Auxiliary diagnosis system for patients with consciousness disorders based on P300 electroencephalogram and transfer learning

Abstract

The application discloses a kind of based on P300 brain electricity and transfer learning's consciousness disorder patient auxiliary diagnosis system, comprising: P300 evocation and acquisition module, for evoking and collecting patient P300 brain electricity data, and the data is preprocessed;Source domain selection module, for calculating similarity index, and screening out the health test data more similar to consciousness disorder patient data as source domain, wherein consciousness disorder patient data is target domain data;Auxiliary diagnosis training module, source domain feature extractor and label discriminator are trained using source domain data, using target domain data and source domain data against migration, adjust source domain feature extractor so that the feature extracted in target domain can be in the same feature space with the feature extracted by source domain;Auxiliary diagnosis prediction module is used for the auxiliary diagnosis of patient according to consciousness disorder patient data, and provides the consciousness level and rehabilitation prediction of patient.The present application uses transfer learning technology and health test data, solves the problem of insufficient patient data and acquisition difficulty.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and biomedicine, and in particular to an auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning. Background Technology

[0002] Disorders of consciousness are sequelae of severe brain injury, characterized by loss of consciousness and cognitive impairment, including coma, vegetative state, minimal consciousness, and locked-in syndrome. Patients with different states of consciousness disorder have varying levels of awareness, and accurate detection of this level is crucial for diagnosis and treatment. Currently, there are no clinical tools to diagnose whether a patient can regain consciousness or to what extent they may recover. Clinically, the detection of consciousness disorder status is limited to physician experience and scale assessments (such as the CRS-R Coma Recovery Scale), which are highly subjective and lead to a high rate of misdiagnosis. Recent studies have shown that the P300 brain-computer interface can directly detect consciousness using the patient's P300 brain signals without relying on behavioral responses, offering significant advantages. However, patients with consciousness disorders have weaker levels of awareness and cannot use brain-computer interface systems as effectively as healthy individuals. Therefore, designing a P300 evoked and decoded system is key to developing a P300-assisted diagnostic system for patients with consciousness disorders.

[0003] Current P300-based auxiliary diagnostic systems for patients with disorders of consciousness fall into two categories: 1. Systems based on statistical features and classifiers: These systems use traditional feature extraction techniques (such as independent component analysis and XDAWN spatial filters) to extract P300 EEG features and then use methods such as support vector machines and Bayesian networks to model and predict P300 EEG. 2. Systems based on deep learning: These systems automatically extract and predict the spatiotemporal features of P300 EEG using deep learning methods such as convolutional neural networks. Overall, existing P300-based auxiliary diagnostic systems for patients with disorders of consciousness require a large amount of labeled data to train the model. Furthermore, due to individual differences—that is, different users have different neural responses to the same stimulus—existing systems require model recalibration when applied to new subjects, which is very inconvenient. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose an auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning. This system breaks through the calibration and data limitations of traditional P300 auxiliary diagnostic systems and makes full use of a large amount of healthy subject data by using transfer learning methods, thereby reducing the system's data dependence and calibration difficulty.

[0005] To achieve the above objectives, the technical solution provided by this invention is: an auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning, comprising:

[0006] The P300 induction and acquisition module is used to induce and acquire P300 EEG data from patients and to preprocess the data.

[0007] The source domain selection module is used to calculate similarity indices and select healthy subject data that are more similar to data from patients with impaired consciousness as the source domain, while data from patients with impaired consciousness is the target domain data.

[0008] The auxiliary diagnostic training module uses source domain data to train a source domain feature extractor and a source domain label discriminator. It uses target domain data and source domain data to perform adversarial transfer, and adjusts the source domain feature extractor so that the features extracted in the target domain are in the same feature space as the features extracted in the source domain.

[0009] The auxiliary diagnostic prediction module is used to assist in the diagnosis of patients with disorders of consciousness based on their data, and to provide predictions of the patient's level of consciousness and recovery.

[0010] Furthermore, the P300 induction and acquisition module includes a P300 induction module, a data loading module, and a data preprocessing module, wherein:

[0011] The P300 induction module uses visual and auditory induction to generate P300 signals, which are then amplified and filtered after being collected by an EEG acquisition device.

[0012] The data loading module can read EEG data from the local machine, including MAT, CNT, and EEG formats, and can also receive EEG data transmitted online from the P300 evoked module.

[0013] The data preprocessing module is used to remove noise from the EEG signal and downsample the EEG signal to reduce the amount of data and computational complexity. At the same time, the EEG signal is divided into several time windows to extract the P300 signal.

[0014] Furthermore, the source domain selection module performs the following operations:

[0015] a. Calculate the probability distribution of healthy subject samples and target domain samples based on spatiotemporal convolutional neural networks:

[0016] A source domain feature extractor E based on a spatiotemporal convolutional neural network was trained using healthy subject sample data. S Source domain label discriminator C based on fully connected networks S :

[0017] F (v) =E S (M v ,Θ S )

[0018] In the formula, M vIt is an N×T matrix representing a sample of healthy subjects v, where v = {1, 2, ..., W}, W is the total number of healthy subjects, N is the sampling frequency of the preprocessed EEG data, T is the number of channels in the EEG data, and Θ S F represents the parameters of the source domain feature extractor. (v) The features extracted by the source domain feature extractor for healthy subject v consist of a 5-layer spatiotemporal convolutional neural network. The first layer is the input layer, the second layer is the spatial convolutional layer with a kernel size of T, which is the number of channels of the EEG data, the third layer is the temporal convolutional layer with a kernel size of 4, the fourth layer is the pooling layer with a filter size of (2,1) and the activation function is Tanh, and the fifth layer is the flattening layer, which compresses the input data from a multidimensional array into a one-dimensional array.

[0019] L (v) =C S (F (v) ,θ S )

[0020] In the formula, θ S L is the parameter of the source domain label discriminator. (v) The result of the source domain label discriminant P300 is the result of the healthy subject v being detected by the source domain label discriminant. The source domain label discriminant consists of a 4-layer fully connected neural network with the activation function being Softmax.

[0021] Using source domain feature extractor E S Source domain label discriminator C S Calculate the probability distribution of each sample in the source and target domains as C. S (E S (M v ,Θ S ),θ S );

[0022] b. Calculate the sample similarity between the source and target domains based on Wasserstein distance:

[0023] According to formula C S (E S (M v ,Θ S ),θ S Calculate the probability distribution of the source domain as P. S The probability distribution of the target domain is P T The Wasserstein distance between each healthy subject in the source domain and the target domain is expressed as:

[0024]

[0025] In the formula, Let v represent the probability distribution of the healthy subject. This represents the Wasserstein distance between the healthy subject v and the target domain. Let v represent the joint probability distribution of the healthy subject v and the target domain, γ represent any possible joint distribution, and (x,y) represent a sample x and y obtained from γ, where x is a sample from the source domain distribution and y is a sample from the target domain distribution. This represents the expected value of the Euclidean distance between x and y;

[0026] Finally, healthy subjects with the smallest Wasserstein distance were selected as source domains.

[0027] Furthermore, the auxiliary diagnostic training module performs the following operations:

[0028] 1) Initialize the target domain feature extractor E T With source domain feature extractor E S Same parameters;

[0029] 2) To better learn the differences between the source and target domains, thereby improving the model's generalization ability and performance, a domain discriminator D is trained simultaneously with the feature extractor and label discriminator. Performance is improved by minimizing the classification loss function. The probability that a sample comes from the target domain is obtained from the domain discriminator D and used as a weighting factor to adjust the training direction of the feature extractor.

[0030] H=D(F,θ d ), F = F S ∪F T

[0031] In the formula, θ d F is the parameter of the domain discriminator D. S For source domain feature extractor E S Extract source domain features, F T For target domain feature extractor E T Extracting features from the target domain, F is F S With F T The union of H represents the discrimination result of the domain discriminator. The domain discriminator D consists of a 3-layer fully connected neural network with LogSoftmax as the activation function.

[0032] Optimize the target domain feature extractor E by minimizing the domain alignment loss. T This makes the feature distributions of the source and target domains more similar, thereby improving the model's generalization ability in the target domain. The expression for the domain alignment loss function is:

[0033]

[0034] In the formula, N is the domain alignment loss between the source and target domains. tIt is the amount of data in the target domain. It is the i-th source domain sample. It is the i-th target domain sample;

[0035] Therefore, the target domain feature extractor E T The total loss is:

[0036] Gradient descent is used to reduce the feature extractor E in the target domain. T The total loss, compared with the backpropagation training target domain feature extractor E T This enables the source domain label discriminator's output to accurately predict the label of the target domain data;

[0037] 3) Adjust the parameters of the domain discriminator D to improve its performance:

[0038] The expression for its loss function is:

[0039]

[0040] In the formula, M s (x) and M t (x) represent the source domain data and the target domain data, respectively. x represents the expectation of the logarithm of the probability that the target domain discriminator D classifies the target domain data as belonging to the target domain. t In order to sample the target domain data, Let x be the expectation of the logarithm of the probability that the source domain data is classified as the target source domain by the domain discriminator D. s For sampling of source domain data;

[0041] The domain discriminator D is trained using gradient descent and backpropagation. The classification error of the domain discriminator D in the target domain is minimized, while the difference between the label discriminators in the source and target domains is maximized, thereby achieving the purpose of domain adaptation.

[0042] 4) By iterating through steps 2) and 3) above, the feature alignment effect is gradually improved, and finally the optimal target domain feature extractor E is obtained. T Source domain label discriminator C S .

[0043] Furthermore, the auxiliary diagnostic prediction module performs the following operations:

[0044] 1) Use the P300 induction module in the P300 induction and acquisition module to induce P300 EEG and amplify and filter it; use the data loading module to receive the P300 induction module online; and use the data preprocessing module to process the P300 data.

[0045] 2) The P300 data input source domain selection module after step 1) is used to select appropriate healthy subject data as the source domain;

[0046] 3) The P300 data processed in step 1) and the healthy subject data selected in step 2) are input into the auxiliary diagnosis training module. The target domain discriminator and the target domain feature extractor are trained iteratively to generate a feature representation similar to the source domain feature representation. The network parameters are adjusted by minimizing the total loss through backpropagation to obtain the optimal model. The trained target domain feature extractor and source domain label discriminator are used to detect the P300 signal of patients with impaired consciousness. The auxiliary diagnosis results, level of consciousness and rehabilitation prediction of the patients are obtained based on the detection accuracy of P300 and clinical assessment.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] 1. Compared with traditional clinical scale methods, using P300 EEG to detect the level of consciousness in patients with consciousness disorders can directly and objectively reflect the level of consciousness in patients, and it requires less of a user and basically does not require long-term training and clinical experience.

[0049] 2. In the source domain selection module and the auxiliary diagnosis training module, P300 EEG data from healthy subjects are used as the source domain training model, which reduces the data dependence of the auxiliary diagnosis system for patients with consciousness disorders and can make full use of a large amount of effective P300 data from healthy subjects.

[0050] 3. In the auxiliary diagnostic training module, the transfer learning method is used to transfer the model, so that the features extracted by the model in the source domain and the target domain are in the same feature space, thus solving the calibration problem of the system. Attached Figure Description

[0051] Figure 1 This is a schematic diagram showing the relationship between the various modules of the system of the present invention.

[0052] Figure 2 This is a flowchart of the system training and prediction process of the present invention.

[0053] Figure 3 This is a schematic diagram of the P300 evoked paradigm used in the system of the present invention. In the diagram, Instructions is the instruction phase, that is, pay attention to the number 6 and count its repetitions; Stimulations is the stimulation phase, where each number button flashes 10 times; Feedback is the feedback phase; and Rest is the rest phase.

[0054] Figure 4This is a schematic diagram of the feature extractor used in the system of the present invention. In the diagram, Spatial Convolution Layer is a spatial convolution layer, Temporal Convolution Layer is a temporal convolution layer, Feature Pooling Layer is a pooling layer, and Flatten is a flattening layer.

[0055] Figure 5 This is a schematic diagram of the migration method used in the system of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to specific embodiments.

[0057] This embodiment discloses an auxiliary diagnostic system for patients with disorders of consciousness based on P300 EEG and transfer learning. It is an auxiliary diagnostic system for patients with disorders of consciousness developed using Python and capable of running on Windows devices. The relationships between the various modules of the system are as follows: Figure 1 As shown, the system training and prediction process is as follows: Figure 2 As shown.

[0058] It includes:

[0059] The P300 induction and acquisition module is used to induce and acquire P300 EEG data from patients and to preprocess the data.

[0060] The source domain selection module is used to calculate similarity indices and select healthy subject data that are more similar to data from patients with impaired consciousness (target domain) as the source domain.

[0061] The auxiliary diagnostic training module uses source domain data to train a source domain feature extractor and a source domain label discriminator. It uses target domain data and source domain data to perform adversarial transfer, and adjusts the source domain feature extractor so that the features extracted in the target domain are in the same feature space as the features extracted in the source domain.

[0062] The auxiliary diagnostic prediction module is used to assist in the diagnosis of patients with disorders of consciousness based on their data, and to provide predictions of the patient's level of consciousness and recovery.

[0063] Specifically, the P300 induction and acquisition module includes a P300 induction module, a data loading module, and a data preprocessing module, wherein:

[0064] The P300 triggering module has the following triggering paradigm: Figure 3 As shown, the P300 signal is induced by the flashing of two numeric buttons and the voice feedback provided by the speaker, and is collected, amplified and filtered using an EEG acquisition device;

[0065] The data loading module can read EEG data from the local machine, including MAT, CNT, and EEG formats, and can also receive EEG data transmitted online from the P300 evoked module.

[0066] The data preprocessing module is used to remove noise from the EEG signal and downsample the EEG signal to reduce the amount of data and computational complexity. At the same time, the EEG signal is divided into several time windows to extract the P300 signal.

[0067] Specifically, the source domain selection module performs the following operations:

[0068] a. Calculate the probability distribution of healthy subject samples and target domain samples based on spatiotemporal convolutional neural networks:

[0069] A source domain feature extractor E based on a spatiotemporal convolutional neural network was trained using healthy subject sample data. S ,like Figure 4 As shown, the source domain label discriminator C based on a fully connected network S :

[0070] F (v) =E S (M v ,Θ S )

[0071] Among them, M v It is an N×T matrix representing a sample of healthy subjects v, where v = {1, 2, ..., W}, W is the total number of healthy subjects, N is the sampling frequency of the preprocessed EEG data, T is the number of channels in the EEG data, and Θ S F represents the parameters of the source domain feature extractor. (v) The features extracted by the source domain feature extractor for healthy subject v consist of a 5-layer spatiotemporal convolutional neural network. The first layer is the input layer, the second layer is the spatial convolutional layer with a kernel size of T, which is the number of channels of the EEG data, the third layer is the temporal convolutional layer with a kernel size of 4, the fourth layer is the pooling layer with a filter size of (2,1) and the activation function is Tanh, and the fifth layer is the flattening layer, which compresses the input data from a multidimensional array into a one-dimensional array.

[0072] L (v) =C S (F (v) ,θ S )

[0073] Where, θ S L is the parameter of the source domain label discriminator. (v) The result of the source domain label discriminant P300 is the result of the healthy subject v being detected by the source domain label discriminant. The source domain label discriminant consists of a 4-layer fully connected neural network with the activation function being Softmax.

[0074] Using source domain feature extractor E S Source domain label discriminator C S Calculate the probability distribution of each sample in the source and target domains as C. S (E S (M v ,Θ S ),θ S );

[0075] b. Calculate the sample similarity between the source and target domains based on Wasserstein distance:

[0076] According to formula C S (E S (M v ,Θ S ),θ S Calculate the probability distribution of the source domain as P. S The probability distribution of the target domain is P T Then the Wasserstein distance between each healthy subject in the source domain and the target domain can be expressed as:

[0077]

[0078] in, Let v represent the probability distribution of the healthy subject. This represents the Wasserstein distance between the healthy subject v and the target domain. Let v represent the joint probability distribution of the healthy subject v and the target domain, γ represent any possible joint distribution, and (x,y) represent a sample x and y obtained from γ, where x is a sample from the source domain distribution and y is a sample from the target domain distribution. This represents the expected value of the Euclidean distance between x and y.

[0079] Finally, healthy subjects with the smallest Wasserstein distance were selected as source domains.

[0080] Specifically, such as Figure 5 As shown, the auxiliary diagnostic training module uses the domain discriminator D to analyze the source domain feature extractor E. S Extracted source domain features and target domain features by feature extractor E T The discriminant results of the extracted features are adjusted based on the target domain features extracted by the feature extractor E. T With the domain discriminator D, the feature representation has better domain invariance, and the target domain feature extractor E... T The extracted features are closer to the source domain. Specifically, the following operations are performed:

[0081] 1) Initialize the target domain feature extractor E T With source domain feature extractor ES Same parameters;

[0082] 2) To better learn the differences between the source and target domains, thereby improving the model's generalization ability and performance, a domain discriminator D is trained simultaneously with the feature extractor and label discriminator. Performance is improved by minimizing the classification loss function. The probability that a sample comes from the target domain is obtained from the domain discriminator D and used as a weighting factor to adjust the training direction of the feature extractor.

[0083] H=D(F,θ d ), F = F S ∩F T

[0084] Where, θ d F is the parameter of the domain discriminator D. S For source domain feature extractor E S Extract source domain features, F T For target domain feature extractor E T Extracting features from the target domain, F is F S With F T The union of H represents the discrimination result of the domain discriminator. The domain discriminator D consists of a 3-layer fully connected neural network with LogSoftmax as the activation function.

[0085] Optimize the target domain feature extractor E by minimizing the domain alignment loss. T This makes the feature distributions of the source and target domains more similar, thereby improving the model's generalization ability in the target domain. The expression for the domain alignment loss function is:

[0086]

[0087] in, N is the domain alignment loss between the source and target domains. t It is the amount of data in the target domain. It is the i-th source domain sample. It is the i-th target domain sample;

[0088] Therefore, the target domain feature extractor E T The total loss is:

[0089] Gradient descent is used to reduce the feature extractor E in the target domain. T The total loss, backpropagation training target domain feature extractor E T This enables the source domain label discriminator's output to accurately predict the label of the target domain data;

[0090] 3) Adjust the parameters of the domain discriminator D to improve its performance:

[0091] The expression for its loss function is:

[0092]

[0093] In the formula, M s (x) and M t (x) represent the source domain data and the target domain data, respectively. x represents the expectation of the logarithm of the probability that the target domain discriminator D classifies the target domain data as belonging to the target domain. t In order to sample the target domain data, Let x be the expectation of the logarithm of the probability that the source domain data is classified as the target source domain by the domain discriminator D. s For sampling of source domain data;

[0094] The domain discriminator D is trained using gradient descent and backpropagation. The classification error of the domain discriminator D in the target domain is minimized, while the difference between the label discriminators in the source and target domains is maximized, thereby achieving the purpose of domain adaptation.

[0095] 4) By iterating through steps 2) and 3) above, the feature alignment effect is gradually improved, and finally the optimal target domain feature extractor E is obtained. T Source domain label discriminator C S .

[0096] Specifically, the auxiliary diagnostic prediction module performs the following operations:

[0097] 1) Use the P300 induction module in the P300 induction and acquisition module to induce P300 EEG and amplify and filter it; use the data loading module to receive the P300 induction module online; and use the data preprocessing module to process the P300 data.

[0098] 2) The P300 data input source domain selection module after step 1) is used to select appropriate healthy subject data as the source domain;

[0099] 3) The P300 data processed in step 1) and the healthy subject data selected in step 2) are input into the auxiliary diagnosis training module. The domain discriminator and the target domain feature extractor are trained iteratively to enable the target domain features to generate feature representations similar to the source domain feature representations. The network parameters are adjusted by minimizing the total loss through backpropagation to obtain the optimal model. The trained target domain feature extractor and source domain label discriminator are used to obtain the auxiliary diagnosis results, consciousness level and rehabilitation prediction of patients with consciousness disorders.

[0100] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. An auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning, characterized in that, include: The P300 induction and acquisition module is used to induce and acquire P300 EEG data from patients and to preprocess the data. The source domain selection module is used to calculate similarity indices and select healthy subject data that are more similar to data from patients with impaired consciousness as the source domain, while data from patients with impaired consciousness is the target domain data. The auxiliary diagnostic training module uses source domain data to train a source domain feature extractor and a source domain label discriminator. It uses target domain data and source domain data to perform adversarial transfer, and adjusts the source domain feature extractor so that the features extracted in the target domain are in the same feature space as the features extracted in the source domain. The auxiliary diagnostic prediction module is used to assist in the diagnosis of patients with disorders of consciousness based on data, and to provide patients' level of consciousness and recovery prediction. The source domain selection module performs the following operations: a. Calculate the probability distribution of healthy subject samples and target domain samples based on spatiotemporal convolutional neural networks: A source domain feature extractor E based on a spatiotemporal convolutional neural network was trained using healthy subject sample data. S Source domain label discriminator C based on fully connected networks S : F (v) =E S (M v ,I S ) In the formula, M v It is an N×T matrix representing a sample of healthy subjects v, where v = {1, 2, ..., W}, W is the total number of healthy subjects, N is the sampling frequency of the preprocessed EEG data, T is the number of channels in the EEG data, and Θ S F represents the parameters of the source domain feature extractor. (v) The features extracted by the source domain feature extractor for healthy subject v consist of a 5-layer spatiotemporal convolutional neural network. The first layer is the input layer, the second layer is the spatial convolutional layer with a kernel size of T, which is the number of channels of the EEG data, the third layer is the temporal convolutional layer with a kernel size of 4, the fourth layer is the pooling layer with a filter size of (2,1) and the activation function is Tanh, and the fifth layer is the flattening layer, which compresses the input data from a multidimensional array into a one-dimensional array. L (v) =C S (F (v) ,i S ) In the formula, θ S L is the parameter of the source domain label discriminator. (v) The result of the source domain label discriminant P300 is the result of the healthy subject v being detected by the source domain label discriminant. The source domain label discriminant consists of a 4-layer fully connected neural network with the activation function being Softmax. Using source domain feature extractor E S Source domain label discriminator C S Calculate the probability distribution of each sample in the source and target domains as C. S (E S (M v ,Θ S ),θ S ); b. Calculate the sample similarity between the source and target domains based on Wasserstein distance: According to formula C S (E S (M v ,Θ S ),θ S Calculate the probability distribution of the source domain as P. S The probability distribution of the target domain is P T The Wasserstein distance between each healthy subject in the source domain and the target domain is expressed as: In the formula, Let v represent the probability distribution of the healthy subject. This represents the Wasserstein distance between the healthy subject v and the target domain. Let v represent the joint probability distribution of the healthy subject v and the target domain, γ represent any possible joint distribution, and (x,y) represent a sample x and y obtained from γ, where x is a sample from the source domain distribution and y is a sample from the target domain distribution. This represents the expected value of the Euclidean distance between x and y; Finally, healthy subjects with the smallest Wasserstein distance were selected as source domains.

2. The auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning according to claim 1, characterized in that, The P300 induction and acquisition module includes a P300 induction module, a data loading module, and a data preprocessing module, wherein: The P300 induction module uses visual and auditory induction to generate P300 signals, which are then amplified and filtered after being collected by an EEG acquisition device. The data loading module can read EEG data from the local machine, including MAT, CNT, and EEG formats, and can also receive EEG data transmitted online from the P300 evoked module. The data preprocessing module is used to remove noise from the EEG signal and downsample the EEG signal to reduce the amount of data and computational complexity. At the same time, the EEG signal is divided into several time windows to extract the P300 signal.

3. The auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning according to claim 2, characterized in that, The auxiliary diagnostic training module performs the following operations: 1) Initialize the target domain feature extractor E T With source domain feature extractor E S Same parameters; 2) To better learn the differences between the source and target domains, thereby improving the model's generalization ability and performance, a domain discriminator D is trained simultaneously with the feature extractor and label discriminator. Performance is improved by minimizing the classification loss function. The probability that a sample comes from the target domain is obtained from the domain discriminator D and used as a weighting factor to adjust the training direction of the feature extractor. H=D(F,θ d ),F=F S ∪F T In the formula, θ d For the parameters of the domain discriminator D, F S For source domain feature extractor E S Extract source domain features, F T For target domain feature extractor E T Extracting features from the target domain, F is F S With F T The union of H represents the discrimination result of the domain discriminator, which is composed of a 3-layer fully connected neural network with LogSoftmax as the activation function. Optimize the target domain feature extractor E by minimizing the domain alignment loss. T This makes the feature distributions of the source and target domains more similar, thereby improving the model's generalization ability in the target domain. The expression for the domain alignment loss function is: In the formula, N is the domain alignment loss between the source and target domains. t It is the amount of data in the target domain. It is the i-th source domain sample. It is the i-th target domain sample; Therefore, the target domain feature extractor E T The total loss is: Gradient descent is used to reduce the feature extractor E in the target domain. T The total loss, compared with the backpropagation training target domain feature extractor E T This enables the source domain label discriminator's output to accurately predict the label of the target domain data; 3) Adjust the parameters of the domain discriminator D to improve its performance: The expression for its loss function is: In the formula, M s (x) and M t (x) represent the source domain data and the target domain data, respectively. x represents the expectation of the logarithm of the probability that the target domain discriminator D classifies the target domain data as belonging to the target domain. t In order to sample the target domain data, Let x be the expectation of the logarithm of the probability that the source domain data is classified as the target source domain by the domain discriminator D. s For sampling of source domain data; The domain discriminator D is trained using gradient descent and backpropagation. The classification error of the domain discriminator D in the target domain is minimized, while the difference between the label discriminators in the source and target domains is maximized, thereby achieving the purpose of domain adaptation. 4) By iterating through steps 2) and 3) above, the feature alignment effect is gradually improved, and finally the optimal target domain feature extractor E is obtained. T Source domain label discriminator C S .

4. The auxiliary diagnostic system for patients with consciousness disorders based on P300 EEG and transfer learning according to claim 3, characterized in that, The auxiliary diagnostic prediction module performs the following operations: 1) Use the P300 induction module in the P300 induction and acquisition module to induce P300 EEG and amplify and filter it; use the data loading module to receive the P300 induction module online; and use the data preprocessing module to process the P300 data. 2) The P300 data input source domain selection module after step 1) is used to select appropriate healthy subject data as the source domain; 3) The P300 data processed in step 1) and the healthy subject data selected in step 2) are input into the auxiliary diagnosis training module. The target domain discriminator and the target domain feature extractor are trained iteratively to generate a feature representation similar to the source domain feature representation. The network parameters are adjusted by minimizing the total loss through backpropagation to obtain the optimal model. The trained target domain feature extractor and source domain label discriminator are used to detect the P300 signal of patients with impaired consciousness. The auxiliary diagnosis results, level of consciousness and rehabilitation prediction of the patients are obtained based on the detection accuracy of P300 and clinical assessment.

Images