Unsupervised Feature Extraction And Classification Of EEG Signals Based On Learning In The Model Space

The emergence of advanced diagnostic tool such as electroencephalogram(EEG)has facilitated real-time collection of large amounts of EEG signals from epileptic patients,which can help to detect epilepsy and monitor signs of seizure.However,manual analysis of large amounts of raw EEG signals by medical specialists is often exhausting and prone to human error,where an effective feature extraction method is crucial to study these raw EEG signals.This paper will present an unsupervised feature extraction method to the study of EEG signal analysis,without the need to extract an extensive number of hand-crafted features from raw EEG signals.Moreover,instead of investigating the abnormality of EEG signals in the original signal space,discrimination between signals is performed in the so-called model space.This paper further extends the concept of learning in the model space,by training a self-predictive model to allow cross-domain classification,which improves the generalization ability of the model.The main work of this dissertation include:1)unsupervised feature extraction of raw EEG signals;in order to replace manual feature extraction with machine-learning-based model,a hybrid approach of combining echo state network(ESN)and principal component analysis(PCA)is adopted to achieve unsupervised feature extraction.Since the raw EEG signals is extremely chaotic and variational among individuals,traditional way of extracting features from different domains has drawbacks,such as the need for domain-specific knowledge,low generalization ability,and long training process.In order to solve the problems,hybrid feature extraction method based on reservoir model is presented here in the study.The experimental results on two public datasets show the robustness and effectiveness of the features extracted by the proposed method,which yielded the highest classification accuracy and produced excellent clustering results.2)cross-domain classification based on transductive transfer learning;it is mandatory to train a patient-specific model with sufficient amount of labeled data,in order to obtain great classification result.However,limited training data available in the medical field,as well as an imbalance of epileptic signals in the training data can affect the training process.Moreover,previous studies on EEG signal identification assume that the data distributions between the training and testing data are identical,which is highly unlikely in reality.In order to tackle these problems,this dissertation adopts a novel method of transductive transfer learning to deal with the drifting distributions.It enables transfer of knowledge between target domain and source domain,by training a self-predictive model to study a linear transformation matrix for conversion of data between domains.The experimental results show that by adding such self-predictive model to the existing framework,successful knowledge transfer between target domain and source domain can be achieved,which improves the generalization ability of the existing model in performing EEG signal classification.The methodologies used in this dissertation were developed based on the concept of learning in the model space,where any time series can be learned by a fitting model,the learnt model should give a better representation to the original time series.Based on such concept,there are two ways of measuring the distance of any two sequences 1)using an adapted Euclidean distance on the readout parameters to calculate the distance between two readout models;2)comparing how well a trained model learned on one sequence fits the other,the distance between the two time-series can be approximated by the prediction error of a model trained on one time series and applied to the other.Based on the concept of learning in the model space,this dissertation performed unsupervised feature extraction and cross-domain classification of EEG signals.