Method, device and storage medium for recognizing attention deficit-related electroencephalogram signals

Abstract

The present invention discloses a method, device and storage medium for recognizing EEG signals associated with attention deficit. The recognition method comprises: obtaining original EEG signals corresponding to multiple types of attention deficit; using the EEG signals corresponding to multiple types of attention deficit as training samples to train a pre-constructed classification model, wherein the classification model comprises a multi-layer convolutional neural network and a multi-layer long short-term memory network connected in series; inputting the original EEG signal to be recognized into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data in combination with the model parameters of the classification model; and determining the generation time period and brain region of the EEG marker signal associated with attention deficit by comparing the generated spectrum graph and EEG topography graph of the gradient data and the original EEG signal. The present method can more fully extract the global and local fine-grained EEG signal features, improve the recognition performance, and locate the EEG marker activity at the same time.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical electroencephalogram signal processing, and in particular, relates to a method for recognizing electroencephalogram signals associated with attention deficit, a recognition device, a computer-readable storage medium and a computer device. Background Art

[0002] Attention is a cognitive process of selecting and focusing on relevant stimuli, and is a very important cognitive ability in people's daily lives. According to the clinical practice guidelines of the American Academy of Pediatrics, attention-deficit / hyperactivity disorder (ADHD) is the most common neurobehavioral disorder in childhood, which can seriously affect children's academic performance, mental health and social interaction. The most widely used ADHD diagnostic criteria are the Diagnostic and Statistical Manual of Mental Disorders, 5th edition (DSM-5) and the International Statistical Classification of Diseases and Related Health Problems, 10th edition (ICD-10). Untreated ADHD children may have language disorders and sleep disorders, while untreated adolescents may develop substance abuse, anxiety disorders, depression, etc. and have a higher risk of risky behaviors. At present, the understanding of children's attention process is limited to the field of medical clinical practice, and the existing assessment methods of children's attention cognitive function are characterized by subjective evaluation. The diagnostic criteria based on child behavioral psychology do play a role in diagnosis, but this subjective method of evaluating children's attention requires comprehensive evaluation of various information and requires experienced doctors to analyze and conclude. This method has the disadvantage of relying on parents' observations and opinions and lacking objective and accurate evaluation. It requires effective objective measurement indicators to reach scientific and accurate judgments. Therefore, early evaluation of children's attention and cognitive abilities can effectively take action to intervene in the growth of children with potential attention deficits. The evaluation method based on neurophysiological data can also serve as a basis for clinicians to diagnose ADHD and provide treatment.

[0003] Although attention deficit disorder in children is a highly prevalent and influential neurodevelopmental disorder, after many studies on its etiology, such as genetic factors, social and psychological factors, and neurophysiological factors, there is still a lack of sufficient evidence for the diagnosis and assessment of children's attention. Pathological connectivity believes that ADHD involves abnormalities in brain networks, including structural damage and reduced functional connectivity. There are no confirmed diagnostic neurobiological markers, and its research findings on brain functional networks lack application in daily clinical practice. Due to its real-time, easy-to-use, non-invasive and low-cost characteristics, electroencephalogram (EEG) is used in some brain neurological studies to analyze the relationship between brain wave activity and cognitive processes. The field of cognitive psychology has provided cognitive measurement information related to attention. The use of EEG technology can provide an understanding of the neural mechanisms related to attention deficits, and provide more information for the development of attention cognitive functions, psychopathology, and intervention.

[0004] Traditional machine learning methods have been used to evaluate the usefulness of functional imaging data, including P300 wave analysis of time domain signals, frequency domain power spectrum analysis, brain functional connectivity analysis and other techniques. The cognitive potential P300 reflects multi-level psychological activities such as sensation, memory, understanding, judgment and emotion, and is closely related to the cognitive processing of high-level psychological activities in the human brain. There have been many studies exploring the development of attention mechanisms through P300 induced experiments. Segalowitz et al. believed that P300 is generated when the stimulus is noticed and can be interpreted as the first major component of controlling attention. Crone believes that the P300 signal is related to executive attention functions, such as concentration and inhibition, making it a suitable target for studying the functional development of the adolescent brain. Barry et al. summarized the relationship between EEG signals in the resting state and ADHD, and believed that the increase in the relative power of theta, the decrease in the relative power of alpha and beta, and the increase in the theta / alpha ratio and the theta / beta ratio are all related to ADHD.

[0005] The existing machine learning methods usually involve researchers actively selecting EEG features to conduct statistical analysis on children's attention deficit data. The EEG features selected include event-related potentials in the time domain, and the amplitude and latency of these signals are calculated to observe the differences in EEG signals among different populations. When calculating frequency domain features, EEG signals are divided into δ, θ, α, β and other signals from low to high according to multiple spectrum segments, so as to further calculate the connection between spectrum energy and corresponding brain area activities. Brain network connection analysis technology is a number of nonlinear features related to the phase synchronization of EEG signals, and the corresponding brain function network is constructed based on electrode channels and phase connection edges. The disadvantages of the above methods are that each study has adopted independent data set experiments and selected feature calculations limited to a single direction of EEG. Some only focus on time domain features, while others focus on frequency domain features. The changes in EEG signals related to attention activities are complex, and the results obtained by different processing methods are very different. Its effect as a clinical neurobiological marker is unreliable and incomplete. There is no independent cognitive psychological activity in the brain area activities of the experimental subjects under specific experimental conditions. The most significant difference activities obtained based on the above EEG signal processing method may represent common activities related to a certain cognitive process, rather than potential activities that distinguish people with attention deficit from healthy people.

[0006] Existing deep learning methods can calculate all the information of a data set and automatically extract features without the need for a complex feature screening process. Using this method to identify and classify attention deficits can achieve good classification accuracy. However, most neural network models only use convolutional neural networks, which are mostly used to process two-dimensional features of images. In its feature extraction layer, the input of each neuron is connected to the local receptive field of the previous layer, and the local features are extracted. Once the local feature is extracted, its positional relationship with other features is also determined. The calculated multiple local feature maps form a network that can share learning weights. When processing EEG signal data, it is usually a compressed time series signal. Since EEG signals include activities of different frequencies, neuronal clusters propagate through rhythmic oscillations, and there is a certain connection between energy activities in different time periods, that is, subsequent EEG activities may be affected by the signals in the previous period. Multi-channel EEG electrodes also reflect the connection between different brain regions. Convolutional neural networks are not conducive to analyzing the dynamic spatiotemporal characteristics of EEG signals.

[0007] Moreover, most of the current deep learning methods can only output the classification index of the corresponding data set, that is, an index similar to the accuracy rate. The process of learning the data set is still a black box for the entire technology. It is impossible to understand which part of the EEG signal is learned during the training process, and it is impossible to determine the cognitive function activities related to attention deficit. EEG signals have multiple features in the time and frequency domain, and multi-channel electrodes represent different brain area activities. These factors lead to the complexity of EEG features. Therefore, it is necessary to visualize the learning process of the network model and its automatically extracted features in order to find specific and concrete EEG activities related to attention function. Summary of the invention

[0008] (I) Technical Problems to be Solved by the Present Invention

[0009] The technical problem solved by the present invention is: how to effectively learn the correlation characteristics of EEG signals in time sequence and how to locate the time period and brain area where EEG marker signals associated with attention deficit are generated.

[0010] (II) Technical solution adopted by the present invention

[0011] A method for identifying attention deficit-related electroencephalogram (EEG) signals, the method comprising:

[0012] Acquire original EEG signals corresponding to multiple types of attention deficits, wherein the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels;

[0013] Using EEG signals corresponding to multiple types of attention deficits as training samples to train a pre-built classification model, the classification model comprising a serially connected multi-layer convolutional neural network and a multi-layer long short-term memory network;

[0014] Inputting the original EEG signal to be identified into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model;

[0015] Generate a first spectrum map and a first electroencephalogram according to the gradient data, and generate a second spectrum map and a second electroencephalogram according to the original electroencephalogram signal to be identified;

[0016] The first spectrum map and the second spectrum map are compared, and the first electroencephalogram and the second electroencephalogram are compared to determine electroencephalogram marker signals associated with attention deficit and the time period and brain region in which the electroencephalogram marker signals are generated.

[0017] Preferably, the specific steps of calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model are:

[0018] The loss function value is calculated according to the following formula:

[0019] The gradient data is calculated according to the following formula:

[0020] Where M represents the number of categories, y c is the symbolic function of the category, y c If the predicted category in the one-hot encoding label of the sample is consistent with the true category, the value is 1, otherwise it is 0, p c It represents the predicted probability that the observed sample belongs to category c, grad represents the gradient data, and w c is the weight of sample i belonging to category c, k represents the true category of the sample, s c is the vector matrix output by this layer of network, where σ(s) is the score vector predicted by this layer of network, i.e., the output feature, y is the true vector, and x is the original sample vector value, i.e., the original EEG signal to be identified.

[0021] Preferably, the identification method further comprises:

[0022] The gradient data is normalized according to the following formula:

[0023]

[0024] Among them, Smap represents the normalized gradient data, min(grad) represents the minimum value of the gradient data, max(grad) represents the maximum value of the gradient data, and ε is the sample accuracy error.

[0025] Preferably, the method of comparing the first spectrum map with the second spectrum map and comparing the first EEG map with the second EEG map to determine the EEG marker signal associated with attention deficit and the generation area and generation time of the EEG marker signal is:

[0026] Comparing the energy distributions on the first spectrum graph and the second spectrum graph to determine an EEG marker signal associated with attention deficit and a time period in which the EEG marker signal is generated;

[0027] The brain region generating the electroencephalogram marker signal is determined according to the energy distribution of the electroencephalogram marker signal on the first electroencephalogram map and the second electroencephalogram map.

[0028] Preferably, the identification method further comprises:

[0029] After determining the time period and brain region where the electroencephalogram marker signal is generated, a time domain waveform diagram of the original electroencephalogram signal to be identified is drawn, and ERP waveform analysis is performed to obtain a neural biomarker for identifying attention deficit disorder.

[0030] The present application also discloses a device for identifying attention deficit-related EEG signals, the device comprising:

[0031] A signal acquisition unit, used to acquire original EEG signals corresponding to multiple types of attention deficits, wherein the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels;

[0032] A model training unit, used to train a pre-built classification model using EEG signals corresponding to multiple types of attention deficits as training samples, wherein the classification model includes a multi-layer convolutional neural network and a multi-layer long short-term memory network connected in series;

[0033] A gradient calculation unit, used for inputting the original EEG signal to be identified into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model;

[0034] A graphics generating unit, used to generate a first spectrum graph and a first electroencephalogram according to the gradient data, and to generate a second spectrum graph and a second electroencephalogram according to the original electroencephalogram signal to be identified;

[0035] The EEG localization unit is used to compare the first spectrum map with the second spectrum map and the first EEG topographic map with the second EEG topographic map to determine the EEG marker signal associated with attention deficit and the time period and brain area where the EEG marker signal is generated.

[0036] The present application also provides a computer-readable storage medium, which stores a program for identifying attention deficit-related EEG signals. When the program for identifying attention deficit-related EEG signals is executed by a processor, the above-mentioned method for identifying attention deficit-related EEG signals is implemented.

[0037] The present application also provides a computer device, which includes a computer-readable storage medium, a processor, and a recognition program for attention deficit-associated EEG signals stored in the computer-readable storage medium. When the recognition program for attention deficit-associated EEG signals is executed by the processor, the above-mentioned method for recognizing attention deficit-associated EEG signals is implemented.

[0038] (III) Beneficial effects

[0039] The present invention discloses a method, device, storage medium and equipment for recognizing attention deficit-related electroencephalogram signals, which have the following technical effects compared with the prior art:

[0040] According to the multi-channel and time-domain continuous characteristics of EEG signal data, the convolutional network layer and the long short-term memory network layer are serially connected to form a classification model, which can avoid excessive convolution compression of spatiotemporal features, more fully extract global and local fine-grained EEG signal features, and improve recognition performance; through the saliency map gradient data, combined with signal time-frequency analysis, the distribution and spatiotemporal position of EEG activities related to specific populations or attention deficit functions can be located, that is, the brain areas and time periods of these activities are determined, which is convenient for observing its effect as a neural marker. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a method for identifying attention deficit-related EEG signals according to Embodiment 1 of the present invention;

[0042] Figure 2 is a schematic diagram of a classification model according to the first embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of multi-channel EEG electrode distribution and event-related potential identification in Example 1 of the present invention;

[0044] Figure 4 The spectrum diagram generated according to the original EEG signals and gradient data of three types of attention deficits in the first embodiment of the present invention;

[0045] Figure 5 The EEG topographic map generated according to the original EEG signals and gradient data of three types of attention deficits in the first embodiment of the present invention;

[0046] Figure 6 This is a principle block diagram of a device for identifying attention deficit-related EEG signals according to Embodiment 2 of the present invention;

[0047] Figure 7 It is a schematic diagram of a computer device according to a fourth embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] Before describing the various embodiments of the present application in detail, the inventive concept of the present application is first briefly described: In the prior art, a deep learning neural network model is used to classify EEG signals of different attention deficits, that is, the attention deficit category corresponding to each EEG signal, but it is not clear which part of the EEG signal is used for classification. At the same time, the current deep learning neural network model is mainly based on a convolutional neural network, which is not conducive to learning the temporal characteristics of the EEG signal and affects the recognition accuracy. To this end, the present application provides a method for recognizing EEG signals associated with attention deficits, which uses a classification model composed of a convolutional neural network and a long short-term memory network to fully learn the spatial characteristics and temporal characteristics of the original EEG signal. After the classification model training is completed, the gradient data is obtained by using the saliency map calculation principle, and the EEG marker signal associated with the attention deficit is extracted according to the gradient data and the original EEG signal, and the time period and brain area of ​​the EEG marker signal are determined, which is conducive to the subsequent targeted analysis of the EEG marker signal.

[0050] Specifically, Figure 1 As shown, the method for identifying attention deficit-related EEG signals of the first embodiment includes the following steps:

[0051] Step S10: acquiring original EEG signals corresponding to multiple types of attention deficits, wherein the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels;

[0052] Step S20: using EEG signals corresponding to multiple types of attention deficit as training samples to train a pre-built classification model, the classification model comprising a serially connected multi-layer convolutional neural network and a multi-layer long short-term memory network;

[0053] Step S30: inputting the original EEG signal to be identified into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model;

[0054] Step S40: generating a first spectrum map and a first electroencephalogram according to the gradient data, and generating a second spectrum map and a second electroencephalogram according to the original electroencephalogram signal;

[0055] Step S50: Compare the first spectrum map with the second spectrum map and compare the first EEG map with the second EEG map to determine the EEG marker signal associated with attention deficit and the generation area and generation time of the EEG marker signal.

[0056] For example, in step S10, 56 EEG electrodes are used to collect EEG signals from different brain regions. The distribution of the EEG electrodes is as follows: Figure 3As shown in A, the specific collection process is the existing technology and will not be described in detail here.

[0057] Furthermore, if Figure 2 As shown, the convolutional neural network of the classification model of the first embodiment of the present invention includes two convolutional layers, namely the shallow Con2d and the deep DepthwiseCon2d. First, 50 convolution kernels of Con2d are used to perform one-dimensional filtering on the 50 time point length signals to extract the time features containing the 200ms window. The feature extraction here is to capture the frequency information at 5Hz (the sampling rate is set to 256Hz), and each window can output the corresponding time domain convolution feature map. Then, a 2-layer deep convolution is used to perform spatial filtering for each input feature map, so each time domain convolution kernel generates two spatial filters to obtain 100 compressed single-channel time domain waveforms. Afterwards, a batch normalization layer is used to make the convolutional network part faster and more stable, an exponential linear unit is used as an activation layer to solve the problems of gradient disappearance and gradient explosion, an average pooling operation is used to downsample the convolution output, and a dropout layer is used to prevent overfitting in training.

[0058] Furthermore, the long short-term memory network consists of two layers of LSTM networks. The single-channel time domain waveform after convolution is input into the first layer of LSTM network, and the input of the second layer of LSTM network is the output time series data of the previous layer. Each LSTM network has 10 units, each of which consists of a memory unit, an input gate, an output gate and a forget gate. The memory unit can memorize the signal value in any time period, and the three gates regulate the inflow and outflow of information to update the signal data. LSTM regulates the information flow in and out of the memory unit through the gate control unit, and updates the hidden state and unit state of the time series.

[0059] Further, in step S20, the classification model of the first embodiment is trained using a public children's attention deficit EEG dataset, and the input data of the model are channels and time points, which are 56 and 385 respectively in this dataset. In the training phase, the input batch size is 256, the learning rate is 0.001, and the training iteration is 300 epochs. The Adam optimizer and the classification cross entropy loss function are used in model compilation. Early stopping and learning rate reduction methods are also used to optimize the training process. In the above two callback methods, the loss parameter of the validation dataset is supervised. If it does not continue to decrease within 5 epochs, the learning rate will be multiplied by 0.5. If the loss no longer decreases within 20 epochs, the training is terminated.

[0060] Exemplarily, the classification model of the first embodiment uses a cross entropy loss function, and the weights are modified according to the number of samples of the data category. The loss calculation formula is:

[0061]

[0062] Where N represents the number of samples, M represents the number of categories, and y ic is the symbolic function of the category, which is selected according to the one-hot heat encoding label of sample i. If the predicted category is consistent with the true category, it takes the value 1, otherwise it takes 0, p ic It represents the predicted probability that the observed sample i belongs to category c.

[0063] Furthermore, conventional convolutional neural networks generally directly perform further convolution compression on the local spatiotemporal features obtained by shallow convolution, deeply extract more abstract features, and input the feature map into a fully connected layer to obtain the final EEG signal classification result. The signal data related to EEG activity involves the influence of the number of channels and the event-related time period. There is no need to use a network with too large a depth to compress its spatiotemporal features. The more graph convolution layers are used, the more the local features extracted by their nodes will tend to converge to the same vector, so that the information of the signal is over-squeezed into an area of ​​fixed size. Therefore, excessive convolution of EEG signal data will lead to a regression in the network's learning ability. In this embodiment, the channel distribution and signal time domain waveform of the EEG signal are obtained by preprocessing the corresponding data set, and the receptive field related parameters of the shallow convolution are adjusted to make the extracted feature space consistent with the number of channels and event-related duration, avoiding excessive compression of information. Figure 3 As shown, Figure 3 A in the figure is the EEG electrode distribution map of 56 channels, and B is the event-related potential identification map of different populations. With reference to these data, the spatiotemporal extraction parameters of the convolutional neural network are set to the same size. For example, 50 spatial channels and a 200ms time window are extracted. The channel depth is generally set to the information interaction between the two channels, and the time window corresponds to the time length of most event-related potentials. It is known that EEG cognitive activities are related to attention functions, including P300 waveforms, CNV negative waves, etc., and the activity range of these waveforms is about 200ms. After the convolutional layer structure is reasonably designed for each event-related potential activity, the time window feature map is input into the long short-term memory network, and the features of a series of waveform change information are further extracted, and the connection between waveforms in different time periods is analyzed, so as to extract information with richer time granularity. The final convolutional neural network architecture designed in this way is more reliable than the learning of directly deepening the convolutional layer.

[0064] Furthermore, in step S30, when the classification performance of the classification model reaches a higher standard, the model training can be stopped, and the original EEG signal to be identified is input into the trained classification model to obtain the output features of the last layer of long short-term memory network. The original EEG signal to be identified here is the EEG signal of a group of people with known attention deficit types. Furthermore, the gradient data is calculated based on the original EEG signal to be identified, the output features and the model parameters of the trained classification model. Specifically, the method of this step is:

[0065] Here, we need to calculate the gradient of the loss function from the output features of the last LSTM network layer of the model, which is also from the input data to the previous layer of the prediction matrix, to obtain the gradient data of the initial input signal data of the model after all network layers are trained, and finally perform the same time domain waveform calculation and spectrum conversion calculation on the gradient data. The specific steps are as follows:

[0066] The loss function value is calculated according to the following formula:

[0067] The gradient data is calculated according to the following formula:

[0068] Where M represents the number of categories, y c is the symbolic function of the category, y c If the predicted category in the one-hot encoding label of the sample is consistent with the true category, the value is 1, otherwise it is 0, p c It represents the predicted probability that the observed sample belongs to category c, grad represents the gradient data, and w c is the weight of the sample belonging to category c, k represents the true category of the sample, s c is the vector matrix output by this layer of network, where σ(s) is the score vector predicted by this layer of network, i.e., the output feature, y is the true vector, and x is the original sample vector value, i.e., the original EEG signal to be identified.

[0069] Furthermore, the identification method further includes normalizing the gradient data according to the following formula:

[0070]

[0071] Among them, Smap represents the normalized gradient data, min(grad) represents the minimum value of the gradient data, max(grad) represents the maximum value of the gradient data, and ε is the sample accuracy error.

[0072] Furthermore, the process of generating the first spectrum map and the first EEG map according to the normalized gradient data, and generating the second spectrum map and the second EEG map according to the original EEG signal to be identified are well known to those skilled in the art and will not be described in detail here. Figure 4 As shown, Figure 4 A in the figure represents the spectrum generated from the original EEG signals of the three types of attention deficits. Figure 4 B in represents the spectrum graph generated based on the three types of attention deficit representation gradient data. Figure 5 As shown, Figure 5 A in the figure represents the EEG topography map generated from the original EEG signals of the three types of attention deficits. Figure 5 B in FIG. 1 represents the EEG topography map generated based on the three types of attention deficit representation gradient data. HC represents healthy people, ADD represents attention deficit disorder people, and ADHD represents attention deficit hyperactivity disorder people.

[0073] In step S50, the method of comparing the first spectrum map with the second spectrum map and comparing the first electroencephalogram with the second electroencephalogram to determine the electroencephalogram marker signal associated with attention deficit and the generation area and generation time of the electroencephalogram marker signal is:

[0074] Compare the energy distribution on the first spectrum graph and the second spectrum graph to determine the EEG marker signal associated with attention deficit and the time period of the EEG marker signal; determine the brain region where the EEG marker signal is generated based on the distribution of the energy distribution of the EEG marker signal on the first EEG topographic map and the second EEG topographic map. When comparing the spectrum graph and the EEG topographic map corresponding to the gradient data and the original EEG signal, if the time period and brain region of the energy distribution of the two are different, the energy distribution corresponding to the gradient data shall prevail. The EEG marker signal here is the signal at the highest energy point of the entire spectrum graph and the EEG topographic map. Based on this, it can be determined that the signal at this point is the signal that the classification model pays the most attention to when identifying EEG signals for attention deficit classification.

[0075] Furthermore, after determining the time period and brain area in which the EEG marker signal is generated, a time domain waveform diagram of the original EEG signal to be identified is drawn, and ERP waveform analysis is performed to obtain a neurobiological marker for identifying attention deficit disorder.

[0076] The method for identifying attention-deficit-related EEG signals disclosed in the first embodiment, based on the multi-channel and time-domain continuity characteristics of EEG signal data, connects the convolutional network layer and the long short-term memory network layer in series to form a classification model, which can avoid excessive convolution compression of spatiotemporal features, and can more fully extract global and local fine-grained EEG signal features to improve recognition performance; through the saliency map gradient data, combined with signal time-frequency analysis, the EEG signals of the entire attention-deficit data set are visualized, the EEG activities that have a major impact on the model classification are displayed, and the distribution and spatiotemporal position of the EEG activities related to specific populations or attention-deficit functions are located, that is, the brain areas and time periods of these activities are determined, so as to facilitate the observation of their effects as neural markers.

[0077] like Figure 6 As shown, this second embodiment also discloses a device for identifying attention deficit-related EEG signals, and the identification device includes a signal acquisition unit 100, a model training unit 200, a gradient calculation unit 300, a graphics generation unit 400 and an EEG positioning unit 500. The signal acquisition unit 100 is used to acquire original EEG signals corresponding to multiple types of attention deficits, and the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels; the model training unit 200 is used to use the EEG signals corresponding to multiple types of attention deficits as training samples to train a pre-constructed classification model, and the classification model includes a multi-layer convolutional neural network and a multi-layer long short-term memory network connected in series; the gradient calculation unit 300 is used to input the original EEG signal to be identified into the trained classification model, obtain the output features of the last layer of the long short-term memory network, and calculate the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model; the graphic generation unit 400 is used to generate a first spectrum map and a first EEG topographic map according to the gradient data, and generate a second spectrum map and a second EEG topographic map according to the original EEG signal to be identified; the EEG localization unit 500 is used to compare the first spectrum map with the second spectrum map and the first EEG topographic map with the second EEG topographic map to determine the EEG marker signal associated with the attention deficit and the time period and brain area in which the EEG marker signal is generated. Among them, for more specific working processes of the signal acquisition unit 100, the model training unit 200, the gradient calculation unit 300, the graphics generation unit 400 and the EEG localization unit 500, please refer to the description of the first embodiment, which will not be repeated here.

[0078] Embodiment 3 also discloses a computer-readable storage medium, which stores a recognition program for attention deficit-related EEG signals. When the recognition program for attention deficit-related EEG signals is executed by a processor, the above-mentioned recognition method for attention deficit-related EEG signals is implemented.

[0079] Furthermore, the fourth embodiment also discloses a computer device, at the hardware level, such as Figure 7As shown, the computer device includes a processor 12, an internal bus 13, a network interface 14, and a computer-readable storage medium 11. The processor 12 reads the corresponding computer program from the computer-readable storage medium and then runs it, forming a request processing device at the logical level. Of course, in addition to the software implementation, one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logical unit, but can also be hardware or logic devices. The computer-readable storage medium 11 stores a recognition program for attention deficit-related EEG signals. When the recognition program for attention deficit-related EEG signals is executed by the processor, the above-mentioned recognition method for attention deficit-related EEG signals is implemented.

[0080] Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, modules of programs or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0081] The specific implementation methods of the present invention are described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and improved without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents. These modifications and improvements should also be within the scope of protection of the present invention.

Claims

1. A method for identifying attention deficit-related EEG signals, characterized in that: The identification method comprises: Acquire original EEG signals corresponding to multiple types of attention deficits, wherein the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels; Using EEG signals corresponding to multiple types of attention deficits as training samples to train a pre-built classification model, the classification model comprising a serially connected multi-layer convolutional neural network and a multi-layer long short-term memory network; Inputting the original EEG signal to be identified into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model; Generate a first spectrum map and a first electroencephalogram according to the gradient data, and generate a second spectrum map and a second electroencephalogram according to the original electroencephalogram signal to be identified; The first spectrum map and the second spectrum map are compared, and the first electroencephalogram and the second electroencephalogram are compared to determine electroencephalogram marker signals associated with attention deficit and the time period and brain region in which the electroencephalogram marker signals are generated.

2. The method for identifying attention deficit-related EEG signals according to claim 1, characterized in that: The specific steps of calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model are: The loss function value is calculated according to the following formula: The gradient data is calculated according to the following formula: Where M represents the number of categories, y c is the symbolic function of the category, y c If the predicted category in the one-hot encoding label of the sample is consistent with the true category, the value is 1, otherwise it is 0, p c It represents the predicted probability that the observed sample belongs to category c, grad represents the gradient data, and w c is the weight of the sample belonging to category c, k represents the true category of the sample, s c is the vector matrix output by this layer of network, where σ(s) is the score vector predicted by this layer of network, i.e., the output feature, y is the true vector, and x is the original sample vector value, i.e., the original EEG signal to be identified.

3. The method for identifying attention deficit-related EEG signals according to claim 2, characterized in that: The identification method further comprises: The gradient data is normalized according to the following formula: Among them, Smap represents the normalized gradient data, min(grad) represents the minimum value of the gradient vector, max(grad) represents the maximum value of the gradient vector, and ε is the sample precision error.

4. The method for identifying attention deficit-related EEG signals according to claim 1, characterized in that: The method of comparing the first spectrum map with the second spectrum map and comparing the first electroencephalogram with the second electroencephalogram to determine the electroencephalogram marker signal associated with attention deficit and the generation area and generation time of the electroencephalogram marker signal is: Comparing the energy distributions on the first spectrum graph and the second spectrum graph to determine an EEG marker signal associated with attention deficit and a time period in which the EEG marker signal is generated; The brain region generating the electroencephalogram marker signal is determined according to the energy distribution of the electroencephalogram marker signal on the first electroencephalogram map and the second electroencephalogram map.

5. The method for identifying attention deficit-related EEG signals according to claim 1, characterized in that: The identification method further comprises: After determining the time period and brain region where the electroencephalogram marker signal is generated, a time domain waveform diagram of the original electroencephalogram signal to be identified is drawn, and ERP waveform analysis is performed to obtain a neural biomarker for identifying attention deficit disorder.

6. A device for identifying attention deficit-related EEG signals, characterized in that: The identification device comprises: A signal acquisition unit, used to acquire original EEG signals corresponding to multiple types of attention deficits, wherein the original EEG signals corresponding to each type of attention deficit include EEG signals collected by EEG electrodes of multiple channels; A model training unit, used to train a pre-built classification model using EEG signals corresponding to multiple types of attention deficits as training samples, wherein the classification model includes a multi-layer convolutional neural network and a multi-layer long short-term memory network connected in series; A gradient calculation unit, used for inputting the original EEG signal to be identified into the trained classification model, obtaining the output features of the last layer of the long short-term memory network, and calculating the gradient data according to the original EEG signal to be identified, the output features and the model parameters of the trained classification model; A graphics generating unit, used to generate a first spectrum graph and a first electroencephalogram according to the gradient data, and to generate a second spectrum graph and a second electroencephalogram according to the original electroencephalogram signal to be identified; The EEG localization unit is used to compare the first spectrum map with the second spectrum map and the first EEG topographic map with the second EEG topographic map to determine the EEG marker signal associated with attention deficit and the time period and brain area where the EEG marker signal is generated.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a recognition program for attention deficit-related EEG signals, and when the recognition program for attention deficit-related EEG signals is executed by a processor, the recognition method for attention deficit-related EEG signals described in any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The computer device includes a computer-readable storage medium, a processor, and a recognition program for attention deficit-associated EEG signals stored in the computer-readable storage medium. When the recognition program for attention deficit-associated EEG signals is executed by the processor, the recognition method for attention deficit-associated EEG signals described in any one of claims 1 to 5 is implemented.

Images