| Epilepsy is a kind of neurological disorders with high prevalence.Childhood epilepsy is more complicated and difficult to diagnose than adult epilepsy because of the immature brain development.Children with central temporal spikes in the electroencephalogram(EEG)will be diagnosed as childhood benign epilepsy with central temporal spikes(BECT)or other types of epilepsy.The analysis of the EEG data and the medical scales facilitates the diagnose and helps the doctor understand overall situation of patients.However,the analysis process is complicate and intricate,it is time and energy consuming.It is necessary to use machine learning methods to achieve facilitating diagnose with higher accuracy and less time-consuming.The main contents of this thesis are as follows:1.A method of EEG signal preprocessing is proposed.It is mainly based on resting state EEG data.The preprocessing analyzes and removes the noise generated during the EEG acquisition,including the noise filtering and artifacts removal.The noise filter removes noise such as the baseline drift and power frequency interference.The artifacts removal is based on EEGLAB and ADJUST plug-in,which removes artificial components such as eye movement artifacts and electromyographic artifacts.The proposed method designed to be fast and straightforward,suitable for researchers with no medical background.2.Classification method based on EEG signals for children diagnosed with BECT or non-BECT is proposed.For children with center temporal spikes,there is high misdiagnose or missed diagnose rate.The work in thesis proposed two classification methods based on the patient’s EEG signals.The first proposed method uses the morphological filter to detect the position of the sharp wave,and extract the features based on the geometric structure of the sharp wave,then combines with random forest(RF)and extreme random forest(ERF).With ten-fold cross-validation,we have 73%accuracy.The second proposed method uses deep neural networks VGG16 and Res Net18.The multi-channel EEG signal is used as the input,and the classification accuracy rates of 91.72% and 90.49% are obtained respectively,which reflects the important significance of the deep learning method in medical facilitating diagnosis.Through comparison,the evaluation indexes of the proposed methods are better than the 72.40% accuracy rate of the existing method.3.The sleep habit is predicted and analyzed based on EEG signals.After the sleep habit scale analysis,we found that children with central temporal spikes have significant differences in sleep from the reference standard.Therefore,two methods are designed to predict and analyze the total score and eight dimensions in the sleep habit scale.The first proposed method uses the principal component analysis(PCA)combined with the gradient boosting regression tree(GBRT)model.After ten-fold cross-validation,the Mean Absolute Error(MAE)is 7.18.Different dimensions in the sleep habit scale contain different sleep behaviors,so the prediction of each dimension can estimate different sleep habits.The second proposed method extracts the Cz channel of the EEG data and converts it into the spectrogram by the short-time Fourier transform(STFT)method,and applies the improved Res Net18 to perform prediction.The improved Res Net18 model treats the score prediction as a multi-classification task,and uses the sum of the product of the probability obtained after the network through the Soft Max function and the corresponding score interval as the predicted value,and the total sleep score get the result of MAE of 0.46.The cross-validation method between individuals is applied,with every patient used as a new test dataset which has not been trained in the model.With cross-validation between individuals the MAE is 5.49.The two proposed methods can effectively predict the patient’s sleep scale score,help to evaluate the sleep habits of patients,assist the clinician to evaluate the clinical behavior of patients and formulate the treatment methods. |
PDF Full Text Request