Section: New Results

Improving BCI user performance and training

Participants: Jelena Mladenovic, Léa Pillette, Thibaut Monseigne, Fabien Lotte 

 

The potential of learning companions: As mentioned before, current BCI training protocols do not enable every user to acquire the skills required to use BCIs. We showed that learning companions were promising tools to increase BCI user experience during training, as well as to increase the performances of users who are more inclined to work in a group. Encouraged by these first results we investigated all the other potential benefits learning companions could bring to BCI training by improving the feedback, i.e., the information provided to the user, which is primordial to the learning process and yet have proven both theoretically and practically inadequate in BCI. From these considerations, some guidelines were drawn, open challenges identified and potential solutions were suggested to design and use learning companions for BCIs [29]. 

 

Active Inference for P300 speller: Brain Computer Interface (BCI) mostly relies, on one hand, on the stability of a person's mental commands, and on the other, on the machine's capacity to interpret those commands. As a person is naturally changing and adapting all the time, the machine becomes less successful in interpreting user's commands. In turn, the machine should be able to predict and minimize undesired user fluctuations. Moreover, it should build bottom-up information about the user through physiological input (EEG observations), and influence the user by providing optimal task (action) to minimize prediction error. A novel neuroscience approach, Active (Bayesian) Inference, is a very generic and flexible computational framework that can predict user intentions through a series of optimal actions and observations. On simulated data, we have shown that Active Inference has great potential to enable the machine to co-adapt with the user, and increase performance levels in a P300 speller BCI. We further tested Active Inference on real data, and show that active inference surpasses the standard algorithms while permitting the implementation of various cases of p300 speller BCI within one single framework [57]. 

 

Towards Congruent Feedback for BCI: Congruent visual environment in MI BCI has been researched in virtual reality, giving a sense of body ownership illusion, and showed to be more robust and improve performance. On the other hand, the effects of a congruent, purely audio environment, have not yet been explicitly explored in BCI. This inspired us to explore the benefits of a task-related (congruent) and synchronised audio feedback which would comply with the user’s imagined movements. We investigate the potential of such an audio feedback congruent to the task, tackling the sensory illusion of presence by providing realistic audio feedback using natural sounds. Our preliminary results show the benefits of a congruent, audio MI feedback of feet (sound of footsteps in gravel) as opposed to no congruent feedback using abstract sound [50]. 

 

Neurofeedback of daytime alertness: Neurofeedback consists in providing a subject with information about his own EEG by means of a sensory feedback (visual, auditory ...) in real-time, in order to enable cognitive learning. In collaboration with SANPSY (Pellegrin Hospital/Univ. Bordeaux), we implemented a complete Neurofeedback solution as a proof of concept that aims to determine the level of effectiveness of Neurofeedback on daytime alertness ability. Indeed, excessive daytime sleepiness (EDS) is a common complaint associated with increased accidental risk. The usual countermeasures such as blue light, caffeine or nap have been shown to be effective but have limitations. With a test on five subjects, preliminary data showed that it was possible to learn how to regulate our own EEG activity with a short number of sessions (8 sessions of 40 min). Clinical trials to confirm these results should be initiated in the course of 2019.