Consciousness disorder electroencephalogram discrimination method and system adaptive to channel deficiency

Abstract

The invention discloses a disturbance of consciousness electroencephalogram discrimination method and system adaptive to channel deletion. According to the method, on the basis of an electroencephalogram classification model of a deep network, automatic discrimination of disturbance of consciousness (MCS and UWS) is carried out with high accuracy according to resting-state electroencephalogram energy of a patient. The model utilizes a time domain double-branch CNN module and a space-time Transform module to learn features in electroencephalogram data hierarchically; and a model architecture with variable channel dimensions is introduced, an inter-channel association learning method fusing electrode position information and a training strategy of channel random discarding are introduced, so that the model can adapt to a scene in which part of channels are missing, and the clinical applicability of the method is improved.

Description

Technical Field

[0001] The present invention relates to a method for discriminating electroencephalogram (EEG) of disorders of consciousness, and specifically to a method and system for discriminating EEG of disorders of consciousness with missing channels based on a CNN-Transformer hybrid model. Background Art

[0002] Persistent disorders of consciousness (pDOC) refer to a pathological state of long-term loss of consciousness, usually including two categories: unresponsive wakefulness syndrome (UWS) and minimally conscious state (MCS). Patients with UWS only retain the wakefulness state but have no conscious content; while patients with MCS have partial consciousness, manifested as being able to comply with simple instructions, utter understandable language, or have purposeful behaviors. Patients with MCS have greater rehabilitation potential compared to patients with UWS. Formulating appropriate treatment strategies requires discriminating the patient type. Traditional methods of clinical discrimination rely on expert experience assessment, usually using the revised coma recovery scale (CRS-R). However, the assessment operation of this stimulus feedback scale is cumbersome, has high requirements for experience, and has low consistency of assessment results and a high misdiagnosis rate. Electroencephalogram (EEG) is an objective electrophysiological index reflecting the brain function state, with the characteristics of being safe, non-invasive, having high resolution, and being convenient and inexpensive. Therefore, EEG is also used to achieve objective discrimination of disorders of consciousness.

[0003] Most of the existing pDoC EEG discrimination methods rely on high-density EEG data covering all brain channels. In the real-world clinical environment, there are often the following situations: (1) skull defects after surgical intervention; (2) the cost of high-density EEG acquisition equipment is high and difficult to provide; (3) electrode displacement and channel signal loss. This means that high-density whole-brain EEG may not be available, thus leading to low applicability of the automatic discrimination methods that rely on this kind of EEG. In addition, many existing methods rely on non-resting-state EEG data such as transcranial magnetic stimulation electroencephalogram (TMS-EEG). The acquisition of non-resting-state EEG requires designing stimulation tasks, has experience requirements for experimental operators, and will bring additional operation interference. Summary of the Invention

[0004] Aiming at the two aspects of problems in the existing methods, the present invention proposes a method for discriminating EEG of disorders of consciousness with missing channels, which is used to discriminate disorders of consciousness (pDoC) on a resting-state EEG configuration that may have missing channels. This method utilizes a deep network model with a convolutional neural network-transformer (CNN-Transformer) architecture, which consists of two parts: (1) a time-domain double-branch CNN module; (2) a spatio-temporal Transformer module based on electrode position encoding. After inputting an EEG data segment and the corresponding channel list into the model, the model can output the classification results of MCS and UWS for this segment.

[0005] The present invention proposes a method for discriminating electroencephalogram (EEG) of patients with disorders of consciousness with missing channels based on a CNN-Transformer hybrid model. The method is based on an EEG classification model with a CNN-Transformer hybrid architecture, which can automatically discriminate the types of disorders of consciousness (MCS and UWS) with high accuracy according to the resting-state EEG of patients. The model uses a time-domain dual-branch CNN and a spatio-temporal Transformer module to hierarchically learn the features in EEG data; and introduces a model architecture with variable channel dimensions, an inter-channel correlation learning method that integrates electrode position information, and a training strategy of randomly discarding channels, enabling the model to adapt to scenarios with partial channel loss, thereby improving the clinical applicability of the method.

[0006] To achieve the above objectives, the specific implementation steps are as follows: Step 1, data acquisition: Use standard-lead EEG data as the data source for model training and type discrimination. The EEG data includes the resting-state EEG of MCS patients and the resting-state EEG of UWS patients. Step 2, data preprocessing: Process the original signals collected by the EEG acquisition device into data suitable for model input. Step 3, construction of the EEG classification model: The model includes a time-domain dual-branch CNN, a spatio-temporal Transformer module, and an MLP classifier. Step 4, model training: Batch input multi-channel EEG segments into the model, and randomly discard some channels before the spatio-domain Transformer processing. Calculate the loss based on the output results and class labels, and adjust the model parameters through backpropagation. Train for multiple rounds until the performance is stable. Step 5, model application: Collect test EEG data, adjust the electrode position encoding of the model according to the EEG channels of the test data, and input the test data segments to obtain the classification results.

[0007] Further, the specific implementation method of step 1 is as follows: 1.1, The EEG signal acquisition is carried out by attaching sensors to the scalp surface of the subjects with disorders of consciousness, recording the data during rest or anesthesia, and transmitting it to the computer for storage in real time. 1.2, Pass the EEG acquisition electrodes, a reference electrode, and a ground wire through the headgear and closely attach them to the scalp of the subject, and apply conductive paste to reduce the resistance coefficient. 1.3, Record the channel electrode coordinates of the EEG data and the types of chronic disorders of consciousness of the subjects.

[0008] Further, the specific implementation method of step 2 is as follows: 2.1, First, downsample the collected EEG data, and then filter out non-EEG frequency bands through a band-pass filtering algorithm, and only retain part of the EEG data with a certain frequency range. 2.2. Segment the complete EEG record of each subject into segments of n seconds each. 2.3. Reshape the segment data to a dimension of 1 × number of time points × number of channels.

[0009] Furthermore, in the time-domain dual-branch CNN, one branch uses standard convolution to capture fine-grained local time-domain features, while the other branch uses dilated convolution with the same number of parameters but an expanded receptive field to obtain a broader time context. Both the standard convolution and the dilated convolution are followed by a batch normalization layer and an ELU activation function. The outputs of the two branches are processed by a 1-D temporal max-pooling layer to reduce the number of time points, then passed through dropout, and finally the features extracted by the two branches are fused by a 1×1 convolutional layer to reduce the dimension of the feature map.

[0010] Furthermore, the processing procedure of the spatio-temporal Transformer module is as follows: First, calculate the electrode position encoding, fuse it with the channel embedding sequence at each time point, process it through the spatial-domain Transformer layer, then perform spatial-domain pooling, concatenate the pooling results along the time dimension, and use the resulting time series as the input to the time-domain Transformer layer. Finally, the class result is output by the MLP classifier.

[0011] Furthermore, the output of the time-domain dual-branch CNN is regarded as a spatio-temporal two-dimensional list of feature embeddings. The spatial-domain Transformer layer will independently process the multi-channel feature embedding sequences at each time point. Each sequence is added to the electrode position encoding and processed by the TransformerEncoderLayer. The results are pooled and concatenated into a one-dimensional time series. Regard the output of the spatial-domain Transformer layer at each time point as a temporal embedding. The time-domain Transformer is processed based on the standard TransformerEncoderLayer, but without adding position encoding and without performing pooling. The processed temporal sequence will be directly used as the input to the MLP classifier, which is finally used to generate the discrimination result.

[0012] Furthermore, calculate the electrode position encoding according to the channel electrode coordinates. The input data contains EEG channels, and each channel is associated with a 3D spatial coordinate representing its physical position. Each dimension is independently normalized to the range [0,1] to form a normalized coordinate matrix , and the formula for the electrode position encoding is shown in Equation (1) below: ​

[0013] Among them, pos represents the serial number of the feature embedding in the sequence, j represents the index of the element in the feature embedding, D is the total dimension size of the feature embedding, λ is a hyperparameter controlling the frequency change, g( ) is a mapping function, represents the remainder operation.

[0014] Furthermore, the specific implementation of step 4 is as follows: 4.1, Input batches of EEG data segments and process them by the time-domain double-branch CNN; 4.2, For the results processed by the time-domain double-branch CNN, discard some channels according to the independent probability p to construct a masking vector ∈ {0, 1} C , each element of the vector indicates whether to discard the channel at time , and the discard values follow the Bernoulli distribution , as shown in formula (2):

[0015] where T is the number of time points and C is the number of EEG channels; 4.3, Input the data into the spatio-temporal Transformer module for processing and output the category prediction results; 4.4, Calculate the classification loss based on the prediction results and the category labels, adjust the model using backpropagation, repeatedly input batches, and train for multiple rounds until the validation performance is stable.

[0016] Furthermore, step 5 specifically includes: 5.1, Select some channels for EEG acquisition according to the clinical conditions and process the EEG data into input segments; 5.2, Adjust the electrode position encoding according to the channel electrode coordinates of the test EEG data; 5.3, Input the test data into the trained model to obtain the prediction results of the consciousness disorder category.

[0017] The present invention also provides a discriminant system for EEG of consciousness disorder adapted to channel loss, including: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for discriminant of EEG of consciousness disorder adapted to channel loss as described in the above technical solution.

[0018] The present invention has the following positive effects and advantages: 1) The present invention uses a CNN-Transformer hybrid architecture model to realize a hierarchical spatio-temporal Transformer for EEG, and can automatically discriminate the type of consciousness disorder with high accuracy based on the resting-state EEG of patients.

[0019] 2) In the present invention, the resting-state EEG model utilizes the prior knowledge of the EEG electrode space to achieve compatibility and adaptation to the EEG data of missing channels. In the face of clinical scenarios such as skull injuries of subjects, low electrode density of acquisition devices, and data bad channels, this discrimination method can still maintain applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a model architecture diagram of the method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To further clarify the technical means adopted by the present invention and the achieved effects, the following will be described in detail with reference to the drawings and specific examples.

[0022] As Figure 1 shown, a method for discriminating EEG of patients with disorders of consciousness adapting to missing channels provided by an embodiment of the present invention specifically works in the following 5 points: Step 1: Data acquisition. In this work, EEG data using the 10-20 standard lead is used as the data source for model training and type discrimination, and the EEG acquisition states include the resting-state EEG of MCS patients and the resting-state EEG of UWS patients.

[0023] Step 2: Data preprocessing. The following processing is performed on the original signals collected by the EEG acquisition device: (1) Channel average reference, downsampling, and band-pass filtering are performed to reduce the influence of noise; (2) The original signals are segmented into 10s segments as the input units of the deep network model.

[0024] Step 3: Model construction. A neural network model is constructed in the following module order: (1) Time-domain dual-branch CNN, including: one branch uses standard convolution to capture fine-grained local time-domain features; while the other branch uses dilated convolution with the same number of parameters but an expanded receptive field to obtain a wider time context. The output of the CNN will be processed by a 1-D time-dimensional max pooling layer to reduce the number of time points. Finally, a 1×1 convolution layer fuses the features extracted by the dual branches and reduces the dimension of the feature map. (2) Spatiotemporal Transformer module. The electrode position encoding is calculated and fused with the channel token sequence at each time point, processed by the spatial domain Transformer, and then spatial domain pooling is performed. The pooled results are concatenated in the time dimension and, as a time series, after being processed by the time domain Transformer, the class results are output by the MLP classifier.

[0025] Step 4: Model training. Input multi-channel EEG segments into the model in batches, and randomly discard some channels before spatial domain Transformer processing. Calculate the loss based on the output results and class labels, and adjust the model parameters through backpropagation. Train for multiple Epochs until the performance on the validation set is stable.

[0026] Step 5: Model application. Collect test EEG data, adjust the electrode position encoding of the model according to the EEG channels of the test data, and input the test data segments to obtain classification results.

[0027] In the said Step 1, its specific steps include: 1.1, EEG signal acquisition is carried out by attaching sensors to the scalp surface of the subjects with disorders of consciousness, recording the data of people during rest or anesthesia, and transmitting it to the computer for storage in real time.

[0028] 1.2, Pass the EEG acquisition electrodes, a reference electrode, and a ground wire through the headgear and closely fit them to the scalp of the subject, and apply conductive paste to reduce the resistance coefficient.

[0029] 1.3, Record the channel electrode coordinates of the EEG data and collect the types of chronic disorders of consciousness of the subjects.

[0030] In the said Step 2, its specific steps include: 2.1, First, downsample the collected EEG data to 250 Hz, and then filter out the non-EEG frequency bands through a band-pass filtering algorithm, only retaining the part of 0.5 - 45 Hz. 2.2, Cut the complete EEG record of each subject into segments of 10 seconds each.

[0031] 2.3, Change the dimension of the segment data to 1 × number of time points × number of channels.

[0032] In the said Step 3, its specific steps include: 3.1, Construct a 1-D convolutional layer of the model along the time dimension, and add a batch normalization layer and an ELU activation layer to stabilize the training.

[0033] 3.2, Construct a 1-D dilated convolutional layer of the model along the time dimension, and add a batch normalization layer to stabilize the training.

[0034] 3.3, Construct a 1-D time dimension max pooling layer of the model to reduce the number of time points while retaining key features, and then add a Dropout layer to stabilize the training. Construct a 1×1 convolutional layer to fuse the features extracted by the dual-branch 1-D convolutional layer and reduce the dimension of the feature map.

[0035] 3.4, Calculate the electrode position encoding according to the channel electrode coordinates. The input data contains EEG brain channels, each channel associated with 3D spatial coordinates representing its physical location For ensuring numerical stability, each dimension is independently normalized to the range of [0, 1], forming a normalized coordinate matrix The formula for electrode position encoding is shown in Equation (1) as follows:

[0036] where pos represents the sequence number of the feature embedding in the sequence (corresponding to the EEG channel), j represents the index of the element in the feature embedding, D is the total dimension size of the feature embedding, and λ is a hyperparameter controlling the frequency variation. The mapping function g( ), usually a trigonometric function, represents the modulo operation.

[0037] 3.5, constructing the spatial domain Transformer layer of the model. The output of the shallow feature extraction module, i.e., the time domain double-branch CNN, can be regarded as a spatio-temporal two-dimensional list of feature embedding tokens. The inter-channel correlation learning, i.e., the spatial domain Transformer layer, will independently process the multi-channel feature Token sequence at each time point. Each sequence will be added with the electrode position encoding, processed by the TransformerEncoderLayer, and the result will be pooled and concatenated into a one-dimensional time sequence.

[0038] 3.6, constructing the time domain Transformer layer of the model. This module regards the output of the inter-channel correlation learning at each time point as a token. The time domain Transformer is based on the standard TransformerEncoderLayer, but without adding position encoding and without pooling. The processed time series will be directly used as the input of the MLP classifier, and finally used to generate the discrimination result.

[0039] In step 4, the specific steps include: 4.1, inputting batches of EEG data segments and processing them by the time domain double-branch CNN of the model 4.2, discarding some channels from the result processed by the time domain double-branch CNN according to the independent probability p. Construct a masking vector ∈ {0, 1} C Each element of the vector indicates whether to discard the channel at time . The discarded values follow the Bernoulli distribution as shown in Equation (2):

[0040] where T is the number of time points and C is the number of EEG channels; 4.3. The input data is processed by the spatio-temporal Transformer module, and the output is the category prediction result.

[0041] 4.4. Calculate the classification loss according to the prediction result and the category label. In the present invention, the cross-entropy classification loss is adopted, and the model is adjusted by backpropagation. Repeatedly input batches and train for multiple rounds until the validation performance is stable.

[0042] In the said step 5, the specific steps include: 5.1. Select some channels for EEG acquisition according to the clinical conditions, and process the EEG data into input segments.

[0043] 5.2. Adjust the electrode position encoding of the classification model according to the channel electrode coordinates of the test EEG data.

[0044] 5.3. Input the test data into the trained model to obtain the prediction result of the level of consciousness disorder category.

[0045] The implementation and effects of the present invention are illustrated by a specific example as follows: 1. Data acquisition. First, paste the electrodes at the specified positions on the scalp of the subject, and apply conductive paste to reduce the impedance between the electrodes and the scalp. After confirming that all impedance tests are qualified, the experimenter starts the EEG device to record. Record the information of the level of consciousness disorder diagnosed by the doctor. The electrode coordinate file of the EEG cap can be extracted by the EEGLAB processing software.

[0046] 2. Data preprocessing. By default, the sampling frequency of the acquired EEG data is 1000 Hz. The sampling rate is adjusted to 250 Hz by the downsampling method, and 30 channels are selected. Next, use the band-pass filtering algorithm to filter out the noise below 0 - 0.5 Hz and above 45 Hz. The continuous recording is segmented into segments of 10 seconds each.

[0047] 3. Build the model. Build a neural network model in the following module order: (1) Time-domain dual-branch CNN. One branch uses standard convolution to capture fine-grained local time-domain features; the other branch uses dilated convolution with the same number of parameters but an expanded receptive field to obtain a wider time context. The CNN output is processed by a 1-D time-dimensional max pooling layer to reduce the number of time points. Finally, a 1×1 convolutional layer fuses the features extracted by the dual branches and reduces the dimension of the feature map. (2) Spatiotemporal Transformer module. Calculate the electrode position encoding, fuse it with the channel token sequence at each time point, process it through the spatial Transformer, and then perform spatial pooling. Concatenate the pooled results in the time dimension. As a time series, after being processed by the time-domain Transformer, the class results are output by the MLP classifier. The specific model structure parameters can be adjusted and optimized according to experience and validation accuracy, but it is recommended to follow the following principles: (1) The 1-D CNN convolution kernel should not be too large and can be set between 25 and 50. (2) The dilation rate of the dilated convolution is recommended to be set to 2. (3) The number of layers of the spatial and time-domain transformers is recommended to be set to 1. (4) The number of heads of the spatial and time-domain transformers is recommended to be set to 1 or 2. (5) The number of feature maps of the CNN can be set to 32 or 64.

[0048] 4. Model training. The model input is a 1×2500×30 feature vector; the output is a 2×1 array, and each value is between 0 and 1 and the sum is 1, representing the probabilities of the types MCS and UWS respectively. The training data can be randomly discarded channels between the shallow feature extraction module and the sk association learning module. In the special case where all channels are discarded, reverse compensation can be performed, and all channels are retained instead. Each time a batch of data is input during training, the loss is calculated based on the prediction result and the list label, and a backpropagation is performed. The specific model training parameters can be adjusted and optimized according to experience and validation accuracy, but it is recommended to follow the following principles: (1) It is recommended to choose the ADAM optimizer that can automatically adjust the learning rate dynamically. (2) The channel discard probability is recommended to be set to 0.2. (3) The training period is recommended to be 50, and it is set to automatically stop early according to the validation error and save the optimal parameters among them. (4) After the model training and optimization are completed, the final performance of the model can be evaluated through the validation dataset, including the test accuracy, model specificity, and model sensitivity.

[0049] 5. Model application. In actual application, incomplete EEG channels can be selected according to the acquisition conditions, and the electrode data information of EEGLAB can also be utilized. After adjusting the electrode position coding, the model can accept EEG data with missing channels relative to the training data, but the following suggestions are provided for EEG acquisition in application: (1) The number of EEG channels collected should be greater than 8. (2) The collected channels should include Fp series electrodes. (3) The electrode density in the temporal lobe and occipital lobe can be relatively reduced.

[0050] Compared with the 30-channel data of the whole brain: After removing a single channel, leaving 29 channels, the average accuracy drops by 0.0016; After removing the Fp electrodes, leaving 27 channels, the accuracy drops by 0.09; Only retaining 8 channels of F and Fp forehead, the accuracy drops by 0.03.

[0051] On the other hand, the embodiment of the present invention also provides a discriminant system for EEG of disorders of consciousness adapted to missing channels, including: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a discriminant method for EEG of disorders of consciousness adapted to missing channels as described in the above technical solution.

[0052] The above embodiments only illustrate the implementation manners of the present invention, and the description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those skilled in the art and ordinary technicians, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A discriminant method for electroencephalogram of disturbance of consciousness adapting to channel loss, characterized in that It includes the following steps: Step 1, data acquisition: The electroencephalogram (EEG) data of standard leads are used as the data source for model training and type discrimination. The EEG data include the resting-state EEG of MCS patients and the resting-state EEG of UWS patients. Step 2, data preprocessing: The original signals collected by the EEG acquisition device are processed into data suitable for model input. Step 3, construction of the EEG classification model: The model includes a time-domain dual-branch CNN, a spatio-temporal Transformer module, and an MLP classifier. Step 4, model training: Multichannel EEG segments are input into the model in batches, and some channels are randomly discarded before the spatio-domain Transformer processing. The loss is calculated based on the output results and the class labels, and the model parameters are adjusted through backpropagation. Training is performed for multiple rounds until the performance is stable. Step 5, model application: Test EEG data are collected, the electrode position encoding of the model is adjusted according to the EEG channels of the test data, and the test data segments are input to obtain the classification results.

2. The EEG discrimination method for disorders of consciousness adapting to channel loss according to claim 1, wherein: The specific implementation method of Step 1 is as follows: 1.1, The EEG signal acquisition is carried out by attaching sensors to the scalp surface of the subjects with disorders of consciousness, recording the data of people during rest or anesthesia, and transmitting them to the computer for storage in real time. 1.2, The EEG acquisition electrodes, a reference electrode, and a ground wire are respectively passed through the headgear and closely attached to the scalp of the subject, and conductive paste is applied to reduce the resistance coefficient. 1.3, Record the channel electrode coordinates of the EEG data and the type of chronic consciousness disorder of the subject.

3. The EEG discrimination method for disturbance of consciousness adapting to channel loss according to claim 1, characterized in that: The specific implementation method of Step 2 is as follows: 2.1, First, downsample the collected EEG data, and then filter out the non-EEG frequency bands through a band-pass filtering algorithm, only retaining part of the EEG data with the frequency in a certain range. 2.2, Cut the complete EEG record of each subject into segments of n seconds. 2.3, Change the dimension of the segment data to 1×time points×number of channels.

4. The EEG discrimination method for disorders of consciousness adapting to channel loss according to claim 1, wherein: One branch in the time-domain dual-branch CNN uses standard convolution to capture fine-grained local time-domain features, while the other branch uses dilated convolution with the same number of parameters but an expanded receptive field to obtain a wider time context. Both the standard convolution and the dilated convolution are followed by a batch normalization layer and an ELU activation function. The outputs of the two branches will be processed by a 1-D maximum pooling layer in the time dimension to reduce the number of time points, and then through dropout processing. Finally, the features extracted by the two branches are fused through a 1×1 convolution layer and the dimension of the feature map is reduced.

5. The EEG discrimination method for disorders of consciousness adapting to channel loss according to claim 1, wherein: The processing process of the spatio-temporal Transformer module is as follows: First, calculate the electrode position encoding, fuse it with the channel embedding sequence at each time point, process it through the spatio-domain Transformer layer, and then perform spatio-domain pooling. The pooled results are concatenated in the time dimension and used as a time series to be processed through the time-domain Transformer layer, and then the class results are output by the MLP classifier.

6. The electroencephalogram discrimination method for disturbance of consciousness adapted to channel loss according to claim 5, characterized in that: The output of the time-domain dual-branch CNN is regarded as a spatio-temporal two-dimensional list of feature embeddings. The spatial-domain Transformer layer will independently process the multi-channel feature embedding sequences at each time point. Each sequence will be added to the electrode position encoding and processed by the Transformer Encoder Layer. The result will be pooled and concatenated into a one-dimensional time sequence. The output of the spatial-domain Transformer layer at each time point is regarded as a temporal embedding. The time-domain Transformer is processed based on the standard Transformer Encoder Layer, but without adding position encoding and without pooling. The processed temporal sequence will be directly used as the input of the MLP classifier and finally used to generate the discrimination result.

7. A discriminant method for electroencephalogram of disturbance of consciousness adapting to channel loss according to claim 5, characterized in that: Calculate the electrode position encoding based on the channel electrode coordinates. The input data contains EEG channels, and each channel is associated with 3D spatial coordinates representing its physical position which are independently normalized to the range [0, 1] for each dimension, forming a normalized coordinate matrix . Then the formula for the electrode position encoding is shown in Equation (1) as follows: Among them, pos represents the serial number of the feature embedding in the sequence, j represents the index of the element in the feature embedding, D is the total dimension size of the feature embedding, λ is a hyperparameter that controls the frequency change, and g( ) is a mapping function. Indicates the remainder operation.

8. The EEG discrimination method for disorders of consciousness adapting to channel loss according to claim 1, wherein: The specific implementation method of step 4 is as follows: 4.1, Input batches of EEG data segments and process them by the time-domain dual-branch CNN. 4.2, the result of the time domain dual-branch CNN processing, discarding some channels according to the independent probability p, and constructing a mask vector ∈ {0, 1} C , each element of the vector Indicates at time Whether to discard the channel , the discarded values follow a Bernoulli distribution , as shown in formula (2): Among them, T is the number of time points, and C is the number of EEG channels. 4.3, The input data is processed by the spatio-temporal Transformer module to output the class prediction result. 4.4, Calculate the classification loss according to the prediction result and the class label, adjust the model with backpropagation, repeatedly input batches, and train for multiple rounds until the verification performance is stable.

9. The electroencephalogram discrimination method for disturbance of consciousness adapted to channel loss according to claim 1, characterized in that: Step 5 specifically includes: 5.1, Select some channels for EEG acquisition according to clinical conditions and process the EEG data into input segments. 5.2, Adjust the electrode position encoding according to the channel electrode coordinates of the test EEG data. 5.3, Input the test data into the trained model to obtain the class prediction result of the disorder of consciousness.

10. An electroencephalogram discrimination system for consciousness disorders adapted to channel loss, characterized in that, Including: A processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for discriminating EEG of disorder of consciousness adapted to channel loss as described in any one of claims 1-9.

Images