Construction method of pain model based on multi-dimensional electroencephalogram feature fusion

Abstract

This application relates to the field of computer technology and discloses a method for constructing a pain model based on multi-dimensional EEG feature fusion. The method includes: acquiring segmented EEG data; extracting four-layer heterogeneous features from the segmented EEG data to construct a high-dimensional feature matrix. The four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling. An algorithmic consensus screening framework is constructed. Using the algorithmic consensus screening framework, a consensus score for the four-layer heterogeneous features corresponding to the high-dimensional feature matrix is ​​determined. Based on the consensus score and the four-layer heterogeneous features, multiple target feature sets and their corresponding scores are determined. Based on the multiple target feature sets and their corresponding scores, a subset of features with the highest scores is determined from the multiple target feature sets. A pain model is constructed based on the feature subset.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method for constructing a pain model based on the fusion of multi-dimensional electroencephalogram (EEG) features and a method for analyzing pain intensity. Background Technology

[0002] Electroencephalography (EEG) can reflect the collective firing patterns of brain neurons and the functional state of brain networks in real time and non-invasively by collecting spontaneous electrophysiological activity signals from the cerebral cortex. Its signal characteristics are directly related to the physiological and pathological changes of the nervous system.

[0003] In related technologies, features such as the spectrum, time domain, and nonlinear dynamics of electroencephalogram (EEG) signals are extracted and combined with machine learning models to determine the degree of pain.

[0004] However, improper selection of the number of features can lead to the model losing key diagnostic information or overfitting, thereby reducing the accuracy of chronic pain severity assessment. Summary of the Invention

[0005] In view of this, this application provides a method for constructing a pain model based on multi-dimensional EEG feature fusion and a method for analyzing pain intensity, in order to solve the technical problem that the number of features for model training relies solely on human experience to set, which can easily reduce the accuracy of pain intensity determination due to improper selection of the number of features.

[0006] Firstly, this application provides a method for constructing a pain model based on multi-dimensional EEG feature fusion. The method includes: acquiring segmented EEG data; extracting four-layer heterogeneous features from the segmented EEG data to construct a high-dimensional feature matrix. The four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling. An algorithm consensus screening framework is constructed, which includes at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm, and extreme gradient boosting algorithm. Using the algorithm consensus screening framework, a consensus score is determined for the four-layer heterogeneous features corresponding to the high-dimensional feature matrix. Based on the consensus score and the four-layer heterogeneous features, multiple target feature sets and their corresponding scores are determined. Based on the multiple target feature sets and their corresponding scores, a subset of features with the highest scores is determined from the multiple target feature sets. A pain model is constructed based on the feature subset.

[0007] The method for constructing a pain model based on multi-dimensional EEG feature fusion provided in this application extracts four layers of heterogeneous features highly correlated with the pathological mechanism of chronic pain from segmented EEG data: spectrum, temporal dynamics, brain functional connectivity, and cross-frequency coupling, and constructs a high-dimensional feature matrix to reduce interference from meaningless redundant features. Then, it constructs an algorithmic consensus screening framework that includes at least one of sparse linear, random forest, gradient boosting decision tree, and extreme gradient boosting algorithms. This framework is used to calculate the consensus score for each feature, objectively quantifying the importance of features to the pain model from an algorithmic perspective, replacing subjective judgments of feature value based on human experience.

[0008] Subsequently, the four-layer heterogeneous features are sorted based on consensus scores, and multiple target feature sets and corresponding scores are determined. Instead of relying on manual experience to fix the number of features, the highest-scoring feature subset is automatically selected from multiple feature sets based on scores. This process combines the objective quantitative results of feature importance to achieve a scientific selection of the number of features. It avoids the problem of losing key features for pain diagnosis due to too few features set manually, and also avoids the problem of model overfitting and reduced judgment accuracy caused by too many features.

[0009] In one possible implementation, acquiring segmented EEG data includes: acquiring a first multi-channel EEG signal; performing sliding segmentation on the first multi-channel EEG signal using a preset time window and overlap step size, and sequentially performing bandpass filtering and notch filtering to obtain a second multi-channel EEG signal; discarding the second multi-channel EEG signal with an amplitude exceeding a preset amplitude to obtain segmented EEG data.

[0010] In one possible implementation, an algorithmic consensus screening framework is used to determine the consensus score of the four heterogeneous features corresponding to the high-dimensional feature matrix. This includes: using the algorithmic consensus screening framework to calculate the original importance score of each heterogeneous feature in the high-dimensional feature matrix; and determining the consensus score of the four heterogeneous features corresponding to the high-dimensional feature matrix based on the original importance score.

[0011] In one possible implementation, the consensus score of the four heterogeneous features corresponding to the high-dimensional feature matrix is ​​determined based on the original importance score, including: performing maximum-minimum processing on the original importance scores output by each algorithm to calculate the target score corresponding to each heterogeneous feature; and using the target score as the corresponding consensus score.

[0012] In one possible implementation, based on consensus scores and four-layer heterogeneous features, multiple target feature sets and their corresponding scores are determined, including: sorting the four-layer heterogeneous features based on consensus scores to obtain a candidate feature list; selecting a target number of features sequentially from the candidate feature list to construct feature subsets; performing cross-validation on the feature subsets to determine the overfitting risk value for each feature subset; and determining multiple target feature sets and their corresponding scores based on the overfitting risk value and a balanced scoring function.

[0013] Secondly, this application provides a method for analyzing pain intensity, which includes: acquiring a multi-dimensional EEG feature set to be processed; and determining the pain intensity based on the multi-dimensional EEG feature set and a pain model.

[0014] Thirdly, this application provides a device for constructing a pain model based on multi-dimensional EEG feature fusion. The device includes: an acquisition module for acquiring segmented EEG data; a first construction module for extracting four-layer heterogeneous features from the segmented EEG data to construct a high-dimensional feature matrix, wherein the four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features, wherein the temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling; a second construction module for constructing an algorithm consensus screening framework, wherein the algorithm consensus screening framework includes at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm, and extreme gradient boosting algorithm; a first determination module for using the algorithm consensus screening framework to determine the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix; a second determination module for determining multiple target feature sets and their corresponding scores based on the consensus scores and the four-layer heterogeneous features; a third determination module for determining the highest-scoring feature subset from the multiple target feature sets based on the scores of the multiple target feature sets; and a third construction module for constructing a pain model based on the feature subset.

[0015] Fourthly, this application provides a computer device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the method for constructing a pain model based on multi-dimensional EEG feature fusion as described in the first aspect or any of its corresponding embodiments, and / or the method for analyzing pain intensity as described in the second aspect or any of its corresponding embodiments.

[0016] Fifthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to execute the method for constructing a pain model based on multi-dimensional EEG feature fusion according to the first aspect or any corresponding embodiment described above, and / or the method for analyzing pain intensity according to the second aspect or any corresponding embodiment described above.

[0017] In a sixth aspect, this application provides a computer program product, including computer instructions for causing a computer to perform the methods described in the first aspect or the second aspect or any corresponding embodiment thereof. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating another method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application.

[0021] Figure 3 This is a flowchart illustrating another method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0022] Figure 4 This is a flowchart illustrating another method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0023] Figure 5 This is a flowchart illustrating another method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0024] Figure 6 This is a flowchart illustrating another method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0025] Figure 7 This is an architecture diagram of a pain model construction system based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0026] Figure 8This is a structural block diagram of a pain model construction device based on multi-dimensional EEG feature fusion according to an embodiment of this application;

[0027] Figure 9 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Based on relevant technologies, chronic pain is a complex neurological disorder, and its objective assessment has long been a clinical challenge. Electroencephalography (EEG), as a non-invasive technique, can reflect the electrophysiological activity of the cerebral cortex and has diagnostic potential. However, current technologies have the following significant limitations:

[0030] The curse of feature dimension: EEG signals are non-stationary and vary greatly from person to person. The massive features extracted by traditional methods contain a lot of redundancy and noise, which makes the model prone to overfitting and has poor generalization ability.

[0031] Feature selection bias: Existing research often relies on a single algorithm (such as LASSO or random forest alone) for feature selection. Because different algorithms have different underlying assumptions, a single algorithm is prone to missing key features or selecting spurious features.

[0032] Lack of system optimization: The selection of the number of features often depends on empirical settings (such as selecting the first 10), and there is a lack of an automated search mechanism based on the balance between model performance and complexity.

[0033] Insufficient interpretability: Most "black box" models cannot explain to doctors why a patient is diagnosed with chronic pain, lacking a biologically meaningful explanation.

[0034] Based on this, this application provides a method for constructing a pain model based on multi-dimensional EEG feature fusion. By extracting four layers of heterogeneous features highly correlated with the pathological mechanism of chronic pain from segmented EEG data—spectral, temporal dynamics, brain functional connectivity, and cross-frequency coupling—and constructing a high-dimensional feature matrix, meaningless redundant feature interference is reduced. Then, an algorithm consensus screening framework is constructed, which includes at least one of sparse linear, random forest, gradient boosting decision tree, and extreme gradient boosting algorithms. The consensus score of each feature is calculated through this algorithm consensus screening framework, objectively quantifying the importance of features to the pain model from the algorithm level, replacing the subjective judgment of feature value by human experience.

[0035] Subsequently, the four-layer heterogeneous features are sorted based on consensus scores, and multiple target feature sets and corresponding scores are determined. Instead of relying on manual experience to fix the number of features, the highest-scoring feature subset is automatically selected from multiple feature sets based on scores. This process combines the objective quantitative results of feature importance to achieve a scientific selection of the number of features. It avoids the problem of losing key features for pain diagnosis due to too few features set manually, and also avoids the problem of model overfitting and reduced judgment accuracy caused by too many features.

[0036] According to an embodiment of this application, a method for constructing a pain model based on multi-dimensional EEG feature fusion is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0037] This embodiment provides a method for constructing a pain model based on multi-dimensional EEG feature fusion, which can be used in computer devices such as computers and servers. Figure 1 This is a flowchart of a method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0038] Step S101: Obtain EEG segmentation data.

[0039] EEG segmentation data indicates high signal-to-noise ratio EEG data segments obtained by preprocessing the original multi-channel EEG signals (filtering, artifact removal, outlier filling, etc.) and cutting them according to a fixed time window and overlap step size. It is the basic data for feature extraction.

[0040] The acquired EEG segment data is preprocessed high signal-to-noise ratio data (not the original EEG signal). The preprocessing process includes: bandpass filtering, notch filtering, ICA artifact removal, amplitude threshold screening, and missing value interpolation filling of the original multi-channel EEG signal. Then, it is slid-cut according to the preset time window and overlap step size to finally obtain standardized EEG segment data. The data format is a time-series numerical sequence, which can be directly used for feature extraction.

[0041] As an example, a 2-second time window plus a 1-second overlap step is used to segment the preprocessed EEG signal to obtain EEG segment data with a single segment duration of 2 seconds and an overlap of 1 second between adjacent segments. In this embodiment, the number of sampling points can be determined by the sampling rate and the time window length (e.g., sampling rate × time window length). For example, each segment of a single channel can contain 2000 sampling points.

[0042] Step S102: Extract four-layer heterogeneous features from the EEG segment data and construct a high-dimensional feature matrix. The four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling.

[0043] The four-layer heterogeneous feature set refers to a set of complementary features extracted from four dimensions of EEG signals, all of which are directly related to the neuropathological mechanisms of chronic pain. These four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling.

[0044] Temporal dynamic characteristics can indicate the temporal variation patterns of EEG signals and the temporal fluctuations in pain patients. Spectral characteristics can indicate the energy distribution across different frequency bands of EEG. Brain functional connectivity characteristics based on regions of interest can indicate signal synchronization between brain regions. Cross-frequency coupling characteristics can indicate synergistic effects between different frequency bands.

[0045] Specifically, based on segmented EEG data, heterogeneous features are extracted from four dimensions: spectrum, temporal dynamics, brain functional connectivity based on regions of interest, and cross-frequency coupling. Each dimension of features focuses on different neuropathological features of chronic pain. All extracted features are organized according to the "sample-feature" dimension to construct a high-dimensional feature matrix.

[0046] In this application, the four-layer heterogeneous features are calculated based on narrowband filtered EEG signals (for both time-domain and nonlinear dynamic features, the original EEG signals are first filtered to the corresponding frequency bands before relevant statistics are calculated). Each frequency band is divided into Delta (1-4 Hz), Theta (4-8 Hz), Alpha (8-14 Hz), Beta (15-30 Hz), Low_Gamma (30-45 Hz), and High_Gamma (55-100 Hz). Furthermore, each of the four heterogeneous features corresponds to the pathological trend changes in specific brain regions of patients with chronic pain. The specific feature indicators used in actual extraction are as follows:

[0047] Frequency domain characteristics: These are the absolute and relative power of each frequency band in a specific brain region, and are the core indicators reflecting the energy distribution of EEG signals.

[0048] Temporal characteristics: The temporal variability (Mean of Temporal SD) and temporal dispersion (SD of Temporal SD) of each frequency band in a specific brain region are calculated based on narrowband filtered signals.

[0049] Cross-frequency coupling characteristics can include coupling between different frequency bands of the same electrode / across electrodes, and coupling types include phase-amplitude coupling, amplitude-amplitude coupling, phase-phase coupling, etc.

[0050] Nonlinear dynamic characteristics: Permutation entropy of each frequency band in a specific brain region, calculated based on narrowband filtered signals, is an indicator reflecting the time series complexity of EEG signals.

[0051] Among the four heterogeneous features selected in this application, permutation entropy and phase-amplitude coupling are the core feature indicators for chronic pain recognition. Permutation entropy reflects the complexity changes of the EEG signal time series of chronic pain patients, while phase-amplitude coupling reflects the abnormal cross-frequency synergistic effect of EEG signals in different frequency bands of patients. The two, combined with the frequency domain and time domain features of specific brain regions and specific frequency bands, form a set of EEG feature systems specifically for chronic pain. The feature selection method of this application effectively improves the specificity and accuracy of the pain model for chronic pain recognition.

[0052] Step S103: Construct an algorithm consensus screening framework. The algorithm consensus screening framework shall include at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm, and extreme gradient boosting algorithm.

[0053] The algorithm consensus screening framework can be a parallel evaluation framework composed of multiple feature selection algorithms, which is used to quantify feature importance through the collaboration of multiple algorithms and avoid the screening bias of a single algorithm.

[0054] The constructed framework is a multi-algorithm parallel evaluation system. The framework includes at least one of the following algorithms: sparse linear algorithm, random forest algorithm, gradient boosting decision tree (GBDT) algorithm, and extreme gradient boosting (XGBoost) algorithm. Preferably, all four algorithms are included. Each algorithm runs independently and is used to quantify the importance score of the feature.

[0055] As an example, the full algorithm framework (optimal) includes four algorithms: LASSO regression (sparse linear algorithm), random forest, gradient boosting decision tree, and extreme gradient boosting (XGBoost), which run in parallel.

[0056] As an example, a dual-algorithm framework is used, which incorporates LASSO regression and random forest algorithms to evaluate both linear and nonlinear features.

[0057] As an example, the single-algorithm framework, which only contains the XGBoost algorithm, is suitable for scenarios with low feature dimensionality and high requirements for screening efficiency.

[0058] As an example, the three-algorithm framework includes Random Forest + Random Forest + XGBoost algorithms, focusing on the importance assessment of nonlinear features and adapting to EEG feature sets with a high proportion of nonlinear features.

[0059] Step S104: Using the algorithm consensus screening framework, determine the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix.

[0060] The consensus score indicates the average importance score of each feature after it has been evaluated and standardized by the algorithm within the framework. The higher the score, the stronger the diagnostic value of the feature for the pain model.

[0061] When determining the consensus score, an algorithmic consensus screening framework can be used.

[0062] As an example, the original importance score of each feature in the high-dimensional feature matrix is ​​first calculated by each algorithm within the framework. Then, the original scores of each algorithm are standardized by min-max (to eliminate the difference in scale between different algorithms and map to the [0,1] interval). Finally, the arithmetic mean of the standardized scores of each feature under all algorithms within the framework is calculated. This average value is the consensus score of the feature.

[0063] In this application, an algorithmic consensus screening framework including LASSO regression, random forest, gradient boosting decision tree, and extreme gradient boosting is used to calculate the original importance score, standardize by min-max, and arithmetically average the above four heterogeneous feature indicators to obtain the consensus score of each feature. The consensus scores are then sorted in descending order from high to low to generate a candidate feature list. This candidate feature list provides a direct feature basis for the subsequent step S105 to determine multiple target feature sets.

[0064] Step S105: Based on consensus scoring and four-layer heterogeneous features, determine multiple target feature sets and the scores corresponding to the target feature sets.

[0065] First, sort all four heterogeneous features in descending order of consensus score. Then, select the top N features after sorting according to different feature quantities to form multiple target feature sets (N is a positive integer, and different N correspond to different target feature sets). At the same time, assign a corresponding score to each target feature set (the score can be the mean or sum of the consensus scores of all features in the target feature set, or the model performance score obtained based on cross-validation, with preferred model performance scores such as AUC and balanced scores).

[0066] In a feasible scenario, three target feature sets are generated according to the formula "feature quantity N=1, 2, 3".

[0067] Target feature set 1 (N=1): {B}, corresponding to a balance score of 0.85;

[0068] Target feature set 2 (N=2): {B, C}, corresponding to a balance score of 0.92;

[0069] Target feature set 3 (N=3): {B, C, A}, corresponding to a balance score of 0.88;

[0070] As an example: Generate two target feature sets by "N=1, 3" (skip N=2);

[0071] Target feature set 1 (N=1): {B}, corresponding to a score (sum of consensus scores) of 1.0;

[0072] Target feature set 2 (N=3): {B, C, A}, corresponding score (mean of consensus score) (1.0+0.475+0.0) / 3≈0.49;

[0073] As an example: The feature set contains 10 features (consensus score ranking: F1(0.98) > F2(0.95) > ... > F10(0.1)), and 5 target feature sets are generated according to N=2, 4, 6, 8, 10. The score of each set is the AUC value of the test set used for model training.

[0074] Target feature set 1 (N=2): {F1, F2}, AUC=0.88; Target feature set 2 (N=4): {F1-F4}, AUC=0.93; Target feature set 3 (N=6): {F1-F6}, AUC=0.95; Target feature set 4 (N=8): {F1-F8}, AUC=0.92; Target feature set 5 (N=10): {F1-F10}, AUC=0.89.

[0075] Step S106: Based on multiple target feature sets and the scores corresponding to the target feature sets, determine the feature subset with the highest score from the multiple target feature sets.

[0076] A feature subset can be understood as the set of features with the highest scores and most suitable for training the pain model selected from multiple target feature sets. It is the final feature basis for model construction.

[0077] Specifically, all target feature sets are scored and ranked, and the target feature set with the highest score is selected as the final feature subset.

[0078] Step S107: Construct a pain model based on a subset of features.

[0079] Using the feature subset selected in step S106 as the core input features and “chronic pain level (mild / moderate / severe)” as the label, a pain model is trained using a machine learning algorithm. In this scheme, the Stacking ensemble learning model is preferred (balancing model accuracy and generalization ability), but a single machine learning algorithm can also be used. After the model is trained, it can be directly used to determine the pain level of the subject to be diagnosed.

[0080] The method for constructing a pain model based on multi-dimensional EEG feature fusion provided in this application extracts four layers of heterogeneous features highly correlated with the pathological mechanism of chronic pain from segmented EEG data: spectrum, temporal dynamics, brain functional connectivity, and cross-frequency coupling, and constructs a high-dimensional feature matrix to reduce interference from meaningless redundant features. Then, it constructs an algorithmic consensus screening framework that includes at least one of sparse linear, random forest, gradient boosting decision tree, and extreme gradient boosting algorithms. This framework is used to calculate the consensus score for each feature, objectively quantifying the importance of features to the pain model from an algorithmic perspective, replacing subjective judgments of feature value based on human experience.

[0081] Subsequently, the four-layer heterogeneous features are sorted based on consensus scores, and multiple target feature sets and corresponding scores are determined. Instead of relying on manual experience to fix the number of features, the highest-scoring feature subset is automatically selected from multiple feature sets based on scores. This process combines the objective quantitative results of feature importance to achieve a scientific selection of the number of features. It avoids the problem of losing key features for pain diagnosis due to too few features set manually, and also avoids the problem of model overfitting and reduced judgment accuracy caused by too many features.

[0082] In one possible implementation, step S101 above includes:

[0083] Step a1: Acquire the first multi-channel EEG signal.

[0084] The first multichannel EEG signal can be understood as the raw EEG signal collected from specific physical electrodes on the subject's scalp, which is a continuous EEG temporal signal without any preprocessing.

[0085] The electrode locations collected were precisely matched to brain regions associated with chronic pain (prefrontal lobe, central region, occipital lobe, parietal lobe, temporal lobe, etc.) in the data file.

[0086] As an example, for subjects with chronic pain, continuous raw EEG signals were collected from physical electrodes on their scalp, including Fp1, Fp2, Cz, O1, O2, FC4, Pz, T8, and P3 (which are completely consistent with the electrodes corresponding to the core features in the document), to obtain the first multi-channel EEG signal.

[0087] Step a2: The first multi-channel EEG signal is segmented by a preset time window and overlap step size, and bandpass filtering and notch filtering are performed sequentially to obtain the second multi-channel EEG signal.

[0088] The preset time window can be understood as the duration of a single segment of EEG data when sliding segments continuous EEG signals.

[0089] The overlap step size can be understood as the duration of the overlap between two consecutive EEG segments. Setting an overlap step size can prevent the loss of the temporal characteristics of EEG signals and improve data utilization.

[0090] The first multichannel EEG signal can be a continuous EEG signal.

[0091] Bandpass filtering can be understood as a signal processing method that filters out effective signals within a specific frequency range in EEG signals and removes low-frequency noise and high-frequency interference outside that range.

[0092] Notch filtering can be understood as a narrowband filtering method that specifically filters out power frequency interference (such as 50Hz) in EEG signals.

[0093] The second multichannel EEG signal can be understood as the EEG signal obtained after the first multichannel EEG signal is processed by sliding segmentation + bandpass filtering + notch filtering.

[0094] This application's embodiment consists of two core sub-steps: sliding segmentation and dual filtering. The processing order is segmentation followed by filtering. The processed second multi-channel EEG signal retains all frequency band features (Delta, Theta, Alpha, Beta, Low Gamma, High Gamma) related to chronic pain in the file, while filtering out irrelevant noise and power line interference. The specific logic is as follows:

[0095] The first multi-channel EEG signal is segmented into several segments of equal duration according to a preset time window and overlap step size to ensure the temporal continuity of the signal. For the segmented EEG signal, the signal range of 0.1-100Hz (covering all pain-related frequency bands in the full coverage file) is selected, and noise outside this range is filtered out. For the bandpass filtered signal, 50Hz power frequency interference is specifically filtered out to further improve the signal quality. Finally, the second multi-channel EEG signal after segmentation and double filtering is obtained.

[0096] As an example, a 2-second preset time window plus a 1-second overlap step is used to slide segment the first multi-channel EEG signal. First, a 0.1-100Hz bandpass filter is performed, and then a 50Hz notch filter is performed to remove power frequency interference, thus obtaining the second multi-channel EEG signal.

[0097] Step a3: Discard the second multichannel EEG signal whose amplitude exceeds the preset amplitude to obtain segmented EEG data.

[0098] The effective data segments in the second multi-channel EEG signal are filtered out by a preset amplitude threshold: EEG segments with amplitude exceeding the preset threshold are judged as invalid data contaminated by artifacts such as electrooculography, electromyography, and body tremors, and are directly discarded; EEG segments with amplitude within the threshold range are judged as valid data, retained and integrated into the final EEG segment data.

[0099] As an example, we set ±100μV as the preset amplitude, traversed all segments of the second multi-channel EEG signal, discarded all contaminated data segments with amplitudes exceeding ±100μV, and retained the valid EEG segments with amplitudes within the threshold range. After integration, we obtained EEG segment data that can be directly used for feature extraction.

[0100] The method for constructing a pain model based on multi-dimensional EEG feature fusion provided in this application first collects the first multi-channel EEG signal of the corresponding brain region for chronic pain core analysis, then uses a preset time window and overlapping step size to slide the segment to ensure that the temporal features of the EEG are not lost, and simultaneously performs bandpass filtering and notch filtering in sequence to effectively filter out noise outside the frequency band and power frequency interference. Finally, abnormal signal segments with amplitude exceeding the preset threshold are discarded. This not only removes invalid data contaminated by artifacts such as electrooculography and electromyography, but also obtains high signal-to-noise ratio and standardized EEG segment data.

[0101] In one possible implementation, step S104 above includes the following steps:

[0102] Step S1041: Using the algorithm consensus screening framework, calculate the original importance score corresponding to each heterogeneous feature in the high-dimensional feature matrix.

[0103] The raw importance score can be understood as the unstandardized value of the feature importance output by each algorithm after independently evaluating a single heterogeneous feature in the high-dimensional feature matrix.

[0104] The high-dimensional feature matrix is ​​input into each algorithm in the consensus screening framework. Each algorithm independently evaluates the contribution of each four-layer heterogeneous feature in the high-dimensional feature matrix to the chronic pain model based on its own operation logic (e.g., sparse linear algorithm based on the absolute value of feature coefficients, random forest based on the reduction of node impurity), and outputs the original importance score corresponding to each feature. This step must ensure that all algorithms in the framework complete the scoring of all heterogeneous features, and the original scores retain the original output form of the algorithm without any standardization processing.

[0105] As an example, the input is a high-dimensional feature matrix with 1000 rows (samples) × 3 columns (features), where the features are the absolute power of HighGamma of Fp2 (feature A), the Delta permutation entropy of Fp2 (feature B), and the Alpha-HighGamma phase-amplitude coupling of Cz (feature C).

[0106] Each algorithm outputs the original importance score:

[0107] LASSO: Feature A = 0.75, Feature B = 0.92, Feature C = 0.88;

[0108] Random Forest Algorithm: Feature A = 15.6, Feature B = 22.8, Feature C = 19.5;

[0109] XGBoost: Feature A=88, Feature B=96, Feature C=91.

[0110] Step S1042: Determine the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix based on the original importance score.

[0111] The original importance scores output by each algorithm within the framework are standardized using Min-Max, mapping scores of different dimensions to a uniform 0-1 range. For each heterogeneous feature, the arithmetic mean of its standardized scores under all algorithms within the framework is calculated, and this mean is the consensus score for that feature. Finally, all four layers of heterogeneous features obtain a unique consensus score, with higher scores indicating greater importance of the feature to the chronic pain model.

[0112] Specifically, step S1042 includes:

[0113] Step b1: Perform maximum-minimum processing on the original importance scores output by each algorithm to calculate the target scores corresponding to each heterogeneous feature.

[0114] For the original importance scores of all heterogeneous features output by each algorithm in the algorithm consensus screening framework, a maximum-minimum processing is performed independently. First, the minimum and maximum values ​​of the original scores of all features under the algorithm are determined. Then, the standardized score (i.e. the target score) of each feature is calculated one by one according to the Min-Max formula. The standardization processes of different algorithms are independent of each other and are not cross-calculated. Finally, the target score of each feature corresponds to the number of algorithms in the framework.

[0115] In one possible implementation, the target score is determined using the following formula:

[0116] ;in, This represents the original feature importance score output by the k-th algorithm.

[0117] Step b2: Use the target score as the corresponding consensus score.

[0118] If the algorithm consensus screening framework contains only one algorithm, the target score of each feature under that algorithm is directly used as the consensus score of the corresponding feature. If the framework contains two or more algorithms, the arithmetic mean of the target scores of each feature under all algorithms must be calculated first, and this mean is the consensus score of that feature; finally, all four layers of heterogeneous features obtain a unique consensus score.

[0119] As an example, a single algorithm framework (XGBoost only):

[0120] Feature 1: Target score = 0.0, consensus score = 0.0;

[0121] Feature 2: Target score ≈ 0.82, consensus score ≈ 0.82;

[0122] Feature 3: Target score = 1.0, consensus score = 1.0;

[0123] Feature 4: Target score ≈ 0.35, consensus score ≈ 0.35;

[0124] Feature 5: Target score ≈ 0.18, consensus score ≈ 0.18.

[0125] As an example, a dual-algorithm framework: LASSO + Random Forest (RF) algorithm:

[0126] Feature A: Target score: LASSO=0.0, RF=0.0, consensus score=(0.0+0.0) / 2=0.0;

[0127] Feature B: Target score: LASSO=1.0, RF=1.0, consensus score=(1.0+1.0) / 2=1.0;

[0128] Feature C: Target score: LASSO≈0.76, RF≈0.54, consensus score = (0.76+0.54) / 2≈0.65.

[0129] The method for constructing a pain model based on multi-dimensional EEG feature fusion provided in this application calculates the original importance scores of the four heterogeneous features in the high-dimensional feature matrix through an algorithm consensus screening framework. Then, the original scores output by different algorithms are subjected to maximum-minimum processing to obtain a standardized target score. Finally, the target score is determined as the consensus score. This method avoids the feature selection bias problem that is prone to occur in single-algorithm screening by using parallel evaluation of multiple algorithms, and eliminates the difference in the dimensions of the original scores of different algorithms through maximum-minimum processing, thus achieving the standardization and unification of feature importance scores.

[0130] In one possible implementation, step S105 above includes the following steps:

[0131] Step S1051: Based on the consensus score, sort the four heterogeneous features to obtain a candidate feature list.

[0132] The candidate feature list can be understood as a feature sequence formed by arranging four heterogeneous features in descending order of consensus scores from high to low. The features in the list are the core features related to chronic pain that have been screened by algorithm consensus.

[0133] All four heterogeneous features are sorted in descending order according to their respective consensus scores (from high to low). Features with the same score can be sorted by feature type / brain region priority to form an ordered list of candidate features.

[0134] Step S1052: Based on the candidate feature list, select the target number of features in sequence to construct a feature subset.

[0135] The target number can be understood as the number of features selected from the candidate feature list, which is any positive integer between 1 and the total number of features in the candidate feature list. Different target numbers correspond to different feature subsets.

[0136] A feature subset can be understood as the set of features formed by selecting the top N features from the candidate feature list according to the target number.

[0137] Based on the candidate feature list, select the top N features of different target quantities in order from 1 to the total number of features in the list to construct multiple feature subsets (one feature subset corresponds to one target quantity). The selection rule is to select from the beginning of the list in order to ensure that the features of the feature subsets are all core features with higher priority in the candidate list.

[0138] Step S1053: Perform cross-validation on the feature subsets to determine the overfitting risk value corresponding to each feature subset.

[0139] For each constructed feature subset, 5-fold cross-validation with subjects as the sample unit was performed. The EEG subject samples corresponding to the feature subset were randomly divided into 5 equal parts. One part of the subjects was selected as the test set and the remaining 4 parts as the training set. The training and validation were repeated 5 times. The AUC of the training set and the AUC of the test set were calculated each time. The average of the 5 results was then taken to obtain the average AUC of the training set (AUC_train_mean) and the average AUC of the test set (AUC_test_mean). Among them, the EEG features of each subject were obtained by arithmetically averaging all of their EEG segment features to obtain a single feature vector, which reflects the stable neurophysiological pattern of the subject in the resting state.

[0140] Finally, the overfitting risk value (Gap) is determined using the following formula:

[0141] The overfitting risk value corresponding to each feature subset is calculated. The smaller the value, the lower the overfitting risk when the feature subset is adapted to the model.

[0142] Step S1054: Based on the overfitting risk value and the balanced scoring function, determine multiple target feature sets and the scores corresponding to the target feature sets.

[0143] ;

[0144] Where α is 0.6-0.8 and Gap_threshold is 0.1, the balanced score of each feature subset is calculated; the feature subset with effective scores (scores 0-1) calculated by the balanced scoring function is defined as the target feature set, and each target feature set corresponds to a unique balanced score.

[0145] In one possible implementation, the balanced scoring function is determined as follows:

[0146] ;in, The area under the curve for the test set. For overfitting gap, These are the weighting coefficients.

[0147] The method for constructing a pain model based on multi-dimensional EEG feature fusion provided in this application provides a candidate feature list by ranking four heterogeneous features highly correlated with chronic pain based on consensus scores. Then, feature subsets are constructed by sequentially selecting different target numbers of features. The overfitting risk value of each subset is quantified through cross-validation. Finally, the target feature set and corresponding score are determined by combining the overfitting risk value and the balanced scoring function. This method not only avoids the drawback of relying on manual experience to set the number of features in existing technologies, but also accurately quantifies the overfitting risk of each feature subset through cross-validation. Finally, the balanced scoring function is used to integrate the model performance and generalization ability for comprehensive evaluation, effectively avoiding the problems of losing key information for pain diagnosis due to too few features and causing model overfitting due to too many features.

[0148] Please refer to Figures 2 to 6 ;in, Figure 2 This is another flowchart illustrating the method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application; Figure 3 This is another flowchart illustrating the method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application. Figure 4 This is another flowchart illustrating the method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application; Figure 5 This is another flowchart illustrating the method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application. Figure 6 This is another flowchart illustrating the method for constructing a pain model based on multi-dimensional EEG feature fusion according to an embodiment of this application.

[0149] Figure 2 This demonstrates a complete technical loop from raw data acquisition to model building and finally, pain level analysis. Specifically, it can include: raw multi-channel EEG signals → EEG signal preprocessing (outputting segmented EEG data) → four-layer heterogeneous feature extraction (outputting a high-dimensional feature matrix) → algorithm consensus screening (outputting a candidate feature list) → three-dimensional optimization of feature quantity (outputting the optimal feature subset) → Stacking integrated pain model construction → input of EEG signals to be analyzed → pain level analysis results.

[0150] For the EEG signal preprocessing process, the first multi-channel EEG signal is acquired, and the continuous EEG signal is slidably segmented using a 2-second time window and a 1-second overlap step. Bandpass filtering of 0.1-100Hz and notch filtering of 50Hz are performed sequentially, and abnormal signal segments with amplitudes exceeding ±100μV are discarded. Finally, standardized EEG segmented data is obtained. The data format is a time-series numerical sequence, which can be directly used for subsequent feature extraction.

[0151] To extract four layers of heterogeneous features, a high-dimensional feature matrix is ​​constructed:

[0152] Spectrum feature module: Label the extracted indicators (absolute power, relative power, Alpha peak frequency, frequency band power ratio), and label the corresponding frequency bands (Delta1-4Hz, Theta4-8Hz, Alpha8-14Hz, Beta15-30Hz, Low Gamma30-45Hz, High Gamma55-100Hz).

[0153] Time dynamic feature module: label the core extraction indicators (time domain volatility, time domain dispersion, Hjorth parameter, permutation entropy), and label "calculated based on narrowband filtered signal";

[0154] ROI-based brain functional connectivity feature module: labeled extraction index (weighted phase lag index wPLI), labeled core brain regions (prefrontal cortex, central cortex, occipital cortex, parietal cortex, temporal cortex) and corresponding physical electrodes (Fp1, Fp2, Cz, O1, O2, etc.).

[0155] Cross-frequency coupling feature module: Labels core extraction metrics (phase-amplitude coupling, PAC), core coupling frequency bands (Alpha-HighGamma), and corresponding brain regions (central region Cz); Output: High-dimensional feature matrix (sample × feature). Within each module, core features and pathological trends of chronic pain are labeled (e.g., permutation entropy - reduced complexity, PAC - enhanced coupling). Modules are connected by dashed lines to reflect the characteristics of "heterogeneous features, mutual complementarity."

[0156] Among them, the sparse linear algorithm (LASSO regression) is labeled with its core principles (L1 regularization, extraction of strongly correlated linear features); the random forest algorithm (RF) is labeled with its core principles (Gini impurity, extraction of nonlinear features); the gradient boosting decision tree algorithm (GBDT) is labeled with its core principles (sequence optimization, node splitting gain); the extreme gradient boosting algorithm (XGBoost) is labeled with its core principles (regularization boosting, node splitting gain); the consensus calculation unit is labeled with "original importance score → Min-Max standardization → arithmetic mean → consensus score"; the output is the algorithm consensus screening framework (which can be directly used for feature importance scoring).

[0157] Regarding the determination of the consensus score: Each algorithm within the framework independently evaluates the importance of each feature in the high-dimensional feature matrix based on its own core principles, outputting a raw importance score. The raw scores of different algorithms have different dimensions and are not standardized. First, the raw importance scores output by each algorithm are processed using Min-Max to map the scores to the [0,1] interval to obtain the target score. Then, based on the number of algorithms within the framework, the target score is determined as the consensus score (single algorithm: target score = consensus score; multiple algorithms: arithmetic mean of target scores = consensus score).

[0158] Regarding the optimization steps: Four major optimization steps (circular closed loop, executed sequentially):

[0159] Feature ranking: Sort by consensus score in descending order → Output candidate feature list;

[0160] Feature subset construction: Select the target number of features sequentially → output multiple feature subsets;

[0161] Cross-validation: 5-fold cross-validation → Calculate overfitting risk value;

[0162] Balanced score calculation: Substitute into the balanced score function → output target feature set + corresponding score; Output: multiple target feature sets and their respective balanced scores.

[0163] Please refer to Figure 7 ;in, Figure 7 This is an architecture diagram of a pain model construction system based on multi-dimensional EEG feature fusion according to an embodiment of this application.

[0164] based on Figure 7 The architecture shown uses the optimal feature subset as the core input feature and "chronic pain severity (mild / moderate / severe)" as the label to construct a Stacking ensemble learning pain model. Specific training and construction requirements are as follows:

[0165] Base learner layer: XGBoost, LightGBM, and Random Forest are selected as base learners. The sample data of the optimal feature subset are input into each base learner. After independent training, each base learner outputs the classification probability value of the degree of chronic pain for the sample.

[0166] Feature concatenation: The classification probability values ​​output by the three base learners are concatenated into a new meta-feature vector, which is then used as the input to the meta-learner;

[0167] Meta-learner layer: Logistic regression or random forest is selected as the meta-learner to learn the meta-feature vector. The bias of the base learner is corrected by weighted fusion, and finally a stable and accurate model prediction result is output.

[0168] Model Validation: The model is validated using an independent test set to ensure its generalization ability. After successful validation, the final Stacking integrated pain model is obtained.

[0169] In one possible implementation, this application also provides a method for analyzing pain levels, the method comprising:

[0170] Step c1: Obtain the multidimensional EEG feature set to be processed.

[0171] The multidimensional EEG feature set to be processed can be understood as the EEG feature set that needs to be analyzed to determine the degree of pain.

[0172] Step c2: Based on a multi-dimensional EEG feature set, the pain level is determined using a pain model.

[0173] After determining the multidimensional EEG feature set to be processed, it is input into the pain model to determine the degree of pain.

[0174] This embodiment also provides a device for constructing a pain model based on multi-dimensional EEG feature fusion. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0175] This embodiment provides a device for constructing a pain model based on the fusion of multi-dimensional electroencephalogram (EEG) features, such as... Figure 8 As shown, the system includes: an acquisition module 801 for acquiring segmented EEG data; a first construction module 802 for extracting four-layer heterogeneous features from the segmented EEG data and constructing a high-dimensional feature matrix, wherein the four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features, wherein the temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling; a second construction module 803 for constructing an algorithm consensus screening framework, wherein the algorithm consensus screening framework includes at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm, and extreme gradient boosting algorithm; a first determination module 804 for using the algorithm consensus screening framework to determine the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix; a second determination module 805 for determining multiple target feature sets and their corresponding scores based on the consensus scores and the four-layer heterogeneous features; a third determination module 806 for determining the highest-scoring feature subset from multiple target feature sets based on the scores of the multiple target feature sets; and a third construction module 807 for constructing a pain model based on the feature subset.

[0176] This embodiment provides a pain intensity analysis device, which includes: a feature set acquisition module for acquiring a multi-dimensional EEG feature set to be processed;

[0177] The pain intensity determination module is used to determine the pain intensity based on a multi-dimensional EEG feature set and a pain model. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0178] In this embodiment, the device for constructing a pain model based on multi-dimensional EEG feature fusion and / or the device for determining the degree of pain are presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0179] This application also provides a computer device having the above-described features. Figure 9 The device shown is for constructing a pain model based on the fusion of multidimensional EEG features.

[0180] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of this application, such as... Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0181] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0182] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0183] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0184] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0185] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 9 Taking the example of a connection between China and Israel via a bus.

[0186] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0187] The computer device also includes a communication interface for communicating with other devices or communication networks.

[0188] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0189] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0190] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for constructing a pain model based on multi-dimensional EEG feature fusion, characterized in that, The method includes: Acquire segmented EEG data; Four-layer heterogeneous features are extracted from the EEG segmentation data to construct a high-dimensional feature matrix. The four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling. An algorithm consensus screening framework is constructed, which includes at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm, and extreme gradient boosting algorithm. Using the aforementioned algorithm consensus screening framework, the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix is ​​determined; Based on the consensus score and the four-layer heterogeneous features, multiple target feature sets and the scores corresponding to the target feature sets are determined. Based on the multiple target feature sets and the scores corresponding to the target feature sets, determine the feature subset with the highest score from the multiple target feature sets; The pain model is constructed based on the aforementioned feature subset.

2. The method for constructing a pain model based on multi-dimensional EEG feature fusion according to claim 1, characterized in that, The acquisition of EEG segment data includes: Acquire the first multichannel EEG signal; The first multi-channel EEG signal is segmented by using a preset time window and overlap step size, and then bandpass filtering and notch filtering are performed sequentially to obtain the second multi-channel EEG signal. The second multi-channel EEG signal with an amplitude exceeding the preset amplitude is discarded to obtain the EEG segmentation data.

3. The method for constructing a pain model based on multi-dimensional EEG feature fusion according to claim 1, characterized in that, Using the aforementioned algorithmic consensus screening framework, the consensus score for the four-layer heterogeneous features corresponding to the high-dimensional feature matrix is ​​determined, including: Using the consensus screening framework of the algorithm, the original importance scores corresponding to each heterogeneous feature in the high-dimensional feature matrix are calculated respectively; The consensus score of the four heterogeneous features corresponding to the high-dimensional feature matrix is ​​determined based on the original importance score.

4. The method for constructing a pain model based on multi-dimensional EEG feature fusion according to claim 3, characterized in that, The consensus score for determining the four-layer heterogeneous features corresponding to the high-dimensional feature matrix based on the original importance score includes: The original importance scores output by each algorithm are subjected to maximum-min processing to calculate the target scores corresponding to each heterogeneous feature. The target score is used as the corresponding consensus score.

5. The method for constructing a pain model based on multi-dimensional EEG feature fusion according to claim 1, characterized in that, Based on the consensus score and the four-layer heterogeneous features, multiple target feature sets and their corresponding scores are determined, including: Based on the consensus score, the four heterogeneous features are sorted to obtain a candidate feature list; Based on the candidate feature list, a feature subset is constructed by sequentially selecting a target number of features. Cross-validation is performed on the feature subsets to determine the overfitting risk value corresponding to each feature subset; Based on the overfitting risk value and the balanced scoring function, multiple target feature sets and their corresponding scores are determined.

6. A method for analyzing pain intensity, characterized in that, The method includes: Acquire the multidimensional EEG feature set to be processed; Based on the multi-dimensional EEG feature set, the pain level is determined using a pain model; wherein the pain model is obtained by the method for constructing a pain model based on multi-dimensional EEG feature fusion as described in any one of claims 1-5.

7. A device for constructing a pain model based on multi-dimensional EEG feature fusion, characterized in that, The device includes: The acquisition module is used to acquire EEG segmentation data; The first construction module is used to extract four-layer heterogeneous features from the EEG segmentation data and construct a high-dimensional feature matrix. The four-layer heterogeneous features include spectral features, temporal dynamic features, brain functional connectivity features based on regions of interest, and cross-frequency coupling features. The temporal dynamic features include permutation entropy, and the cross-frequency coupling features include phase-amplitude coupling. The second building module is used to build an algorithm consensus screening framework, which includes at least one of the following: sparse linear algorithm, random forest algorithm, gradient boosting decision tree algorithm and extreme gradient boosting algorithm. The first determining module is used to determine the consensus score of the four-layer heterogeneous features corresponding to the high-dimensional feature matrix using the algorithm consensus screening framework. The second determining module is used to determine multiple target feature sets and the scores corresponding to the target feature sets based on the consensus score and the four-layer heterogeneous features; The third determining module is used to determine the feature subset with the highest score from the multiple target feature sets based on the multiple target feature sets and the scores corresponding to the target feature sets; The third construction module is used to construct the pain model based on the feature subset.

8. An electronic device, characterized in that, include: A processor and a memory connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to cause the processor to perform the method for constructing a pain model based on multidimensional EEG feature fusion as described in any one of claims 1-5 and / or the method for analyzing pain intensity as described in claim 6.

9. A non-transitory computer-readable storage medium storing computer instructions, the computer instructions being configured to cause the computer to execute the method for constructing a pain model based on multidimensional EEG feature fusion as described in any one of claims 1-5 and / or the method for analyzing pain intensity as described in claim 6.

10. A computer program product comprising a computer program that, when executed by a processor, implements the method for constructing a pain model based on multidimensional EEG feature fusion according to any one of claims 1-5 and / or the method for analyzing pain intensity according to claim 6.

Images