Research On Brain State Pattern Recognition Based On Attention Neural Network

The human brain is a complex nonlinear time-varying system whose dynamic response processes to different cognitive tasks comprise the observable brain activity.A large number of functional magnetic resonance imaging(f MRI)studies in recent years have attempted to identify the task state of the subject from brain activity,and these efforts have deepened our understanding of brain mechanisms.The brain’s task response process encompasses multiple dimensions,including temporal,spatial,and brain-area interactions,and the complexity of cognitive tasks at different levels varies,as does the difficulty of identification.However,previous studies often extract task features from only a single dimension and use fixed-length f MRI data to identify different cognitive tasks,which greatly limits the recognition performance and application scenarios of the model.To solve the above problems,this paper proposes a study on brain state pattern recognition based on multi-attentional neural networks,which extracts f MRI task features of different dimensions by introducing multiple attentional modules,and determines the length of data needed for different tasks by using a dynamic network structure.The main research of this paper is as follows:1.To address the problem of f MRI signal task feature extraction,this paper proposes a feature extraction method based on a multi-attention mechanism,which uses temporal attention module,spatial attention module and self-attention module used to extract taskrelated features.The proposed model was evaluated in this study using f MRI data of 1200 subjects from the Human Connectome Project(HCP)on an emotion task.The study achieved 99.51% accuracy in classifying both conditions of the emotion task.To further investigate the interpretability of the attention module,this study conducted a visual analysis of the attention weights and conducted an ablation experiment on the attention module.The results show that the attention module can learn biologically meaningful brain representations and thus improve classification accuracy.2.To address the problem of determining the length of data required for task recognition,this paper proposes a dynamic neural network-based approach to brain state recognition by introducing a dynamic network structure for determining the visual length required to recognize different samples.This study incorporated 21 cognitive states from six tasks(emotional,motor,verbal,working memory,social,and relational)in the HCP dataset,and f MRI data from 1111 subjects were used to evaluate the proposed model.The results showed that this study achieved 93.57% accuracy in the identification of the 21 task states.The results of the analysis of attention weights further demonstrated the effectiveness of the attention module.In addition,this study found that the introduction of a dynamic network structure can effectively reduce the length of data required for recognition,and the decoding accuracy can still reach 91% at an average length of 6.2s.The statistical results show that the data length required for recognition is related to the complexity of the task,i.e.,complex cognitive tasks(social,relational,working memory)require longer data length than simple cognitive tasks(verbal,emotional,motor).In this paper,we use deep learning algorithms to construct task-state f MRI data decoding models in the above two works to extract task-related features from multiple perspectives,and then use these features to identify task conditions as well as to determine the data length required for task recognition.The results show that the model proposed in this paper improves the accuracy of task recognition and can also effectively decode the data length required for task recognition,which provides two more effective methods for task state This provides two more effective models for task state recognition and provides a basis for the implementation of brain-computer interface systems.