Application Of Deep Learning Methods In EEG Signal Analysis And Classification

In recent years,brain computer interface(BCI)systems are making great progress in many applications such as motor function rehibiliation,human interaction and so on.Brain computer interfaces are designed to build direct pathways between computers and brains,which are used to transfer brain activity signals such as electroencephalography(EEG)to computer systems for processing and analyzing.EEG-based brain com-puter interfaces are one of the non-invasive brain computer interfaces that are commonly used due to the fine temporal resolution,ease of use and portability.However,the non-invasive EEG signals have apparent deficiencies including the low spatial resolution and the low signal-to-noise ratio.EEG signals refer to the electrical activities in the cerebral cortex.EEG is a common method for the diagnosis,treatment and rehibil-iation of epilepsy,stroke and other diseases.For example,stroke patients can use EEG-based brain computer interfaces to complete motor imagery tasks and get the feedback from systems to stimulate the corresponding muscles and restore the lost motor functions.Moreover,patients can replace lost functions by controlling mechanical arms and wheelchairs directly using brain computer interfaces.In neuroscience and cognitive science,EEG signals and event-related potentials are often used to analyze the brain activities of subjects in experiments.Traditional EEG processing methods consist of temporal filtering,spatial filtering,princi-ple component analysis,independent component analysis and so on.Power spectral density and common spatial pattern algorithms are frequently used to extract features from preprocessed EEG data.Classification algorithms are applied on the extracted features in traditional machine learning architectures.In our paper,different classifiers are included to test the classification accuracies of our proposed architectures.The classi-fiers are composed of k-nearest neighbors,support vector machine,random forest,naive bayes and multilayer perceptron.Deep learning aims to build the hierarchical neural networks to model high-level representations from raw data.Compared to traditional machine learning methods,which only extract low level features using hand-craft assumption,our proposed method attempts to automatically model high-level representations from raw data by using hierarchical model architectures,with complex structures or otherwise,composed of multiple non-linear transformations.Deep learning techniques have achieved great success in various tasks,especially in computer vision,speech and natural language processing.Deep learning applications in brain computer interface and EEG recognition tasks are still in an experimental phase,whose applicability requires verifi-cation.Attempts have been made in some motor imagery and emotion recognition tasks.In this paper,we propose an end-to-end deep learning based architecture for recognizing EEG signals,by applying some pop-ular deep models such as convolutional neural networks and recurrent neural networks.Deep forest model is also applied which envolves ensemble learning and cascaded architecture.After the preprocessing steps involving noise rejection and decorrelation,we can use the preprocessed data directly to train the model and get the corresponding output.Compared to shallow models,deep models can represent data in a more hi-erarchical way,which helps the model to perform better in generalization ability.Multi-frame architecture and recurrent neural networks can be utilized to capture long term dependencies in EEG series and enhance the recognition and classification accuracy.Specifically,we focus on the comparison of traditional machine learning methods and deep learning methods on the classification tasks of several EEG datasets.The first one is the dataset of BCI Competition Ⅲ Dataset Ⅴ which contains the cued mental imagery EEG with 3 classes.The second is the motor imagery EEG dataset we collected from 5 stroke patients after the operation,which is more difficult to find common patterns for recognition.Finally,we use the intracranial EEG dataset provided by Kaggle which contains long duration recordings collected from patients suffering from epilepsy.The results systematically validate the superior performance of our method despite the limited size of datasets.