A method and system for cross-time and space alignment and prediction analysis of electroencephalogram data

Abstract

The application discloses a kind of electroencephalogram data cross space-time alignment prediction analysis method and system.The method is by collecting EEG sequence, using continuous diffusion prediction network, combined with spatial interpolation technique and multilayer perceptron to capture the directional information of brain activity.Through convolutional neural network, the space-time evolution of electroencephalogram signal is simulated, the brain state is mapped back to EEG signal using convolutional decoder, and the disease probability is predicted by classifier.The application effectively simulates the continuity of electroencephalogram signal, and provides a new technical means for early diagnosis of mental and psychological diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of electroencephalogram data cross space-time alignment prediction analysis method, also simultaneously relates to corresponding electroencephalogram data cross space-time alignment prediction analysis system, belong to medical care information science technical field. BACKGROUND

[0002] In recent years, electroencephalogram data (EEG) as a kind of non-invasive neural signal monitoring means, is widely used in the research and diagnosis of mental and psychological diseases, such as Alzheimer's disease, depression and anxiety disorders etc.These diseases are usually accompanied by abnormal changes in brain activity, and EEG can provide high time resolution neural signals to help reveal the functional state of the brain.The propagation and interaction of electroencephalogram signals between different regions of the brain are continuous, rather than isolated;And the evolution of electroencephalogram signals over time is a continuous dynamic change.Therefore, simulating the continuity of electroencephalogram activity from both spatial and temporal dimensions can improve the effectiveness of electroencephalogram data analysis and achieve prediction of mental and psychological diseases such as dementia.However, due to the complex mechanism of mental and psychological diseases and significant individual differences, the analysis of EEG data faces many challenges, such as signal noise interference, inconsistent data distribution between individuals and across time, etc.These problems increase the difficulty of modeling the signal, especially in tasks that require cross-time and space data alignment and prediction, the effectiveness of traditional methods is often unsatisfactory.

[0003] Existing electroencephalogram data analysis techniques mainly fall into the following categories.Methods based on feature extraction and traditional statistical models rely on hand-crafted features such as time-domain features (mean, variance), frequency-domain features (power spectral density, frequency band energy) and nonlinear features (entropy, phase synchronization, etc.), and then use linear regression, support vector machines (SVM) and other models for classification and prediction.However, this type of method is highly dependent on domain knowledge, difficult to capture complex signal patterns, and has poor generalization, unstable performance in different individuals and environments.In addition, these methods are usually static modeling, and cannot reflect the dynamic characteristics of electroencephalogram signals.Methods based on time series models, such as autoregressive models (AR), autoregressive moving average models (ARMA) and hidden Markov models (HMM), analyze electroencephalogram signals by capturing temporal dependencies, suitable for continuous time series modeling.However, they have limited ability to handle complex electroencephalogram signals, which are non-stationary signals, and are sensitive to electrode noise or physiological noise.

[0004] Deep learning-based models have been widely applied in recent years, including convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), and graph convolutional networks (GCN). These models can automatically extract spatio-temporal features end-to-end and achieve certain success in classification, regression, and anomaly detection tasks. However, these methods have difficulties in aligning across time and individuals, and usually divide signals into discrete segments for processing, ignoring the continuity of signals. Graph model-based brain network analysis models the spatial relationship of electroencephalogram signals as a graph structure, such as a functional connectivity network or a structural connectivity network, and analyzes it through graph algorithms or graph neural networks (GNN). However, these methods usually assume that the network structure is static, making it difficult to handle time-varying network topology, and the computational cost is high when processing large-scale multi-channel data.

[0005] In summary, existing technologies have made progress in different tasks, but there are still problems such as difficulties in aligning across space and time and insufficient dynamic capture of continuity. SUMMARY

[0006] The primary technical problem to be solved by the present application is to provide a method for predicting and analyzing electroencephalogram data across space and time.

[0007] Another technical problem to be solved by the present application is to provide a system for predicting and analyzing electroencephalogram data across space and time.

[0008] To achieve the above technical purposes, the present application adopts the following technical solutions:

[0009] According to a first aspect of an embodiment of the present application, a method for predicting and analyzing electroencephalogram data across space and time is provided, comprising the following steps:

[0010] S1. Randomly sampling one or more EEG sequences from a collected EEG data training set;

[0011] S2. Constructing a continuous diffusion prediction network and training it according to the EEG sequences;

[0012] S3. Using the trained continuous diffusion prediction network to predict whether a subject has a specific mental or psychological disease.

[0013] Preferably, the step S2 comprises the following sub-steps:

[0014] S201. Establishing a continuous electroencephalogram field based on the EEG sequences through spatial interpolation;

[0015] S202. Using a multi-layer perceptron to capture the direction information of brain changes at each time from the entire EEG sequence;

[0016] S203, calibrate continuous electroencephalic field based on direction information, and establish initial electroencephalic state;

[0017] S204, explicitly simulate brain process through spatial diffusion process and direction information correction, deduce electroencephalic signal evolution, and predict brain state of next time step;

[0018] S205, based on electro signal obtained from brain state of each time step, build classifier, and output probability of suffering from specific mental and psychological disease;

[0019] S206, update parameters of continuous diffusion prediction network by using cyclic iteration training mode until model converges, and complete training of continuous diffusion prediction network.

[0020] Preferably, the classifier adopts a one-dimensional convolution layer to transform electro signal of multiple channels to one channel, and adopts full connection network and Softmax activation function to map electro signal sequence to specific probability distribution of whether suffering from mental and psychological disease.

[0021] Preferably, in the step S203, calibrating continuous electroencephalic field based on direction information includes combining predicted brain state h t and direction information e t to obtain combined brain state r t to reflect brain state of current time t+1.

[0022] Preferably, in the step S204, the initial brain state Z t established in the step S203 is used to estimate electroencephalic signal transmission of next time.

[0023] Preferably, the calculation process of the initial electroencephalic state Z t is as follows:

[0024] z t =s t +K t ⊙(r t -s t ), K t =σ(f conv (s t ||r t )),

[0025] r t =tanh(w h *h t +w e *e t +b),

[0026] where || is a concatenation operation, σ is a sigmoid function, tanh is a hyperbolic tangent activation function, f conv is a convolutional layer, w h , e and b are learnable parameters, S t is a continuous state electroencephalography field, K t is a Kalman gain, h t is a predicted brain state at previous time t, e t is a direction information extracted from the entire EEG sequence.

[0027] wherein preferably, the direction information e t is a direction information of brain changes at each time captured from the entire EEG sequence by a multilayer perceptron.

[0028] wherein preferably, in the step S204,

[0029] diffusion displacement [u i , u j ] = M(z t [p i , p j ]);

[0030] diffusion difference Ah of the electroencephalography signal = z' t - z t ,

[0031] z' t [p i , p j ] = z t [p i - u i , p j - u j ],

[0032] wherein [u i , u j ] is a diffusion displacement of the electroencephalography signal at any position [p i , p j ] in the continuous space estimated from the initial brain state z t by a convolutional neural network M;

[0033] z' t is a diffused electroencephalography field, each brain region position [p i , p j ] in the diffused electroencephalography field z' t is a position [p i - u i , p j - u j ] of the initial electroencephalography state z tThe diffusion is obtained.

[0034] wherein preferably, in the step S204, the direction information of the current time step of the brain electrical activity is inferred from the complete EEG sequence to correct the diffusion difference Ah:

[0035]

[0036] wherein h t+1 is the estimated brain electrical state at time step t+1 from the initial brain electrical state z t denotes a convolutional network, e t+1 denotes the direction information at time step t+1.

[0037] According to a second aspect of the embodiments of the present application, a brain electrical data spatio-temporal alignment prediction analysis system is provided, comprising a processor and a memory; wherein the memory is coupled with the processor, and is configured to store a computer program, which, when executed by the processor, causes the processor to implement the brain electrical data spatio-temporal alignment prediction analysis method described above.

[0038] Compared with the prior art, the present application has the following technical features: 1) the discrete electrode channel data is used to simulate the real brain electrical state, which improves the effect of brain electrical data analysis; 2) the initial brain electrical state is constructed by spatial interpolation and direction information calibration, the brain electrical state is improved by combining the dense prediction result of the previous time step and the overall trend of the brain electrical direction information, which enhances the time dependence, and the direction information ensures the global continuity; 3) the diffusion difference between the current time step and the subsequent time step is estimated, and the direction information of the brain electrical state is introduced from the entire brain electrical sequence to reduce the uncertainty of the estimation, thereby further constraining the time continuity and simulating the real brain electrical state change process. Therefore, the present application realizes the spatio-temporal alignment, and can perform brain electrical data spatio-temporal alignment prediction analysis. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A flowchart of the brain electrical data spatio-temporal alignment prediction analysis method provided by the first embodiment of the present application is shown in the figure;

[0040] Figure 2 A continuous diffusion prediction network for implementing the brain electrical data spatio-temporal alignment prediction analysis method in the embodiments of the present application is shown in the figure;

[0041] Figure 3 A structure diagram of the brain electrical data spatio-temporal alignment prediction analysis system provided by the second embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0042] ​The technical content of the present application will be described in detail below in combination with the drawings and specific embodiments.

[0043] Referring to Figure 1 and Figure 2 , the technical concept in the embodiment of the present application is: using an interpolation method to fill in the data gap between electrodes to generate a continuous electroencephalogram field; using diffusion difference estimation to improve the electroencephalogram state of the current time step based on the prediction result of the previous time step to construct a state transition model in the time sequence. Therefore, the continuity of brain movement is simulated from the spatial and temporal angles, the EEG data analysis effect is improved, and more accurate prediction of mental and psychological diseases (including but not limited to dementia, etc.) is realized.

[0044] First embodiment

[0045] As Figure 1 shown, the method for cross-spatiotemporal alignment and prediction analysis of electroencephalogram data provided by the first embodiment of the present application comprises at least the following steps:

[0046] S1, randomly sampling one or more EEG sequences from the collected electroencephalogram data training set;

[0047] S2, constructing a continuous diffusion prediction network and training according to the EEG sequence;

[0048] S3, using the trained continuous diffusion prediction network to predict whether the subject has a specific mental and psychological disease.

[0049] In an embodiment of the present application, a sequence of electroencephalogram data (EEG) is denoted as wherein N past L time steps of signals of N channels are included, denotes a real set. N represents the number of electrodes, i.e. the number of sensors used in electroencephalogram recording; L represents the number of time points, i.e. the length of the recorded time sequence.

[0050] As Figure 2As shown, the continuous diffusion prediction network for implementing the brain electrical data cross-time and space alignment prediction analysis method includes a collection module, a continuous calibration initialization module, a diffusion difference estimation module, a convolution decoder, and a classifier. The collection module first randomly samples one or more EEG sequences from the EEG data training set, which contains signals of multiple channels at different time points. The continuous calibration initialization module establishes a continuous electroencephalogram field based on the EEG sequence through a spatial interpolation technique. The observed data of the electrodes are projected into a two-dimensional grid, and the values of the grid points are estimated by interpolation of the nearest neighbor electrodes, so as to fill in the data gaps between the electrodes and generate a continuous electroencephalogram field. The diffusion difference estimation module uses a multilayer perceptron to capture the direction information of the brain changes at each time from the entire EEG sequence. These direction information reflects the trend of brain activity over time, which is crucial for understanding brain function and predicting changes in electroencephalogram signals. In the continuous diffusion prediction network, the convolution decoder is used to map the brain state back to the electrical signal of each electrode channel. It estimates the diffusion displacement of the electroencephalogram signal at any position in the continuous space from the initial brain state through a convolutional neural network, and infers the evolution of the electroencephalogram signal by explicitly simulating the brain process through diffusion difference estimation. Finally, based on the electrical signal obtained from the brain state at each time step, a classifier is constructed to output the probability of the subject suffering from a specific mental and psychological disease. The classifier uses a one-dimensional convolution layer to transform the electrical signals of multiple channels to one channel, and maps the electrical signal sequence to the probability distribution of whether suffering from mental and psychological diseases through a fully connected network and a Softmax activation function. The above modules are connected to each other through data flow and control flow to form a complete cross-time and space alignment prediction analysis network. The output of each module is used as the input of the next module to form a closed loop, ensuring the continuity and accuracy from the original EEG data to the final mental and psychological disease prediction.

[0051] In an embodiment of the present application, constructing the continuous diffusion prediction network includes the following sub-steps:

[0052] S201, the continuous calibration initialization module establishes a continuous electroencephalogram field based on the EEG sequence through spatial interpolation.

[0053] Specifically, the continuous calibration initialization module projects the observed data of each electrode into a two-dimensional grid according to the coordinate position of the electrode. For example, the value of each grid point in the two-dimensional grid is estimated by interpolation based on the nearest 10 electrodes, and the reciprocal of the distance between the electrode and the grid point is used as the weight for weighted summation. The two-dimensional grid generated by interpolation can be regarded as the initial continuous state electroencephalogram field S t .

[0054] S202, the direction information of the brain changes at each time is captured from the entire EEG sequence by using a multilayer perceptron.

[0055] Specifically, considering the global perspective, the entire EEG sequence and the coordinate information of each electrode, the time information of each time point are taken as auxiliary information Input into the multi-layer perception to capture the direction information e of the brain change at each time point t , which can be expressed by the formula:

[0056]

[0057] S203, calibrate the continuous electroencephalogram field based on the direction information and establish an initial electroencephalogram state.

[0058] The interpolation method based on distance reciprocal weight cannot fully represent the spatiotemporal complexity of the brain. Therefore, a Kalman filter is used to calibrate the continuous electroencephalogram field in step S201. In this example, first, the predicted brain state h t at the previous t time point is combined with the direction information e t extracted from the entire EEG sequence t ,

[0059] r t = tanh(w h *h t +w e *e t +b),

[0060] where tanh is the hyperbolic tangent activation function, w h , w e and b are learnable parameters. A learnable convolution is used to approximate the nonlinear Kalman filter. The learnable Kalman filter determines the weight when updating the continuous electroencephalogram field S t . K t is the Kalman gain, which combines the continuous state electroencephalogram field S t and the electroencephalogram state measurement r t . Based on the Kalman gain K t , the initial electroencephalogram state Z t can be calculated:

[0061] z t = s t + K t ⊙(r t -s t ), K t = σ(f conv (s t ||r t )),

[0062] where || is the concatenation operation, σ is the sigmoid function, and f conv is the convolution layer.

[0063] S204, using a diffusion difference estimation module, explicitly simulating the brain process through a spatial diffusion process and direction information correction to infer the evolution of the electroencephalogram signal and predict the brain state at the next time step.

[0064] The diffusion difference estimation module estimates the electroencephalogram signal transmission at the next time step according to the initial brain state z t established in the previous step. Specifically, a convolutional neural network M is used to estimate the diffusion displacement [u t , u i ] of the electroencephalogram signal at any position [p j , p i ] in the continuous space from the initial brain state z j . This process can be represented as [u i , u j ] = M(z t [p i , p j ]).

[0065] Based on the estimated electroencephalogram signal transmission, the brain process can be simulated. The diffusion difference Ah of the electroencephalogram signal can be represented as:

[0066] Ah = z' t - z t ,

[0067] z' t [p i , p j ] = z t [p i - u i , p j - u j ],

[0068] In the above formula, z' t is the diffused electroencephalogram field. Considering the influence of the brain process, each brain region position [p i , p j ] in the diffused electroencephalogram field z' t is obtained by diffusing the initial electroencephalogram state z t at position [p i - u i , p j - u j ]. The diffusion displacement [u i , u j ] is obtained from the motion estimation network M described above.

[0069] However, the single time step diffusion process is difficult to guarantee the continuity of the entire brain electrical process. Therefore, the diffusion difference estimation module infers the direction information of the brain electrical change at the current time from the complete EEG sequence collected to correct the diffusion difference. The complete diffusion difference estimation process can be written as follows:

[0070]

[0071] In the above formula, h t+1 is the brain state at time step t+1 estimated from the initial brain electrical state z t . represents a simple convolutional network, e t+1 represents the direction information at time step t+1.

[0072] Finally, the diffusion difference estimation module maps the brain state h t+1 back to the electrical signal of each electrode channel through the convolutional decoder D. This process can be described as

[0073] S205, based on the electrical signal obtained from the brain state at each time step, a classifier is constructed to output the probability of suffering from a specific mental and psychological disease (such as dementia)

[0074] In an embodiment of the present application, the classifier is implemented by a small network, which uses a one-dimensional convolutional layer to transform the electrical signals of multiple channels to one channel, and uses a fully connected network and a Softmax activation function to map the electrical signal sequence to the probability distribution of whether dementia. This embodiment only takes dementia as an example to represent mental and psychological diseases, but this does not constitute a limitation of the present application, and the present application can also be applied to other mental and psychological diseases, such as depression.

[0075] S206, update the continuous diffusion prediction network parameters using a cyclic iterative training method until the model converges, and complete the training of the continuous diffusion prediction network.

[0076] In an embodiment of the present application, the training process uses an iterative generation method to predict the brain electrical state at the next time step based on the information at the current time step, and then predicts the dementia probability based on all time step brain electrical states. The loss function in the training process is as follows:

[0077]

[0078] Where N is the number of samples, and C is the number of categories. is the probability that the model predicts that sample i belongs to category j, and y ij is the true label (usually one-hot encoded, but not limited to this).

[0079] Next, in step S3, the continuous diffusion prediction network that has been trained through the foregoing steps is used to analyze the EEG data of the subject to predict whether the subject has a specific mental and psychological disease. Specifically, this step involves inputting the EEG sequence of the subject into the well-trained continuous diffusion prediction network. The network first initializes a continuous electroencephalographic field through the continuous calibration initialization module using a spatial interpolation technique. This process fills in the data gaps between electrodes and provides a basis for spatial continuity for subsequent time series analysis.

[0080] Then, the multilayer perceptron in the above network captures the directional information of brain changes at each time point in the EEG sequence, which is crucial for understanding the trend of brain activity and predicting changes in electroencephalographic signals. Subsequently, the diffusion difference estimation module combines the prediction results of the previous time step and the directional information to estimate the diffusion displacement of the electroencephalographic signal in the continuous space through a convolutional neural network, simulate the brain process, infer the evolution of the electroencephalographic signal, and predict the brain state at the next time step.

[0081] The convolutional decoder then maps these brain states back to the electrical signals of each electrode channel, restoring or reconstructing the signals. Finally, the classifier transforms the electrical signals of multiple channels to 1 channel through a 1-dimensional convolutional layer based on the electrical signals obtained from the brain state at each time step, and maps the electrical signal sequence to the probability distribution of whether the subject has a specific mental and psychological disease using a fully connected network and a Softmax activation function. In this way, the classifier outputs a probability value representing the likelihood of the subject having a specific mental and psychological disease, thereby achieving prediction of the mental and psychological health status of the subject.

[0082] It should be noted that the EEG data spatio-temporal alignment prediction analysis method based on the continuous diffusion prediction network provided by the embodiments of the present application achieves spatio-temporal alignment through the following technical means:

[0083] 1) Spatial interpolation is used: the continuous calibration initialization module performs spatial interpolation between electrodes to create an initial spatially continuous electroencephalographic state. This interpolation method can fill in the data gaps between electrodes and generate a continuous electroencephalographic field, providing a spatial basis for temporal alignment. Because the neural activity of the brain is transmitted through the connections between neurons, this process is continuous in space, so by simulating this spatial diffusion process, the true pattern of brain activity can be more accurately captured.

[0084] 2) Incorporating the prediction results of the previous time step to infer the evolution of the electroencephalogram signal: The diffusion difference estimation module utilizes the prediction results of the previous time step to improve the electroencephalogram state at the current time step. This step takes into account the dependency between adjacent time points in the time series, i.e., the current state is largely influenced by the previous state, thereby enhancing the temporal dependency. Because the prediction results of the previous time step and the observation data of the current time step are considered, a state transition model in the time series can be constructed, so this helps to capture the dynamic changes of the electroencephalogram signal over time (temporal continuity); because the causal relationship and temporal dependency in the time series are combined, the electroencephalogram state at the next time step can be more accurately predicted.

[0085] 3) Introducing directional information: The diffusion difference estimation module incorporates electroencephalogram directional information, which reflects the trend of the electroencephalogram signal over time. By incorporating directional information into the model, global continuity can be ensured, i.e., maintaining consistent evolution direction in the entire time series, thereby achieving temporal alignment. According to the law of neural activity of the brain, directional information reflects the directionality of neural activity propagation, providing additional clues about the evolution trend of the electroencephalogram signal, which helps to correct and optimize the prediction model, which is crucial for understanding brain function and predicting changes in the electroencephalogram signal.

[0086] 4) Using a Kalman filter to calibrate the interpolation results: The Kalman filter is a recursive filter that can combine prediction and observation data to estimate the state of a dynamic system. The Kalman filter can fuse the prediction results of the previous time step and the current observation data (electroencephalogram signal), thereby achieving more accurate alignment in time. Moreover, the present embodiment uses a learnable convolutional layer to approximate the nonlinear Kalman filter, which allows the model to learn how to best combine the predicted brain state and directional information to update the current electroencephalogram state. This learning mechanism enables the model to adapt to different data characteristics and changes, further improving the accuracy of temporal alignment.

[0087] In summary, the continuous calibration initialization module achieves temporal alignment of electroencephalogram data through spatial interpolation, consideration of temporal dependency, introduction of directional information, and application of the Kalman filter. Based on the spatial diffusion process and directional information, the uncertainty in prediction can be reduced, and the stability and reliability of the prediction of the evolution of the electroencephalogram signal can be improved.

[0088] In addition, the spatial diffusion process ensures the consistency of the electroencephalogram signal in space, allowing for consideration of the interaction and influence between different brain regions, capturing the synergy and information exchange between different brain regions, thereby analyzing and predicting the functional state changes of the brain as a whole, avoiding fragmented or static analysis methods, thereby better adapting to individual differences and reducing the impact of noise, maintaining high prediction accuracy in different individuals and different conditions.

[0089] To verify the actual effect of the electroencephalogram data cross-time and space alignment prediction analysis method based on the continuous diffusion prediction network provided by the embodiments of the present application, the inventors conducted experiments on real EEG data. Specifically, the data set includes 36 Alzheimer's disease (AD) patients, 23 frontotemporal dementia (FD) patients, and 29 healthy control (HC) subjects. The data set has 19 channels, and the original sampling rate is 500 Hz. The sampling rate of each test is reduced to 256 Hz, and it is segmented into non-overlapping 1-second samples and 256 time stamps, while discarding any samples shorter than 1 second. This process produces 69,752 samples. For the splitting of the training set, the validation set and the test set, 60%, 20% and 20% of the total number of subjects and their corresponding samples are respectively allocated to the training set, the validation set and the test set.

[0090] The experimental results are shown in Table 1:

[0091] Table 1: Comparison of experimental results

[0092] Method Accuracy Transformer 50.47％ Autoformer 45.25％ FEDformer 46.30％ PatchTST 44.37％ Nonformer 49.95％ The present invention 51.46％

[0093] As can be seen from Table 1, the present application achieves better performance in predicting the accuracy rate of dementia patients in real EEG data sets.

[0094] Second embodiment

[0095] Based on the above-mentioned electroencephalogram data cross-time and space alignment prediction analysis method, the third embodiment of the present application provides an electroencephalogram data cross-time and space alignment prediction analysis system. As shown in Figure 3 The system includes one or more processors and a memory. The memory is coupled to the processor and is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the electroencephalogram data cross-time and space alignment prediction analysis method as described in the above embodiments.

[0096] The processor is configured to control the overall operation of the system to complete all or part of the steps of the above-mentioned electroencephalogram data cross-time and space alignment prediction analysis method. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a field programmable logic gate array (FPGA), an application specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. The memory is configured to store various types of data to support the operation of the system, which can include, for example, instructions for operating any application or method on the system, and application-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, etc.

[0097] In one example embodiment, the system can be specifically implemented by a computer chip or entity, or by a product having certain functions, for executing the above-mentioned electroencephalogram data cross-time and space alignment prediction analysis method, and achieving the technical effects consistent with the above-mentioned method. A typical embodiment is a computer, a server. Specifically, the computer can be, for example, a personal computer, a laptop computer, an in-vehicle human-machine interaction device, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0098] In another example embodiment, the present application also provides a computer readable storage medium including program instructions, which, when executed by a processor, implement the steps of the electroencephalogram data cross-time and space alignment prediction analysis method in any one of the above-mentioned embodiments. For example, the computer readable storage medium can be the above-mentioned memory including program instructions, which can be executed by the processor of the system to complete the above-mentioned electroencephalogram data cross-time and space alignment prediction analysis method, and achieve the technical effects consistent with the above-mentioned method.

[0099] It should be noted that the above-mentioned embodiments are only illustrative. The technical solutions of each embodiment can be combined, and the order of each step can be changed, all of which are within the scope of protection of the present application.

[0100] The electroencephalogram data cross-time and space alignment prediction analysis method and system provided by the present application are described in detail above. Any obvious modification made by a person skilled in the art without departing from the essential content of the present application will constitute an infringement of the patent right of the present application and will bear the corresponding legal responsibility.

Claims

1. A method for cross-temporal and spatial alignment prediction analysis of electroencephalogram (EEG) data, characterized in that... Includes the following steps: S1. Randomly sample one or more EEG sequences from the collected EEG data training set; S2. Construct a continuous diffusion prediction network and train it based on the EEG sequence; S3. Using a trained continuous diffusion prediction network, predict whether a subject suffers from a specific mental illness. Step S2 includes the following sub-steps: S201. The continuous calibration initialization module establishes a continuous electroencephalogram (EEG) field based on the EEG sequence through spatial interpolation. S202. Use a multilayer perceptron to capture directional information of brain changes at each moment in the entire EEG sequence; S203. Based on directional information, calibrate the continuous EEG field and establish the initial EEG state; S204. Using the diffusion difference estimation module, the brain process is explicitly simulated through spatial diffusion process and directional information correction, the evolution of EEG signals is inferred, and the brain state at the next time step is predicted. S205. Based on the electrical signals obtained from the brain state at each time step, construct a classifier and output the probability of having a specific mental illness. S206. Update the parameters of the continuous diffusion prediction network using a cyclic iterative training method until the model converges, thus completing the training of the continuous diffusion prediction network.

2. The method for cross-temporal alignment prediction and analysis of EEG data as described in claim 1, characterized in that: The classifier uses a single 1D convolutional layer to transform electrical signals from multiple channels to a single channel, and employs a fully connected network and a Softmax activation function to map the electrical signal sequence to a probability distribution of whether or not a person suffers from a specific mental illness.

3. The method for cross-temporal alignment prediction and analysis of EEG data as described in claim 1, characterized in that: In step S203, calibrating the continuous electroencephalogram (EEG) based on direction information includes combining the brain state h predicted at the previous time t. t and directional information e t The brain state r after merging was obtained t This reflects the brain state at the current moment t+1.

4. The method for cross-temporal alignment prediction and analysis of EEG data as described in claim 3, characterized in that... In step S204, the initial brain state Z established in step S203 is utilized. t It estimates the transmission of brain signals in the next moment.

5. The method for cross-temporal alignment prediction and analysis of EEG data as described in claim 4, characterized in that... The initial EEG state Z t The calculation process is as follows: Z t =s t +K t ⊙(r t -s t ),K t =σ(f conv (s t ||r t )), r t =tanh(w h *h t +w e *e t +b), Where || represents the connection operation, σ is the sigmoid function, tanh is the hyperbolic tangent activation function, and f conv For convolutional layers, w h w e b and S are learnable parameters. t For continuous state EEG fields, K t For Kalman gain, h t e is the brain state predicted at the previous time t. t This refers to the directional information extracted from the entire EEG sequence.

6. The method for cross-temporal alignment prediction analysis of EEG data as described in any one of claims 1 to 5, characterized in that: The direction information e t It is the directional information of brain changes at every moment captured by the multilayer perceptron from the entire EEG sequence.

7. The method for cross-temporal alignment prediction analysis of EEG data as described in any one of claims 3 to 5, characterized in that... In step S204, Diffusion displacement [u] i u j ]=M(z t [p i ,p j ]); The diffusion difference of EEG signals Δh=z′ t -z t , z′ t [p i ,p j ]=z t [p i -u i ,p j -u j ], Among them, [u i ,u j To utilize a convolutional neural network M to analyze the initial brain state z t Estimate any position in continuous space [p] i p j The diffusion displacement of the EEG signal; z′ t The diffused brain electric field, the diffused brain electric field z′ t Location of each brain region in [p] i p j ] is determined by the position [p] i -u i p j -u j The initial EEG state z t Obtained through diffusion.

8. The method for cross-temporal alignment prediction analysis of EEG data as described in claim 7, characterized in that... In step S204, the directional information of the EEG changes at the current moment is inferred from each complete EEG sequence to correct the diffusion difference Δh: Among them, h t+1 For the initial EEG state z t The estimated EEG state at time step t+1 is obtained. Represents a convolutional network, e t+1 This indicates the direction information at time step t+1.

9. A spatiotemporal alignment prediction and analysis system for electroencephalogram (EEG) data, characterized in that... It includes a processor and a memory; wherein the memory is coupled to the processor and is used to store a computer program, which, when executed by the processor, enables the processor to implement the EEG data cross-temporal alignment prediction analysis method according to any one of claims 1 to 8.

Images