Research On Construction Method Of Consciousness Evaluation System Based On Nonlinear Theory And Machine Learning

The assessment of conscious state is an important issue in field of neuroscience and clinical practice.Although many theories have been proposed to measure the state of consciousness,such as the application of entropy and information integration theory in the depth of general anesthesia.However,these methods are usually proved to be effective in one specific application scenario.Therefore,in order to verify the effectiveness of different theoretical methods,assessing states of consciousness needs to consider different conscious states comprehensively.Consciousness can be divided into two dimensions:wakefulness and awareness.However,these theories are only describe the assessment of consciousness from one single dimension.At present,there is no study to propose a systematic framework of quantify the variable consciousness from two dimensions.Therefore,this study introduces four conscious states: coma,general anesthesia,minimally conscious state,and normal wakefulness.And we construct the evaluation system of consciousness from two dimensions of wakefulness and awareness.First of all,through EEG preprocessing of forehead in four different conscious states,the obtained EEG signals are analyzed from time and frequency.The results show that the amplitude of EEG signals in conscious state is lower than that in other three conscious states,while the spectrum energy in comatose state is mainly concentrated in under frequency of 8Hz.Then,this paper implements feature extraction from two aspects of power spectral density and non-linear methods.The power spectral density includes five sub-band.The non-linear methods includes permutation entropy,sample entropy,permutation Lempel-Ziv complexity and detrended fluctuation analysis.The results show that the permutation entropy has a significant difference between the four states.In the awareness dimension of consciousness,the sample entropy and permutation Lempel-Ziv complexity are used as features to distinguish effectively,while in the wakefulness dimension of consciousness,the spectral power density of gamma and permutation entropy are used to distinguish effectively.Then,this paper uses three machine learning classification algorithms: a geneticalgorithm to optimize support vector machine,random forest and BP neural network,and we combine the classification model with the features to distinguish four states.The results show that the combination of non-linear methods as features is the best way to classify,and the combination with genetic algorithm to optimize support vector machine will achieve the highest classification accuracy,that is 92.3%.Finally,in order to construct the evaluation "scale" of consciousness from wakefulness and awareness,this paper introduces permutation entropy,sample entropy,permutation Lempel-Ziv complexity and detrended fluctuation analysis as independent variables into the multiple linear regression model.It is proved that the combination of multiple linear regression model,non-linear theory and machine learnings is an effective way to evaluate two dimensions of consciousness.