Improved Recognition of Error Related Potentials through the use of Brain Connectivity Features Huaijian Zhang, Ricardo Chavarriaga, Mohit Kumar Goel, Lucian Gheorghe, and Jos´ e del R. Mill´ an Abstract— Brain error processing plays a key role in goal- directed behavior and learning in human brain. Directed transfer function (DTF) on EEG signal brings unique features for discrimination between correct and error cases in brain- computer interface (BCI) system. We describe the first applica- tion of brain connectivity features for recognizing error-related signals in non-invasive BCI. EEG signal were recorded from 16 human subjects when they monitored stimuli moving in ei- ther correct or erroneous direction. Classification performance using waveform features, brain connectivity features and their combination were compared. The result of combined features yielded highest classification accuracy, 0.85. In addition, we also show that brain connectivity at theta band around 200ms after stimuli carry highly discriminant information between error and correct trials. This paper provides evidence that the use of connectivity features improve the performance of an EEG based BCI. I. INTRODUCTION Error related brain activity has been studied in last decade with great interest for its crucial role in goal-directed behav- ior and learning [1], [2]. Electrophysiological recordings and fMRI studies suggest that error processing involves the dorsal anterior cingulate cortex (ACC) and the medial prefrontal cortex (PFC) [2]. Furthermore, specific brain interaction patterns after presentation of erroneous stimuli have been reported by studies using fMRI and Stereoelectroencephalog- raphy (SEEG) signals, in particular, a network comprising the anterior cingulate cortex and other neural sources, including dorsolateral prefrontal cortex, parietal lobe, medial temporal lobe, and thalamus [1], [3], [4]. In particular, it has been reported that error-related poten- tials can be detected in scalp EEG recordings with human subjects. Typically they consist of an error-related negativity (ERN) located in frontocentral areas, followed by an error positivity (EP) in centroparietal areas. Because of its key role in human brain function, error related potentials have been proposed as input for non-invasive Brain-Computer Interfaces (BCI). Single trial detection of error potential has been applied to monitor erroneous stimulus [5] or during interaction with external devices [6] [7]. So far, all these systems used only waveform or spectral information as *This study was supported by Nissan Motor Co. Ltd., and carried out under the “Research on Brain Machine Interface for Drivers” project. M.K.Goel is supported by the Swiss National Science Foundation, Project 200021-120293. H. Zhang, R. Chavarriaga, M. K. Goel, and J. d. R. Mill´ an are with the Defitech Foundation Chair in Non-Invasive Brain-Machine In- terface, Center for Neuroprosthetics, ´ Ecole Polytechnique F´ ed´ erale de Lausanne (EPFL), Switzerland. L. Gheorghe is from Mobility Ser- vices Laboratory, Nissan Research Center, Nissan Motor Co., Japan. huaijian.zhang@epfl.ch jose.millan@epfl.ch discriminant features. We propose that channel interaction and phase differences can also be used as features for classification in BCI systems [8], [9]. We assess the use of directed transfer function (DTF) as a feature extraction method that reflects directional connectivity across multiple channels. This method, which is an extension of Granger causality from pairwise variables [10], allows the estimation of the directed information transfer between multi-variables. It has been applied to compute the brain connectivity in several areas, including localization of epileptic foci [11] and memory information processing [12]. In this study, EEG signal was recorded when human subjects monitored the direction of a moving square. Classifi- cation performance between correct and erroneous movement direction using DTF features was evaluated, and the char- acteristics of connectivity features in temporal, frequency domains and brain regions were assessed. II. METHODS A. Experimental protocols In the experiment, subjects were seated in front of a computer screen, placed at about 50cm from their eyes. At the beginning of the experiment, 11 empty white squares arranged horizontally are shown. An orange target square appeared randomly to either leftmost or rightmost position. It was followed by a blue cursor square in the central location. Then, the cursor moved towards the target square with 80% probability every 2s. The cursor kept on moving until reaching the target where it changed color to green. If the target was not reached before 40s, the trial was stopped. During the experiment, subjects were requested to minimize eyes blinking and continuously focus their eyes on the next position of the cursor until 0.5s after the cursor reaching the new position. The trials that moved towards the target were considered as correct trials, whereas movements in the opposite direction were considered as error trials. For each subject, more than 100 error trials and more than 400 correct trials were performed. 10 subjects (3 females, age 26 ± 2.40) were included in the experiment. The data from this experiment is referred to as dataset1 in the following sections. In addition, we also report results on 6 subjects (1 female, age = 27.83 ± 2.23) from a previous study using a similar protocol (here denoted dataset2) [5]. B. Directed Transfer Function We used directed transfer function (DTF) to estimate the brain information interaction between EEG channels. The DTF method is based on multivariate autoregressive (MVAR)