Abstract
CYBATHLON is an international championship where people with severe physical disabilities compete with the aid of state-of-the-art assistive technology. In one of the disciplines, the BCI Race, tetraplegic pilots compete in a computer game race by controlling an avatar with a brain-computer interface (BCI). This competition offers a perfect opportunity for BCI researchers to study long-term training effects in potential end-users, and to evaluate BCI performance in a realistic environment. In this work, we describe the BCI system designed by the team Mirage91 for participation in the CYBATHLON BCI Series 2019, as well as in the CYBATHLON 2020 Global Edition. Furthermore, we present the BCI’s interface with the game and the main methodological strategies, along with a detailed evaluation of its performance over the course of the training period, which lasted 14 months. The developed system was a 4-class BCI relying on task-specific modulations of brain rhythms. We implemented inter-session transfer learning to reduce calibration time, and to reinforce the stability of the brain patterns. Additionally, in order to compensate for potential intra-session shifts in the features’ distribution, normalization parameters were continuously adapted in an unsupervised fashion. Across the aforementioned 14 months, we recorded 26 game-based training sessions. Between the first eight sessions, and the final eight sessions leading up to the CYBATHLON 2020 Global Edition, the runtimes significantly improved from 255 ± 23 s (mean ± std) to 225 ± 22 s, respectively. Moreover, we observed a significant increase in the classifier’s accuracy from 46 to 53%, driven by more distinguishable brain patterns. Compared to conventional single session, non-adaptive BCIs, the inter-session transfer learning and unsupervised intra-session adaptation techniques significantly improved the performance. This long-term study demonstrates that regular training helped the pilot to significantly increase the distance between task-specific patterns, which resulted in an improvement of performance, both with respect to class separability in the calibration data, and with respect to the game. Furthermore, it shows that our methodological approaches were beneficial in transferring the performance across sessions, and most importantly to the CYBATHLON competitions.
Highlights
Learning to control an application with a brain-computer interface (BCI) is a tale of two learners (Pfurtscheller and Neuper, 2001; Perdikis et al, 2020)
The adaptive classifier exhibited large biases toward feet but in subsequent games the biases became smaller and the mean probability was closer to 25%. This indicates that the adaptive normalization technique could compensate for distribution shifts between the calibration paradigm and the games within approximately one game run. In this single case study, we applied inter-session transfer learning and unsupervised adaptation techniques to fit a 4-class BCI based on mental imagery to a user with tetraplegia with the goal of participating in a competition called BCI Race (Riener, 2016; Novak et al, 2017)
We decided to fix the features and mental tasks early after a few screening sessions so that we could collect a large dataset with the identical experimental protocol and BCI system. While this allowed us to identify training effects, we cannot rule out that other mental tasks or feature types, or advanced feature selection algorithms (Ang et al, 2008) might have led to higher classification accuracies and, in turn, to stronger training effects for our pilot. This longitudinal study has allowed us to gain valuable insights into long-term user training with a four class BCI based on mental imagery
Summary
Learning to control an application with a brain-computer interface (BCI) is a tale of two learners (Pfurtscheller and Neuper, 2001; Perdikis et al, 2020). In the case of oscillatory BCIs that detect power modulations in brain rhythms associated with distinct mental tasks (Wolpaw et al, 2002), ideally both the brain/user and the computer learn. Given a set or stream of data, machine learning is applied to detect and track user-specific patterns associated with mental tasks. Initial models are typically obtained from userand session-specific calibration data (Pfurtscheller et al, 1999; Guger et al, 2001; Pfurtscheller and Neuper, 2001), previous sessions (Perdikis et al, 2018), or sometimes other users (Kobler and Scherer, 2016; Zanini et al, 2018). Long-term user training with BCIs can be applied in the context of rehabilitation of stroke patients (Ang et al, 2009; Mrachacz-Kersting et al, 2016; Mane et al, 2020)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
