AUTHOR=Moly Alexandre , Aksenov Alexandre , Martel Félix , Aksenova Tetiana 

TITLE=Online adaptive group-wise sparse Penalized Recursive Exponentially Weighted N-way Partial Least Square for epidural intracranial BCI

JOURNAL=Frontiers in Human Neuroscience

VOLUME=Volume 17 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2023.1075666

DOI=10.3389/fnhum.2023.1075666

ISSN=1662-5161

ABSTRACT=Motor Brain-computer interfaces (BCIs) create new communication pathways between the brain and the external effectors for severely motor-impaired subjects. The control of complex effectors such as robotic arm or exoskeleton is generally based on real time decoding of high-resolution neural signal. High dimensional and noisy brain signal brings challenges to overcome such as the decoding model generalization ability, a high computational load etc. Identification of sparse decoders may allow addressing the problems. Sparsity promoting penalization is a common approach to obtain sparse solution. Generally, BCI features are naturally structured and grouped according to spatial (electrodes), frequencies and temporal dimensions. Group-wise sparsity, i.e. the setting to zero the model coefficients within a such group simultaneously, may be beneficial for reducing the computational time/memory and for data transfer. Additionally, online closed loop decoder adaptation (CLDA) is known to be efficient procedure for BCI decoders training allowing taking into account the neuronal feedback. In the article, algorithm for online closed-loop group-wise sparse multilinear decoders training is proposed. Namely, L_p-Penalized Recursive Exponentially Weighted N way Partial Least Square (PREW-NPLS) was explored for three types of sparsity promoting penalization L_p,p=0,0.5,1. The algorithms were tested offline in a pseudo-online manner for features grouped in spatial dimension. Epidural ECoG dataset recorded during long-term BCI experiments of virtual avatar control (left / right hand 3D translation) by a tetraplegic was used for comparison study. Novel algorithms highlighted comparable or better decoding performance to conventional REW-NPLS, achieved with sparse models. The proposed algorithms are compatible with real time CLDA.