Epileptic EEG Processing And Applications Based On Machine Learning

Epilepsy is a widespread and serious brain disease.EEG examination is an important method in the diagnosis and treatment of epilepsy.While it still relies on manual reading and analysis by epileptologists in clinical practice,researchers aim to automatically process and analyze epilepsy-related EEG signals by machine learning.In the field of epileptic EEG research,problems of interest include feature extraction and selection for epileptic EEG activities,combination of different epileptic EEG components for joint analysis,and learning without enough labeled data to detect epileptic activity,localize epileptic foci,and assess co-occurring dysfunction.In this thesis,we studied different application problems of epileptic EEG and proposed corresponding automatic EEG processing methods based on machine learning and signal processing methods.Firstly,to solve the problem of epileptic EEG identification,an automatic epileptic EEG classification method based on multi-domain feature extraction was proposed.Using this proposed method,we can distinguish the epileptic EEG,especially the interictal EEG from healthy EEG,so as to diagnose the patients who may have epilepsy.It used temporal and frequency domain analysis,nonlinear analysis and onedimensional local pattern recognition method to extract epileptic EEG features.A gradient energy operator and a one-dimensional local speed pattern were proposed to better represent the EEG signal of activity measured during seizure free intervals.After manual selection,a genetic algorithm was used to select the obtained hybrid features;then the ensemble AdaBoost classifier was used to classify epileptic EEG in a variety of classification conditions.Classification results on the dataset developed by University of Bonn show that the proposed method can be used to classify normal EEG,interictal EEG,and seizure activity with only a few features.Compared with the related research using the same data set,the proposed method can obtain an equally satisfactory classification accuracy while the feature amount was reduced by 61-95%.The classification of interictal and normal EEGs is particularly important as seizure EEG data may not be captured in EEG recordings of potentially epileptic patients requiring diagnosis.Using our proposed method,the correct classification rate of interictal and normal EEG reaches 99%,which effectively helps epilepsy diagnosis without recording of seizures.Secondly,to automatically detect the high frequency oscillations(HFO),which is an important physiological marker of epileptic focus area,we proposed an algorithm based on visual features based on manual visual detection methods and non-intuitive multi-domain features extracted by signal processing techniques.Therefore,we can automatically detect the HFOs in long-term iEEG signals without tags and locate the epileptic focal zone for patients with refractory epilepsy who are going to have resection surgeries.A coastline index based signal to noise ratio feature was proposed to better characterize HFO events.An artefact removal method based on a neighboring environment reference was proposed to eliminate the interference of continuous oscillatory activity in detected sporadic short HFO events.The proposed method was developed as a MatLab-based HFO detector to automatically detect HFOs in multichannel,long-term iEEG signals.The performance of our detector was evaluated on iEEG recordings from epileptic mice and patients with intractable epilepsy.More than 90%of the HFO events detected by this method were confirmed by experts,and the average missed-detection rate was less than 10%.Compared with recent related research,the proposed method achieved a synchronous improvement of sensitivity and specificity,and balanced low false-alarm rate and high detection rate.As an auxiliary tool,our detector can greatly improve the efficiency of clinical experts in inspecting HFO events during the diagnosis and treatment of epilepsy.Thirdly,to solve the problem of using epileptic activity propagation networks for the localization and resection of the epileptic focal areas represented by seizure onset zone(SOZ),high-frequency oscillation propagation networks and seizure propagation networks were combined as the integrated epilepsy propagation network.A semisupervised graph convolutional network based on pseudo-labelling and confidence coefficients was proposed to process this integrated epilepsy propagation network and localize the SOZ.In this method,in order to locate the SOZ,a high-frequency oscillation propagation network based on an unsupervised graph convolutional network was proposed with its core nodes as pseudo-labels of the SOZ;a seizure propagation network based on a local spatial averaging method was proposed with its propagation start nodes as the pseudo-labels of the SOZ.Using the iEEG records of patients with epilepsy to verify our proposed integrated epilepsy propagation network based SOZ localization method,the brain region localized by our proposed method covers 90%of the clinically diagnosed SOZ.Compared to seizure propagation networks,highfrequency oscillatory propagation networks,or other quantitative SOZ detection methods,the combined epilepsy propagation network based on graph convolutional networks was more strongly related with SOZ.Therefore,the epilepsy propagation network combined high-frequency oscillation propagation networks and seizure propagation networks may provide a better reference for epilepsy lesion resection surgery.Finally,to deal with the issue of cognitive impairment in patients with epilepsy,we comprehensively analyzed their P300 signal during cognitive and working memory tasks and spike components during slow wave sleep.Then the effect of epilepsy on P300 signal amplitude and latency can be verified.According to the abnormal P300 characteristics of patients with epilepsy,a transfer-learning-based multi-domain feature combined classification method was proposed to improve the recognition rate of P300 signals in patients with epilepsy.A stronger correlation was shown between the P300 classification accuracy and the epilepsy period than that between the P300 amplitude and the epilepsy period.A spike detection method combined multiple features was proposed to automatically detect the number of spikes in slow wave sleep EEG.It was also found that the spike frequency during slow wave sleep of patients with epilepsy detected using this detector had a significant negative correlation with the patients’P300 amplitude and P300 classification accuracy in cognitive and working memory tasks,which can be helpful and informative for the treatment of epilepsy-related cognitive impairment and effective in exploring the relationship between epilepsy and cognitive impairment.