Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Neural correlates of decoding and learning to control a covert speech brain-computer interface
Poster B7 in Poster Session B, Tuesday, October 24, 3:30 - 5:15 pm CEST, Espace Vieux-Port
Kinkini Bhadra1, Anne Lise Giraud1,2*, Silvia Marchesotti1*; 1Department of Fundamental Neurosciences, University of Geneva, Geneva, Switzerland, 2Institut Pasteur, Université Paris Cité, Inserm, Institut de l’Audition, F-75012 Paris, France
Disruption in speech production can have a devastating effect on patients and their caregivers in terms of quality of life. Brain-Computer interfaces (BCI) have the potential to revolutionize communication in such patients by providing alternative communication channels (by real-time decoding of speech directly from the remaining intact brain areas) or rehabilitation solutions (by exploiting neural plasticity mechanisms using BCI-feedback). Although recent studies confirm the possibility of decoding covert speech (i.e. imagined speech) from pre-recorded intracranial neurophysiological signals, current efforts focus on collecting vast amounts of data to train classifiers, rather than exploring how the brain can adapt to improve BCI control. In addition, neural signals related to covert speech production are weak as compared to overt speech, and might be subjected to intra-individual differences in the ability to perform the imaginary task. In this study we addressed speech-BCI controllability from a neurophysiological point of view by training 15 healthy participants to operate a BCI based on electroencephalography (EEG) signals during a binary syllable imagery task for 5 consecutive days. We calibrated the decoder based on offline data (i.e. without feedback) using a Random Forest classifier and then applied it to classify the EEG signals in real-time alongside providing the participant with a visual feedback (online session). We investigated whether BCI-control can be improved with training and characterize the evolution of the underlying neural correlates, both in terms of changes in EEG power during syllable imagery and in the neural features used for real-time classification. We found a significant linear improvement in BCI control performance from day 1 to 5. This improvement was found in 11 out of 15 participants and was proportional to the average BCI-control performance. Next, we tested the classifier features’ dynamics over the 5 training days. We found a significant decrease in discriminating between the two syllables in low-frequencies (2-10 Hz) over bilateral temporal and frontal regions, together with an increase in high-frequencies (52-66 Hz) over the left fronto-temporal regions. Furthermore, training was accompanied by widespread power increase in theta and low-gamma bands. The change in BCI performance was specifically related to an increase in frontal theta and left temporal gamma power. Overall, our results show that neural features can adapt to improve covert-speech BCI performance. More specifically, we observed focal high gamma oscillations to be involved in multiple aspects of BCI control, including overall discrimination between the syllables (bilateral temporal), feature evolution (fronto-temporal), and neural changes occurring during the 5 training days (left temporal). Low-frequency oscillations below 10 Hz were more distributed and contributed to feature evolution and learning to control the BCI system. These results show that learning to operate a covert-speech BCI involves dynamical changes in both low- and high-frequency neural correlates and provides solid neurophysiological grounds to improve current speech-BCI systems. Moreover, future BCI applications will require a combination of machine and human learning to reach optimal controllability. Improvements on the user side could effectively compensate for the limited success in accurately decoding imagined speech as compared to attempted speech.
Topic Areas: Language Production, Speech-Language Treatment