Method for training brain network collaborative classification model
By pre-processing and slicing processing of EEG signals, a connection matrix is built and a brain network collaborative classification model is trained, which solves the problem of low classification accuracy of brain network signals, and achieves more accurate capture of brain region synergistic changes and improves classification accuracy.
Patent Information
- Application Number
- CN202510449814.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-08-05
AI Technical Summary
The accuracy of signal classification of brain networks in the prior art is low, making it difficult to accurately explore the dynamic changes of brain networks.
By preprocessing and slicing the pre-acquisitioned EEG signals, a connection matrix is constructed, and a collaborative classification model is input to the brain network for training, and the model parameters are adjusted until the matching rate reaches the specified threshold.
It significantly improves the classification accuracy of EEG signals in the task and enhances the ability to identify task-related EEG signals characteristics.
Smart Images
Figure CN120429705A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a training method for a brain network collaborative classification model. Background Art
[0002] Autonomous Sensory Meridian Response (ASMR) is a unique physiological phenomenon, typically triggered in sensitive individuals by external auditory or visual stimuli. This response often stems from low-frequency, repetitive auditory stimuli, such as whispers, rustling sounds, and tapping. These stimuli trigger a uniquely pleasurable experience, described as a "tingling sensation" or a slight "tremor," often accompanied by a sense of physical and mental relaxation. The sensation of Autonomous Sensory Meridian Response (ASMR) typically begins around the head and may spread along the body to the neck and even the spinal region. Although the mechanisms of ASMR are not yet fully understood, many studies have suggested that it may be related to the nervous system's sensitivity to specific external stimuli. ASMR experiences can reduce stress, improve mood, and even have a positive impact on sleep quality, and therefore have received increasing attention in the fields of mental health and mood regulation in recent years.
[0003] Raw scalp electroencephalogram (EEG) signals are an important tool for reflecting brain function by recording surface neural activity from the scalp. However, signal processing is challenging. EEG signals are typically divided into distinct frequency bands, each associated with specific brain functions. Delta waves (0.5-4 Hz) typically predominate during deep sleep and are associated with brain tissue repair and recovery. Theta waves (4-7 Hz) are associated with light sleep, relaxation, and emotional memory processing. Alpha waves (8-13 Hz) reflect a state of wakefulness yet relaxation, particularly pronounced with eyes closed, and are associated with attention regulation and resting states. Beta waves (13-30 Hz) are associated with focus, alertness, and information processing, while gamma waves (30 Hz and above) are involved in higher-level cognitive functions such as sensory integration and memory processing. Overall, EEG signals are challenging to process and analyze due to their low signal-to-noise ratio, multi-band overlap, and susceptibility to interference. Effective techniques are needed to extract features that more accurately reveal the functional states they reflect.
[0004] EEG microstates refer to a pattern of electrical activity in the brain that remains stable for a short period of time, and can be regarded as the basic unit of the functional network when the brain is working. Each microstate usually lasts for tens to hundreds of milliseconds, and during this period, the spatial distribution of electrical signals corresponding to different brain regions remains basically unchanged. Different microstates may correspond to different brain functions, such as perception, memory, or attention regulation. Studies have found that the duration, frequency of occurrence, and switching mode of microstates can reflect the working state of the brain and even help identify some pathological characteristics.
[0005] Therefore, how to improve the accuracy of brain network signal classification and accurately explore the dynamic changes of brain networks has become a technical problem that needs to be solved urgently. Summary of the Invention
[0006] (1) Technical issues to be resolved
[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a training method for a brain network collaborative classification model, which solves the technical problem of low accuracy in brain network signal classification.
[0008] (2) Technical solution
[0009] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:
[0010] In a first aspect, an embodiment of the present invention provides a training method for a brain network collaborative classification model, comprising:
[0011] S100, preprocessing the EEG signals of at least 16 channels collected in advance for each user to obtain the preprocessed EEG signals and attribute information of the EEG signals in a specified frequency band;
[0012] S200, slicing the preprocessed EEG signal and removing interference signals to obtain valid EEG signal segments and attribute information of the EEG signal segments; the attribute information of the EEG signal segments includes: stimulus classification information and brain region location information;
[0013] S300, constructing a connection matrix for training a brain network collaborative classification model based on each EEG signal segment and the brain region location information of the segment;
[0014] S400. Input the constructed connection matrix into the brain network collaborative classification model to obtain the output result, match the stimulation classification information of each output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
[0015] The training method of the brain network collaborative classification model proposed in the embodiment of the present invention preprocesses and slices the pre-collected EEG signals, constructs a connection matrix, and inputs the connection matrix into the brain network collaborative classification model for training. Compared with the existing technology, it can more accurately capture the collaborative changes between brain regions under task states, enhance the classification model's ability to recognize task-related EEG signal features, and achieve the technical effect of significantly improving the classification accuracy of EEG signals in tasks.
[0016] Optionally, the S100 includes:
[0017] A 48Hz to 52Hz notch filter is used to remove the power frequency interference in the EEG signal of each channel, and a 0.5Hz to 100Hz bandpass filter is used to remove the noise of the EEG signal of each channel to obtain a preprocessed EEG signal of 4Hz to 49Hz. The attribute information of the EEG signal of each channel includes: the brain area location information of the EEG signal and the stimulation classification information of each time period of the EEG signal.
[0018] Optionally, the S200 includes:
[0019] S201, dividing the pre-processed EEG signal of each channel into designated time segments according to the time axis to obtain multiple non-overlapping EEG signal segments,
[0020] S202, removing predetermined artifact interference signals from all multiple non-overlapping EEG signal segments to obtain valid EEG signal segments;
[0021] Among them, the brain area location information of the valid EEG signal segments corresponding to the O1 channel, O2 channel, Oz channel, Cz channel, P5 channel, P6 channel, PO3 channel, and PO4 channel is the visual network location;
[0022] The brain region location information of the valid EEG signal segments corresponding to the C3 channel, C4 channel, P5 channel, P6 channel, F5 channel, and F6 channel is the auditory network location;
[0023] The brain region location information of the valid EEG signal segments corresponding to the Fz channel, Cz channel, Pz channel, C3 channel, C4 channel, P5 channel, and P6 channel is the frontoparietal network location.
[0024] Optionally, the S300 includes:
[0025] For all valid EEG signal segments in the same brain region and in the same segmentation time period, the Pearson correlation coefficient is used to calculate the linear correlation between any two electrodes in the brain region, and the connection matrix of the brain region is constructed.
[0026] Specifically, the signal data points in the effective EEG signal segment corresponding to electrode X form an electrical segment sequence X = [x1, x2, ..., x n ]; the signal data points in the effective EEG signal segment corresponding to electrode Y constitute the electrode segment sequence Y = [y1, y2, ..., y n ], its linear correlation r XY for,
[0027]
[0028] Where n is the number of signal data points collected in each EEG signal segment, x i is the i-th signal data point collected in the electrode segment sequence X, y i is the i-th signal data point collected in the electrode segment sequence Y, is the average value of the signal data points in the electrode segment sequence X; is the average value of the signal data points in the electrode segment sequence Y;
[0029] For each brain region, based on the r between any two electrodes in each brain region XY , construct the connection matrix of each brain region;
[0030] Get the connection matrix C of the visual network position 视觉 , the connection matrix C of the auditory network position 听觉 , the connection matrix C of the frontoparietal network position 额顶 .
[0031] Optionally, before S400, the method further includes:
[0032] Constructing a brain network collaborative classification model,
[0033] The brain network collaborative classification model includes: an input layer that receives a connection matrix, a convolutional layer corresponding to each input layer, a pooling layer corresponding to each convolutional layer, a splicing layer corresponding to all pooling layers, a fully connected layer corresponding to the splicing layer, and an output layer;
[0034] There are three input layers, which are used to receive the connection matrix of the visual network position, the connection matrix of the auditory network position, and the connection matrix of the frontoparietal network position respectively;
[0035] The convolution layer and pooling layer of each input layer process the features of the connection matrix of the input layer in sequence and output the processed features;
[0036] The concatenation layer is used to concatenate the features output by all pooling layers and output them through the fully connected layer and the output layer.
[0037] Optionally, the S400 includes:
[0038] The connection matrices used to train the brain network collaborative classification model include: the connection matrices of the three brain regions of users who are sensitive to ASMR, and the connection matrices of the three brain regions of users who are not sensitive to ASMR. The users are equal in gender and have no neurological diseases.
[0039] The stimulus classification information of the output results includes: audio-visual stimulus classification information, auditory stimulus classification information, white noise stimulus classification information and resting state classification information.
[0040] Optionally, the S201 includes:
[0041] The EEG signal is divided into N 2.5s segments according to the acquisition time axis.
[0042]
[0043] Among them, T is the total duration of the EEG signal after preprocessing, Indicates rounding down;
[0044] Each EEG signal segment z j (τ)=z(t j +τ), where j = 1, 2, ..., N represents the jth slice, t j =(j-1)×2.5, which represents the start time of the jth slice, and τ∈[0,2.5) represents the relative time within the slice.
[0045] Optionally, the brain network collaborative classification model is a DenseNet model.
[0046] In a second aspect, an embodiment of the present invention provides a training device for a brain network collaborative classification model, comprising:
[0047] A data processing module is configured to pre-process at least 16 channels of EEG signals collected from each user to obtain pre-processed EEG signals and EEG signal attribute information within a specified frequency band; and to slice the pre-processed EEG signals and remove interference signals to obtain valid EEG signal segments and EEG signal segment attribute information; the EEG signal segment attribute information includes stimulus classification information and brain region location information;
[0048] A connection matrix acquisition module is used to construct a connection matrix for training a brain network collaborative classification model based on each EEG signal segment and the brain region location information of the segment;
[0049] The model training module is used to input the constructed connection matrix into the brain network collaborative classification model, obtain the output result, match the stimulation classification information of the output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
[0050] In a third aspect, an embodiment of the present invention provides an electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program in the memory, specifically performing the steps of the training method of the brain network collaborative classification model described in the first aspect above.
[0051] (3) Beneficial effects
[0052] The beneficial effect of the present invention is that the training method of the brain network collaborative classification model of the present invention pre-processes and slices the pre-collected EEG signals, constructs a connection matrix, and inputs the connection matrix into the brain network collaborative classification model for training. Compared with the existing technology, it can more accurately capture the collaborative changes between brain regions under task states, enhance the classification model's ability to recognize task-related EEG signal features, and achieve the technical effect of significantly improving the classification accuracy of EEG signals in tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 Schematic diagram of a flow chart of a training method for a brain network collaborative classification model provided in an embodiment;
[0054] Figure 2 This is a position distribution diagram of electrode channels used in a training method for a brain network collaborative classification model provided in an embodiment.
[0055] Figure 3 Schematic diagram of the framework of the DenseNet model in the embodiment. DETAILED DESCRIPTION
[0056] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.
[0057] Convolutional Neural Network (CNN) is a deep learning model that is widely used in the fields of image and signal processing. It effectively extracts spatial features from data through convolution operations, and uses a multi-layer structure to gradually abstract and express the features at a high level. The advantage of CNN is that it can automatically learn and extract features without the need to manually design complex feature extraction algorithms, so it performs well in pattern recognition and classification tasks. CNN is very suitable for feature extraction for spatial distribution analysis of EEG microstates. By combining the experience of spatiotemporal analysis of microstates, the embodiment of the present invention constructs a knowledge-driven model based on convolutional neural networks to explore the distribution characteristics and dynamic change laws of microstates, thereby revealing the intrinsic mechanism of specific brain functional activities such as ASMR activities in a deeper level.
[0058] The embodiment of the present invention is based on scalp EEG signals, divides different brain area networks according to the prior knowledge of EEG microstates, and constructs a classification model based on the collaborative input of different brain area networks to extract effective network synergy information; its purpose is to combine the spatiotemporal analysis experience of microstates with the convolutional neural network model to construct a knowledge-driven deep learning classification model, accurately explore the dynamic change laws of brain networks, and thus deeply reveal the intrinsic mechanism of specific brain functional activities (such as ASMR).
[0059] Example 1
[0060] like Figure 1 As shown, this embodiment provides a training method for a brain network collaborative classification model, including:
[0061] S100 , pre-processing the EEG signals of at least 16 channels collected in advance for each user to obtain the pre-processed EEG signals of a specified frequency band and attribute information of the EEG signals.
[0062] Specifically, if Figure 2 As shown, according to the international 10 / 10 lead system, the electrode channels that can be selected are: Fp1, Fp2, F5, F6, C3, C4, P5, P6, PO3, PO4, O1, O2, Fz, Cz, Pz, Oz, and the acquisition rate of the EEG signal is 1200 Hz.
[0063] It should be noted that this step is to remove power frequency interference and noise and obtain the frequency band corresponding to the feature that needs attention.
[0064] S200, slicing the preprocessed EEG signal and removing interference signals to obtain valid EEG signal segments and attribute information of the EEG signal segments; the attribute information of the EEG signal segments includes: stimulus classification information and brain area location information.
[0065] It should be noted that removing interference signals means removing abnormal EEG signal fragments caused by artifacts such as eye movements and electromyography to ensure signal quality. In particular, the EEG signal fragments of the two electrode channels Fp1 and Fp2, which are severely affected by electrooculography, are removed.
[0066] S300: Based on each EEG signal segment and the brain region location information of the segment, a connection matrix for training a brain network collaborative classification model is constructed.
[0067] S400. Input the constructed connection matrix into the brain network collaborative classification model to obtain the output result, match the stimulation classification information of each output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
[0068] Among them, the brain network collaborative classification model is a pre-built model used to classify EEG signals.
[0069] The training method of the brain network collaborative classification model proposed in this embodiment preprocesses and slices the pre-collected EEG signals, constructs a connection matrix, and inputs the connection matrix into the brain network collaborative classification model for training. Compared with the existing technology, it can more accurately capture the collaborative changes between brain regions under task states, enhance the classification model's ability to recognize task-related EEG signal features, analyze EEG microstates, and derive cross-modal interaction relationships between different brain region networks, thereby achieving the technical effect of significantly improving the classification accuracy of EEG signals in tasks.
[0070] In addition, compared with the prior art, the classification model trained by the method of this embodiment can explore the synergistic relationship between hidden brain regions in the task state, and significantly improve the accuracy of the classification results of EEG signals in the task.
[0071] Each step is described in detail below.
[0072] Step S100 includes:
[0073] A 48Hz to 52Hz notch filter is used to remove the power frequency interference in the EEG signal of each channel, and a 0.5Hz to 100Hz bandpass filter is used to remove the noise of the EEG signal of each channel to obtain a preprocessed EEG signal of 4Hz to 49Hz. The attribute information of the EEG signal of each channel includes: the brain area location information of the EEG signal and the stimulation classification information of each time period of the EEG signal.
[0074] It should be noted that this step uses a 48Hz to 52Hz hardware filter to eliminate power frequency interference, a 0.5Hz to 100Hz band-pass filter to obtain the relevant frequency bands of the EEG signal and suppress noise interference, and the collected EEG signal is band-pass filtered, selecting the frequency band of 4Hz to 49Hz. The preprocessed EEG signal covers theta waves (4Hz-7Hz), alpha waves (8Hz-13Hz), beta waves (13Hz-30Hz), and gamma waves (30Hz-49Hz, gamma waves are usually 30Hz-80Hz, and only part of the gamma wave frequency band is covered here) to ensure that the signal contains ASMR-specific responses.
[0075] Step S200 includes:
[0076] S201, dividing the pre-processed EEG signal of each channel into designated time segments according to the time axis to obtain multiple non-overlapping EEG signal segments,
[0077] Specifically, the EEG signal is divided into N 2.5s segments according to the acquisition time axis.
[0078]
[0079] Among them, T is the total duration of the EEG signal after preprocessing, Indicates rounding down;
[0080] Each EEG signal segment z j (τ)=z(t j +τ), where j = 1, 2, ..., N represents the jth slice, t j =(j-1)×2.5, represents the starting time of the jth slice, and τ∈[0,2.5) represents the relative time within the slice.
[0081] S202: Remove predetermined artifact interference signals from all multiple non-overlapping EEG signal segments to obtain valid EEG signal segments.
[0082] Among them, the brain area location information of the valid EEG signal segments corresponding to the O1 channel, O2 channel, Oz channel, Cz channel, P5 channel, P6 channel, PO3 channel, and PO4 channel is the visual network location;
[0083] The brain region location information of the valid EEG signal segments corresponding to the C3 channel, C4 channel, P5 channel, P6 channel, F5 channel, and F6 channel is the auditory network location;
[0084] The brain region location information of the valid EEG signal segments corresponding to the Fz channel, Cz channel, Pz channel, C3 channel, C4 channel, P5 channel, and P6 channel is the frontoparietal network location.
[0085] It should be noted that step S202 involves segmenting the EEG signals into different brain regions based on the characteristics of the EEG microstate network, using different electrodes to represent different brain regions. Based on the electrode locations and brain region divisions of the three brain regions (visual, auditory, and frontal-parietal) on which EEG microstate changes depend, the corresponding electrode data is intercepted to accurately extract the dynamic characteristics of EEG microstate changes.
[0086] Step S300 includes:
[0087] For all valid EEG signal segments in the same brain region and in the same segmentation time period, the Pearson correlation coefficient is used to calculate the linear correlation between any two electrodes in the brain region, and the connection matrix of the brain region is constructed.
[0088] Specifically, the signal data points in the effective EEG signal segment corresponding to electrode X form an electrical segment sequence X = [x1, x2, ..., x n ]; the signal data points in the effective EEG signal segment corresponding to electrode Y constitute the electrode segment sequence Y = [y1, y2, ..., y n ], its linear correlation r XY for,
[0089]
[0090] Where n is the number of signal data points collected in each EEG signal segment (more specifically, each EEG signal segment is 2.5s and the sampling rate is 1200Hz, so the number of signal data points n collected in each EEG signal segment is 3000), x i is the i-th signal data point collected in the electrode segment sequence X, y i is the i-th signal data point collected in the electrode segment sequence Y, is the average value of the signal data points in the electrode segment sequence X; is the average value of the signal data points in the electrode segment sequence Y.
[0091] For each brain region, based on the r between any two electrodes in each brain region XY , construct the connection matrix of each brain region; obtain the connection matrix C of the visual network position 视觉 , the connection matrix C of the auditory network position 听觉 , the connection matrix C of the frontoparietal network position 额顶 .
[0092] The dimension of each connection matrix is m×m. Specifically, the connection matrix C of the visual network position is 视觉 The dimension is 8×8, and the connection matrix C of the auditory network position 听觉 The dimension is 6×6, and the connection matrix C of the visual network position额顶 The dimension is 7×7; for each brain region, the connection matrix is used to represent the connection relationship between different electrodes, providing a feature matrix for subsequent classification tasks.
[0093] It should be noted that in step S400, the connection matrix used to train the brain network collaborative classification model includes: the connection matrix of the three brain regions of users who are sensitive to ASMR, and the connection matrix of the three brain regions of users who are not sensitive to ASMR, and the users are of equal gender and have no neurological diseases; the output results of each stimulus classification information include: audiovisual stimulus classification information, auditory stimulus classification information, white noise stimulus classification information and resting state classification information. The diversity of the connection matrix improves the accuracy of model classification and optimizes model performance.
[0094] Example 2
[0095] The training method provided in Example 1 is applicable to various brain network collaborative classification models. This example uses the DenseNet network based on dense connections and convolutions as a classification model for specific explanation.
[0096] Before step S400, the method of this embodiment further includes:
[0097] Constructing a brain network collaborative classification model, such as Figure 3 As shown, the brain network collaborative classification model includes: an input layer that receives the connection matrix respectively, a convolution layer corresponding to each input layer, a pooling layer corresponding to each convolution layer and a splicing layer corresponding to all pooling layers, a fully connected layer corresponding to the splicing layer, and an output layer;
[0098] There are three input layers, which are used to receive the connection matrix of the visual network position, the connection matrix of the auditory network position, and the connection matrix of the frontoparietal network position respectively;
[0099] The convolution layer and pooling layer of each input layer process the features of the connection matrix of the input layer in sequence and output the processed features;
[0100] The concatenation layer is used to concatenate the features output by all pooling layers and output them through the fully connected layer and the output layer.
[0101] Specifically, the connection matrix of each brain region is processed by its own branch to obtain eigenvectors F1, F2, and F3, each with a dimension of (d, 1). The three eigenvectors are concatenated into a new eigenvector F total , with a dimension of (3d, 1), and is used as the input of the fully connected layer and the output layer for classification tasks.
[0102] In addition, an ablation experiment can be performed after step S400 of Example 1 and Example 2 to traverse the combinations between different brain area positions and verify the effect of the whole-brain input to seek the difference in contribution between different brain areas and obtain the optimal model configuration. Specifically, before the ablation experiment, the feature matrix with whole-brain electrode information processed by step S200 is used as the direct input of the model, and the classification result of this whole-brain input is used as the benchmark for performance evaluation. Then, an ablation experiment is performed to explore the impact of inputs from different brain area combinations on the classification results, including the combination of visual network position and auditory network position input, the combination of visual network position and frontal-parietal network position input, the combination of auditory network position and frontal-parietal network position input, and the case where the three are input separately. All of the above maintain a parallel input mode. By comparing the classification results of these different input combinations, the contribution of different brain area combinations is analyzed, the model effect is evaluated, and the optimal model configuration is selected for further analysis.
[0103] Example 3
[0104] This embodiment provides a training device for a brain network collaborative classification model, comprising:
[0105] A data processing module is configured to pre-process at least 16 channels of EEG signals collected from each user to obtain pre-processed EEG signals and EEG signal attribute information within a specified frequency band; and to slice the pre-processed EEG signals and remove interference signals to obtain valid EEG signal segments and EEG signal segment attribute information; the EEG signal segment attribute information includes stimulus classification information and brain region location information;
[0106] A connection matrix acquisition module is used to construct a connection matrix for training a brain network collaborative classification model based on each EEG signal segment and the brain region location information of the segment;
[0107] The model training module is used to input the constructed connection matrix into the brain network collaborative classification model, obtain the output result, match the stimulation classification information of the output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
[0108] Example 4
[0109] This embodiment provides an electronic device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program in the memory, specifically performing the steps of the training method of the brain network collaborative classification model described in Examples 1 and 2.
[0110] The above embodiments are further explained below in conjunction with specific application scenarios.
[0111] 1. Object selection part.
[0112] 63 undergraduate participants, including 35 males and 28 females, were selected using the ASMR assessment questionnaire. The questionnaire is used to screen for people who are sensitive to ASMR. Among all participants, 33 were in the experimental group, who were sensitive to ASMR, and 30 were in the control group, who were not sensitive to ASMR. All participants were in good health and had no brain damage, visual impairment, or other central nervous system diseases. The age distribution of the participants ranged from 18 to 24 years old, with an average age of 22.79 years old. Participants were told to get enough sleep the day before the experiment.
[0113] 2. In the implementation and data processing part, the following steps 1 to 7 are specifically performed for each subject.
[0114] Step 1: During the preparation phase, the subject's scalp is treated. First, confirm that the subject has no scalp wounds or other health problems that may affect the experiment. Then use a degreasing detergent to clean the scalp and blow-dry the hair to remove excess oil, dirt and hair at the contact points. Then, according to the international standard 10 / 10 system, measure and mark the length of the central axis meridian of the head to find the position of the Cz point. Align the Cz lead of the electrode cap with the mark and wear the electrode cap correctly. Subsequently, inject an appropriate amount of conductive paste under each electrode and measure to ensure that the impedance of each electrode position is less than 30KΩ to ensure the high signal-to-noise ratio requirement of the EEG. The electrode positions placed according to the 10 / 10 system are Fp1, Fp2, F5, F6, C3, C4, P5, P6, PO3, PO4, O1, O2, Fz, Cz, Pz, Oz. In addition, according to the 10 / 10 system, the reference electrode is located on the left earlobe and the ground electrode is located on the forehead. The specific electrode placement positions are as follows: Figure 2As shown. After the preparation phase, participants were asked to sit in a non-sliding chair with a backrest, facing a computer screen at a distance of approximately 0.6-0.7 meters. Participants were instructed to remain silent with their eyes open, waiting for a prompt on the computer screen. A cross would first appear on the screen for 2 seconds, after which four ASMR-related paradigms would begin to play in a pseudo-sequential manner. All audio and video clips lasted 40 seconds and had the same sound intensity. Four paradigms were designed to trigger ASMR and serve as controls. A sand-cutting trigger video served as a video paradigm, while its audio-only counterpart served as an audio paradigm. For additional control, two additional paradigms were added: pre-generated Gaussian white noise and a paradigm in which the patient remained at rest without any stimulation. After each paradigm, participants rested for one minute. The EEG signal was sampled at a 1200 Hz rate, and active electrodes connected to electrode cables transmitted EEG signals to a g.HIamp amplifier via an activebox. The hardware setup uses a notch filter between 48Hz and 52Hz and a bandpass filter between 0.5Hz and 100Hz to remove power frequency interference and suppress noise. Finally, a computer connected to a g.HIamp signal amplifier stores, displays, and analyzes the collected EEG signals. The g.HIamp signal amplifier driver is based on C++ and uses the API of an open-source data acquisition toolkit. After data acquisition, statistical analysis of the EEG microstates revealed that three types of microstates exhibited significantly different characteristics in ASMR-sensitive individuals compared to controls. The results showed significant differences in the distribution, frequency, and duration of these three microstates in ASMR-sensitive individuals. These three microstates correspond to the auditory, visual, and frontoparietal networks, respectively. These three networks contribute to the different responses of ASMR-sensitive individuals. Therefore, a classification model architecture based on the interaction between these three networks was proposed to extract features.
[0115] Step 2: Use bandpass filtering to obtain EEG signals in different frequency bands. Generally speaking, EEG signals are divided into several frequency bands: delta (0.5Hz-4Hz), theta (4Hz-7Hz), alpha (8Hz-13Hz), beta (13Hz-30Hz), and gamma (30Hz-80Hz). ASMR has been shown to respond to different frequency bands. This time, a filter frequency band of 4-49Hz was selected to filter out irrelevant noise.
[0116] Step 3: Slice the EEG signal z(t) obtained after step 2 one by one and remove the signal segments interfered by artifacts. The specific method is to divide the EEG signal z(t) into multiple 2.5-second signal segments along the time axis. Assuming that the total duration of the EEG signal processed in step 2 is T, the number of signal segments after slicing is
[0117]
[0118] in Indicates rounding down, each slice signal segment can be expressed as z j (τ)=z(t j +τ), where j = 1, 2, ..., N represents the jth slice, t j =(j-1), where j=1, 2, ..., N represents the jth slice, t j =(j-1)×2.5 represents the starting time of the jth slice, τ∈[0,2.5) represents the relative time within the slice, and through this slicing method, the continuous EEG signal is converted into a discrete signal segment set {z1(τ), z2(τ), ..., z N (τ)}, while removing the EEG signal segments of the two channels, Fp1 and Fp2, which have the most severe electrooculographic effects. Each EEG signal segment is then screened to remove abnormal segments caused by eye movements, electromyography, or other artifacts to ensure signal quality for subsequent analysis.
[0119] Step 4: Use electrodes of different channels to characterize the regional distribution of EEG activity. According to the electrode positions of the visual, auditory, and frontoparietal networks on which EEG microstate changes depend, the corresponding EEG signals are intercepted for each EEG signal segment. Specifically, the visual network position corresponds to the following electrode channels: O1 channel, O2 channel, Oz channel, Cz channel, P5 channel, P6 channel, PO3 channel, and PO4 channel. The auditory network position corresponds to the following electrode channels: C3 channel, C4 channel, P5 channel, P6 channel, F5 channel, and F6 channel. The frontoparietal network position corresponds to the following electrode channels: Fz channel, Cz channel, Pz channel, C3 channel, C4 channel, P5 channel, and P6 channel. Through the correspondence between electrodes and brain regions, the dynamic characteristics of different brain functional networks in EEG microstate changes can be accurately extracted and analyzed.
[0120] Step 5. Use the Pearson correlation coefficient to measure the connection relationship between each electrode in each brain region, and construct a connection matrix as a feature matrix. The Pearson correlation coefficient is a statistical indicator used to measure the degree of linear correlation between two variables. Its value range is between -1 and 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation. For example, the signal data points in the valid EEG signal segment corresponding to electrode X constitute the brain segment sequence X of electrode X = [x1, x2, ..., x n ]; the signal data points in the effective EEG signal segment corresponding to electrode Y constitute the electrode segment sequence Y = [y1, y2, ..., y n ], its linear correlation r XY for,
[0121]
[0122] Where n is the number of signal data points collected in each EEG signal segment, x i is the i-th signal data point collected in the electrode segment sequence X, y i is the i-th signal data point collected in the electrode segment sequence Y, is the average value of the signal data points in the electrode segment sequence X; is the average value of the signal data points in the electrode segment sequence Y.
[0123] Before calculation, the input is an m×n matrix, where m is the number of channels containing electrodes at each brain region location, and n is the number of signal data points collected in each EEG signal segment. After calculation, the output is an m×m connection matrix R, where each element R ij Represents the linear correlation between electrodes i and j. For each segmented time period, the linear correlation between different electrode channels is calculated and used as a feature matrix for subsequent processing.
[0124] Step 6: In the input layer, the connection matrices of three different brain regions are first input into the brain network collaborative classification model in parallel, corresponding to the connection relationship of the visual network, frontoparietal network and auditory network respectively. The three connection matrices are C 视觉 、C 额顶 、C 听觉 , the dimension of each connection matrix is m×m, where m is the number of channels of electrodes contained in each brain region. These connection matrices are processed as three parallel input streams through their respective branches (such as convolution, pooling, etc.), and each branch finally outputs a feature vector F1, F2, F3, and each feature vector has a dimension of (d, 1). Then, the three feature vectors are concatenated in the first dimension to form a new feature vector F total , whose dimension is (3d, 1). The spliced F total The feature vector serves as the input for the subsequent fully connected layer and output layer, resulting in the classification result. Next, a separate (16, 16)-dimensional feature matrix of the entire brain was input as a control, and the feature extraction performance of different models was compared. The DenseNet model achieved the highest accuracy, and was therefore selected as the final feature extraction model. The results of the model comparison are shown in Table 1.
[0125] Table 1 Comparison of classification results of various models when three brain regions are input in parallel
[0126]
[0127] In step 7, the feature matrix containing whole-brain electrode information processed in step 3 is used as the direct input for the model and used as a benchmark for the classification task. The classification results of this whole-brain input serve as a benchmark for evaluating other input combinations. Pairwise combinations of the visual, auditory, and frontoparietal network locations from step 6 are performed, while retaining the separate input method for each brain region. All inputs are input in parallel, consistent with the architecture in step 6. These different input combinations include inputs from the visual and auditory networks, inputs from the visual and frontoparietal networks, inputs from the auditory and frontoparietal networks, and inputs from the visual, auditory, and frontoparietal networks, respectively. By comparing the classification results of these input combinations, we explore the impact of different input methods on classification performance and thus evaluate the effectiveness of the model driven by microstate network knowledge. The specific comparative results are shown in Table 2, which comprehensively analyzes the contribution of different input methods to model performance. Comparing the feature extraction performance of different models shows that the DenseNet model achieves the highest accuracy, thus validating the comparative results in step 6.
[0128] Table 2 Comparison of classification results of each model when two brain region positions are combined and a single brain region position is input
[0129]
[0130] In summary, the brain network collaborative classification model can be used to effectively explore the collaborative relationship between brain regions in EEG signals. Compared with the existing technology, this embodiment divides different brain area networks by introducing prior knowledge of EEG microstates, and then constructs a classification model of collaborative input, which can more accurately capture the collaborative changes between brain regions under task states. This method not only enhances the classification model's ability to recognize task-related EEG signal features, but also significantly improves the classification accuracy of EEG signals in tasks, and is therefore conducive to promoting applications in EEG-based neurorehabilitation, intelligent interaction and other fields.
[0131] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0132] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.
[0133] It should be noted that, in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third etc. is for convenience only and does not indicate any order. These words may be understood as part of the component name.
[0134] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.
[0135] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments after learning the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0136] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.
Claims
1. A training method for a brain network collaborative classification model, characterized in that: include: S100, preprocessing the EEG signals of at least 16 channels collected in advance for each user to obtain the preprocessed EEG signals and attribute information of the EEG signals in a specified frequency band; S200, slicing the preprocessed EEG signal and removing interference signals to obtain valid EEG signal segments and attribute information of the EEG signal segments; The attribute information of the EEG signal segment includes: stimulus classification information and brain region location information; S300, constructing a connection matrix for training a brain network collaborative classification model based on each EEG signal segment and the brain region location information of the segment; S400. Input the constructed connection matrix into the brain network collaborative classification model to obtain the output result, match the stimulation classification information of each output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
2. The method according to claim 1, characterized in that The S100 includes: A 48Hz to 52Hz notch filter is used to remove the power frequency interference in the EEG signal of each channel, and a 0.5Hz to 100Hz bandpass filter is used to remove the noise of the EEG signal of each channel to obtain a preprocessed EEG signal of 4Hz to 49Hz. The attribute information of the EEG signal of each channel includes: the brain area location information of the EEG signal and the stimulation classification information of each time period of the EEG signal.
3. The method according to claim 1, characterized in that The S200 includes: S201, dividing the pre-processed EEG signal of each channel into designated time segments according to the time axis to obtain multiple non-overlapping EEG signal segments, S202, removing predetermined artifact interference signals from all multiple non-overlapping EEG signal segments to obtain valid EEG signal segments; Among them, the brain area location information of the valid EEG signal segments corresponding to the O1 channel, O2 channel, Oz channel, Cz channel, P5 channel, P6 channel, PO3 channel, and PO4 channel is the visual network location; The brain region location information of the valid EEG signal segments corresponding to the C3 channel, C4 channel, P5 channel, P6 channel, F5 channel, and F6 channel is the auditory network location; The brain region location information of the valid EEG signal segments corresponding to the Fz channel, Cz channel, Pz channel, C3 channel, C4 channel, P5 channel, and P6 channel is the frontoparietal network location.
4. The method according to claim 3, characterized in that The S300 includes: For all valid EEG signal segments in the same brain region and in the same segmentation time period, the Pearson correlation coefficient is used to calculate the linear correlation between any two electrodes in the brain region, and the connection matrix of the brain region is constructed. Specifically, the signal data points in the effective EEG signal segment corresponding to electrode X constitute an electrode segment sequence X = [x1, x2, ..., x n ]; the signal data points in the effective EEG signal segment corresponding to electrode Y constitute the electrode segment sequence Y = [y1, y2, ..., y n ], its linear correlation r XY for, Where n is the number of signal data points collected in each EEG signal segment, x i is the i-th signal data point collected in the electrode segment sequence X, y i is the i-th signal data point collected in the electrode segment sequence Y, is the average value of the signal data points in the electrode segment sequence X; is the average value of the signal data points in the electrode segment sequence Y; For each brain region, based on the r between any two electrodes in each brain region XY , construct the connection matrix of each brain region; Get the connection matrix C of the visual network position 视觉 , the connection matrix C of the auditory network position 听觉 , the connection matrix C of the frontoparietal network position 额顶 .
5. The method according to claim 4, characterized in that Before S400, the method further includes: Constructing a brain network collaborative classification model, The brain network collaborative classification model includes: an input layer that receives a connection matrix, a convolutional layer corresponding to each input layer, a pooling layer corresponding to each convolutional layer, a splicing layer corresponding to all pooling layers, a fully connected layer corresponding to the splicing layer, and an output layer; There are three input layers, which are used to receive the connection matrix of the visual network position, the connection matrix of the auditory network position, and the connection matrix of the frontoparietal network position respectively; The convolution layer and pooling layer of each input layer process the features of the connection matrix of the input layer in sequence and output the processed features; The concatenation layer is used to concatenate the features output by all pooling layers and output them through the fully connected layer and the output layer.
6. The method according to claim 5, characterized in that The S400 includes: The connection matrices used to train the brain network collaborative classification model include: the connection matrices of the three brain regions of users who are sensitive to ASMR, and the connection matrices of the three brain regions of users who are not sensitive to ASMR. The users are equal in gender and have no neurological diseases. The stimulus classification information of the output results includes: audio-visual stimulus classification information, auditory stimulus classification information, white noise stimulus classification information and resting state classification information.
7. The method according to claim 3, characterized in that The S201 includes: The EEG signal is divided into N 2.5s segments according to the acquisition time axis. Among them, T is the total duration of the EEG signal after preprocessing, Indicates rounding down; Each EEG signal segment z j (τ)=z(t j +τ), where j = 1, 2, ..., N represents the jth slice, t j =(j-1)×2.5, represents the starting time of the jth slice, and τ∈[0,2.5) represents the relative time within the slice.
8. The method according to any one of claims 1 to 7, characterized in that: The brain network collaborative classification model is a DenseNet model.
9. A training device for a brain network collaborative classification model, characterized in that: include: A data processing module is used to pre-process the EEG signals of at least 16 channels collected in advance for each user to obtain the pre-processed EEG signals and attribute information of the EEG signals in a specified frequency band; and to slice the pre-processed EEG signals and remove interference signals to obtain valid EEG signal segments and attribute information of the EEG signal segments; The attribute information of the EEG signal segment includes: stimulus classification information and brain region location information; A connection matrix acquisition module is used to construct a connection matrix for training a brain network collaborative classification model based on each EEG signal segment and the brain region location information of the segment; The model training module is used to input the constructed connection matrix into the brain network collaborative classification model, obtain the output result, match the stimulation classification information of the output result with the stimulation classification information of the EEG signal segment in the input connection matrix, and if the matching result is less than the specified threshold, adjust the parameters and architecture of the brain network collaborative classification model until the matching rate is greater than or equal to the specified threshold, thereby obtaining a trained brain network collaborative classification model.
10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program in the memory, specifically performing the steps of the training method of the brain network collaborative classification model described in any one of claims 1 to 8.