Hybrid Approach in Recognition of Visual Covert Selective Spatial Attention based on MEG Signals S.A. Hosseini, Student Member, IEEE, M.-R. Akbarzadeh-T., Senior Member, IEEE, M.-B. Naghibi-Sistani Dept. of Electrical Engineering, Center of Excellence on Soft Computing and Intelligent Information Processing, Ferdowsi University of Mashhad, Mashhad, Iran Emails: {Sa.hosseyni@stu.um.ac.ir, Akbazar@um.ac.ir, Mb-naghibi@um.ac.ir} Abstract— This paper proposes a reliable and efficient method for recognition in two different orientations (either left or right) by Magnetoencephalograph (MEG) signals. The brain activities are measured using different approaches with different spatial and temporal resolutions. The MEG signals are usually used for brain–computer interface (BCI) applications due to high temporal resolution. The MEG signals were recorded from different brain regions of four different human subjects during visual covert selective spatial attention task. The hybrid method proposes pre-processing; feature extraction by Hurst exponent, Morlet wavelet coefficients, and Petrosian fractal dimension; normalization; feature selection by p-value; and classification by support vector machine (SVM) and fuzzy support vector machine (FSVM). The results show that the proposed method can predict the location of the attended stimulus with a high accuracy of 91.62% and 92.28% for two different orientations with SVM and FSVM, respectively. Finally, these methods can be useful for BCI applications based on visual covert selective spatial attention. Keywords— attention; magnetoencephalograph; brain– computer interface; cognitive system. I. INTRODUCTION The brain receives many inputs from different sensory systems, but the capacity of information processing is limited, therefore, it cannot fully process all the objects or stimuli at any given time [1]. Attention due to its considerable influence on many brain activities is one of important topics in psychology and cognitive science. It has many medical applications such as in rehabilitation, attention deficit hyperactivity disorder and autism, as well as many applications in brain–computer interface (BCI), decision making, security, transportation and robotics [2]-[9]. There are several mechanisms for attention in the brain. One of the most used mechanistic models for the attention is the Sohlberg and Mateer model [10]. According to this model, there are various types of attention such as focused, sustained, selective, alternating, and divided. Visual selective attention refers to the limitations of human processing capacity in multiple simultaneous stimuli [11]-[14]. Therefore, attention are needed to allow the organism to select for further processing information that is currently task-relevant, while ignoring other information that is not relevant [1]. Attention can be distinguished into overt and covert orienting [15]. “Overt attention is the act of selectively attending to a location over others by moving the eyes to point in that direction. Covert orienting is the act to mentally shifting one's focus without moving one's eyes” [15]-[17]. Attention of other categories can classified as feature [18],[19], spatial [20], and object-based [21]. Evidence for spatial selection comes mostly from spatial cueing studies [22]. In these researches, spatial attention is varied by pre- cueing the location where the target stimulus is likely to appear [1]. Recent studies have manifested the potential of brain signals as an attractive alternative to BCI. A successful BCI system provides a direct control pathway from brain to external devices such as computers, wheelchairs, or robotic limbs [23]. During such experiments, human subjects are asked to perform mind processes such as cube rotation, press a mouse button, move one finger, or count the number of stimulations. In these systems, suitable brain signals serve as control signals for output external devices. These signals are usually categorized according to their invasive or non-invasive approaches with a particular emphasis on attention applications [24]. The invasive techniques consist of implanted multi- electrode grids in the motor cortex of paralyzed patients [25], pre-motor cortex of monkeys [26], or parietal motor command regions [27], electrocorticograms (ECoG) [28], action potentials (APs) from nerve cells [29],[30] and synaptic and extracellular local field potentials (LFPs) [31],[32]. The non- invasive techniques use signals such as electroencephalography (EEG) [33], magnetoencephalography (MEG) [34]-[36] near-infraRed spectroscopy (NIRS) measuring cortical blood flow [37], functional near-infraRed spectroscopy (fNIRS) [38], functional magnetic resonance imaging (fMRI) [39] and optical imaging [40]. Invasive methods often lead to efficient BCI systems, but they have inherent technical difficulties such as the risk associated with surgical implantation of electrodes. Therefore, the non-invasive techniques such as EEG and MEG signals are generally preferred. Cohen first introduced MEG signals in 1968 [41]. To record MEG signals, an array of gradiometers is placed over the head-surface. These signals typically contain frequencies of up to 100Hz. Therefore, they can be divided into several frequency-ranges such as Delta (0.1-4Hz), Theta (4-8Hz), Alpha (8–14Hz), Beta (12-30Hz) and Gama (30- 100Hz). This research is supported by the cognitive sciences and technologies council in the Iran.