Abstract— Brain computer interfaces (BCI) are used for communication and rehabilitation. One of the main categories of BCI techniques is motor imagery based BCI (MI-BCI). A large number of studies have focused on machine learning approaches to optimize MI-BCI performance. However, enhancement of MI-BCI through provision of optimized feedback modalities has not received equal attention. Motor imagery and motor execution activate almost the same area of the brain. During motor skills performance, a combination of proprioceptive and direct visual feedback (PDVF) is provided. Thus, we hypothesized that MI-BCI that receives PDVF outperforms the traditional MI-BCI, which only uses indirect visual feedback (IVF). We studied 8 healthy subjects performing MI through (i) IVF and (ii) PDVF. We used 8 channel electroencephalogram (EEG) signals and extracted features using an autoregressive model and classified MIs using linear regression. On average, PDVF increased the accuracy of MI performance by 11%, and improved information transfer rate (ITR) by more than two times. In conclusion, using PDVF appears to improve MI-BCI performance according to the studied metrics, making this approach potentially more reliable. Keywords—EEG; motor learning; brain-computer interfaces; motor imagery; information transfer rate, accuracy I. INTRODUCTION Brain-computer interface (BCI) technology has established the foundation for the human brain to communicate with machines directly. Motor imagery (MI) based BCI (MI-BCI) that relies on the rhythm changes occur within the sensorimotor area of the brain during MI [1], is one of the main BCI paradigms. In non-invasive MI-BCI, the brain activity during MI is recorded using EEG [2], functional magnetic resonance imaging [3] (fMRI), or near infrared spectroscopy (NIRS) [4]. Among the aforementioned techniques, EEG is the most practical and affordable technique and thus, the most commonly exploited modality in non-invasive MI-BCI applications. One of the challenges of MI-BCI is its rather low accuracy and information transfer rate (ITR). This drawback limits the dissemination of MI-BCIs for widespread application. Provision of optimum feedback is believed to improve MI- BCI performance metrics [5]. Proprioceptive feedback, visual feedback, or different combinations thereof are among the most common feedback modalities in MI-BCIs [6]. While S. Darvishi, M. Baumert, and D. Abbott are with the Centre for Biomedical Engineering, School of Electrical and Electronic Engineering, The University of Adelaide, SA 5000 (phone:+618-8313-4115; e-mail: sam.darvishi@adelaide.edu.au, mathias.baumert@adelaide.edu.au, derek.abbott@adelaide.edu.au). M.C. Ridding is with the Robinson Institute, School of Pediatrics and Reproductive Health, The University of Adelaide, SA 5005 (michael.ridding@adelaide.edu.au). visual feedback is mostly supplied via cursor position update on a monitor [7], proprioceptive feedback has been provided using either orthoses [8] or robots [9]. Nijboer et al. [10], investigated suitability of auditory feedback for MI-BCI, and found its performance comparable with indirect visual feedback (IVF). Ramos-Murguialday et al. [11], applied concurrent proprioceptive and direct visual feedback (PDVF) as a feedback modality in MI-BCI restorative applications. PDVF showed increased accuracy of subject response to MI compared to either no feedback or sham feedback. However, they did not compare PDVF with other feedback modalities. Motor execution and motor imagery of a particular task, activate almost the same area of the brain [12]. Thus, in search for optimization of feedback modality for MI-BCI we surveyed different feedback types in motor learning. Enough repetition of a movement, followed by feedback, results in motor learning in healthy subjects. Intrinsic feedback is realized through proprioceptive and/or visual sensory inputs as a result of the performed motor task. Extrinsic (augmented) feedback, however, is provided artificially by an external agent to enhance the motor learning outcomes; an example of this are athletes who learn new moves via auditory feedback from the coach [13]. When augmented feedback is added to intrinsic feedback, it improves the retention and motor learning outcomes by provision of knowledge of performance and/or knowledge of result [14]. In contrast to motor learning, there is no muscle activation during motor imagery and, therefore, no source of feedback. As a consequence, an external actuator is required to supply extrinsic feedback in MI-BCI setups. Provision of IVF through updating the cursor position on a monitor is currently the most ubiquitous feedback modality in BCI applications [11]. This type of feedback provision might be quite effective for some BCI applications, such as in the P300-based Speller [15]. However, considering the outcomes of motor learning studies on feedback modalities [16], IVF may not be as effective in MI-BCI because it lacks intrinsic (direct) feedback to close the sensorimotor loop. By contrast, PDVF, in addition to the augmented feedback of IVF, provides intrinsic visual and proprioceptive feedback. While, PDVF provides feedback that is closest to motor learning, supplying IVF via updating cursor position on a monitor remains the most prevalent feedback modality in MI- BCI setups. Recently, Lotte et al [17], suggested that current BCI training approaches that use IVF were suboptimal and need to be improved. Thus, to investigate alternative feedback modalities for MI-BCIs we compared two similar BCI designs that used either IVF or PDVF in eight BCI-naive subjects. According to our results, PDVF seems to be superior to the traditional IVF that promotes the application of PDVF to make MI-BCIs more efficient and accurate. Does Feedback Modality Affect Performance of Brain Computer Interfaces? Sam Darvishi, Member IEEE, Michael C. Ridding, Derek Abbott, Fellow IEEE, Mathias Baumert, Senior Member IEEE. 7th Annual International IEEE EMBS Conference on Neural Engineering Montpellier, France, 22 - 24 April, 2015 978-1-4673-6389-1/15/$31.00 ©2015 IEEE 232