Research On Detection Method For EEG-Based Cognitive Load

In human-computer interaction systems,cognitive load has been proven to be highly correlated with system interaction efficiency.The user's cognitive load will be accurately measured whether it is beyond its acceptable range.This allows the system to intelligently adjust its interaction strategy to achieve the best human-computer interaction performance.The EEG-based cognitive load recognition system can get more objective results as it is closer to the place where consciousness is generated.However,the signal-to-noise ratio of EEG signals is relatively low,and there are differences between individuals.These factors make it difficult to accurately decode the cognitive load.At the same time,the feature-based cognitive load recognition system requires manual design of features,which also increases the workload of technical personel related to the brain-computer interface system.In addition,it is one of the important tasks of neuroscientists to explore the temporal,frequency and spatial features related to cognitive load is one of the important tasks of neuroscientists.Currently researchers usually select electrodes or rhythms for data analysis based on prior knowledge,which may lose the best component features.Therefore,we conduct three studies in response to the mentioned above issues:1.Cognitive load detection based on covariance matrix features.In order to confirm the task-related brain areas,time-frequency analysis of the cognitive load EEG data based on the delayed-match-to-sample task.The results show that the frontal lobe and parietal lobe are brain areas related to cognitive load.The covariance features of the cognitive load EEG signals have been extracted for detecting cognitive load.Discriminate projection features in the tangent space related to the subject.Then the cognitive load across the subjects is measured.The results show that the accuracy of cognitive load detection within the subjects is 53%(significantly higher than the chance level,33%),and the cross-subject detection accuracy can be improved(an increase of 6%)based on the tangent space mapping framework.The results show that the covariance matrix can indeed be used to detect the cognitive load detection.2.Cognitive load detection based on neural network.In order to realize end-to-end cognitive load detection,we propose a model based on convolutional neural network for cognitive load detection.The model uses the original EEG signal as input,and introduces the priori EEG signal extraction process in the network structure design,eliminating the need to manually extract EEG features.The model has the ability to extract task-related frequency features,spatial distribution and time distribution.The results show that the accuracy of cognitive load detection within the subjects is 56% which is better than the existing end-to-end neural network model.The proposed neural network model further improves the performance of cognitive load detection.3.Visualization and interpretability of neural networks based on model parameters and gradients.In order to explain the task-related features extracted by the neural network the parameters of batch normalization layer are interpreted as the contribution degree of the corresponding rhythms and the saliency maps corresponding to the rhythm are interpreted as the temporal and spatial importance of the corresponding rhythms.Based on the proposed model in cognitive load detection tasks,the relative importance of frequency components,spatial positions and temporal nodes are obtained in a data-driven manner.The results show that the parietal lobe and the prefrontal lobe are the task-related brain areas obtained by neural network training,and the alpha,beta and gamma rhythms are the relevant frequency bands for cognitive load detection.