Abstract—Brain-computer interface (BCI) provides patients suffering from severe neuromuscular disorders an alternative way of interacting with the outside world. The P300-based BCI is among the most popular paradigms in the field and most current versions operate in synchronous mode and assume participant engagement throughout operation. In this study, we demonstrate a new approach for assessment of user engagement through a hybrid classification of ERP and band power features of EEG signals that could allow building asynchronous BCIs. EEG signals from nine electrode locations were recorded from nine participants during controlled engagement conditions when subjects were either engaged with the P3speller task or not attending. Statistical analysis of band power showed that there were significant contrasts of attending only for the delta and beta bands as indicators of features for user attendance classification. A hybrid classifier using ERP scores and band power features yielded the best overall performance of 0.98 in terms of the area under the ROC curve (AUC). Results indicate that band powers can provide additional discriminant information to the ERP for user attention detection and this combined approach can be used to assess user engagement for each stimulus sequence during BCI use. I. INTRODUCTION rain-computer interface (BCI) is an emerging and rapidly growing research area that enables the development of systems that bypass the brain’s normal communication pathways of nerve and muscle, allowing the brain to communicate directly with the external world. Clinical applications of BCIs are targeted for patients suffering from severe neuromuscular disorders to provide them with an alternative way of interacting with the world. The P300-based BCI is among the most popular paradigm in the field due to its ease of use, high performance and reliable signal it can offer [1]. Classification of signals in BCI relies on the P300 event-related potential (ERP) that is elicited in the oddball paradigm. In a P300-based BCI, sequences of visual or audio stimuli are presented to the users of the BCI who are asked to focus their attention on the occurrence of rare target stimuli among more frequent non-targets. The P300 waves are generated by the brain after the user recognized the target stimuli. These P300 waves are then detected and translated to perform actions such as turning on a switch in environmental control, or choosing a letter during a spelling task [1]. Manuscript received March 15, 2012. Y. Liu, H. Ayaz, A. Curtin and B. Onaral are with the Drexel University, School of Biomedical Engineering, Science and Health Systems, Philadelphia, PA 19104 USA. P. Shewokis is with the College of Nursing and Health Professions, Drexel University, Philadelphia, PA 19104 , USA Phone: 215-571-3709; Fax: 215-571-3718; e-mail: (yl565, ayaz, abc48, shewokis, banu.onaral)@drexel.edu. Research on the P300-based BCI has been mainly focused on three areas: 1) Stimulus presentation paradigm, which concerns designing different presentation modes such as row/column, single cell and checkerboard for faster communication speed, better ERP signal to noise ratio or to enhance user experience. 2) Feature extraction and classification algorithms, which involves developing algorithms to translate EEG signals into actions more accurately, and 3) novel applications. A detailed review of the current status and future directions of P300-based BCI research can be found in [1]. Most BCI systems proposed in the literature are synchronized, requiring that users must follow the pace of the BCI system and that they have no control of when to start and stop using the BCI. A more practical BCI should allow users to interact with it in an asynchronous manner. To achieve this goal of interactive asynchrony, it is critical to determine whether or not the user is attending to the BCI system. An asynchronous P300-based BCI was first proposed by Zhang et al [2] and studied by various other groups [3-6]. These studies can be grouped into two categories according to the methods applied for user attention detection: those based on statistical analysis of the P300 wave amplitude features, such as [2-4]; and those based on a hybrid BCI, such as in [5] where the steady state visually evoked potentials (SSVEP) paradigm were applied in conjunction with the P300 paradigm, and in [6] where the event-related desynchronization (ERD) paradigm was employed. In this study, we investigated using band powers as features for user attention detection. It has been long known in the literature that ERP is related to the rhythmic activity of the brain [7-10]. However, most of the studies investigated rhythmic activity using time-frequency analysis during the time course of P300 activity. It is still unknown whether band powers are capable of characterizing whether a user is attending (engaged actively) to the P300-based BCI. To investigate this band power active engagement, we first compared the delta (0.5-4Hz), theta (4-8Hz), alpha (8- 13Hz) and beta (13-30Hz) activities of subjects when they were actively engaged in the P300-based BCI to those times when they were not engaged. Secondly, we used the rhythmic activities of selected bands and channels as feature for user attention classification. II. MATERIALS AND METHODS A. Participants Ten right-handed volunteers (ages between 20 to 24 years) participated in the study but one subject was excluded Detection of attention shift for asynchronous P300-based BCI Yichuan Liu, Hasan Ayaz, Adrian Curtin, Patricia A. Shewokis, Banu Onaral B 34th Annual International Conference of the IEEE EMBS San Diego, California USA, 28 August - 1 September, 2012 3850 978-1-4577-1787-1/12/$26.00 ©2012 IEEE