|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Brain Computer Interfaces
CoAdapt P300 speller: optimized flashing sequences and online learning
Participants : Maureen Clerc, Théodore Papadopoulo, Loïc Mahé.
Our work in Brain Computer Interfaces was centered around the visual P300 speller system: a virtual keyboard allowing to type words by detecting the P300 wave, an automatic deflection of the central component of the electric potential, occurring approximately 300 ms after the apparition of an intermittent and rare event, on which the user's attention is focussing. The idea behind the P300 speller is very simple: the system displays series of stimuli (flashes), over the keyboard elements, and detects whether or not the EEG recorded after each flash contains a P300. Its implementation is not so simple, because of the low amplitude of the P300 compared to the background EEG, and of the inter-subject variability of this signal.
The advantage of this system is not to require any training on the part of the user. However the BCI system has to be trained to detect the P300 component from the background EEG: this is done through a calibration phase.
We developped a new method to reduce the calibration phase, with a transfer learning method called “mixture of experts” (MOE). The MOE classifier makes its predictions by averaging the decisions from a pre-recorded database of classifiers coming from other recording sessions (with other subjects) [31] . The decisions were made by using an evidence accumulation scheme, which updated the prediction at every flash of the keyboard [17] .
Part of this work has been implemented in the software: CoAdapt P300 Stimulator.
P300 speller: clinical feasibility study with Amyotrophic Lateral Sclerosis
Participants : Maureen Clerc, Théodore Papadopoulo, Loïc Mahé, Asya Metelkina, Violaine Guy [Nice University Hospital] , Claude Desnuelle [Nice University Hospital] .
From September 2013 to July 2014, we were very involved in running an experiment with the Centre de Référence Maladies Neuromusculaires et SLA (CRMN/SLA) of Nice University Hospital. This study, partly funded by “Assocation pour la Recherche sur la Sclérose Latérale Amyotrophique”, was conducted on 20 patients, who routinely come to be examined at the hospital. Each patient came for 3 sessions where he/she was allowed to use the P300 speller, after being equipped with an electro-encephalography cap, and watching a video explaining the modus operandi of the P300 and on their role in the study.
The P300 speller system has been organized in a way to make it relatively easy to deploy in a clinical setting: it involves only one laptop, and requires limited intervention from the caregiver. The most intricate operation is to position the EEG headset and ensure a correct impedance (below 5 k) for all electrodes.
Each session consisted of three blocks, after the initial calibration phase: a copy spelling task of two ten-letter words, a free spelling task of approximately twenty characters, and an optional block of free use of the system for writing. Finally, the patient was asked to answer a questionnaire. This study intends to investigate the feasibility of setting up and using the P300 speller, from an operational point of view at the hospital. Translational studies of this type are extremely important for the adaptation the BCI systems to the target patient populations, and a large-scale usability study for the P300 speller has never been done before in France.
BCI Challenge: A spell on you if you cannot detect errors!
Participants : Maureen Clerc, Théodore Papadopoulo, Jérémie Mattout [Centre de Recherche en Neurosciences de Lyon, INSERM] , Emmanuel Maby [Centre de Recherche en Neurosciences de Lyon, INSERM] .
We have proposed an international BCI Challenge on decoding Error Potential signals. The winners will be announced at the 7th International IEEE/EMBS Conference on Neural Engineering, Montpellier, in April 2015. The Challenge was open on the Kaggle platform on Nov 14, 2014 and will close on Feb 24, 2015, see: website . At this date the competition is still open, and it has so far attracted 212 participants forming 181 teams.
In the P300 speller paradigm (see above) and in other BCI where a discrete feedback can be presented to the user, the EEG evoked response to the feedback can be recorded and processed online in order to evaluate whether the item selection was correct or not. This decision, if reliable, could then be used to improve the BCI performance by implementing some error correction strategy. In this competition, participants are asked to submit an Error Potential detection algorithm, capable of detecting the erroneous feedbacks online and to generalize across subjects (transfer learning).
The data used in this competition was acquired in the scope of the CoAdapt ANR project.