EEG based mental tasks recognition for neurofeedback systems

• Main Author: Wang, Qiang
• Other Authors: Olga Sourina
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: http://hdl.handle.net/10356/56302
• Institution: Nanyang Technological University

Electroencephalography (EEG) is the time series recording of electrical activities along the scalp, which allows monitoring the brain activities with high temporal resolution. EEG techniques get more attention due to novel EEG devices which are portable and mobile. These EEG devices allow the user to setup EEG-based application systems easily within seconds. Another reason that puts EEG techniques under the spotlight is that applications based on these techniques are successfully applied in both medical applications and en- tertainment. EEG-based monitoring the state of the user’s brain functioning and giving her/him the visual/audio/tactile feedback is called neurofeedback technique, and it could allow the user to train the corresponding brain functions. It could provide an alternative way of treatment for some psychological disorders such as Attention Deficit Hyperactivity Disorder (ADHD), where concentration function deficit exists, Autism Spectrum Disorder (ASD), or dyscalculia where the difficulty in learning and comprehending the arithmetic exists. In our work, a novel method for multifractal analysis of EEG signals named generalized Higuchi fractal dimension spectrum (GHFDS) is proposed first for quantification of the self-similarity of EEG signals and applied in mental arithmetic tasks recognition from EEG signals. Then, we investigate different method for ocular artifacts detection. An experiment for different types of ocular movements is designed and implemented for 5 subjects with Emotiv device. Two types of ocular movements such as slow blinking and fast blinking can be recognized from one EEG channel with high accuracy (81.98%). Real-time mental task recognition is studied. Experiment for arithmetic mental task and relax task is implemented for 10 subjects with Emotiv device. GHFDS is applied in mental arithmetic task recognition from EEG signals. The usage of the proposed fractal dimension spectrum of EEG signal in combination with other features can improve the mental arithmetic task recognition accuracy in both multi-channel and one-channel subject-dependent algorithms up to 97.92% and 85.81% correspondingly. Based on the channel ranking, four channels (F8, F3, O2, AF3) were chosen which gave the accuracy up to 97.11%. A mental arithmetic task experiment for 8 subjects is designed and implemented with different levels of mental workload with Emotiv device. The classification accuracy for multi-level mental arithmetic task recognition is 53.94% for 6 class (baseline and 5 different levels of mental arithmetic tasks) and 74.87% for 4 classes (baseline and 3 different levels of mental arithmetic tasks) with 14 EEG channels. A well-known benchmark EEG database was also used for data analysis and testing. 7 subjects participated in the experiment for two sessions. In each session, five mental tasks such as Relax (Baseline), Counting (Count), Letter Composition (Letter), Multiplication (Math), and Rotation (Rotate) are repeated for five trials. EEG signals for 10 seconds period from C3, C4, P3, P4, O1, O2 and EOG channels were recorded in each trial. Different types of features extracted from EEG signals are compared based on the mental tasks classification accuracy using SVM classifier. These features include relative powers in different EEG bands, autoregressive (AR) coefficients, Higher Order Crossings (HOC), statistical features, entropy, and GHFDS. Among these features, statistical features can obtain the best classification accuracy. However, for different subjects, the feature importance rank could be different. It is found that GHFDS features could also be significant for some subjects. Random forests (RF) method is used for supervised feature selection. The num- ber of features used in the SVM classifier was significantly reduced while the classification accuracy slightly increased. Then, the channel importance is calculated according to the importance of the features belonging to that channel where the features importance was defined by RF feature selection process. Channel from occipital lobe (O2) had the highest importance rank as this lobe is responsible for visual perception. Real-time mental task recognition was also applied in neurofeedback system in our work. Neurofeedback games with 3D virtual scenes for relax training and work performance training were proposed and implemented.