The Construction Of Deep Model For Consciousness Level Classification And Analysis Of Feature Visualization

Consciousness disorders are classified into unresponsive syndromes and microconscious states,and clarifying the patient type helps to promote wake-up treatment.However,the current assessment methods based on traditional machine learning models have difficulty in extracting accurate features and poor classification due to the lack of a priori knowledge.EEG contains rich information of brain activity.Therefore,this paper constructs a highly accurate consciousness-level deep classification model based on EEG with adaptive feature extraction,and enhances its interpretation by combining with visualization scheme.1)In this paper,we propose an architecture for visualizing the level of consciousness classification model that incorporates the input set of EEG feature maps as well as the deep neural network structure.Two types of EEG feature maps are considered for the input set: 1)a brain network connectivity matrix characterized by phase-locked values of multichannel intrinsic band EEG signals,and 2)a matrix of EEG periodic and non-periodic components characterized by parametrically separated power spectra.The convolutional neural network is chosen as the binary classification model of UWS and MCS,and the gradient-weighted class activation mapping technique is used to visualize the classification results of the model,which makes the model interpretable.For this model architecture,this paper designs a clinical experiment to collect resting EEG data from DOC patients,constructs two types of input sets for the consciousness level classification model based on the retrospective data,and validates and analyzes the model classification effect.2)Structural enhancement of the original input dataset using four approaches:spectral clustering,weighting coefficients,rearrangement by anatomical brain regions,and rearrangement by functional networks,improved the ability of the CNN structure to handle non-Euclidean inputs and increased the accuracy of consciousness level classification.The results showed that the data with Alpha in the fixed band had the best classification results,with a correct rate of 82.7% when rearranged by anatomical brain regions and 87.2% when rearranged by functional networks.Visual analysis of the model classification results revealed that connections between anatomical brain regions(frontal-parietal,frontal-occipital)and within functional networks(DMN and DAN)dominated the classification results for the level of consciousness,and that these connection characteristic areas were significantly correlated with the clinical behavioral CSR-R scores of consciousness disorders.3)The model classification results using the EEG power spectrum feature map as the input set showed that the periodic component was the main component of the power spectrum to characterize consciousness,and excluding the interference of the nonperiodic component could effectively improve the accuracy of the classification of consciousness level using the power spectrum.The visualization results showed that the alpha band and the beta low frequency of the period component of the power spectrum are the "high activation bands" for the classification of consciousness level from the mean perspective;however,from the individual perspective,the "high activation bands" are specific to different patients,i.e.,there are Individualized bands exist.In order to compare with the input set of fixed-band brain network,the brain network matrix of individualized frequency bands is extracted as the input set in this paper.The model accuracy can reach 87.5% with this input set,and the training speed and classification effect can be improved by using the threshold filtering method to optimize the personalized input set,and the final model accuracy can reach 90.3%.