Robust asynchronous control of ERP-Based brain-Computer interfaces using deep learning

Abstract

Background and Objective. Brain-computer interfaces (BCI) based on event-related potentials (ERP) are a promising technology for alternative and augmented communication in an assistive context. However, most approaches to date are synchronous, requiring the intervention of a supervisor when the user wishes to turn his attention away from the BCI system. In order to bring these BCIs into real-life applications, a robust asynchronous control of the system is required through monitoring of user attention. Despite the great importance of this limitation, which prevents the deployment of these systems outside the laboratory, it is often overlooked in research articles. This study was aimed to propose a novel method to solve this problem, taking advantage of deep learning for the first time in this context to overcome the limitations of previous strategies based on hand-crafted features. Methods. The proposed method, based on EEG-Inception, a novel deep convolutional neural network, divides the problem in 2 stages to achieve the asynchronous control: (i) the model detects user’s control state, and (ii) decodes the command only if the user is attending to the stimuli. Additionally, we used transfer learning to reduce the calibration time, even exploring a calibration-less approach. Results. Our method was evaluated with 22 healthy subjects, analyzing the impact of the calibration time and number of stimulation sequences on the system’s performance. For the control state detection stage, we report average accuracies above 91% using only 1 sequence of stimulation and 30 calibration trials, reaching a maximum of 96.95% with 15 sequences. Moreover, our calibration-less approach also achieved suitable results, with a maximum accuracy of 89.36%, showing the benefits of transfer learning. As for the overall asynchronous system, which includes both stages, the maximum information transfer rate was 35.54 bpm, a suitable value for high-speed communication. Conclusions. The proposed strategy achieved higher performance with less calibration trials and stimulation sequences than former approaches, representing a promising step forward that paves the way for more practical applications of ERP-based spellers.