Feature Analysis And Classification Of States Of Consciousness In Brain Injuries Through EEG Signals

EEG records the oscillations of brain electric potential and provides a very large-scale measure of brain source activity.EEG monitoring can register functional and physiological changes within the brain.EEG signals also can be used to investigate clinical problems involving:monitoring alertness,coma,and brain death;locating areas of damage following a head injury,stroke,and tumour;investigating mental disorders,sleep disorders,and epilepsy.Brain injuries often suffer from cognitive impairment,disorders of consciousness,motor impairment and other neurological disorders.Some of these disorders,if not given timely treatment can lead to significant medical complications that can slow recovery and interfere with treatment interventions.With the development of computer science and neuroscience,more researchers explore computer-aided approaches to study EEG signals and their correlation with our brain function and disease.This thesis focuses on studying EEG signals from brain injuries,extracting various kinds of features of EEG and employing machine learning methods to improve the detection of clinical disorders through EEG monitoring.The main work and contributions of this thesis are listed as follows:(1)We investigate the phase synchrony index in differentiating the states of consciousness(wakefulness,somnolence,stupor,light coma,middle coma,and deep coma)in stroke patients.We propose the quantitative EEG measure,phase synchrony index of the left and right hemispheres(PLI-LR),and evaluate whether this measure can facilitate the assessment of consciousness in stroke patients through three experiments.The first experiment includes 82 patients with ischemic stroke.We explore the correlation between PSI-LR and level of consciousness in ischemic stroke patients.Also,other existing quantitative EEG features are analyzed to compare with PLI-LR.In the second experiment,EEG signals recorded from 27 cerebral haemorrhage patients are analyzed to explore whether the phase synchrony index is effective in distinguishing the states of consciousness in cerebral haemorrhage patients.In this experiment,a variety of quantitative EEG features are extracted and a linear regression model based on multiple quantitative EEG features is built for the correlation analysis.In the third experiment,14 EEG recordings from four ischemic stroke patients,who have experienced more than two types of conscious disturbance,are studied to complement the analysis of the longitudinal value of the phase synchrony index.In the last experiments,a weighted sum of six quantitative EEG features is analyzed on its correlation with the level of consciousness and a linear regression model is built for the correlation analysis.(2)We developed an ensemble of support vector machine(EOSVM)to address the issues from the imbalanced dataset.The common problem in classification on the imbalanced dataset is that the classifier always biases to the majority class in the dataset.However,in EOSVM,the overall datasets are aggregated and then split into several balanced subsets.These balanced subsets are fed into different support vector machines(SVMs),respectively.Each SVM makes its own decision for a prediction,which classified the corresponding subset of the dataset into wakefulness or a DoC state.After that,multiple predictions from these SVMs are fused through a voting rule to make the final prediction.The experiment includes 147 ischemic stroke patients.Nine different quantitative EEG features are extracted to input into the classifier EOSVM for distinguishing different states of consciousness in ischemic stroke patients.The experimental results show that EOSVM can diagnose 10%of DoC patients in ischemic stroke with an accuracy of 94.44%,sensitivity of 94.44%and specificity of 100.00%.(3)We propose a framework to combine EEG microstates and machine learning to monitor and detect disorders of consciousness(DoC)in stroke patients.We firstly design experiments to joint EEG signals from 147 ischemic stroke patients to one signal to extract exactly same microstate maps for every subject through clustering,thus the classification can be performed later.Then the statistical parameters of these microstates are calculated and their correlations with the level of consciousness are analyzed.Finally,we input statistical parameters of EEG microstates to the classifier EOSVM as features to identify the subjects with DoC and wakefulness.Our EOSVM classification results revealed that when EEG data is clustered into six microstate maps,the classification could diagnose about 37.5%of DoC patients in ischemic stroke patients with an accuracy of nearly 100%.(4)We introduce a new connectivity index,and improved other five connectivity indexes with nested cosine functions.Then,the approach classifies brain-injured patients into DoC and wakefulness classes through classifier EOSVM.It is further applied to a dataset of 607 patients with brain injuries.Our classification results show that the EOSVM classifier with the new connectivity index has achieved the best classification performance among 12 connectivity indexes.They have diagnosed in and out 35.14%of patients correctly,implying a saving of 35.14%of resources that would otherwise be consumed in conventional clinical examinations.The accuracy,sensitivity,and specificity of the classification have reached 98.21%,100%,and 95.79%,respectively.(5)We developed an ensemble of AdaBoost(EoAdaBoost)to detect DoC in brain injuries.The experiment includes 648 brain injuries with a higher quality of EEG signals(more channels and higher sampling frequency).Six connectivity indexes are extracted and the combination of them are inputted to EoAdaBoost to detect DoC in brain injuries.The experimental results showed that the combined brain functional connectivity indexes could diagnose 88.86%of DoC in brain injuries with an accuracy(92.04%),sensitivity(93.26%),specificity(90.97%)and Fl-Score(0.92)of all higher than 90%.