candidate feature sets, a classification performance evaluation function is calculated using cross-validation. and will The candidate feature set with the highest value and the number of features that meets the complexity constraint is determined as the core feature subset.

[0070] S32: Train the machine learning model using the CPO (Coral Porcupine Optimization) algorithm. Input the subset of core EEG signal features obtained in S31 into the machine learning tri-classification model to construct a tri-classification model for identifying the feature control group HC, the first feature classification group noMHE, and the second feature classification group MHE. The machine learning model in this embodiment includes a Support Vector Machine (SVM) model and an Extreme Gradient Boosting (XGBoost) model; preferably, the CPO-SVM and the CPO-XGBoost models are used.

[0071] To improve the performance of a machine learning tri-class classification model, this invention employs the Crowned Porcupine Optimization Algorithm (CPO) to globally optimize the key hyperparameters of the machine learning model. Let the set of parameters to be optimized be: ;in, This represents the number of parameters to be optimized. For the extreme gradient boosting model XGBoost, the parameters to be optimized include tree depth, learning rate, subsampling ratio, column sampling ratio, minimum leaf node weight, and regularization parameter; for SVM, the parameters include penalty parameter. Kernel function parameters and kernel function type. The optimization objective of the CPO optimization algorithm for the hog porcupine is to optimize the parameter space. Finding the optimal set of parameters To optimize the validation performance of the machine learning three-class classification model, specifically: ; in, The function is a performance evaluation function for a machine learning three-class classification model, preferably a weighted combination of the macro average F1 score and the overall accuracy. The optimal set of parameters; The parameters that allow the objective function to reach its maximum value.

[0072] The performance evaluation function for a machine learning three-class classification model is: , ; in, In the model with parameters The recognition performance at that time is the macro-average of the F1 scores for each category, which is used to characterize the recognition performance of the classification model for samples of each category; θ represents the classification accuracy when the model parameter is θ, which characterizes the overall correctness of the classification model in classifying all samples. To identify performance weighting coefficients, ; This is the weighting coefficient for classification accuracy. ,and .

[0073] Let the first During the nth iteration The candidate solutions are The update process is then represented as: ; in, For the first During the nth iteration There are 10 candidate solutions; For the first During the nth iteration There are 10 candidate solutions; The parameter search increment is generated based on the current optimal solution, population distribution information, and random perturbations.

[0074] The optimal parameter combination is obtained through multiple rounds of iteration. Based on this, a machine learning model optimized by the CPO (Crowned Porcupine) optimization algorithm was established. Figure 5As shown, during the machine learning model training phase, a subset of core features is input into the machine learning three-class classification model to construct a machine learning three-class classification model for identifying three states: feature control group HC, first feature classification group noMHE, and second feature classification group MHE. The machine learning model consists of a support vector machine (SVM) model and an extreme gradient boosting model (XGBoost), with the CPO-SVM model or CPO-XGBoost model being preferred. To improve the model's classification performance, the Crowned Porcupine Optimization (CPO) algorithm is used to globally optimize the key hyperparameters of the machine learning model. Through multiple rounds of iteration, the optimal parameter combination is searched, enabling the model to achieve a better performance on the validation set. The model optimized by the Crowned Porcupine Optimization (CPO) algorithm can reduce the subjectivity brought about by empirical parameter tuning and improve classification accuracy, robustness, and generalization performance.

[0075] S33: Output three posterior probabilities and construct a classification evaluation index for continuous EEG signals. The machine learning model optimized by the Crowned Porcupine (CPO) algorithm in S32 is applied to EEG signal classification, outputting the posterior probabilities of belonging to three states: the feature control group (HC), the first feature classification group (noMHE), and the second feature classification group (MHE). These are denoted as: Feature Control Group Posterior Probability Posterior probability of the first feature classification group Posterior probability of the second feature classification group And satisfy: .

[0076] Based on three pre-defined categories of EEG signal features, the first EEG signal feature category C0, the second EEG signal feature category C1, and the third EEG signal feature category C2 are assigned values ​​of 0, 1, and 2, respectively, to construct a continuous EEG signal classification and evaluation index. for: , ; in, As a classification and evaluation index for continuous electroencephalogram (EEG) signals; The posterior probability of the characteristic control group; The posterior probability of the first feature classification group; is the posterior probability of the second feature classification group.

[0077] The above formula outputs a classification and evaluation index for continuous EEG signals. A higher value indicates that the EEG signal is closer to the MHE state of the second feature classification group; continuous EEG signal classification assessment index The smaller the value, the closer the EEG signal is to a healthy state; continuous EEG signal classification assessment index This is the core quantitative indicator that is ultimately output by this invention.

[0078] S4: Determine the dual thresholds and complete the three-level classification of EEG signals. Determine the first threshold based on continuous evaluation metrics from samples in the training or validation set. Second threshold ,satisfy: Based on the classification and evaluation indicators of continuous EEG signals determined in S3 With the first threshold Second threshold Based on this relationship, the risk of EEG signals is classified into three levels: low risk, medium risk, and high risk. This tiered control of associated medical devices is then implemented, and the categorized EEG signals are stored in a database and displayed on the terminals of the associated medical devices. Specifically: When continuous EEG signal classification assessment index At that time, it was determined to be low risk, corresponding to the characteristic control group HC, which is also low risk, corresponding to the first EEG signal feature category C0 in S1. A standard operating mode command was sent: the associated medical device executed the default operating procedure, completing EEG signal acquisition, preprocessing, and basic feature analysis with standard parameters, employing standard data storage and transmission strategies, performing complete EEG feature calculations, ensuring efficient and stable operation of the device, and outputting and displaying the risk level.

[0079] When the classification assessment index of continuous EEG signals meets When the risk level is determined to be medium, corresponding to the first feature classification group noMHE, it corresponds to the second EEG signal feature category C1 in S1. An enhanced acquisition mode instruction is sent: EEG monitoring increases the cache capacity of key data, retains the original signal and intermediate processing results, calculates the C1 EEG feature, provides a higher quality data foundation for subsequent analysis, and outputs and displays the risk level.

[0080] When continuous EEG signal classification assessment index If the risk level is high, corresponding to the second feature classification group MHE, it is considered high-risk. Corresponding to the third EEG signal feature category C2 in S1, a deep analysis mode command is sent: EEG monitoring switches to its highest performance state, automatically backs up recent historical data for secondary verification, and simultaneously enables dual data backup both locally and in the cloud to ensure the integrity and accuracy of critical data, and outputs and displays the risk level.

[0081] Output auxiliary assessment results. Classify and assess continuous EEG signals using relevant indicators. The risk level, category label, and three-class posterior probabilities are output to a display terminal, database, or clinical decision support system for screening, early warning, and auxiliary assessment of EEG signal states. The final output of this invention includes: continuous EEG signal classification assessment indicators. Two stratified thresholds and Low-risk, medium-risk, or high-risk assessment results; category label features corresponding to the risk level: control group HC, first feature classification group noMHE, and second feature classification group MHE.

[0082] like Figure 6 As shown, after obtaining the machine learning model optimized by the CPO algorithm, the EEG signal features of the EEG signal to be tested are input into the model, and the posterior probability of its belonging to three states—the control group HC, the first feature classification group noMHE, and the second feature classification group MHE—is output. Then, based on the assignment relationship between the three EEG signal states, a continuous EEG signal classification evaluation index is constructed. The smaller the continuous EEG signal classification evaluation index, the closer the EEG signal is to the healthy control state; the larger the continuous EEG signal classification evaluation index, the closer the EEG signal is to the second feature classification group MHE state. Further, based on the sample index distribution in the training or validation set, a first threshold and a second threshold are determined, and the EEG signal is divided into three risk levels: low risk, medium risk, and high risk. Low risk corresponds to the control group HC, medium risk corresponds to the first feature classification group noMHE, and high risk corresponds to the second feature classification group MHE. By combining continuous indicators with dual-threshold stratification, this invention can perform discrete classification judgments and also reflect the continuous change trend of EEG signals, improving the interpretability and application value of the auxiliary evaluation results.

[0083] The second aspect of this invention proposes an EEG signal data classification and processing system based on a multi-dimensional feature representation method. The system includes: an EEG signal acquisition and processing module, a multi-dimensional feature extraction module, a feature fusion and screening module, a model optimization and training module, a continuous index construction module, a signal hierarchical output module, and a data management module. The system is used to execute the above method and is deployed in an EEG acquisition device, an EEG analysis workstation, a server, or an edge computing terminal.

[0084] The EEG signal acquisition and processing module is used to acquire multi-channel EEG signal data and perform quality control and preprocessing on the raw EEG signals. Its function is to improve the quality of EEG signals and reduce the impact of noise and artifacts on multi-dimensional feature extraction. Specifically, it includes: an EEG data receiving unit, a filtering unit, a bad lead identification unit, an artifact removal unit, a rereference processing unit, and an effective data segmentation unit. The EEG data receiving unit receives multi-channel EEG signal data from the EEG acquisition device and records information such as sampling frequency, channel name, acquisition duration, and sample number. The filtering unit performs power frequency notch filtering and bandpass filtering on the EEG signals. The system employs several techniques: a wave generator to remove power frequency interference, low-frequency drift, and high-frequency noise; a bad lead identification unit to identify abnormal leads based on indicators such as channel variance, signal amplitude range, and channel correlation, and to repair them using neighboring channel interpolation or spherical interpolation; an artifact removal unit to remove electrooculography, electromyography, and motion artifacts through independent component analysis or other artifact recognition methods; a rereference processing unit to convert EEG signals into average or specified references to reduce the impact of reference electrode differences on subsequent feature extraction; and an effective data segmentation unit to divide the preprocessed EEG signals into multiple effective time periods, providing standardized input for subsequent feature calculations.

[0085] The multi-dimensional feature extraction module is used to extract spectral features, nonlinear complexity features, microstate features, and EEG signal network features from the preprocessed EEG signals. Its function is to characterize the frequency domain changes, nonlinear complexity changes, sub-second topological dynamic changes, and brain region connectivity changes of EEG signals from multiple dimensions. Specifically, it includes: a spectral feature extraction unit, a nonlinear complexity feature extraction unit, a microstate feature extraction unit, and a brain network feature extraction unit. The spectral feature extraction unit is used to calculate the absolute power and relative power of each channel in the Delta, Theta, Alpha, Beta, and Gamma frequency bands. The nonlinear complexity feature extraction unit performs empirical mode decomposition and Hilbert-Huang spectrum calculation on the preprocessed EEG signal, and constructs the intrinsic mode IMF component weights based on the energy proportion of intrinsic mode IMF components and mode stability coefficients. It further calculates the Hilbert-Huang state entropy (HHSE) and Hilbert-Huang response entropy (HHRE). The microstate feature extraction unit calculates the global field power and topological stability, and determines candidate microstate peak points based on these parameters. Microstate clustering is then performed on the candidate microstate peak points to obtain microstate templates and statistical features. The brain network feature extraction unit calculates multi-band dwPLI functional connectivity, constructs time window stability coefficients based on the mean and standard deviation of connectivity strength within each time window, corrects the stability of the functional connectivity matrix, and further extracts graph theory features such as node clustering coefficients, node efficiency, global efficiency, modularity, small-world property, and rich club coefficients.

[0086] The feature fusion and filtering module is used to concatenate, standardize, and filter various features output by the multi-dimensional feature extraction module. Its purpose is to reduce feature redundancy and improve model training efficiency and generalization performance. Specifically, it includes a feature concatenation unit, a feature standardization unit, a redundant feature removal unit, and a core feature output unit. The feature concatenation unit concatenates spectral features, nonlinear complexity features, microstate features, and brain network features into a unified multi-dimensional EEG signal feature vector. The feature standardization unit standardizes each dimension of features based on the mean and standard deviation of features in the training set to eliminate differences in the dimensionality of each feature. The redundant feature removal unit uses recursive feature elimination, feature importance ranking, or cross-validation methods to filter core features and remove redundant and low-contribution features. The core feature output unit outputs the filtered core EEG signal feature subset and transmits it to the model optimization and training module.

[0087] The model optimization and training module is used to build an EEG signal classification model and globally optimize the model's hyperparameters using the CPO optimization algorithm. Its purpose is to reduce reliance on manual parameter tuning and improve the model's classification performance, robustness, and cross-sample generalization performance. Specifically, it includes: a training data receiving unit, a model building unit, a CPO parameter optimization unit, a model validation unit, and an optimal model output unit. The training data receiving unit receives a subset of core EEG signal features and their corresponding class labels. The model building unit constructs SVM, XGBoost, random forest, or combined classification models. The CPO parameter optimization unit iteratively optimizes the model's hyperparameters within a preset parameter search space to find the optimal parameter combination that maximizes the model's performance evaluation function. The model validation unit evaluates model performance through cross-validation or independent validation sets, calculating metrics such as accuracy, macro-average F1 score, sensitivity, and specificity. The optimal model output unit outputs the CPO-optimized machine learning three-class classification model.

[0088] The continuous index construction module is used to construct continuous EEG signal classification evaluation indices based on the three-class posterior probabilities output by the optimized machine learning model. Its function is to convert discrete classification probabilities into continuous numerical indices, improving the interpretability of EEG signal classification results. Specifically, it includes: a posterior probability receiving unit, a class assignment unit, a continuous index calculation unit, and an index calibration unit. The posterior probability receiving unit is used to receive the model output. , and The category assignment unit is used to classify the features of the first EEG signal. Second category of EEG signal characteristics and the third category of EEG signal features The values ​​were assigned to 0, 1, and 2 respectively. The continuous index calculation unit is used to calculate the classification and evaluation index of continuous EEG signals according to the following formula. The index calibration unit is used to calibrate the index based on the distribution of the training or validation set. Perform calibration to adapt it to the data collected from each batch or device.

[0089] The signal stratification output module is used to output signal stratification results based on continuous EEG signal classification evaluation indicators and dual thresholds. Its function is to convert the model calculation results into easily understandable engineering output results. Specifically, it includes: a threshold determination unit, a stratification judgment unit, a result display unit, and a report generation unit. The threshold determination unit is used to determine the stratification results based on the training set or validation set. Distribution determines the first threshold Second threshold The stratification judgment unit is used to judge the sample to be processed. and , The system identifies the relationship between the signals and outputs low-risk, medium-risk, or high-signal stratification results. The results display unit shows the three posterior probabilities, continuous EEG signal classification assessment indicators, dual thresholds, and stratification results. The report generation unit generates an EEG signal-assisted assessment report, which includes sample number, acquisition parameters, effective data length, multi-dimensional feature summary, model output probability, continuous indicators, and signal stratification results.

[0090] The data management module is used to store and manage EEG signal data, feature data, model parameters, and output results. Its function is to enable the system's calculation process to be traceable and verified, and to facilitate subsequent model updates and engineering deployment. Specifically, it includes a raw data storage unit, a preprocessing data storage unit, a feature data storage unit, a model parameter storage unit, and a result recording unit. The raw data storage unit is used to store the raw EEG signal data and acquisition parameters. The preprocessing data storage unit is used to store the preprocessed EEG signal data and preprocessing logs. The feature data storage unit is used to store spectral features, nonlinear complexity features, microstate features, brain network features, and a subset of core features. The model parameter storage unit is used to store the optimal model parameters obtained by CPO optimization. The result recording unit is used to store the three posterior probabilities, continuous EEG signal classification evaluation indicators, and signal stratification results for each sample.

[0091] like Figure 7The diagram illustrates the single-sample processing and output effect of the EEG signal data classification and processing system of this invention. For the EEG signal to be processed, the system first receives an input sample X, which is multi-channel EEG signal data. Subsequently, it undergoes preprocessing, multi-dimensional feature extraction, CPO optimization model processing, and continuous index construction. Preprocessing reduces the impact of power line interference, electrooculogram (EOG) artifacts, electromyogram (EMG) artifacts, and abnormal leads on the quality of the EEG signal. Multi-dimensional feature extraction comprehensively characterizes the EEG signal from multiple levels, including spectral features, nonlinear complexity features, microstate features, and brain functional network features. The CPO optimization model outputs the posterior probabilities of the sample belonging to the three states: HC, noMHE, and MHE, based on the selected core features. The diagram exemplifies the model output results, where the posterior probability for HC is 0.11, for noMHE it is 0.28, and for MHE it is 0.61, with a sum of 1 for all three probabilities. A continuous EEG signal classification and evaluation index I is constructed based on three types of posterior probabilities. In the example, I=0.75, indicating that the EEG signal to be processed is closer to the high-risk category in the feature space. As shown in the figure, this invention does not only output a single category label, but can simultaneously output three types of posterior probabilities, continuous evaluation index, and risk level, making the EEG signal classification and processing results more intuitive, quantifiable, and interpretable.

[0092] like Figure 8 The figure shows the effect of using the EEG signal data classification and processing system of this invention. The horizontal axis represents the low-risk HC group, the medium-risk noMHE group, and the high-risk MHE group, respectively; the vertical axis represents the continuous EEG signal classification assessment index I; each scatter point represents the continuous assessment result obtained after processing an EEG signal sample by the system of this invention, and the horizontal short line represents the central tendency of the index for the corresponding category of samples. Figure 8 It can be seen that the continuous assessment indicators of the low-risk HC group samples are generally at a low level, mainly distributed below the first threshold T1; the continuous assessment indicators of the noMHE medium-risk group samples are in the middle region, mainly distributed between the first threshold T1 and the second threshold T2; the continuous assessment indicators of the MHE high-risk group samples are generally at a high level, mainly distributed above the second threshold T2. By setting the first threshold T1 and the second threshold T2, this invention can divide the continuous EEG signal classification assessment indicators into three intervals: low risk, medium risk, and high risk, realizing the conversion from continuous indicators to three-level risk stratification results. The figure further illustrates that the continuous assessment indicators output by this invention have obvious distinguishability between samples of different categories, and can better reflect the trend of EEG signal characteristics changing from low-risk to medium-risk and high-risk states, thereby improving the stability and auxiliary assessment value of EEG signal classification processing results.

[0093] In summary, this embodiment demonstrates that the EEG signal classification and evaluation method proposed in this invention, based on multi-channel EEG, can extract EEG signal classification-related information from multiple levels, including frequency domain, complexity, microstate, and EEG signal classification network topology, based on resting-state EEG signal data. Furthermore, it constructs a continuous EEG signal classification and evaluation index and a dual-threshold three-level signal stratification mechanism using a CPO-optimized machine learning model, thereby providing objective, quantitative, and intelligent auxiliary evaluation of EEG signals. This invention can be used for early screening and risk warning in EEG signal monitoring devices and has promising application prospects.

[0094] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for classifying and processing electroencephalogram (EEG) signal data based on multi-dimensional feature representation, characterized in that: It includes: S1: Preprocess the multi-channel EEG signal data to obtain preprocessed EEG signal data. Classify EEG signal features based on pre-labeled EEG signal feature patterns; S2: Extract multidimensional EEG signal features that can characterize the classification of EEG signals; Set the power spectral density to Determine the frequency band of EEG signal data Internal bandwidth power To obtain the relative power Extracting spectral features from electroencephalogram (EEG) signals; The marginal spectrum is obtained by performing a Hilbert transform on the intrinsic mode components. Construct intrinsic mode component weights The weighted feature marginal spectrum is obtained. Determine the Hilbert yellow state entropy Hilbert's Yellow Response Entropy Extracting the nonlinear complexity features of EEG signal data; Based on global field power and topological stability, construct microstate peak evaluation parameters. ; Filtering the set of candidate microstate peaks Extracting microstate features from electroencephalogram (EEG) signal data; The preprocessed EEG signals were divided into time windows, and the stability coefficient of each time window was determined. The functional connection strength is then corrected to obtain the functional connection strength. The functional connectivity matrices corrected for each frequency band are fused to obtain the final EEG signal network matrix. Extract features from the EEG signal classification network; S3: Based on the multidimensional EEG signal features extracted in step S2, construct EEG signal feature vectors, perform standardization and screening, use the Crowned Porcupine optimization algorithm to optimize the machine learning model, output three classes of posterior probabilities, and construct a continuous EEG signal classification evaluation index. ; S4: Determine the first threshold based on the continuous evaluation index of the sample. Second threshold ; Based on the classification and evaluation indicators of continuous EEG signals determined in step S3 With the first threshold Second threshold The system performs three-level classification of EEG signals, controls associated medical devices at different levels, and stores the classified EEG signals in a database for display on the associated medical device display terminal.

2. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: Step S2 is as follows: S21: Based on the power spectral density is Determine the bandwidth and power of EEG signals in different frequency bands. To obtain the relative power ; S22: Perform Hilbert transform on the intrinsic mode components; construct component weights. The weighted feature marginal spectrum is obtained. We obtain Hilbert's yellow state entropy. Hilbert's Yellow Response Entropy ; S23: Construct microstate peak evaluation parameters based on global field power and topological stability. ; Filter the set of candidate microstate peak points Extract average duration, average occurrence rate, coverage, and average global field power; S24: Divide the preprocessed EEG signal into time windows and determine the stability coefficient of the time windows. The functional connection strength is then corrected to obtain the functional connection strength. By fusing the functional connectivity matrix, the final EEG signal network matrix is ​​obtained. .

3. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: In step S22, the intrinsic mode component weights are constructed. The weighted feature marginal spectrum is obtained. Specifically: ； ； in, For the first The weighted feature marginal spectrum of each channel; For the first The first channel The weights of each intrinsic mode component; For the first The first channel Each intrinsic mode component at frequency Marginal spectrum at the location; The frequency of the brainwave signal; This represents the total number of intrinsic modal components. Number the frequency bands for EEG signal data; For the first The first channel The energy percentage of each intrinsic mode component; For the first The first channel Modal stability coefficients of each intrinsic modal component; For the first The first channel Modal stability coefficients of each intrinsic modal component; Index for summing intrinsic modal components; For the first The energy of each intrinsic mode component; This is the parameter for correcting the denominator.

4. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: In step S23, microstate peak evaluation parameters are constructed based on global field power and topological stability. ; Filtering the set of candidate microstate peaks Specifically: ； ； in, For evaluating the peak value of the microstate; This is the topological stability adjustment coefficient; For the first Topological stability at each sampling point; This is the set of candidate microstate peak points; For microstate peak evaluation parameters The Quantile threshold; Quantile index for peak evaluation parameters of microstate; ; This refers to the time sampling point number in the characteristic time spectrum.

5. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: In step S24, the EEG signal network matrix The method for obtaining it is as follows: ； ； ； in, This is the stability coefficient for the time window; For the first The first channel and the first The first channel in the The average connection strength across multiple time windows in a frequency band; For the first The first channel and the first The first channel in the Standard deviation of connection strength across multiple time windows in a frequency band; For the first The first channel and the first The first channel in the Stability correction function connection strength on each frequency band; For the first The first channel and the first The strength of EEG signal connectivity between channels; For the first Weights of each frequency band; Total number of frequency bands; This is a frequency band index.

6. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: In step S1, the brain signal feature categories are divided into three categories: first brain signal feature category C0, second brain signal feature category C1, and third brain signal feature category C2, which correspond to the feature control group HC, the first feature classification group noMHE, and the second feature classification group MHE identified by the machine learning three-classification model.

7. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: Step S3 is as follows: S31: Concatenate the features extracted in step S2 to form a multidimensional EEG signal feature vector. Standardization processing is performed to form a multidimensional EEG signal feature vector. ; S32: Train the machine learning model optimized using the Crowned Porcupine optimization algorithm; input the subset of core features of the EEG signal obtained in step S31 into the machine learning three-class classification model to construct the machine learning three-class classification model in the parameter space. Finding the optimal set of parameters This ensures that the machine learning three-class classification model achieves optimal validation performance. S33: Apply the optimized machine learning model from step S32 to EEG signal classification, and output the posterior probability of the feature control group. Posterior probability of the first feature classification group Posterior probability of the second feature classification group .

8. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: The performance evaluation function for the machine learning three-class classification model in step S32 is: ， ； ； in, This is a performance evaluation function for a machine learning three-class classification model. The optimal set of parameters; The parameters that allow the objective function to reach its maximum value; In the model with parameters Real-time recognition performance; The classification accuracy is given by the model parameter θ. To identify performance weighting coefficients; This is the weighting coefficient for classification accuracy.

9. The method for classifying and processing EEG signal data based on multi-dimensional feature representation according to claim 1, characterized in that: The three-level classification of EEG signals in step S4 is as follows: When continuous EEG signal classification assessment index When the risk is low, the first EEG signal feature category C0 in step S1 is identified, the category label is feature control group HC, and a standard working mode instruction is sent to the associated medical device. When the classification assessment index of continuous EEG signals meets When the risk is determined to be medium, corresponding to the second EEG signal feature category C1 in step S1, the category label is the first feature classification group noMHE, and an enhanced acquisition mode command is sent to the associated medical device. When continuous EEG signal classification assessment index If the risk is high, it corresponds to the third EEG signal feature category C2 in step S1, with the category label being the second feature classification group MHE, and a deep analysis mode command is sent to the associated medical device.

10. A brain signal data classification and processing system for the brain signal data classification and processing method based on multi-dimensional feature representation as described in claim 1, characterized in that: It includes: The system includes an EEG signal acquisition and processing module, a multi-dimensional feature extraction module, a feature fusion and screening module, a model optimization and training module, a continuous index construction module, a signal hierarchical output module, and a data management module. The EEG signal acquisition and processing module is used to acquire multi-channel EEG signal data and perform quality control and preprocessing on the raw EEG signals. Its function is to improve the quality of EEG signals and reduce the impact of noise and artifacts on multi-dimensional feature extraction. The multi-dimensional feature extraction module is used to extract spectral features, nonlinear complexity features, microstate features, and EEG signal network features from the preprocessed EEG signals. Its role is to characterize the frequency domain changes, nonlinear complexity changes, sub-second topological dynamic changes, and brain region connectivity changes of EEG signals from multiple dimensions. The feature fusion and filtering module is used to concatenate, standardize and filter various features output by the multi-dimensional feature extraction module. Its function is to reduce feature redundancy and improve model training efficiency and generalization performance. The model optimization and training module is used to build an EEG signal classification model and to globally optimize the model hyperparameters through optimization algorithms. Its purpose is to reduce the dependence on manual parameter tuning and improve the model's classification performance, robustness and cross-sample generalization performance. The continuous index construction module is used to construct continuous EEG signal classification evaluation indexes based on the three types of posterior probabilities output by the optimized machine learning model. Its function is to convert discrete classification probabilities into continuous numerical indexes, thereby improving the interpretability of EEG signal classification results. The signal stratification output module is used to output signal stratification results based on continuous EEG signal classification evaluation indicators and dual thresholds. Its function is to convert the model calculation results into easily understandable engineering output results. The data management module is used to store and manage EEG signal data, feature data, model parameters and output results. Its function is to enable the system's calculation process to be traced and verified, and to facilitate subsequent model updates and engineering deployment.

Images