Leveraging anatomical information to improve transfer learning in brain–computer interfaces

Abstract

Objective Brain-computer interfaces (BCIs) represent a technology with the potential to rehabilitate a range of traumatic and degenerative nervous system conditions but require a time-consuming training process to calibrate. An area of BCI research known as transfer learning is aimed at accelerating training by recycling previously recorded training data across sessions or subjects. Training data, however, is typically transferred from one electrode configuration to another without taking individual head anatomy or electrode positioning into account, which may underutilize the recycled data. Approach We explore transfer learning with the use of source imaging, which estimates neural activity in the cortex. Transferring estimates of cortical activity, in contrast to scalp recordings, provides a way to compensate for variability in electrode positioning and head morphologies across subjects and sessions. Main Results Based on simulated and measured EEG activity, we trained a classifier using data transferred exclusively from other subjects and achieved accuracies that were comparable to or surpassed a benchmark classifier (representative of a real-world BCI). Our results indicate that classification improvements depend on the number of trials transferred and the cortical region of interest. Significance These findings suggest that cortical source-based transfer learning is a principled method to transfer data that improves BCI classification performance and provides a path to reduce BCI calibration time.

Intra- and Inter-subject Variability in EEG-Based Sensorimotor Brain Computer Interface: A Review

“…A transfer learning method is worthwhile if the subjects share non-stationarities that can be modeled in an intersubject context, but ineffective if the subjects exhibit unlike non-stationarities (Samek et al, 2013). The term inter-subject associativity refers to potential inter-subject BCI performance predictors, which could be incorporated into BCI design to augment transfer learning (Kang and Choi, 2014;Wronkiewicz et al, 2015;Saha et al, 2017aSaha et al, ,b, 2019. Source-space analysis for detecting inter-subject associative EEG channels can improve SMR-based BCI performance (Wronkiewicz et al, 2015;Saha et al, 2017aSaha et al, , 2019.…”

“…The term inter-subject associativity refers to potential inter-subject BCI performance predictors, which could be incorporated into BCI design to augment transfer learning (Kang and Choi, 2014;Wronkiewicz et al, 2015;Saha et al, 2017aSaha et al, ,b, 2019. Source-space analysis for detecting inter-subject associative EEG channels can improve SMR-based BCI performance (Wronkiewicz et al, 2015;Saha et al, 2017aSaha et al, , 2019. For example, the classification accuracies for two different subject pairs are 90.36 ± 5.59% and 63.21 ± 8.43%, respectively, suggesting not both subject pairs can be used to achieve a good performance (Saha et al, 2019).…”

“…Recent advances in deep learning methods demonstrate a potential application that alleviates intraand inter-subject variability in BCI settings (Chiarelli et al, 2018;Fahimi et al, 2018). Meanwhile, recent studies suggest that the quantification of inter-subject associativity could be equally important to increase the efficacy of exclusively machine learning-based transfer learning strategies for covariate shift adaptation (Kang et al, 2009;Kang and Choi, 2014;Wronkiewicz et al, 2015;Saha et al, 2017bSaha et al, , 2019Perdikis et al, 2018).…”

Decoding Motor Imagery through Common Spatial Pattern Filters at the EEG Source Space

“…Since the anatomy used for the forward model is common among all subjects and the selected ROIs are common among all subjects, we would like to check the potential of the algorithm in transfer learning between subjects. There is a study supporting that transfer learning between different subjects by means of source space can achieve higher average single-trial classification accuracy than with a conventional method [ 54 ]. Beyond the BCIC IV 2a dataset that is a common ground for the evaluation of methods decoding multiple MI, we aim to evaluate the improved method on dataset we compiled for the CSI: Brainwave project, containing EEG data of healthy or subjects with spinal cord injury performing multiple motor imagery mainly of the upper limbs [ 36 , 55 ].…”

Empirical Comparison of Distributed Source Localization Methods for Single-Trial Detection of Movement Preparation

“…Due to the necessity of additional constraints and the fact that the EEG and its underlying sources are not fully understood yet, new SLMs are continuously developed (Pizzagalli, 2007 ; Grech et al, 2008 ; Becker et al, 2015 ). For BCIs, mainly the well-known, most established SLMs have been applied, like minimum norm (Noirhomme et al, 2008 ; Besserve et al, 2011 ; Edelman et al, 2014 ; Wronkiewicz et al, 2015 , 2016 ), weighted minimum norm (Qin et al, 2004 ; Babiloni et al, 2007 ; Kamousi et al, 2007 ; Cincotti et al, 2008 ; Yuan and He, 2009 ; Goel et al, 2011 ; Edelman et al, 2015 , 2016 ), standardized low resolution electromagnetic tomography (Congedo et al, 2006 ; Lotte et al, 2009 ; Handiru et al, 2017 ), local autoregressive average (Menendez et al, 2005 ; Poolman et al, 2008 ) and beamformer methods (Grosse-Wentrup et al, 2009 ; Ahn et al, 2012 ). However, a comparison of different distributed SLMs has rarely been reported so far.…”