Recursive N-Way PLS for the adaptive multisession calibration of ECoG based BCI system
-
1
CEA-LETI CLINATEC, France
The goal of Brain Computer Interfaces (BCI) is the translation of the brain electrical activity into commands for external devices. CLINATEC® BCI project aims at improving the quality of life of tetraplegic subjects by allowing them to interact with their environment through the control of effectors with multiple degrees of freedom, such as a 4-limb exoskeleton. A fully implantable device WIMAGINE® (Mestais et al., 2015) for chronic measurement and wireless transmission of Electrocorticograms (ECoG) data, as well as a full body exoskeleton EMY® (Perrot et al., 2013) have been developed within the framework of this project. One of the crucial limitations for the practical application of BCIs is the instability of the characteristics of the recorded signals. This may be caused by external as well as internal reasons such as the brain variability and plasticity. Whereas the complete recalibration of the BCI system is time- and labor consuming, an efficient procedure of the adjustment of the existing BCI decoder is required. Moreover, stable decoder which is not sensitive to session-to-session variation is highly desirable.
Multi-way analysis based on the tensor-data representation, combined with Partial Least Squares (PLS) family approaches, was reported as an efficient way for calibration of BCIs. Recursive N-Way Partial Least Squares (RNPLS) regression was proposed recently (Eliseyev and Aksenova, 2013) for adaptive calibration of NPLS decoder. It allows adjustment of the preliminary calibrated model to mild data changes. In the present study, RNPLS was tested to be applied for multi-session BCI decoder calibration. It is compared with the state of art NPLS algorithm (Bro, 1996) on the publically available database (http://neurotycho.org), providing ECoG recordings of the Japanese monkeys (Shimoda et al., 2012). 10 recordings corresponding to the monkey denoted as “B” were chosen. Each recording was split into training and testing parts: 10 minutes for the training and 5 minutes for the testing. NPLS was independently applied to each training recording. 10 NPLS models were generated, one per each recording. Contrary to the NPLS, the RNPLS algorithm was sequentially applied for all the training recordings and resulted in one “universal” model.
The resulted NPLS models were tested: (1) on the testing recording from the same with training session; (2) on the testing recordings from other sessions. The RNPLS model was applied to the testing recordings of all 10 sessions. In the case (1), the prediction quality of the RNPLS model was not significantly different in comparison with the NPLS models. Root Mean Squares Error (RMSE) for RNPLS and NPLS were 0.71±0.10 and 0.70±0.08, respectively (p=0.89). In the case (2), when the training and testing sets are from the different sessions, the RNPLS model outperformed the NPLS ones: 0.71±0.10 vs. 0.80±0.13, p=0.03 (the improvement reached 72.1%, 12.7% in average). Thus, whereas the NPLS models are considerably related to the recording on which they were identified, the RNPLS model could be efficiently applied to each recording.
The experiments demonstrated the interest of the RNPLS method in BCIs. Due to its possibility to treat data flows, the method allows not only the handling of huge datasets, but also adjustment of the model in the time-varying and non-stationary processes.
Acknowledgements
The BCI project was supported by French National Research Agency (ANR-Carnot Institute), Fondation Motrice, Fondation Nanosciences, Fondation de l’Avenir, and Fondation Philanthropique Edmond J. Safra.
References
Bro R 1996 Multiway calidration. Multilinear PLS Journal of Chemometrics 10 47-61
Eliseyev A and Aksenova T 2013 Recursive N-way partial least squares for brain-computer interface PloS one 8 e69962
Mestais C, Charvet G, Sauter-Starace F, Foerster M, Ratel D and Benabid A L 2015 WIMAGINE®: Wireless 64-channel ECoG recording implant for long term clinical applications IEEE TNSRE 23
Perrot Y, Verney A, Morinière B and Garrec P 2013 EMY: Full-body Exoskeleton. In: ACM SIGGRAPH Emerging Technologies, (Anaheim, US)
Shimoda K, Nagasaka Y, Chao Z C and Fujii N 2012 Decoding continuous three-dimensional hand trajectories from epidural electrocorticographic signals in Japanese macaques Journal of neural engineering 9 036015
Keywords:
brain-computer interface (BCI),
adaptive algorithms,
Tensor-based approach,
electrocorticography (ECoG),
partial least squares (PLS)
Conference:
German-Japanese Adaptive BCI Workshop, Kyoto, Japan, 28 Oct - 29 Oct, 2015.
Presentation Type:
Poster presentation
Topic:
Adaptive BCI
Citation:
Eliseyev
A and
Aksenova
T
(2015). Recursive N-Way PLS for the adaptive multisession calibration of ECoG based BCI system.
Front. Comput. Neurosci.
Conference Abstract:
German-Japanese Adaptive BCI Workshop.
doi: 10.3389/conf.fncom.2015.56.00018
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
30 Sep 2015;
Published Online:
04 Nov 2015.
*
Correspondence:
Dr. Andrey Eliseyev, CEA-LETI CLINATEC, Grenoble, France, eliseyev.andrey@gmail.com