CN-122000024-A - Asynchronous BCI identification system based on multi-brain-area feature fusion of action observation paradigm

CN122000024ACN 122000024 ACN122000024 ACN 122000024ACN-122000024-A

Abstract

An asynchronous BCI identification system based on multi-brain region feature fusion of an action observation paradigm comprises a data preprocessing and anti-leakage data dividing unit, a multi-brain region feature collaborative extraction and weighting unit, an integrated decoding unit and a dynamic decision and output unit. The invention utilizes the motion observation paradigm to construct the BCI, not only can induce brain activities related to motion preparation, but also provides possibility for fusing forehead lobe area characteristics reflecting attention level, and the motion is presented through a frame rate reduction technology, so that steady-state motion visual evoked potential can be induced in a visual area at the same time. By integrating the multidimensional and complementary electroencephalogram characteristics of forehead leaves (cognition/attention), movement areas (movement intention) and vision areas (visual stimulus phase-locked response), the asynchronous BCI with more comprehensive information, more stable discrimination and stronger anti-interference capability is expected to be constructed, so that the reliability of the asynchronous BCI in medical application is improved.

Inventors

ZHANG XIN
LANG YUHANG
ZENG FUKANG
TANG HONGMEI
HU GUIYU
Yang Zongchuan

Assignees

重庆大学

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (10)

1. An asynchronous BCI identification system based on multi-brain region feature fusion of an action observation paradigm is characterized by comprising a data preprocessing and anti-leakage data dividing unit, a multi-brain region feature collaborative extraction and weighting unit, an integrated decoding unit and a dynamic decision and output unit; the data preprocessing and leakage-proof data dividing unit is used for synchronously acquiring the original multichannel electroencephalogram signals and the stimulation marks; the data preprocessing and leakage-proof data dividing unit is used for preprocessing an original multi-channel electroencephalogram signal, dividing the preprocessed electroencephalogram signal into corresponding BCI states based on the stimulation mark, and obtaining an electroencephalogram signal data set in different BCI states; The multi-brain region feature collaborative extraction and weighting unit performs multi-brain region feature extraction on the brain electrical signals in the brain electrical signal data sets under different BCI states to obtain an original feature vector; the multi-brain region feature collaborative extraction and weighting unit weights the original feature vector to obtain a weighted feature vector, takes the weighted feature vector as input and takes the BCI state as output to construct a training set and a testing set; the integrated decoding unit respectively trains a plurality of base classifiers by utilizing a training set, and generates an integrated classification model based on all trained base classifiers; The dynamic decision and output unit optimizes the base classifier and the integrated classification model based on the test set, and performs BCI state judgment on the real-time electroencephalogram signals by utilizing the optimized integrated classification model.
2. The asynchronous BCI identification system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the preprocessing comprises filtering, removing noise.
3. The asynchronous BCI identification system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the BCI states comprise an idle state, a control state.
4. An asynchronous BCI recognition system based on multi-brain region feature fusion of motion observation paradigm according to claim 3, characterized in that the stimulus signature comprises 1, 0; When the state corresponding to the original multichannel electroencephalogram signal is in an idle state, the stimulus mark is 0; when the state corresponding to the original multichannel EEG signal is a control state, the stimulation mark is 1.
5. The asynchronous BCI recognition system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the raw feature vectors include prefrontal region cognitive features, visual region steady state response features, motor region motor intent features; The forehead lobe zone cognitive features comprise approximate entropy features of theta bands, approximate entropy features of alpha bands, approximate entropy features of beta bands, sample entropy features of theta bands, sample entropy features of alpha bands, sample entropy features of beta bands, power ratio beta/theta of beta bands to theta bands, power ratio beta/alpha of beta bands to alpha bands, and ratio beta/(alpha + theta) of beta bands to sum of theta bands and alpha band powers; The visual zone steady-state response characteristics comprise a maximum value of typical correlation analysis coefficient statistical characteristics, an average value of typical correlation analysis coefficient statistical characteristics, a standard deviation of typical correlation analysis coefficient statistical characteristics and signal-to-noise ratio characteristics; the motion region motion intent features include a spatial filter feature of the θ band, a spatial filter feature of the α band, a spatial filter feature of the β band.
6. The asynchronous BCI identification system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the step of obtaining weighted feature vectors is as follows: A1, splicing and fusing original feature vectors to obtain a multi-brain region feature vector; A2, evaluating the multi-brain region feature vector by adopting a Relief algorithm introducing category weight coefficients to obtain a normalized feature weight vector, wherein the method comprises the following steps of: a21, recording a multi-brain region feature vector as a sample, and initializing the weight of each original feature vector in the multi-brain region feature vector to be zero; A22, randomly selecting one sample, and dividing the rest samples into a same-state sample set and a different-state sample set; the same-state sample set comprises samples with the same BCI state as the selected samples; The different state sample set comprises samples with different BCI states from the selected samples; a23, assigning category weight coefficients to the same-state sample set and the different-state sample set based on the sample number of the same-state sample set and the different-state sample set; a24, searching nearest neighbor samples of the selected samples in the same-state sample set and the different-state sample set respectively; A25 updates the weight of each original feature vector based on nearest neighbor samples in the in-state sample set and the out-of-state sample set in combination with the category weight coefficient as follows: (1) in the formula, Representing the original feature vector; Weights representing the original feature vectors before updating; the weight of the original characteristic vector after the update is represented; representing class weight coefficients; Representing the selected sample; Representing nearest neighbor samples in the abnormal state sample set; Representing nearest neighbor samples in the same-state sample set; Representing a sample And sample In the original feature vector Absolute value of the difference between the eigenvalues; Representing a sample And sample In the original feature vector Absolute value of the difference between the eigenvalues; a26, judging whether the ending condition is met, if so, carrying out normalization processing on the weights of all the original feature vectors to reach normalized feature weight vectors, and if not, returning to the step A22; the ending condition comprises reaching a maximum number of iterations; A3, weighting the original feature vector based on the normalized feature weight vector to obtain a weighted feature vector.
7. The asynchronous BCI identification system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the base classifier comprises a random forest, a limiting gradient boost, an extreme random tree.
8. The asynchronous BCI identification system based on multi-brain region feature fusion of action observation paradigm of claim 1, wherein the method for generating the integrated classification model comprises a performance weighted probability fusion strategy and a double-layer stacking integration strategy; The performance weighted probability fusion strategy obtains the weight of each base classifier by normalizing the comprehensive scores of each base classifier, and obtains the output probability of the integrated classification model by weighting and summing the prediction probability of each base classifier based on the weight of each base classifier; the composite score is as follows: (2) in the formula, A composite score representing the base classifier; 、、 weights of F1 score, specificity, AUC value are respectively expressed, and + + =1; An F1 score representing the base classifier; Representing the specificity of the base classifier; AUC values representing the base classifier; The steps of generating an integrated classification model using a dual layer stack integration strategy are as follows: c1, cross-verifying the training set by using all the base classifiers to obtain the prediction probability corresponding to each base classifier; C2 takes the prediction probability of all the base classifiers as meta-features, takes the meta-features as input and the BCI state as output, trains the secondary classifier, and obtains an integrated classification model; the secondary classifier includes a random forest, a limit gradient boost, and an extreme random tree.
9. The asynchronous BCI identification system of multi-brain region feature fusion based on motion observation paradigm of claim 1, wherein the steps of optimizing the base classifier and integrating the classification model are as follows: B1 defines an evaluation function and initializes a weight coefficient 、、 ; The evaluation function is as follows: (3) in the formula, Representing a threshold value; Representing an evaluation function; 、、 are all the weight coefficients of the two-dimensional space model, and a+b +c=1; an F1 score representing a threshold; Specificity representing the threshold; recall representing a threshold; B2 to maximize the evaluation function Dynamically searching to obtain an optimal threshold value for a target; b3 judging the condition of the current threshold value If yes, go to step B4, if no, the current threshold is the optimal decision threshold, wherein Is a preset tolerance value; B4 increasing the weight coefficient Is returned to step B2.
10. The asynchronous BCI identification system based on multi-brain region feature fusion according to any one of claims 3 or 9, wherein the real-time electroencephalogram signal is input into the optimized integrated classification model to obtain the probability of the BCI state being in the control state, and if the probability of the BCI state being in the control state is greater than the optimal decision threshold, the BCI state is determined to be in the idle state.

Description

Asynchronous BCI identification system based on multi-brain-area feature fusion of action observation paradigm Technical Field The invention relates to the technical field of asynchronous brain-computer interfaces, in particular to an asynchronous BCI identification system based on multi-brain-area feature fusion of an action observation paradigm. Background Cerebral stroke is a disease which causes brain functional loss due to brain blood supply obstruction, can cause dysfunctions such as movement, sensation, language and the like, is most common, and according to the report of world health organization, about 80% of cerebral stroke patients have limb functional disorder with different degrees, wherein more than 60% of patients still have hand functional disorder after entering a chronic period and cannot independently live. Traditional passive rehabilitation means are inefficient and difficult to mobilize the aggressiveness of the patient. Brain-computer interface (BCI) based on motor imagery (Brain-Computer Interface, BCI) provides possibility for active rehabilitation, and by decoding the motor intention of a patient, the Brain-computer interface drives exoskeleton, functional electrical stimulation and other devices to form an intention-perception-feedback closed loop, so that the neural plasticity is effectively promoted. Currently, BCI-driven rehabilitation systems mainly employ a synchronous paradigm, requiring a patient to perform tasks such as motor imagery within a fixed time window set by the system. The model has the defects that firstly, the model is asynchronous with spontaneous movement intention of a patient, the training process is unnatural, the active participation degree of the patient is not high, and secondly, the psychological fatigue is easy to cause in a fixed task period, the attention is lowered, and the training effect is influenced. Asynchronous BCI can be autonomously determined by a patient to trigger the time, is more in line with man-machine interaction logic, and has great potential in rehabilitation application. Currently, the difficulty of asynchronous BCI is to detect the motion intention of the user reliably in real time from the continuous electroencephalogram signal, and divide it into a "control state" (with motion intention) and an "idle state" (without motion intention). Most asynchronous BCIs face the problem of high false alarm rate, namely 'idle state' is misjudged as 'control state', and under a rehabilitation scene, the high false alarm rate can cause the false triggering of external equipment when a patient is unconscious, interfere with a training process and even bring potential safety hazard. The existing asynchronous BCI solution is insufficient in coping with the challenges of high false alarm rate (1) based on independent logic switches, such as a two-step system depending on SSVEP or P300, although false triggering is reduced, additional cognitive load and operation complexity are increased, and the asynchronous BCI solution is not friendly to cerebral apoplexy patients possibly accompanied with cognitive impairment, (2) based on an autonomous state detection method of direct decoding of motor cortex signals, due to the complexity and the variability of 'idle state' brain electricity and the heterogeneity of individual nerve responses of the patients, the specificity of a classification model is difficult to improve, so that the false alarm rate is high, and the overall robustness of the system is insufficient. Disclosure of Invention The invention aims to provide an asynchronous BCI identification system for multi-brain region feature fusion based on an action observation paradigm, which comprises a data preprocessing and anti-leakage data dividing unit, a multi-brain region feature collaborative extraction and weighting unit, an integrated decoding unit and a dynamic decision and output unit. The data preprocessing and leakage-proof data dividing unit is used for synchronously acquiring the original multichannel electroencephalogram signals and the stimulation marks. The data preprocessing and leakage-proof data dividing unit is used for preprocessing an original multi-channel electroencephalogram signal, dividing the preprocessed electroencephalogram signal into corresponding BCI states based on the stimulation marks, and obtaining an electroencephalogram signal data set in different BCI states. And the multi-brain region feature collaborative extraction and weighting unit performs multi-brain region feature extraction on the brain electrical signals in the brain electrical signal data sets under different BCI states to obtain an original feature vector. The multi-brain region feature collaborative extraction and weighting unit weights the original feature vector to obtain a weighted feature vector, takes the weighted feature vector as input and takes the BCI state as output to construct a training set and a testing set. The integrated decoding uni