Method and device for identifying brain regions related to wakefulness and awareness based on EEG feature attribution

Abstract

The present application discloses a method and device for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution. The method includes: obtaining EEG datasets under different wakefulness and awareness states; performing preprocessing and feature extraction to obtain an EEG feature dataset; constructing a classification model, training the classification model using the EEG feature dataset and the corresponding state label to obtain a trained wakefulness state classification model and an awareness state classification model; obtaining target EEG data, and performing preprocessing and feature extraction to obtain target EEG features; grouping to obtain multiple groups of brain region feature groups, calculating the importance between each group of brain region feature groups and the wakefulness state classification model or the awareness state classification model through a feature attribution algorithm, and determining the brain region corresponding to the brain region feature group as a brain region related to the wakefulness state or the awareness state based on the importance. The present application can not only distinguish the wakefulness and awareness states of the subjects, but also identify the brain regions related to the wakefulness and awareness states.

Description

Technical Field

[0001] The present application relates to the technical field of identifying brain regions related to wakefulness and awareness states, and in particular to a method and device for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution. Background Art

[0002] In the clinical field of disorders of consciousness, consciousness is typically assessed across two dimensions: wakefulness and awareness. Wakefulness refers to an individual's state of being awake and responsive to external stimuli, while awareness refers to an individual's awareness of their environment and self. Analyzing a patient's wakefulness and awareness states and their correlation with brain activity is crucial. Identifying brain regions associated with wakefulness and awareness is one area of ​​analysis in this field. This identification can aid in designing treatment plans for disorders of consciousness that target these areas for stimulation and can also reduce the cost of EEG analysis by minimizing signal acquisition from irrelevant brain regions.

[0003] Existing methods can identify EEG features that are highly correlated with consciousness disorders and record the corresponding brain regions. The technical path is: (1) EEG acquisition and preprocessing. EEG signals of subjects in different consciousness states are collected and preprocessed such as artifact removal; (2) Feature extraction. Features are extracted from EEG signals through algorithms such as time-frequency analysis and convolutional networks. (3) Correlation analysis. Various correlation coefficients between EEG features and the severity of consciousness disorders are calculated, and finally highly correlated features are found and their corresponding brain regions are recorded. Existing methods under this path are limited to the single data domain of consciousness disorders, and the relevant brain regions of the wakefulness and awareness states are not identified separately during analysis. Summary of the Invention

[0004] The embodiments of the present application provide a method and device for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution, which can not only distinguish the subject's wakefulness and awareness states, but also identify the brain regions related to the wakefulness and awareness states.

[0005] The present application provides a method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution, including:

[0006] Obtain EEG datasets in different states of wakefulness and awareness;

[0007] Preprocessing and feature extraction are performed on the EEG data in the EEG dataset to obtain an EEG feature dataset;

[0008] Constructing a classification model, and training the classification model using the EEG feature dataset and the state labels corresponding to the EEG feature dataset to obtain a trained wakefulness state classification model and an awareness state classification model;

[0009] Acquiring target EEG data, and performing preprocessing and feature extraction on the target EEG data to obtain target EEG features;

[0010] The target EEG features are grouped to obtain multiple brain region feature groups, and the importance between each brain region feature group and the wakefulness state classification model or each brain region feature group and the awareness state classification model is calculated through a feature attribution algorithm. Based on the importance, it is determined that the brain region corresponding to the brain region feature group is a brain region related to the wakefulness state or a brain region related to the awareness state.

[0011] Furthermore, in the above-mentioned method for identifying brain regions related to the awakening state based on EEG feature attribution, the preprocessing and feature extraction of the EEG data in the EEG dataset includes:

[0012] performing a channel average reference operation, a downsampling operation, and a bandpass filtering operation on the EEG data in sequence to obtain noise-reduced EEG data;

[0013] Segmenting each EEG signal in the EEG data into a plurality of signal segments of preset time lengths;

[0014] A plurality of signal features are extracted from each channel time series signal in the signal segment and the corresponding signal features are concatenated to obtain an EEG feature data set.

[0015] Furthermore, in the above-mentioned method for identifying brain regions related to the state of awakening and awareness based on EEG feature attribution, the step of grouping the target EEG features to obtain multiple groups of brain region feature groups includes:

[0016] Determining the channel corresponding to the target EEG feature;

[0017] The target EEG features are grouped according to the brain regions where the channels are located to obtain multiple brain region feature groups.

[0018] Furthermore, the above-mentioned method for identifying brain regions related to the awakening state based on EEG feature attribution, wherein the importance between each group of brain region feature groups and the awakening state classification model or each group of brain region feature groups and the awareness state classification model is calculated by a feature attribution algorithm, and based on the importance, the brain region corresponding to the brain region feature group is determined as a brain region related to the awakening state or a brain region related to the awareness state, including:

[0019] Replacing the eigenvalues ​​in any brain region feature group with random sampling of the eigenvalues ​​to obtain a marginal distribution sample set after removing the brain region feature group;

[0020] removing each feature group in turn according to each order of the full permutation, and calculating the marginal contribution of the feature group according to the classification accuracy of the arousal state classification model or the awareness state classification model during the removal process;

[0021] A Shapley contribution value is calculated according to the marginal contribution, and the Shapley contribution value is determined as the importance.

[0022] Furthermore, the above-mentioned method for identifying brain regions related to the awakening state based on EEG feature attribution further includes: calculating the importance by a first formula, wherein the first formula is:

[0023]

[0024] in, For importance, is the brain region feature group, D is the set of all brain region feature groups, is the number of brain region feature groups, To remove feature groups The marginal distribution sample set after , is the classification accuracy of the wakefulness state classification model or the awareness state classification model.

[0025] Furthermore, the above-mentioned method for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution, wherein the classification model is trained using the EEG feature dataset and the state labels corresponding to the EEG feature dataset to obtain the trained wakefulness state classification model and awareness state classification model, includes:

[0026] Dividing the EEG feature dataset into an awakening state dataset and an awareness state dataset;

[0027] Inputting the wakefulness state dataset into the classification model to obtain a predicted first classification result, calculating a loss function based on the first classification result and the wakefulness state dataset, setting a learning rate, training generations, and an optimizer, and iteratively training the classification model based on the loss function to obtain a trained wakefulness state classification model;

[0028] The awareness state dataset is input into the classification model to obtain a predicted second classification result. The loss function is calculated based on the second classification result and the awareness state dataset, the learning rate, training generations and optimizer are set, and the classification model is iteratively trained based on the loss function to obtain a trained awareness state classification model.

[0029] Furthermore, the above-mentioned method for identifying brain regions related to the state of wakefulness and awareness based on EEG feature attribution further includes:

[0030] After training the wakefulness state classification model and the awareness state classification model using the training set and the test set, the model classification performance is evaluated using the test set, and the classification accuracy of the wakefulness state classification model and the awareness state classification model is obtained based on the model classification performance results.

[0031] Furthermore, in the above-mentioned method for identifying brain areas related to wakefulness and awareness states based on EEG feature attribution, the different wakefulness and awareness states include at least the resting state, anesthesia state and resting state of subjects with chronic consciousness disorders of healthy subjects.

[0032] Furthermore, in the above-mentioned method for identifying brain areas related to the awakening and awareness state based on EEG feature attribution, the multiple signal features include at least permutation entropy, spectral entropy, fractal dimension and frequency band power.

[0033] The present application also provides a device for identifying brain regions related to wakefulness and awareness based on EEG feature attribution, including:

[0034] An acquisition module is used to obtain EEG data sets under different states of wakefulness and awareness;

[0035] A preprocessing and feature extraction module, configured to perform preprocessing and feature extraction on the EEG dataset to obtain an EEG feature dataset;

[0036] A classification module is used to construct a classification model, and train the classification model using the EEG feature data set and the state labels corresponding to the EEG feature data set to obtain a trained wakefulness state classification model and an awareness state classification model;

[0037] The acquisition module is also used to acquire target EEG data, and perform preprocessing and feature extraction on the target EEG data to obtain target EEG features;

[0038] An identification module is used to group the target EEG features to obtain multiple brain region feature groups, calculate the importance between each brain region feature group and the wakefulness state classification model or each brain region feature group and the awareness state classification model through a feature attribution algorithm, and determine based on the importance whether the brain region corresponding to the brain region feature group is a brain region related to the wakefulness state or a brain region related to the awareness state.

[0039] Identification of brain regions related to wakefulness and awareness states The present application provides a method and device for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution. The present application obtains a wakefulness state classification model and a awareness state classification model through training, and then extracts and groups the collected target EEG data. The importance between each group of brain region feature groups and the wakefulness state classification model or each group of brain region feature groups and the awareness state classification model is calculated through a feature attribution algorithm. Based on the importance, the brain region corresponding to the brain region feature group is determined to be a brain region related to the wakefulness state or a brain region related to the awareness state. The present application can not only distinguish between the wakefulness and awareness states of the subjects, but also identify the brain regions related to the wakefulness state and the awareness state. The identification of related brain regions can help the design of diagnostic and treatment equipment related to anesthesia or consciousness disorders. Secondly, in the process of feature extraction of brain region data, nonlinear correlation and the interaction between multiple features are taken into account, which can improve the accuracy of identifying brain regions related to wakefulness and awareness states. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The following detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings will make the technical solutions and other beneficial effects of the present application apparent.

[0041] Figure 1 Flowchart of a method for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution provided in an embodiment of the present application.

[0042] Figure 2 Another flowchart of the method for identifying brain areas related to the awakening state based on EEG feature attribution provided in an embodiment of the present application.

[0043] Figure 3 A schematic diagram of the structure of a device for identifying brain areas related to wakefulness and awareness states based on EEG feature attribution provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0045] Existing methods can identify EEG features that are highly correlated with consciousness disorders and record the corresponding brain regions. The technical path is: (1) EEG acquisition and preprocessing. EEG signals of subjects in different consciousness states are collected and preprocessed such as artifact removal; (2) Feature extraction. Features are extracted from EEG signals through algorithms such as time-frequency analysis and convolutional networks. (3) Correlation analysis. Various correlation coefficients between EEG features and the severity of consciousness disorders are calculated, and finally highly correlated features are found and their corresponding brain regions are recorded. Existing methods under this path have three shortcomings. First, they are limited to the single data domain of consciousness disorders. During the analysis, the relevant brain regions of wakefulness and awareness states are not identified separately. Second, there is a lack of analysis of nonlinear correlations. Many commonly used correlation coefficients can only measure linear correlations between variables. Third, the interaction between multiple features is not considered. Existing methods analyze each feature separately and do not consider the multivariate association between multiple features in the same brain region and consciousness states.

[0046] Existing methods do not decouple the two dimensions of wakefulness and awareness for separate analysis, and ignore nonlinear correlations and interactions between multiple features. To address the above issues, the present invention provides a method and device for identifying brain regions related to wakefulness and awareness based on EEG feature attribution.

[0047] See also Figure 1 and Figure 2 , Figure 1 This is a flowchart of a method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution provided in an embodiment of the present application. Figure 2 Another flowchart of a method for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution provided in an embodiment of the present application, which is applied to an electronic device, includes the following steps:

[0048] S1, obtain EEG data sets under different states of wakefulness and awareness.

[0049] EEG signal acquisition involves attaching sensors to the surface of the human scalp, recording data while the subject is at rest or under anesthesia, and transmitting it to a computer for real-time storage. Before collecting EEG data, the EEG acquisition electrodes, a reference electrode, and a ground wire are threaded through a headgear to securely fit the subject's scalp. Conductive paste is then applied to reduce resistivity.

[0050] Among them, the subject states that need to be collected (that is, different states of wakefulness and awareness) include the resting state, anesthesia state of healthy subjects and the resting state of subjects with chronic consciousness disorders. The health and anesthesia conditions of the subjects are recorded at the same time.

[0051] S2, preprocessing and feature extraction are performed on the EEG data in the EEG dataset to obtain an EEG feature dataset.

[0052] In one embodiment, step S2 includes the following steps:

[0053] S21, performing channel average reference operation, downsampling operation and bandpass filtering operation on the EEG data in sequence to obtain the EEG data after noise reduction.

[0054] Specifically, the EEG data in the EEG dataset were subjected to a channel average reference operation, and then the EEG data were downsampled to 250 Hz. The non-EEG frequency bands were then filtered out using a band-pass filtering algorithm, leaving only the 0.5-45 Hz portion.

[0055] S22, dividing each EEG signal in the EEG data into a plurality of signal segments of preset time lengths.

[0056] Each EEG signal (the complete EEG record of each subject) was divided into 10S signal segments.

[0057] S23, extracting multiple signal features from each channel time series signal in the signal segment and concatenating the corresponding signal features to obtain an EEG feature vector. The set of all EEG feature vectors is an EEG feature dataset.

[0058] For each channel time series signal of each EEG segment, EEG features are extracted. Finally, the EEG feature vectors of each EEG segment are concatenated and the collection of all EEG feature vectors is the EEG feature dataset.

[0059] Among them, the multiple signal features include at least permutation entropy, spectrum entropy, fractal dimension, frequency band power and other features.

[0060] In one embodiment, it is also necessary to label the EEG data in the EEG dataset with state labels. The specific steps are as follows:

[0061] S24, marking the subject state corresponding to each group of EEG feature vectors according to the state record during acquisition.

[0062] S25, according to the rule that "the healthy resting state is high arousal and high awareness, the healthy anesthesia state is low arousal and low awareness, and the consciousness disorder resting state is high arousal and low awareness", the arousal and awareness states corresponding to each group of eigenvectors are marked by the subject's state.

[0063] S26, divide the dataset into training set, validation set and test set in a ratio of 3:1:1. The training set and validation set are used for model training and optimization, and the test set is used for model testing.

[0064] S3, building a classification model, training the classification model through the EEG feature dataset and the state labels corresponding to the EEG feature dataset, and obtaining a trained wakefulness state classification model and an awareness state classification model.

[0065] The classification model is based on an adaptive network and consists of a four-layer network structure. The first layer is the input layer, which receives the EEG feature vector. The middle two layers are hidden layers, and the last layer is the output layer. The output dimension of the output layer is 2×1, corresponding to the high or low state of consciousness.

[0066] Step S3 trains the classification model using the EEG feature dataset and the state labels corresponding to the EEG feature dataset to obtain a trained wakefulness state classification model and an awareness state classification model, specifically including the following steps:

[0067] S31, dividing the EEG feature dataset into an awakening state dataset and an awareness state dataset.

[0068] Specifically, the EEG feature dataset is divided into an awakening state dataset and an awareness state dataset according to the state labels corresponding to the EEG feature dataset.

[0069] S32: Input the wakefulness state dataset into the classification model to obtain a predicted first classification result, calculate the loss function based on the first classification result and the wakefulness state dataset, set the learning rate, training algebra, and optimizer, iteratively train the classification model based on the loss function, and obtain a trained wakefulness state classification model.

[0070] S33, input the awareness state dataset into the classification model to obtain the predicted second classification result, calculate the loss function based on the second classification result and the awareness state dataset, set the learning rate, training algebra and optimizer, iteratively train the classification model based on the loss function, and obtain a trained awareness state classification model.

[0071] Among them, the awakening state dataset and the awareness state dataset belong to the training set and validation set.

[0072] The classification model must be able to distinguish between high and low consciousness but confuse unconsciousness caused by anesthesia and impaired consciousness. The loss function for classification performance is the cross-entropy loss calculated between the classification model output and the state label. The loss function for confusion performance is the adaptation loss calculated between the hidden layer feature representation and the classification model output after inputting a pair of data from the anesthesia and impaired consciousness datasets. The sum of these two losses is the loss function for classification model training.

[0073] After training the wakefulness and awareness classification models using the training and test sets, the test set was used to evaluate the model classification performance, including test accuracy, model specificity, and model sensitivity. The classification accuracy of the wakefulness and awareness classification models was calculated using test accuracy, model specificity, and model sensitivity.

[0074] S4, obtaining target EEG data, and performing preprocessing and feature extraction on the target EEG data to obtain target EEG features.

[0075] It should be explained that steps S1-S3 are the process of model training, and steps S4-S5 are the process of model application. Among them, the target EEG data is the EEG data to be identified. The process of preprocessing and feature extraction of the target EEG data to obtain the target EEG features is similar to the process of preprocessing and feature extraction of the EEG data to obtain the EEG feature dataset in step S2, and will not be described in detail here.

[0076] S5. Group the target EEG features to obtain multiple brain region feature groups. Calculate the importance between each brain region feature group and the wakefulness state classification model or each brain region feature group and the awareness state classification model through a feature attribution algorithm. Based on the importance, determine whether the brain region corresponding to the brain region feature group is a brain region related to the wakefulness state or a brain region related to the awareness state.

[0077] In one embodiment, grouping the target EEG features to obtain multiple brain region feature groups in step S5 includes the following steps:

[0078] S51, determining the channel corresponding to the target EEG feature;

[0079] S52: Group the target EEG features according to the brain regions where the channels are located to obtain multiple brain region feature groups.

[0080] Since each EEG feature is extracted from the EEG data of a certain channel, the brain area where this channel is located is the corresponding brain area of ​​this target EEG feature.

[0081] In step S5, calculating the importance between each group of brain region feature groups and the arousal state classification model or each group of brain region feature groups and the awareness state classification model through a feature attribution algorithm, and determining the brain region corresponding to the brain region feature group as a brain region related to the arousal state or a brain region related to the awareness state based on the importance includes the following steps:

[0082] S51, replacing the eigenvalues ​​in any brain region feature group with random sampling of the eigenvalues ​​to obtain a marginal distribution sample set after removing the brain region feature group.

[0083] S52 , removing each feature group in turn according to each order of the full arrangement, and calculating the marginal contribution of the feature group according to the classification accuracy of the arousal state classification model or the awareness state classification model during the removal process.

[0084] S53, calculate the Shapley contribution value based on the marginal contribution, and determine the Shapley contribution value as the importance.

[0085] Specifically, the set of all brain region feature groups is recorded as , a certain brain region feature group is recorded as . The eigenvalues ​​of the brain region feature group are replaced by random sampling of the corresponding eigenvalues, which can be removed. Sample set of marginal distribution after brain region feature grouping . and remove Compared with the previous classification model, the classification accuracy The descent on the marginal distribution sample set is The marginal contribution to classification. Each brain region feature group can be removed in turn according to each order of the full permutation, and the marginal contribution of the brain region feature group in the removal process can be finally obtained. Shapley contribution value, that is, the SAGE (attribution algorithm) importance of brain region feature groups . It can be expressed by the following formula:

[0086]

[0087] in, Importance of SAGE, is the brain region feature group, D is the set of all brain region feature groups, is the number of brain region feature groups, To remove feature groups The marginal distribution sample set after , It is the classification accuracy of the wakefulness state classification model or the awareness state classification model.

[0088] The following example illustrates the implementation steps of the method for identifying brain regions related to the state of wakefulness and awareness based on EEG feature attribution:

[0089] 1. Data Collection. First, electrodes are attached to designated locations on the subject's scalp, and conductive paste is applied to reduce the impedance between the electrodes and the scalp. After confirming that all impedance tests are qualified, the EEG device is activated to begin recording. For resting-state data collection, the subject is required to sit quietly for several minutes without task interference. For anesthesia data collection, a professional anesthesiologist will deeply anesthetize the subject, and the time it takes for the subject to enter and recover from anesthesia will be recorded.

[0090] 2. Data preprocessing and feature extraction. The default frequency of the collected EEG data is 1000Hz. The sampling rate is adjusted to 250Hz using the downsampling method, and 20 channels are selected. Next, a bandpass filtering algorithm is used to filter out noise between 0-0.5Hz and above 45Hz. The continuous recording is divided into signal segments of 10 seconds. On each channel of each signal segment, its permutation entropy, sample entropy, and fractal dimension are calculated; a discrete Fourier transform is performed to convert the channel signal to the frequency domain, and its spectral entropy is calculated; and the power of each frequency band is calculated based on the brain rhythm frequency band. Finally, the 1176 features on each channel of each EEG segment are spliced ​​together as the EEG feature vector of this signal segment.

[0091] 3. Labeling of wakefulness and awareness states. Each signal segment records the specific information of the subject, including whether the subject has a chronic disorder of consciousness and whether the subject is in an anesthetized state. According to common experimental results in clinical practice, it can be considered that: healthy resting subjects are in a high-awakening and high-awareness state, healthy anesthetized subjects are in a low-awakening and low-awareness state, and consciousness-disordered resting subjects are in a high-awakening and low-awareness state. According to the above rules, each signal segment can be labeled whether it is in a low-awareness state and whether it is in a low-awakening state. By splicing the EEG feature set with the state labeling, we can obtain the wakefulness state dataset and the awareness state dataset. The dataset is divided according to the subjects, with the dataset of one-fifth of the subjects as the validation set and the dataset of the other one-fifth of the subjects as the test set.

[0092] 4. Model construction and training. The input of the classification model is a 1176×1 EEG feature vector. The sizes of the intermediate representations of the two hidden layers are 64×1 and 16×1 respectively. The output of the classification model is a 2×1 array. Each value is between 0 and 1 and the sum is 1, representing the probability of the state being high and low respectively. During training, a pair of data from different data sets is input each time, and the adaptation loss is calculated based on the hidden layer and output layer vectors of the two. The specific model structure can be adjusted and optimized based on experience and verification accuracy. The following principles can be followed: (1) The model needs to have a hidden layer for transfer learning to align the intermediate representation; (2) Considering the stability of the feature attribution results, a dropout layer is required to make the model use as many features as possible, and the dropout probability is preferably 0.5 to make the possible network structure as diverse as possible. Training parameter requirements: (1) The optimizer recommends the ADAM optimizer that can dynamically adjust the learning rate. In order to ensure the stability of the feature attribution results, a larger weight_decay needs to be set so that the model can use as many features as possible; (2) The training cycle is 100, and it is set to automatically stop early based on the verification error and save the optimal parameters. (3) After model training and optimization, the final performance of the model can be evaluated using a test dataset, including test accuracy, model specificity, and model sensitivity.

[0093] 5. Calculation of the importance of brain region feature groups and identification of related brain regions. Each brain region feature is extracted from the signal of a certain channel, and the prefix of the channel name in the 10-20 system indicates its region. Therefore, the channel prefix can be used to divide the features into 8 groups: F, CP, P, FC, O, TP, C, and T. To save computing time, the importance of feature groups can be calculated using the SAGE algorithm library implemented in Python, and the Shapley value can be estimated by random sampling. By inputting the grouping information, target EEG features, and the wakefulness state classification model or the awareness state classification model into the SAGE algorithm, the importance of each feature group and its error range can be obtained. Different brain regions are sorted according to the importance of the corresponding feature group. The top k brain regions in the sort are the k brain regions identified to be most relevant to the wakefulness or awareness state.

[0094] The present application obtains a wakefulness state classification model and a consciousness state classification model through training, and then extracts and groups the collected target EEG data. The importance between each group of brain region feature groups and the wakefulness state classification model or each group of brain region feature groups and the consciousness state classification model is calculated through a feature attribution algorithm. Based on the importance, the brain region corresponding to the brain region feature group is determined to be a brain region related to the wakefulness state or a brain region related to the consciousness state. The present application can not only distinguish the wakefulness and consciousness states of the subjects, but also identify the relevant brain regions of the wakefulness state and the consciousness state. The identification of relevant brain regions can provide assistance in the design of relevant diagnostic and treatment equipment in the field of anesthesia or consciousness disorders. Secondly, in the process of feature extraction of brain region data, nonlinear correlation and the interaction between multiple features are taken into account, which can improve the accuracy of identifying brain regions related to the wakefulness and consciousness states. Finally, in the training process of the classification model, the EEG data sets of anesthesia and consciousness disorders are used for training at the same time, which can realize the classification of the wakefulness state and the consciousness state separately.

[0095] According to the method described in the above embodiment, this embodiment will be further described from the perspective of a device for identifying brain areas related to the state of awakening awareness based on EEG feature attribution. The device for identifying brain areas related to the state of awakening awareness based on EEG feature attribution can be implemented as an independent entity or integrated into an electronic device, which can be a terminal, server or other device, wherein the terminal can include a tablet computer, a laptop computer, a personal computer (PC), a micro processing box, or other devices.

[0096] See also Figure 3 , Figure 3 The present invention specifically describes an apparatus for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution, which is applied to electronic devices. The apparatus for identifying brain regions related to wakefulness and awareness states based on EEG feature attribution may include:

[0097] An acquisition module is used to obtain EEG data sets under different states of wakefulness and awareness;

[0098] A preprocessing and feature extraction module, configured to preprocess and extract features from the EEG data in the EEG dataset to obtain an EEG feature dataset;

[0099] A classification module is used to construct a classification model, and train the classification model using the EEG feature data set and the state labels corresponding to the EEG feature data set to obtain a trained wakefulness state classification model and an awareness state classification model;

[0100] The acquisition module is also used to acquire target EEG data, and perform preprocessing and feature extraction on the target EEG data to obtain target EEG features;

[0101] An identification module is used to group the target EEG features to obtain multiple brain region feature groups, calculate the importance between each brain region feature group and the wakefulness state classification model or each brain region feature group and the awareness state classification model through a feature attribution algorithm, and determine based on the importance whether the brain region corresponding to the brain region feature group is a brain region related to the wakefulness state or a brain region related to the awareness state.

[0102] During specific implementation, the above modules and / or units can be implemented as independent entities, or can be arbitrarily combined to be implemented as the same or several entities. The specific implementation of the above modules and / or units can refer to the previous method embodiments. The specific beneficial effects that can be achieved can also be found in the beneficial effects in the previous method embodiments, which will not be repeated here.

Claims

1. A method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution, characterized in that: The method comprises: Obtain EEG datasets in different states of wakefulness and awareness; Preprocessing and feature extraction are performed on the EEG data in the EEG dataset to obtain an EEG feature dataset; Constructing a classification model, and training the classification model using the EEG feature dataset and the state labels corresponding to the EEG feature dataset to obtain a trained wakefulness state classification model and an awareness state classification model; Acquiring target EEG data, and performing preprocessing and feature extraction on the target EEG data to obtain target EEG features; The target EEG features are grouped to obtain multiple brain region feature groups, and the importance between each brain region feature group and the arousal state classification model or each brain region feature group and the awareness state classification model is calculated using a feature attribution algorithm, and based on the importance, the brain region corresponding to the brain region feature group is determined to be a brain region related to the arousal state or a brain region related to the awareness state; Among them, the importance between each group of brain region feature groups and the wakefulness state classification model or each group of brain region feature groups and the awareness state classification model is calculated by a feature attribution algorithm, including: replacing the characteristic values ​​in any brain region feature group with random sampling of the characteristic values ​​to obtain a marginal distribution sample set after removing the brain region feature group; removing each feature group in turn according to each order of the full arrangement, and calculating the marginal contribution of the feature group according to the classification accuracy of the wakefulness state classification model or the awareness state classification model during the removal process; obtaining a Shapley contribution value based on the marginal contribution calculation, and determining the Shapley contribution value as the importance.

2. The method for identifying brain regions related to the state of awakening and awareness based on EEG feature attribution according to claim 1, characterized in that: The preprocessing and feature extraction of the EEG data in the EEG dataset includes: performing a channel average reference operation, a downsampling operation, and a bandpass filtering operation on the EEG data in sequence to obtain noise-reduced EEG data; Segmenting each EEG signal in the EEG data into a plurality of signal segments of preset time lengths; A plurality of signal features are extracted from each channel time series signal in the signal segment and the corresponding signal features are concatenated to obtain an EEG feature data set.

3. The method for identifying brain regions related to the state of wakefulness and awareness based on EEG feature attribution according to claim 2, characterized in that: The grouping of the target EEG features to obtain multiple brain region feature groups includes: Determining the channel corresponding to the target EEG feature; The target EEG features are grouped according to the brain regions where the channels are located to obtain multiple brain region feature groups.

4. The method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 1, characterized in that: The method further includes calculating the importance using a first formula, where the first formula is: in, For importance, is the brain region feature group, D is the set of all brain region feature groups, is the number of brain region feature groups, To remove feature groups The marginal distribution sample set after , is the classification accuracy of the wakefulness state classification model or the awareness state classification model.

5. The method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 1, characterized in that: The classification model is trained using the EEG feature dataset and the state labels corresponding to the EEG feature dataset to obtain a trained arousal state classification model and an awareness state classification model, including: Dividing the EEG feature dataset into an awakening state dataset and an awareness state dataset; Inputting the wakefulness state dataset into the classification model to obtain a predicted first classification result, calculating a loss function based on the first classification result and the wakefulness state dataset, setting a learning rate, training generations, and an optimizer, and iteratively training the classification model based on the loss function to obtain a trained wakefulness state classification model; The awareness state dataset is input into the classification model to obtain a predicted second classification result. The loss function is calculated based on the second classification result and the awareness state dataset, the learning rate, training generations and optimizer are set, and the classification model is iteratively trained based on the loss function to obtain a trained awareness state classification model.

6. The method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 1, characterized in that: The method further comprises: After training the wakefulness state classification model and the awareness state classification model using the training set and the test set, the model classification performance is evaluated using the test set, and the classification accuracy of the wakefulness state classification model and the awareness state classification model is obtained based on the model classification performance results.

7. The method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 1, characterized in that: The different states of awakening awareness include at least a resting state of healthy subjects, an anesthesia state, and a resting state of subjects with chronic consciousness disorders.

8. The method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 2, characterized in that: The plurality of signal features include at least permutation entropy, spectrum entropy, fractal dimension and frequency band power.

9. A device for identifying brain regions related to wakefulness and awareness based on EEG feature attribution, wherein the device is used to implement the method for identifying brain regions related to wakefulness and awareness based on EEG feature attribution according to claim 1, characterized in that: include: An acquisition module is used to obtain EEG data sets under different states of wakefulness and awareness; A preprocessing and feature extraction module, configured to preprocess and extract features from the EEG data in the EEG dataset to obtain an EEG feature dataset; A classification module is used to construct a classification model, and train the classification model using the EEG feature data set and the state labels corresponding to the EEG feature data set to obtain a trained wakefulness state classification model and an awareness state classification model; The acquisition module is also used to acquire target EEG data, and perform preprocessing and feature extraction on the target EEG data to obtain target EEG features; An identification module is used to group the target EEG features to obtain multiple brain region feature groups, calculate the importance between each brain region feature group and the wakefulness state classification model or each brain region feature group and the awareness state classification model through a feature attribution algorithm, and determine based on the importance whether the brain region corresponding to the brain region feature group is a brain region related to the wakefulness state or a brain region related to the awareness state.

Images