Neural Network-based Three-Class Motor Imagery Classification Using Time-Domain Features for BCI Applications Mahyar Hamedi, Sh-Hussain Salleh, Alias Mohd Noor Center for Biomedical Engineering, Universiti Teknologi Malaysia Johor Bahru, Malaysia hamedi.mahyar@ieee.org, hussain@fke.utm.my Iman Mohammad-Rezazadeh David Geffen School of Medicine. Semel Institute for Neuroscience and Human Behavior, University of California, Los Angeles & Center for Mind and Brain, University of California, Davis, USA irezazadeh@ucdavis.edu Abstract—Many studies have reported the usefulness of motor imagery (MI) electroencephalogram (EEG) signals for Brain Computer Interface (BCI) systems. MI has been broadly characterized by the average of event-related changes of brain activity at specific frequency bands; but, temporal features of EEG have rarely been considered to identify different mental states of BCIs’ users. Additionally, complex classification techniques may have been proposed to enhance the accuracy of system but they may cause a notable delay during online applications. This paper investigated the application of neural network-based algorithms to classify three-class MIs by utilizing EEG time-domain features. Integrated EEG (IEEG) and Root Mean Square (RMS) features were extracted from EEG signals. Then, Multilayer Perceptron and Radial Basis Function Neural Networks were employed to classify the features. The discrimination ratio of such features were examined and compared through different classifiers. Moreover, the robustness of classifiers was investigated and compared. The results of this study indicated that RMS was more capable than IEEG for characterizing MI movements and RBF was more accurate and faster than MLP. The effectiveness of IEEG and RMS features and the performance of MLP and RBF classifiers were compared with Willison Amplitude (WAMP) feature and support vector machine (SVM) classifier respectively. This study proved that WAMP and SVM were more efficient for classification of MI tasks in both terms of accuracy (88.96%) and training time (0.5 second); however, considerable difference was not observed since RBF performed as fast as SVM with only about 3% less accuracy. Keywords—Brain Computer Interface; Electroencephalogram; Motor Imagery; Time-Domain Feature; Classification. I. INTRODUCTION Brain Computer Interface (BCI) aims to create a new communication pathway that can translate sever disabled intentions into control signals to drive a rehabilitation device or a neuroprosthesis [1]. Various BCI systems using EEG signals have been introduced among which Motor Imagery (MI)-based is one of the most promising [2] especially for paralyzed patients and asynchronous BCIs [3]. An efficient pattern recognition-based BCI system includes several key components where feature extraction and classification are the most challenging ones. EEG signals are non-stationary and therefore extraction of accurate and discriminative features is necessary to represent the underlying mental tasks. Moreover, features have a direct impact in determining the performance of a classifier; thus, inaccurate features may lead to poor classification ratio and computational complexity [4]. In EEG-based BCIs, feature extraction methods are categorized into time-domain, frequency-domain, time-frequency analysis, and spatial-domain methods. Autoregressive (AR) modeling is the most commonly used in the time-domain analysis of MIs. Quantification of Event- Related Synchronization/ Desynchronization (ERS/ERD) phenomenon is one of the popular techniques for characterizing EEG power spectra. Morlet Wavelet transform, Empirical Mode Decomposition (EMD), Hilbert-Huang Transform (HHT), short-term Fourier transform (STFT) are examples of Time-Frequency methods that have been proposed in this area. Spatial feature extraction method includes Principal Component Analysis (PCA), Independent Component Analysis (ICA), Common Spatial Pattern (CSP) and Surface Laplacian Derivation (SLD). In addition to feature extraction method, classification algorithm also plays a dominant role in achieving reasonable performance for MI-based BCI systems. Numerous linear and non-linear algorithms have been suggested in this area among which Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), K-Nearest Neighbor [5] are widely used. Literature reviews indicate that researchers are still looking for more effective and robust techniques in order to boost the system performance by employing more complex feature extraction and classification algorithms regardless of how much computational cost may be added to overall system. Nonetheless, a practical and effective BCI system requires a reliable trade-off between its outputs accuracy and interface complexity. Considering this fact, recently, several simple time-domain features Mean Absolute Value (MAV), Maximum value (MAX), Simple Square Integral (SSI), Willison Amplitude (WAMP), Waveform Length (WL) were evaluated 2014 IEEE Region 10 Symposium 978-1-4799-2027-3/14/$31.00 ©2014 IEEE 204