Data-point and Feature Selection of Motor Imagery EEG Signals for Neural Classiication of Cognitive Tasks in Car-Driving Anuradha Saha, Amit Konar, Pratyusha Das Electronics and Telecommunication Engineering Jadavpur University Kolkata-700032, India anuradha.nsec@gmail.com, konaramit@yahoo.co.in, pratyusharj@gmail.com Basabdatta Sen Bhattacharya School of Engineering University of Lincoln Lincoln, UK bbhattacharya@lincoln.ac.uk Atulya K. Nagar Department of Math and Computer Science Liverpool Hope University Liverpool, UK nagara@hope.ac.uk. Abstract- This paper proposes novel algorithms for data-point and feature selection of motor imagery electroencephalographic signals for classifying motor plannings involved in car- driving including braking, acceleration, let steering control and right steering control. Variants of neural network classiiers such as linear support vector machines, and kernel-based support vector machines including radial basis function kernel, polynomial kernel and hyperbolic kernel have been applied to classify the various cognitive tasks. Experimental inding reveals that the proposed data-point and feature selection technique altogether provides better classiication accuracies (more than 88%) for all cognitive tasks in comparison with using factor analysis for data point reduction and feature selection. It is also observed that power spectral density and discrete wavelet transform features are selected among the list of electroencephalographic features for holding the top two rank values for cognitive task classiication during car-driving. From the experimental result, it is conirmed that support vector machines with radial basis function along with power spectral density outperforms the remaining feature-classiier pairs in terms of average classiication accuracy. Keywords- data-point selection, feature selection, suppot vector machine neural network, skewness, dff erential evolution, motor imagey, electroencephalography. I. INTRODUCTION Car-driving is a complex task involving several motor activities in dynamic environments. Motor activities such as acceleration, braking, steering right control and steering let control during driving includes both motor planning and motor execution. Recent studies r I H 61 provide a number of methods to classiy drivers' cognitive states during driving to avoid traic fatalities. From the previous literature, it can be conIrmed that besides timely motor execution, driver needs to plan correct motor intensions for a particular driving instance, and hence correct classiIcation of motor intensions or planning, i.e. let hand motor imagination during let-hand 978-1-4799-1959-8/15/$31.00 @2015 IEEE steering control, right hand motor imagination during right hand steering control, let leg motor imagination during braking, right hand motor imagination during acceleration rom EEG signals are very important concen. This paper, too attempts to classiy motor imagery signals for above motor imagination tasks, which in tun gives a clear indication of drivers' cognitive failures during motor planning phase. The present problem is therefore solved by selecting data-point and features of motor imagery EEG signals, and thus minimizing the misclassiIcation rate due to psychological hindrances or cognitive failures. The Irst novelty of the present paper is to provide a new approach to data-point reduction. Data point reduction is an important issue in motor imagery-based classiIcation problem in order to select one unique class-representative rom a large set of data points (trials). Motor imagery signals, being nondeterministic by nature, does not offer the unique features extracted rom several trials of the motor imagery EEG signals captured rom the same subject for the similar cognitive task. Therefore, we need to identiy the ideal class representative of each data point representing a feature vector of Ixed dimension. Previous literature [7, 8] reveal the use of a very popular technique of data-point reduction using PCA. We, here identiy one unique representative for each motor imagery class by determining skewness of data-points acquired for that class. The second novelty, as has been addressed in the paper, is feature selection. An EEG patten is described by its feature. Feature extraction and selection, therefore are considered as the important steps in motor imagery EEG signal processing. Here, we select well-known EEG feature extraction techniques including time- (Hjorth parameters, Adaptive autoregressive parameters), requency- (Power spectral density, or in short PSD), and time-requency (Discrete wavelet transform, or in short DWT) domain techniques to extract motor imagery features rom acquired EEG signal depending on the intended