Research On Decoding Method Of Fine Motor Imagery Of Unilateral Upper Limb Based On EEG

The brain computer interface technology based on motor imagery is crucial for stroke rehabilitation,as it can convert the brain’s motor imagery intentions into control commands.At present,research on upper limb motor imagery mainly focuses on left and right hand motor imagery.However,controlling the unilateral upper limb rehabilitation device to perform different actions from this can lead to inconsistent motor intentions and the actions of the end effector,i.e.a cognitive disconnect.Therefore,decoding the fine motor imagination task of unilateral upper limb movements is crucial for restoring the upper limb motor ability of stroke patients.Because the regions activated by the motor imagery tasks of different movements of the unilateral upper limb have similar representations on the cerebral cortex,and the EEG signals have non-stationary characteristics,it is a great challenge to detect different motor imagery tasks of the same limb.This article aims to study the decoding of fine motor imagery in unilateral upper limb movements based on EEG signals.The specific work is as follows:(1)This article explores the brain mechanisms of fine motor imagination based on single channel and brain networks,providing support for decoding method design.Using the FCSI algorithm to evaluate the importance of different frequency bands,brain regions,and brain connections in distinguishing between idle state and motor imagery tasks,as well as the importance of distinguishing between imaginary palm grasping and imaginary elbow joint bending tasks.The results showed that the frequency bands with high distinguishability between idle state and motor imagery tasks were beta,gamma,and the entire motor imagery frequency band of 8-30Hz;The brain regions involved are mainly sensory motor cortex,prefrontal lobe,central parietal lobe and parietal lobe;Frequency domain features are more important;The effective brain connections are mainly concentrated in the prefrontal and frontal lobes.The distinguishing frequency bands between imaginary palm grasping and imaginary elbow joint bending are mainly alpha and 8-30 Hz containing alpha;The involved brain regions are mainly concentrated in the sensory motor cortex and the central parietal lobe;Time domain features are more important;The effective brain connections are mainly concentrated in the channel connections between C5,C3,C1,the central frontal lobe,and the central parietal lobe.(2)There is a lack of corresponding interpretable feature extraction and classification methods for fine motion imagination decoding,based on the key frequency band of research work(1),a multi feature fusion based upper limb fine motion decoding method is proposed.Multiple features include extracting energy in the time domain,spectral distance in the frequency domain,and brain network features.Brain network feature extraction includes three ways: degree and eccentricity in graph theory,all subjects jointly select effective brain connections,and subjects independently select effective brain connections.In addition,they decode the fine movements of upper limbs of two information fusion strategies at feature level and decision level on multiple dimensions.The results show that on multiple classifiers,the average classification accuracy of time-frequency domain features is the highest at 61.07%;The average accuracy of brain network features is the highest at 60.44%,and it was found that the correlation between channels is more important than synchronization(p<0.05).In terms of information fusion,the average accuracy obtained by decision level fusion is the highest at 63.78%;The average accuracy obtained by feature layer fusion is the highest at 64.76%,which is 3.69% and 4.32%higher than the highest average classification accuracy obtained by using timefrequency features or brain network features alone.(3)In response to the limitations of EEG feature extraction methods that are difficult to learn all the information of the original signal,this paper proposes an upper limb fine action decoding method based on the Multi-view convolutional attention learning model(MCAL)based on research work(1)and(2).This method includes a raw EEG information learning module and a EEG channel coupling relationship learning module based on correlation coefficients.The results showed that the MCAL model achieved average accuracy of 75.78% and 73.07% on both self collected and publicly available datasets,respectively.Compared to other existing models on the public datasets,there is a 4.39% improvement(p<0.05).